Académique Documents
Professionnel Documents
Culture Documents
or explicitly l= 1,2,...,N:
N 3 N
fl(t) = 2 Z I H l l ( W l , t) IX/Aw fj(i)(t) = 2 E E II-I]m(wl't) lv/-~-~
1=1 m=l 1=1
Since the 4~'s are independent random variables performed on the basis of eqn (14), instead of using eqn
distributed uniformly over the interval [0,27r], the (24). The simulation formula corresponding to eqn (14)
expected value appearing in eqn (31) is different from is:
zero only when:
fj.(t) = Aj(t)gy(t), j = 1,2, 3 (39)
m I = m2 = m and I l = l2 = l (32)
where gy(t) is the simulation of the stationary stochastic
Under the conditions of eqn (32), eqn (31) can be written vector process g°(t) having mean value equal to zero
as: and cross-spectral density matrix shown in eqn (16). It
3 N
should be mentioned that the simulation of stationary
and ergodic stochastic vector processes can be per-
Ryk(t , t + 7.) = 2 E Z IHi m(O)t' t) l
m=l l=1 formed with great computational efficiency using the
Fast Fourier Transform (FFT) technique, as described
x [Hkm(Wl, t+W) lAo) by Deodatis. 25
× COS[O)/7. -1- Ojrn(O)l, l) -- Okra(O) h t -']- 7-)]
Comments on computational efficiency
(33)
Taking into account eqns (19), (22) and (23), the At this juncture, it should be pointed out that it is not
expression for Rjk(t, t + 7) in eqn (33) can be written possible to take advantage of the F F T technique when
in the limit as AO) ---. 0 and N ~ o0 (while keeping in using the (non-stationary) simulation formula shown in
mind that Wu = NAO) is constant and that the elements eqn (24), in contrast to the corresponding formula for
of the cross-spectral density matrix S°(O), t) are zero for simulation of stationary stochastic vector processes. 25
[O)] > O)u at any time instant t) in the following way: This is due to the fact that the coefficients
[nym(o)l, t) [ V / ~ in the double summation of eqn (24)
are now functions of both frequency and time. However,
Rjk(t't + 7-) = J ~ 3my~/lHJ
= l m(O)'t)l'Hk'(O)'t + this should not be of any great concern computationaUy,
X e i[wr+Oy"(w' t)--Okm(W' t+7")] d w (34)
since in most cases of practical interest the non-
stationary stochastic vector process J)°(t), j = 1,2, 3 is
Using now the polar form representation of the elements limited to relatively short durations by the modulating
of the H(o), t) matrix (eqn (20)), eqn (34) is written as: functions Aj(w, t), j = 1,2, 3 (e.g. ground motion
acceleration time histories). A case of non-stationary
Rjk(t,t + 7-) = Z Hjm(O),t)Hi*m(O),t + r)ei~" dO) stochastic vector processes where the F F T technique can
-oo m=l be used in the simulation formula is that of uniformly
(35) modulated processes (eqn (39)). At this juncture, it
should be noted that Li and Kareem 22 have proposed
Then, using the decomposition shown in eqn (17) in a methodology to simulate non-stationary vector
conjunction with eqns (4) and (5), it is straightforward processes taking advantage of FFT.
to show that:
Rjk(t,t + 7") = I f Aj(w,t)Ak(W,t + 7-) The simulation algorithm presented earlier in this
paper can generate sample functions of a general non-
x v/Sy(O)) Sk(o)) l"jk(w) e i~r dw, stationary stochastic vector process with evolutionary
power, according to a prescribed non-stationary cross-
j,k=1,2,3, j ~ k (37)
spectral density matrix. A very important application is
Comparing finally eqns (9) and (10) to eqns (36) and simulation of earthquake ground motion time histories.
