See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267845411

Real Time Conditional Simulation of Earthquake Ground Motion

Article

CITATIONS READS

12 48

2 authors, including:

Tadanobu Sato
Southeast University (China)
156 PUBLICATIONS 1,557 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tadanobu Sato on 29 April 2015.

The user has requested enhancement of the downloaded file.

27
Earthquake Engineering
Earthquake
and Engineering
Engineering
Seismology
and Engineering Seismology, Vol. 1, No. 1 27
Volume 1, Number 1, September 1999, pp. 27–38

Real Time Conditional Simulation of

Earthquake Ground Motion
Tadanobu Sato 1) Hirofumi Imabayashi 2)

1) Professor, Disaster Prevention Research Institute, Kyoto University.

2) Engineer, Toda Cooperation.

ABSTRACT
A method is presented for simulating earthquake ground motions at
local-field points under the condition that the recorded motions are
specified at several locations and the stochastic characteristics of the field
also are designated. The simulated motions at observation points coincide
with observed records. The state equation expressing the spatial and
temporal field of earthquake ground motion compatible with the modeled
stochastic field is derived by using the autoregressive process. The
Kalman filtering technique is used to identify the best adaptive estimator of
a stochastic field from observations made at discrete spatial and temporal
points. To take into account the nonstationarity and inhomogeneity of
earthquake ground motions, the algorithm to identify the stochastic
characteristics of the field from observed earthquake motions is also
developed and introduced to the conditional simulation algorithm. This
analytical method is very promising for conditional simulation of
earthquake motion at unobserved locations.

INTRODUCTION ground motions based on the dense

strong motion arrays have been formed,
Earthquake ground motions vary in the stochastic characteristics of the
time and space. Various methods for space-time correlation of earthquake
simulating space-time correlated ground motions are modeled by the function of
motion have been proposed. The one the separation distance between two
most widely used is a probabilistic points, the traveling time of the wave, the
approach based on the cross-correlation frequency, and the wave number. Waves
function or cross-spectrum. The theory simulated by the random field theory,
of random fields and its application to however, are only sample waves.
digital simulation of earthquake ground Conditional simulation methods must
motions are well established [1]. As be devised that allows realization of a
substantial databases of earthquake sample field at an unobserved location
28 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

which satisfies the properties of a in which wt is the system noise with co-
stochastic field and is compatible with variance matrix Qt = E [ ( w t - w t ) ( w t - w t )T ]
measured values at observed locations. and w t the mean value of noise at time t.
To obtain this kind of simulation a When the posterior best estimator of Zt–1
method that used closed form solutions of
is given by Zˆ t -1 , the apriori best
conditional probability functions for
estimator of zt is calculated as
Fourier coefficients [2], nonstochastic
conditional simulation [3], and appli- ˆ
zt = - F Z t -1 + G w t (5)
cation of the Kriging method [4,5] has
been developed. In these analyses, as all
Defining the difference between zt–k and
time histories of propagating waves at
z t -k for arbitrary k
observation points must be known apriori,
are not applicable to the real-time ~
z t - k = z t -k - zt -k (6)
conditional simulation of earthquake
motions. The use of observed real-time
The covariance matrix of zt is given by
earthquake motions obtained from
existing array observation systems to
Mt = E [ ~
zt ~
zTt ] = F Pt -1 FT + G Qt GT (7)
simulate earthquake motions at several
unobserved locations on real-time is main
concern of this paper. in which Pt –1 is the system covariance
matrix at time t – 1. If the earthquake
motion is assumed to be a stochastic field
that is homogeneous in time and
SPACE-TIME EQUATION OF stationary in space, this matrix is defined
EARTHQUAKE MOTION as

The temporal and spatial fields of ~ ~

Pt -1 = E [ Z t -1 ZTt-1 ]
earthquake motion are assumed to be
expressed by a spatially correlated é R(0) R( -1) L R(1 - q ) ù
autoregressive process ê R(1) R(0) L R(2 - q ) úú
=ê (8)
ê M M O M ú
z t = - A1 z t -1 - A 2 z t - 2 - L - A q z t -q + G w t ê ú
ë R(q - 1) R(q - 2) L R(0) û
(1)
~
in which Z t -1 = Z t -1 - Zˆ t -1 , Zˆ t -1 is the
in which zt is a vector composed of n-
posterior best estimator of Zt –1, and R(p)
dimensional earthquake motion at time t
the cross correlation matrix with dimen-
z t = { g1(t ), g 2 (t ), L, gm (t ) } T sion (n ´ n) defined by
(2)
R( p ) = E [ ~
z t -k ~
zTt -l ] , (p = l - k) (9)
When we define the following vector

The coefficient matrices Ai of the auto-

Z t -1 = { zTt-1, zTt-2 , L, zTt -q } T (3)
regressive process can be calculated by
Eq. (1) can be rewritten use of the cross correlation function of zt.
Subtracting Eq. (5) from Eq. (1) and
z t = -F Z t -1 + G w t , F = [ A1, A 2 , L, A q ] taking its transverse then multiplying by
~
z t -k (k = 1, 2, …, q ) from the left;
(4)
 Sato, Imabayashi: Real time conditional simulation of earthquake ground motion 29

E[~
z t -k ~
zTt ] St = 0 (14)

= -E [ ~
z t -k ~
zTt-1 ] A1T - E [ ~
z t -k ~
zTt-2 ] A T2 The Kalman filtering technique [7] is used
to obtain the posterior best estimator of
-L- E [ ~
z t -k ~
zTt -q ] ATq (10)
earthquake motion at time t conditioned
by the observation equation. Because
Taking into account Eq. (9), Eq. (10) is
the apriori best estimator and its
transferred to the following simultaneous
covariance matrix are given by Eqs. (5)
equation to define the coefficient matrices
and (7), the posterior best estimator of zt
of the autoregressive process in Eq. (1)
and its covariance matrix Pt are

é R(0) R(-1) L R(1 - q ) ù é A1T ù éR(1)ù zˆ t = zt + K t ( y t - H zt ) (15)

ê R(1) ê Tú ê
R(0) L R(2 - q )úú ê A 2 ú = êR(2)ú
ú
-ê
ê M M O M ú ê M ú ê M ú Pt = Mt - K t H Mt (16)
ê ú ê Tú ê ú
ëR(q - 1) R(q - 2) L R(0) û ëê A q ûú ëR(q )û
(11) in which Kt is a Kalman gain at time t and
is defined as
Once the cross correlation matrix defined
by Eq. (9) and the observed time series of K t = Mt HT ( H Mt HT + St )-1 (17)
earthquake motion from time t – q to t – 1
are given, earthquake motion at time t Partitioning the matrix Mt into the
can be simulated by Eq. (1). observed and unobserved parts

é Mt ; m , m Mt ; m , n -m ù
Mt = ê ú (18)
CONDITIONAL SIMULATION OF ëMt ; n -m , m Mt ; n -m , n -m û
EARTHQUAKE MOTION
and substituting the H matrix given in Eq.
Consider the case in which the total (13) and Eq. (18) into Eqs. (16) and (17),
number of points for which the earth- the Kalman gain and covariance matrix Pt
quake ground motion needed is n, but yield
records of earthquake motion are
é I ù
obtained only at m (m < n) observation Kt = ê ú (19)
points as defined by êëMt ; n -m , m Mt-;1m , m úû

y t = { g1(t ), g 2 (t ), L, gm (t ) } (12) Pt =
é0 0 ù
The observation equation then is ê0 M T -1 ú
ë t ; n -m , n -m - Mt ; m , n -m M t; m , m M t ; m , n -m û
yt = [ H ] zt , H=[I 0] (13) (20)

in which H is the observation matrix with Substituting Eqs. (19) and (20) into Eq.
the dimension (m ´ n) and I the unit (15) the best posterior estimator of
matrix with the dimension (m ´ m). This earthquake motion is
is a special case in that the observation
vector yt and the vector zt to be estimated ì zˆ t ; m ü
zˆ t = í ý
have the same physical value [6]. The îzˆ t ; n -m þ
covariance matrix of the observation
ì yt ü
noise, St, is therefore assumed to be the =í T -1 ý (21)
z
î t ; n -m + M M ( y - z )
t; m þ
zero matrix t ; m , n -m t; m , m t
30 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