(37), it is evident that: In such a case, acceleration, velocity, or displacement
time histories can be generated at several locations on
Rjk(t, t + 7-) = ROk(t, t + 7-), j, k = 1,2, 3 (38)
the ground surface according to a target cross-spectral
density matrix. This will be demonstrated in the first two
Special case: uniformly modulated non-stationary numerical examples in the following section. The only
stochastic vector process drawback in this approach is that it is often preferable to
work with time histories that are compatible with
For the special case of a uniformly modulated non- prescribed response spectra, rather than with prescribed
stationary stochastic vector process, simulation can be power spectral density functions (cross-spectral density
154 G. Deodatis
matrix). An obvious reason for this preference is that it Table 1. Iterative scheme to simulate response spectrum
is much easier and more reliable to find a response compatible acceleration time histories at three points on the
ground surface
spectrum specified for a given location, rather than a
power spectral density function. For example, design
codes usually provide response spectra as a function of
local site conditions. Read input data:
t> target acceleration response spectra: RSAj(w); j = 1,2,3
In order to address this issue, a methodology is complex cohereaace functions: Fj/,(w); j , k = 1,2,3; j ~ k
modulating functions: Ai(t)! j = 1, 2, 3
proposed now to simulate seismic ground motion
compatible with the following three prescribed quan-
initialize power spectral density functions Sl(t0), S2(to), S3(w ) I
tifies: (i) response spectra, (ii) complex coherence func- by setting them equal to n (noa-zero) constant value
tions, and (iii) modulating functions. Although the over the entire frequency range
I
methodology is presented in the following for accelera-
.___~ Generate gl(t), g2(t), 93(4) as a stationary, tri-variate, stochastic vector process
tions, it is straightforward to modify it to accommodate rI with cross-spectral density matrix shown in Eq. (16). Refer to Eq. (39).
velocities or displacements. The choice of accelerations
is done solely for demonstration purposes. Compute acceleration time histories as: fl(t) = At(t)gl(t), f2(t) = A2(t)g2(t),
and fa(t) = A3(t)g3(t). Refer to Eq. (39).
According to the proposed methodology, the accele-
ration time histories at three points on the ground Compute acceleration response spectra RSA(ID(w), RSA(I*)(w), RSA(Ia)(w),
surface are considered to be a uniformly modulated, tri- corresponding to fl(t), f2(t), fs(t), respectively.
x[/ u/_
03 2 2 2 ~
1.1
2
+ 4 ~ j ~-~fj. 1
'
literature that can be used for Sj(~), 7jk(W) and Aj(w, t).
The next step is to select numerical values for the
parameters appearing in eqns (42)-(45). Starting from
%j and (gj, j = 1,2, 3 appearing in eqn (42), the values
suggested by Ellingwood and Batts 36 for three different
j = 1,2,3 (42) soil conditions are used in this study:
where Soj is a constant determining the intensity of Point 1: Rock or stiff soil conditions:
acceleration at point j, OJgj and (gj c a n be thought of as
some characteristic frequency and damping ratio of the Wgl = 87rrad/s, (gl = 0.60 (46a)
ground at point j, and wfj and (fj a r e filtering parameters Point 2: Deep cohesionless soils:
for point j.
The Harichandran-Vanmarcke model 35 is chosen to Wg2 = 57r rad/s, (g2 = 0.60 (46b)
describe the (stationary) coherence functions 7jk(~O), Point 3: Soft to medium clays and sands:
j, k = 1,2, 3, j ~ k between fj°(t) and f°(/):
O)g3 = 2.47rrad/s, (g3 = 0.85 (46c)
7 j k ( w ) = A e x p [ - - ~2(jk (1 -- A +etA)] The filtering parameter a;fj (eqn (42)) is set equal to 0.10
of the corresponding Wgj value, while the other filtering
+ (1 - A) exp [ - 2~jk (1 - A + aA)] parameter (fj (eqn. (42)) is set equal to the correspond-
L ing (gj value, following the recommendation by Hindy
j , k = 1,2,3, j ~ k (43) and Novak: 37
where ~jk is the distance between points j and k, O(a)) is Point 1: Rock or stiff soil conditions:
the frequency-dependent correlation distance: ~fl = 0.87rrad/s, (fl = 0.60 (47a)
Point 2: Deep cohesionless soils:
= k 1 + \w0: j (44)
wf2 = 0.57rrad/s, (f2 = 0-60 (47b)
and A, a, k, ~0 and b are model parameters. Point 3: Soft to medium clays and sands:
Since the acceleration time historiesfj°(t),j = 1,2, 3 are
modeled as a uniformly modulated non-stationary sto- ~;f3 = 0.247rrad/s, (f3 = 0"85 (47c)
chastic vector process, the modulating functions Aj(~, t), Finally, the last parameter appearing in eqn (42), Soj,
j = 1,2, 3 will be functions of time only. The Bogdanoff- j = 1,2, 3, is computed so that the standard deviation of
Goldberg-Bernard model 34 is used for this purpose: the Kanai-Tajimi part of the (stationary) power spectral
a l ( ~ , t) = .41(t) = altexp(-a2t) for t > 0 (45a) density function is equal to 100 cm/s 2 for all three points
1, 2 and 3:
A2(w , t) = A2(t )
S01 = 62.3 cm2/s 3, S02 = 99.7 cm2/s 3,
a t -- exp --a2 t -- for t > (2__~1 803 = 184.5cm2/s 3 (48)
0 for0< / < ~2..~1 For the various parameters appearing in eqns (43) and
V (44), the values estimated by Harichandran and Wang 38 by
(45b) analyzing data from the SMART-1 seismograph array will
be used in this study for demonstration purposes:
A3(a;, t) = A3(t)
A = 0.626, a = 0"022, k = 19700m,
/al t- exp -a2 t -~ 1 fort>~31 _
"~-
o
~ )
c:l
~,-
o-
i
2'0 ,'0 6'o ~, ,~o
Frequencyw(rad/sec) o
110 1'5
Fig. 2. The power spectral density functions at points 1, 2 and Time (see)
3, for Example 1: Sl(w)~ rock or stiff soil conditions;
S2(w)---~deep cohesionless soils; and S3(w)~soft
medium clays and sands.