These results show that the posterior best in which fpn are mutually independent
estimators of earthquake motion at uniformly distributed variables over – p to
observation points are identical to the p and
observed values and that the components
of covariance matrix become zero. 2n p
wn = n × Dw = (26)
To obtain sample earthquake motion T
at an unobserved point we must add a
sample fluctuation to the posterior best in which T is the duration of earthquake
estimator of earthquake motion. The motion. H ip (wn ) Dw is obtained by fac-
sample field is simulated by toring the cross spectrum Sij (dij , wn)
z t ; n -m = zˆ t ; n -m + e (22) defined in the field as

in which e is the sample error function é S11(d11, wn )Dw L S1m (d ij , wn )Dw ù

ê ú
vector simulated by multiply-correlated ê M O M ú
process with a zero mean vector, and êSm1(d11, wn )Dw L Smm (d ij , wn )Dwú
ë û
covariance Pt;n–m,n–m being defined by
é H11
*
(wn ) Dw L 0 ù
ê ú
Pt ; n -m , n -m =ê M O M ú
ê H * ( w ) Dw L H * ( w ) Dw ú
= E [ (z t ; n -m - zˆ t ; n -m ) ( z t ; n -m - zˆ t ; n -m )T ] ë m1 n mm n
û

= Mt ; n -m , n -m - MTt ; m , n -m Mt-;1m , m Mt ; m , n -m (23) éH11(wn ) Dw L H m1( wn ) Dw ù

ê ú
×ê M O M ú (27)
By factoring matrix Pt;n–m,n–m the error ê 0 L H mm ( wn ) Dw ú
ë û
sample function is

e = Lt N , Pt ; n -m , n -m = L t LTt (24) and qip is calculated by

ìï Im [H ip ( wn ) Dw ] üï
in which N is an n – m dimensional qip = tan -1 í (28)
ý
sample noise vector generated by the ïî Re [H ip ( wn ) Dw ] þï
standard normal distribution density
function. The cross spectrum is assumed to be
given by the equation

STOCHASTIC MODEL OF S ij (d ij , wn ) Dw

EARTHQUAKE MOTION æ - i wn d ij
= ST ( wn ) Dw × exp çç
ö æ - awn |d ij |ö
÷ × exp ç ÷
c ÷ ç 2p c ÷
è ø è ø
We assume that earthquake motion at (29)
point i and time t is expressed by the
cross-correlated m stochastic process ui(t ) in which dij is the relative distance
(i = 1, 2, …, m) [1] as between points i and j, c the propagating
wave speed, and ST(wn) the discretized
ui (t ) homogeneous and stationary power
i N spectrum density function at the circular
= åå H ip ( wn Dw ) 2 cos { wn t + qip + f pn } frequency of wn and is defined by the
p =1 n =1 following equation, provided the discrete
(25) Fourier cos and sin series amplitudes of
 Sato, Imabayashi: Real time conditional simulation of earthquake ground motion 31

an and bn are given at a certain point in

the concerned field,

an2 + bn2
ST (wn ) Dw = (30)
2

To make the conditional simulation of

earthquake ground motions, the cross
correlation function must be defined in
the field obtained by inverse Fourier
transform of the cross spectrum;

1 ¥
R ij (dij , t) =
2 ò -¥
Sij (dij , w) exp(i wt) dw
N
1 æ -a wn |dij |ö
=
2 å (a
n =1
2
n + bn2 ) exp çç
è 2p c
÷
÷
ø
ìï æ dij ö ïü
× cos í wn çç t - ÷ý (31)
ïî è c ÷ø ïþ

in which t is the relative time delay of

Fig. 1 Sample earthquake ground mo-
earthquake motion between points i and j.
tions simulated by Eq. (25)
Equation (31) is intended only to
illustrate the procedure, but it does show
some characteristics of the empirical
correlation functions obtained from dense
accelerograph array data. Figure 1
shows the simulated sample fields of a
one-dimensional wave propagation
with velocity of c =1000m/s and a =
0.02 from Eq. (25). Each point is
separated by 100m. In the simulation
earthquake ground motion at the left edge
(a) time history
point is assumed to be given and its
discrete Fourier spectrum amplitude
calculated (Fig. 2), after which the
waveforms of the remaining points can be
simulated.

IDENTIFICATION OF
AUTOREGRESSIVE PROCESS

Identification of Observed Parts

(b) Fourier spectrum
So far we have assumed that the
temporal and spatial stochastic charac- Fig. 2 Input earthquake motion into the
teristics are defined in the concerned time space field
32 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

field apriori. Because of nonstationarity ak , t = [ I ] ak , t -1 gk (t ) = Ht ak , t + vk , t (36)

and nonhomogeneity of earthquake
ground motions these stochastic charac- in which
teristics must be identified from observed Ht = - { zTt-1, zTt-2 , L, zTt -q }
motions.
When the time history of earthquake The algorithm of the Kalman filtering for
ground motions at all observed stations Eq. (36) are formulated as developed in [8]
are assumed to be given, Eq. (4) is
(1) Initial conditions:
rewritten in the form of observation
equation ˆ k , 0 = a0 ,
a Pk , 0 = P0 , R = R0, t =0

é A1T ù (2) Time step increase: t = t + 1

ê Tú
A
{ g1(t ), g 2 (t ), L, gm (t ) } = Ht ê 2 ú + v t (32) (3) Predicted state estimation:
ê M ú
ê Tú ˆ k , t -1,
ak , t = a Mk , t = Pk , t -1
ëê A q ûú t
(4) Kalman gain:
in which vt is the observation noise and Kk , t = Mk , t HTt (Ht Mk , t HTt + R )-1
the observation matrix is defined by
(5) Corrected state estimation:
Ht = ZTt-1 = - { zTt-1, zTt-2 , L, zTt -q } (33)
ˆ k , 0 = ak , t + Kk , t (gk (t ) - Ht ak , t )
a

in which zt is a vector composed of m- (6) Estimation of error covariance matrix:

dimensional observed earthquake motion
Pk , t = Mk , t - Kk , t Ht Mk , t
at time t ; zt = { g1(t ), g2(t ), …, gm(t ) }. The
state space description of the coefficient
Through the steps from (1) to (6) the
matrices of the autoregressive process is
posterior best estimator of coefficient
given by
matrices of autoregressive process is
é A1T ù é A1T ù calculated.
ê Tú ê Tú
êA 2 ú = [ I ] êA 2 ú (34) Identification of Unobserved Parts
ê M ú ê M ú
ê Tú ê Tú To identify the coefficient matrices of
êë A q úû êë A q úû t -1
autoregressive process related to
unobserved parts the cross-correlation
To apply the Kalman filtering
function of the concerned field must be
algorithms for identifying autoregressive
given. Because we assume the form
process the coefficient matrix appeared in
expressed by Eq. (31) for cross-correlation
Eqs. (32) and (34) is rearranged as a
function, identification of the parameters
matrix composed of column vectors ak,t (k
a and c from observed earthquake ground
= 1, 2, …, m)
motions are enough to define the cross-
é A1T ù correlation function. Once this function
ê Tú identified Eq. (11) is rewritten as
ê A 2 ú = [a a 2, t L am , t ] (35)
ê M ú 1, t
ê Tú RA=r (37)
êë A q úû
in which R, A and r are the matrices
The state space and observation equa- rearranged their components and
tions can be rewritten partitioned into the observed and
 Sato, Imabayashi: Real time conditional simulation of earthquake ground motion 33