to Fig. 4. Modulating functions At(t), A2(t) and .43(t),for
Example 1.
158 G. Deodatis
O
}~ li i~["l l ~ l J Point I--+ rock or stiff soil conditions
~J
• • °
E3 o+
•z , ~,
0 10 20
Q 0
ta
-----'-r "v-----
o I+o (s a0
Time (see)
Fig. 5. Generated sample function for the acceleration at points 1, 2 and 3, over a length equal to 18.85 s (Example 1).
general shape after some loss of coherence described by from t = 0 up to t = 18.85s (the duration of the time
the (stationary) coherence functions shown in Fig. 3. histories shown in Fig. 5).
It is therefore obvious from Figs 5 and 6 that the
proposed algorithm is able to simulate non-stationary Example 2
ground motion time histories that are spatially corre-
lated according to a given coherence function, include Unlike the first example where ground motion time
the wave propagation effect and are non-homogeneous histories were modeled as a uniformly modulated non-
in space (or equivalently they correspond to different stationary stochastic vector process, in this example they
frequency contents). are modeled as a non-stationary stochastic vector
Finally, the ensemble auto-/cross-correlation function process with amplitude and frequency modulation.
Rjk(t , t + "r) is computed from 1000 sample functions at This means that both the amplitude and the frequency
time instant t = 3.14s and plotted as a function of r in content of ground motion change as a function of time
Fig. 7 versus the target auto-/cross-correlation function (in the first example, it was only the amplitude of ground
R°k(t, t + "r). As can be seen in Fig. 7, the agreement motion that was varying with time). As in the previous
between R°k(t, t + 7-) and Rjk(t, t + "r) is very good, example, sample functions will be generated according
especially in the vicinity of the dominant peak of the to a target cross-spectral density matrix.
auto-/cross-correlation functions. Some small differ- Figure 9 displays an acceleration record from the 1964
ences can be observed away from this dominant peak Niigata earthquake involving both amplitude and fre-
in Fig. 7. However, these differences disappear when the quency content variation as a function of time. Specifi-
ensemble auto-/cross-correlation function Rjk( t , t + -c) is cally, this record shows an abrupt change of its frequency
computed from 100 000 sample functions, as can be seen content between approximately 8 and 10 s. During this 2 s
in Fig. 8 where R?k(t , t + T) and Rjk(t, t + -c) practically period, the frequency content is transformed from one
coincide. containing a relatively broad band of frequencies to one
It should be noted that the agreement between containing essentially a single low frequency. It is believed
R°k(t, t + r) and R/k(t, t + 7-) observed in Figs 7 and 8 that this phenomenon is due to soil liquefaction. The
for time instant t = 3.14 s is typical for any time instant objective of this example will be to reproduce the general
Non-stationary stochastic vector processes 159
A
i A -- , I f~ i~. Point 1 -* rock or stiff soil conditions
s - /V V wv vv v \
,,q
°
I
315 ,~ ) 415 510 515 610
/\
,v
frequency and amplitude variation characteristics of the The three components of the tri-variate non-stationary
acceleration record shown in Fig. 9. stochastic vector process describing the acceleration time
The acceleration time histories at three points on the histories at points 1, 2 and 3 (see Fig. 10) are denoted by
ground surface are again considered to be a tri-variate f]°(t),f°(t) andf3°(t), respectively. The mean value of the
non-stationary stochastic vector process. The configu- process is equal to zero (see eqn (1)), while the elements
ration of the three points is shown in Fig. 10, indicating of its cross-spectral density matrix (see eqn (3)) are given
that it is not necessary for the points to be on a straight by the expressions shown in eqns (40) and (41). It
line, or to be equidistant (compare to previous example). should be pointed out, however, that since no wave
For simplicity, no wave propagation will be considered in propagation is considered in this example, the wave
this example. The simulated earthquake ground motion at propagation term exp[--i(w~jk/V)] in eqn (41) should be
points 1, 2 and 3 will have the following characteristics: set equal to unity (v ~ c~).