unobserved parts as regressive process are determined by

solving Eq. (11) and assigning a = 0.02
é R oo R ou ù é A oo A ou ù ér 0 ù and c = 1000m/sec for Eq. (31). The
R = ê uo ú A = ê uo ú r = ê uú
ëR R uu û ëA A uu û ër û results shown in Fig. 3 are in good
(38) agreement with the sample motions
shown in Fig. 1.
The partitioned matrix Aoo can be
When the stochastic characteristics,
identified by using observed earthquake
i.e., the cross-correlation function of the
ground motions as mentioned in the
field, and earthquake motions at several
previous section. The components of
points in Fig. 1 are given up to time t – 1
auto-regressive matrices related to
as observed motions, the motions at the
unobserved points can be determined by
intermediate points at time t are inter-
solving following equations
polated from Eq. (21). The results shown
in Fig. 4(a) agree well with the sample
m ´q (n -m )´q ü
åR oo
ij A oo
jl + åR ou
ij A uo o
jl = ril ï
ï
motions shown in Fig. 1. Figure 4(b) is
j =1 j =1
ý (39) the interpolated earthquake ground
m ´q (n -m )´q

å uo
R kj A oo
jl + å uu
R kj A uo uï
jl = rkl ï
motions for the case of a = 0.1 and c =
1500m/s. As the value of a increases
j =1 j =1 þ
the distortion of propagating wave be-
m ´q (n -m )´q ü comes large. Even such a case the
å
j =1
R oo oo
ij A js + åRj =1
ou
ij A uu o
js = ris ï
ï developed interpolation algorithm works
m ´q (n -m )´q ý (40)
well.
å uo
R kj A ou
js + å uu
R kj A uu
js = r u ï
ks ï
j =1 j =1 þ

in which (l = 1, 2, …, m), (s = m + 1, m +
2, …, n), (i = 1, 2, …, m ´ q ) and (k = 1,
2, …, (n – m) ´ q ). Solving Eq. (40) the
unknown values of A ou uu
jl and A jl can be
determined. As can be seen from Eq. (39)
the number of unknown values is (n – m) ´
q but the number of equation is n ´ q .
We can apply the least square estimation
algorithm to define the most appropriate
values of A uo
jl .

NUMERICAL EXAMPLES

To verify the applicability of the Fig. 3 Earthquake ground motions

autoregressive process for simulating simulated by the autoregressive
multi-correlated earthquake motions, we process defined by Eq. (4) with no
assume that the sample earthquake system noise. The coefficient
motions shown in Fig. 1 are observed at matrix of AR process are obtained
all points up to time t – 1 and that by solving Eq. (11) provided the
earthquake motions at time t are cross-correlation function of the
simulated by Eq. (4) with no system noise. concerned field is given by Eq.
The coefficient matrices of the auto- (31)
34 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

(a) the case of c = 1000m/s and a = 0.02 ´ 2p

(b) the case of c = 1500m/s and a = 0.1 ´ 2p

Fig. 4 Interpolated earthquake ground motions (thin lines) when earthquake motions
(thick lines) at points expressed by · are given. For the comparison the
sample earthquake motions simulated by Eq. (25) are shown by broken lines.

Increasing the number of observation sider a realistic two-dimensional field,

points improves the accuracy of the in which an unobserved point is lo-
posterior estimation of earthquake cated at the center of the field and one
motion at an unobserved point. Con- observation point (No. 4) is added to the
 Sato, Imabayashi: Real time conditional simulation of earthquake ground motion 35

already distributed three observation quake motion comes close to the exact
points (No. 1, 2, 3) as shown in Fig. 5. time history as can be seen in Fig. 6(c).
The side length of each grid is 100m.
Earthquake motion propagating in the
45° direction to horizontal axis with a
wave velocity of 1000m/sec and a = 0.02
is assumed, and sample motions are
simulated at five points including
unobserved point (Fig. 5). The resulting
time history for the unobserved point
becomes the exact time history when the
posterior estimation of earthquake
motion is introduced. The interpolated
motion at the unobserved point obtained
from records providing three observation
points is shown in Fig. 6(a) and the time
history that takes into account the record
for a newly added observation point is in
Fig. 6(b). As the number of observation Fig. 5 Location of observation points
points increases, the estimated earth- and an unobserved point