The Clough-Penzien acceleration spectrum shown in
(i) Points 1, 2 and 3 correspond to the same local eqn (42) is selected again to model the power spectral
soil conditions. density functions Sj(~), j = 1,2, 3 of the acceleration
(ii) The frequency content of the acceleration time time historiesfj°(t), j = 1,2, 3, respectively. However, in
histories at these three points will change with order to account for the variation of the frequency
time as will be described in the following, in order content with time, the characteristic frequency and
to capture the unique characteristics of the damping ratio of the ground are considered now to vary
record shown in Fig. 9. as a function of time as follows:
(iii) Tbe acceleration time histories at points 1, 2 and
15.56rad/s f o r 0 _ < t < 4 . 5 s
3 will be correlated according to a specified
coherence function. ~gj(t) = 27'12tp3 - 40"68t2 + 15"56
(iv) Finally, the amplitude variation as a function of for 4.5 < t < 5.5 s, j = 1 , 2 , 3
time of the acceleration time histories at these 2.0rad/s for t _> 5.5s
three points will be described by prescribed
modulating functions Aj(u3, t), j = 1,2, 3. (53)
160 G. Deodatis
O
g O
O
tM
... ,. . . . . . ..
..... o
0
0
I
O
g. Rss -* dotted line --* dotted line
(D
• ' ' .-"%...,t ,.. ,.._ . . _ . ".."...', . .. "'.... . ' .. "' . '"-...f",'.. ""_-"..,.'. ..
r (see) r (see)
Fig. 7. Ensemble auto-/cross-correlation functions [R/~(t, t + r)] computed from 1000 sample functions versus the corresponding
targets [R°k(t, t + r)]. Both R j k ( t , t + r) and Ryk(t, t + r) are plotted at time instant t = 3.14s as a function of r.
7rWgj(t)[2~gj(t)+~]' j= 1,2,3
for 4.5 < t < 5.5s, j - - 1,2,3
0.015 for t >_ 5.5s (56)
(54) The H a r i c h a n d r a n - V a n m a r c k e model shown in eqns
(43) and (44) is chosen again to describe the coherence
where tp = t - 4.5 s and
functions 7jk(•), j, k = 1,2, 3; j # k between fj°(t) and
~fj(t) = O'lwgj(t), (fj(t) = (gj(t), j = 1,2, 3
f°(t). The model p a r a m e t e r s A, a, k, w0 and b are
assigned the values f r o m the previous example shown in
(55) eqn (49).
The B o g d a n o f f - G o l d b e r g - B e r n a r d model is selected
The expressions for Wgj(t) and (g/(t) are taken f r o m again for the m o d u l a t i n g functions:
Deodatis and Shinozuka 39 and describe a sudden d r o p
in the values o f the characteristic frequency and Aj(~,t)=Aj(t)=altexp(-a2t), j=1,2,3 (57)
d a m p i n g ratio during the one-second period f r o m with model p a r a m e t e r s al and a2 set equal to
t = 4.5 to 5.5 s. The constant determining the intensity
o f the acceleration is also going to be a function o f time, al = 0.680 and a 2 = 1/4 (58)
so that the standard deviation o f the K a n a i - T a j i m i part so that the m a x i m u m value o f Aj(t), j = 1,2, 3 occurs at
o f the spectrum is equal to ~r = 100 cm/s 2 at every time time t = 4 s and is equal to unity.
Non-stationary stochastic vector processes 161
O O
oo. 804
04
O
O O
O" O
oo
-X F- "
i
80
l
~
R~ss --, continuous line R~s2 ~ continuous line
---, dotted line O z --' dotted line
oo
tO
O
O
r (sec)
Fig. 8. Ensemble auto-/cross-correlation functions [Rjk(!, t + T)] computed from 100 000 sample functions versus the corresponding
targets [R°k(t, t + T)]. Both Rjk(t, t + r) and Ryk(t, t + T) are plotted at time instant t = 3.14s as a function of T.
131.8
0.0
-131.8 . . . . . . . .
0.0 6.79 13.57 20.36 27.15 33.93
Time (sec)
Fig. 9. Acceleration record from the 1964 Niigata earthquake.
162 G. Deodatis
Point 2 ,S~j(w,=t7sec)(dividedby170)
s e c )
[8-
Poin t (/Y"
_ 50 m - - - ~ ' ~ o i nrlt l
Fig. 10. Configuration of points 1, 2 and 3 on the ground
surface, for Example 2. Points 1, 2 and 3 correspond to the
same local soil conditions.
o-
1'0 2'0 ~, ,~ ~
Point 1
0 15 20
Point 2
Co,
0 10 20
C o oO4 Point 3
tl
..%o
b 5 io 15 2'o
Time (sec)
Fig. 12. Generated sample function for the acceleration at points 1, 2 and 3, over a length equal to 20 s (Example 2).
(ii) The acceleration time histories at these three The acceleration response spectra specified by the
points will be correlated according to a pre- Uniform Building Code 4° are selected for the three
scribed coherence function and they will reflect points in Fig. 13. For the purposes of this numerical
the wave propagation effect according to a example, the peak ground acceleration is set equal to
specified velocity of wave propagation. 200 cm/s2 and the corresponding Uniform Building Code
(iii) Finally, the amplitude variation as a function of (UBC) acceleration response spectra RSAj(CO), j = 1,2, 3
time of the acceleration time histories at these are plotted in Fig. 14. It is reminded that sample functions
three points will be described by prescribed of ground motion time histories will be generated to be
modulating functions Aj(co, t), j = 1,2, 3. compatible with these UBC response spectra.
The Abrahamson model 32 is chosen to describe the
The three components of the tri-variate non-stationary (stationary) coherence functions ~jk(CO), j , k = 1,2,3,
stochastic vector process describing the acceleration j ~ k between fj°(t) a n d f ° ( t ) :
time histories at points 1, 2 and 3 (see Fig. 13) are
1
denoted by fj0 (t), f o (t) and f3° (t), respectively. "yjk(co) = CO 6
73
k
/
I~ 50 m -,--'- 50 m ,-1
x tanh J c3 ((jk)
CO CO2
L1 + ~c4(~jk) + 4-~ cT(~jk)
Fig. 13. Configuration of points l, 2 and 3 on the ground
surface along the line of main wave propagation, for Example + [4"80 - c3((j~)] exp c6(~jk)~-~ + 0"35 ,
3. Point 1 corresponds to rock and stiff soils (UBC Type 1),
point 2 corresponds to deep cohesionless or stiff clay soils
(UBC Type 2), and point 3 corresponds to soft to medium
clays and sands (UBC Type 3). j , k = 1,2,3, j Ck (60)
164 G. Deodatis
ff,- Since the acceleration time histories j~0 (t), j = 1,2, 3 are
5% damping modeled as a uniformly modulated non-stationary
stochastic vector process, the modulating functions
A/(~v,t), j = 1,2,3 will be functions of time only. In
order to control the duration of strong ground motion,
& the model suggested by Jennings et ai.4! is selected for
the modulating functions A/(t), j = 1,2,3. Figure 16
plots AI (t) and indicates the numerical values chosen for
ge- all the parameters (A2(t) and Aa(t) have the same form
as Al(t), but they are shifted by ~21/v and ~31/v,
respectively, to include the wave propagation effect in
the same way as in eqn (45)).
The generation of sample functions of the acceleration
i o time histories at the three points shown in Fig. 13 will be
.< 0'.5 1'.o 5'.0 1~.o d.o ,oko performed using the iterative scheme shown in Table 1
F~'equencyw (rad/sec) (refer to the section entitled 'Simulation of seismic
ground motion compatible with prescribed response
Fig. 14. Uniform Building Code acceleration response spectra spectra'). According to Table 1, the power spectral
RSAj(~v),j = 1,2, 3, assigned to points 1, 2 and 3 (see Fig. 13), density functions Sy(~v),j = 1,2, 3 must be initialized by
respectively. Point 1 corresponds to rock and stiff soils (UBC
Type 1), point 2 corresponds to deep cohesionless or stiff clay setting them equal to a constant value over the entire
soils (UBC Type 2), and point 3 corresponds to soft to medium frequency range. This value can be selected arbitrarily
clays and sands (UBC Type 3) (Example 3). and in this example is set equal to:
where ~jk is the distance between points j and k, and Sj(o3) : 100 cm2/s 3, j = 1,2, 3 (68)
3.95 while the upper cut-off frequency wu and the value of N
c3(~jk) -- (1 + 0"0077~yk + 0"000023 ~j2k) (see eqn (27)) are set equal to:
+ 0"85 exp{--0-000 13~yk} (61) ~vu = 128 rad/s and N = 128 (69)
The simulation is performed at 6144 time instants, with
a time step At = 3.07 x 10 -3 s, over a length equal to
0.4 1 3 6144 x 3.07 x 1 0 - 3 = 18-85s. One sample function for
1+ the acceleration at points 1, 2 and 3 (see Fig. 13),
denoted by fl(t), f2(t) and j~(t), respectively, is gene-
rated after 10 iterations and displayed in Fig. 17.
CONCLUSIONS
the wave propagation effect according to a velocity of
wave propagation v = 2000m/s; (iii) their amplitude A spectral-representation-based simulation algorithm
variation as a function of time according to the modu- was used to generate sample functions of a non-
lating function plotted in Fig. 16. In this example, stationary, multi-variate stochastic process with evolu-
however, the main objective is to generate acceleration tionary power, according to its prescribed non-station-
time histories that will be compatible with prescribed ary cross-spectral density matrix. If the components of
response spectra. For this purpose, the acceleration the vector process correspond to different locations in
response spectra RSA(g)(w), j = 1,2, 3 computed using space, then the process is also non-homogeneous in
the ground motion time histories shown in Fig. 17 space (in addition to being non-stationary in time). The
should be compared with the target UBC response ensemble cross-correlation matrix of the generated
spectra RSAj(w), j = 1,2,3 plotted in Fig. 14. This sample functions is identical to the corresponding
comparison is carried out in Fig. 18, where it can be seen target and the generated sample functions are Gaussian
that the 10 iterations performed to obtain the sample in the limit as the number of terms in the frequency
function shown in Fig. 17 are enough for an excellent discretization of the cross-spectral density matrix
match at every frequency. approaches infinity.
0
~-. o. Point 1 --~ rock and stiffsoils ( U B C Type I)
~ ,,e •
0 5 10 15 20
Time (sec)
Fig. 17. Generated sample function for the acceleration at points 1, 2 and 3, after 10 iterations, over a length equal to 18"85s
(Example 3).
166 G. Deodatis
if-.
5% damping ~q 5% damping
;-4
t)
;'4
Fig. 18. Acceleration response spectra RSA (fJ)(w), j = 1,2, 3 computed using the generated acceleration time histories shown in Fig.
17 versus the UBC acceleration response spectra RSAj(w), j = 1,2, 3.
and H. H. E. Leipholz, University of Waterloo Press, 28. Priestley, M. B., Evolutionary spectra and non-stationary
Waterloo, 1974, pp. 277-86. processes. Jr. Royal Statistical Society, Series B, 27 (1965)
12. Deodatis, G. & Shinozuka, M., Simulation of seismic 204-37.
ground motion using stochastic waves. J. Engng Mech., 29. Priestley, M. B., Non-linear and Non-stationary Time Series
ASCE, 115 (1989) 2723-37. Analysis, Academic Press, New York, 1988.
13. Yamazaki, F. & Shinozuka, M., Digital generation of non- 30. Gasparini, D. & Vanmarcke, E. H., Simulated earthquake
Gaussian stochastic fields. J. Engng Mech., ASCE, 114 motions compatible with prescribed response spectra.
(1988) 1183-97. Technical Report, Department of Civil Engineering,
14. Grigoriu, M., On the spectral representation method in Massachusetts Institute of Technology, Publication No.
simulation. J. Probabilistic Engng Mech., 8 (1993) 75-90. R76-4, 1976.
15. Shinozuka, M., Stochastic fields and their digital simula- 31. Hao, H., Oliveira, C. S. & Penzien, J., Multiple-station
tion, Stochastic Methods in Structural Dynamics, eds G. I. ground motion processing and simulation based on
Schu~ller and M. Shinozuka, Martinus Nijhoff Publishers, SMART-1 array data. Nuclear Engng and Design, 111
Dordrecht, 1987, pp. 93-133. (1989) 293-310.
16. Shinozuka, M. & Deodatis, G., Stochastic process models 32. Abrahamson, N. A., Spatial variation of multiple support
for earthquake ground motion. J. Probabilistic Engng inputs. Proc. 1st U.S. Seminar on Seismic Evaluation and
Mech., 3 (1988) 114-23. Retrofit of Steel Bridges, University of California at
17. Shinozuka, M. & Deodatis, G., Simulation of stochastic Berkeley, San Francisco, 18 October, 1993.
processes by spectral representation. Appl. Mech. Rev., 33. Clough, R. W. & Penzien, J., Dynamics of Structures,
ASME, 44 (1991) 191-204. McGraw-Hill, New York, 1975.
18. Shinozuka, M. & Deodatis, G., Simulation of multi- 34. Bogdanoff, J. L., Goldberg, J. E. & Bernard, M. C.,
dimensional Gaussian stochastic fields by spectral repre- Response of a simple structure to a random earthquake-
sentation. Appl. Mech. Rev., ASME, 49 (1996) 29-53. type disturbance. Bulletin of the Seismological Society of
19. Gersch, W. & Yonemoto, J., Synthesis of multi-variate America, 51 (1961) 293-310.
random vibration systems: a two-stage least squares 35. Harichandran, R. S. & Vanmarcke, E. H., Stochastic
ARMA model approach. J. Sound and Vibration, 52 variation of earthquake ground motion in space and time.
(1977) 553-65. J. Engng Mech., ASCE, 112 (1986) 154-74.
20. Mignolet, M. P. & Spanos, P.-T. D., Recursive simulation 36. Ellingwood, B. R. & Batts, M. E., Characterization of
of stationary multi-variate random processes. Part I. J. earthquake forces for probability-based design of nuclear
Appl. Mech., 54 (1987) 674-80. structures. Technical Report BNL-NUREG-51587,
21. Mignolet, M. P. & Spanos, P.-T. D., Recursive simulation NUREG/CR-2945, Department of Nuclear Energy,
of stationary multi-variate random processes. Part II. Brookhaven National Laboratory, Prepared for Office of
J. Appl. Mech., 54 (1987) 681-7. Nuclear Regulatory Research, U.S. Nuclear Regulatory
22. Li, Y. & Kareem, A., Simulation of multivariate Commission, 1982.
nonstationary random processes by FFT. J. Engng 37. Hindy, A. & Novak, M., Pipeline response to random
Mech., ASME, 117 (1991) 1037-58. ground motion. J. Engng Mech., ASCE, 106 (1980) 339-
23. Li, Y. & Kareem, A., Simulation of multi-variate random 60.
processes: hybrid DFT and digital filtering approach. J. 38. Harichandran, R. S. & Wang, W., Effect of spatially
Engng Mech., ASCE, 119 (1993) 1078-98. varying seismic excitation on surface lifelines. Proc. Fourth
24. Ramadan, O. & Novak, M., Simulation of spatially U.S. National Conf. on Earthquake Engineering, Palm
incoherent random ground motions. J. Engng Mech., Springs, 20-24 May 1990, pp. 885-94 (Vol. 1).
ASCE, 119 (1993) 997-1016. 39. Deodatis, G. & Shinozuka, M., Auto-regressive model for
25. Deodatis, G., Simulation of ergodic multi-variate sto- non-stationary stochastic processes. J. Engng Mech., ASCE,
chastic processes. J. Engng Mech., ASCE, 122 (1996). 114 (1988) 1995-2012.
26. Gurley, K. & Kareem, A., On the analysis and simulation 40. International Conference of Building Officials, Uniform
of random processes utilizing higher order spectra and Building Code. Vol. 2, 1994.
wavelet transforms. Proc. Second International Conf. on 41. Jennings, P. C., Housner, G. W. & Tsai, N. C., Simulated
Computational Stochastic Mechanics, Athens, Greece, earthquake motions. Technical Report, Earthquake Engin-
1995, Balkema, pp. 315-24. eering Research Laboratory, California Institute of Tech-
27. Spanos, P.-T. D. & Zeldin, B., Random field simulation nology, 1968.
using wavelet bases. Proc. ICASP-7 Conf., Paris, France,
10-13 July 1995, pp. 1275-83.