(c) comparison between the

(a) the case of three points (b) the case of four points exact ground motion (1)
observations (·) observations (·) and interpolated ones
(2: from case (a),
3: from case (b))

Fig. 6 Increase in interpolation accuracy of earthquake ground motion at an

unobserved point as the number of observation points increases
36 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

When the stochastic characteristics observed points. For conditional inter-

of the field are not given apriori we must polation of earthquake ground motions
identify the cross-correlation function the components of the coefficient
and the coefficient matrices of the matrices of autoregressive process related
autoregressive process from the observed to unobserved points must be calculated
earthquake motions. To demonstrate by Eqs. (39) and (40). The accuracy of
the application of the algorithm developed calculated results is shown in Fig. 10 in
previously, sample earthquake motions which the abscissa is the exact values
are simulated and we select three of them and the ordinate the calculated values.
as observed earthquake motions (Fig. 7). Because all calculated components lie on
Examples of identified cross-correlation the line with 45° to abscissa, we can
functions and components of coefficient confirm the algorithm to identify the
matrices of autoregressive process are autoregressive process only from
shown in Figs. 8 and 9, respectively. In observed data. Once the coefficient
Fig. 8 the thick line is the identified matrices of autoregressive process are
correlation function and thin line is the identified we can interpolate earthquake
function used to simulate sample motions at unobserved points from
earthquake motions. Figure 9 is time observed motions. Figure 11 shows the
histories of the identified autoregressive ability of the proposed method.
coefficient matrix components related to

Fig. 7 Sample earthquake motions (left) and earthquake motions assumed to be

observed (right)

Fig. 8 Examples of identified cross-correlation function (thick line) from ground

motions and one used to simulate sample earthquake motions (thin line)
 Sato, Imabayashi: Real time conditional simulation of earthquake ground motion 37

Fig. 9 Examples of identifying process of AR coefficient matrix components

Fig. 10 Comparison of the identified AR Fig. 11 Interpolated earthquake ground

coefficient matrix components motions (thin line) by using
with the exact ones identified AR matrix components
when earthquake motions are
assumed to be given at three
points expressed by ·
 38 Earthquake Engineering and Engineering Seismology, Vol. 1, No. 1

CONCLUSION 3. Kameda H. and Morikawa H. (1992).

“An interpolating stochastic process for
We have proposed a stochastic simulation of conditional random
interpolation method for the conditional fields,” Probabilistic Engineering
simulation of earthquake motion. It Mechanics, Vol. 7, No. 4, pp. 243–254.
combines the methodology of the multi- 4. Vanmarck E.H. and Fenton G.A. (1991).
autoregressive process for simulating “Conditional simulation of local fields of
earthquake motions and the Kalman earthquake ground motion,” Journal of
filtering technique in which the state Structural Safety, 10, pp. 247–264.
vector, including unknown parameter, is 5. Hoshiya M. and Murayama O. (1993).
updated based on observations. This “Stochastic interpolation of earthquake
method provides the best estimator of wave propagation,” Proceedings of 6th
earthquake motions at unobserved points International Conference on Structural
that coincide with sample motions at Safety and Reliability, Innsbruck,
August.
observed points. Its application to
earthquake ground motion simulation is 6. Hoshiya M. and Yoshida I. (1993).
illustrated by several examples. “Identification of conditional stochastic
Gaussian field,” Submitted to ASCE,
Journal of Engineering Mechanics
Division, dated September.
REFERENCES 7. Kalman R.E. (1960). “A new approach
1. Vanmarcke E.H. (1983). Random to linear filtering and prediction
Fields, MIT Press, Massachusetts. problems,” Journal of Basic
Engineering, Vol. 82, No. 1, pp. 35–45.
2. Kawakami H. (1992). “Simulation of
space-time variation of earthquake 8. Sato T. and Toki K. (1990). “Active
ground motion using a recorded time control of seismic response of
history,” Proceedings of Japan Society of structures,” Journal of Intelligent
Civil Engineers, No. 441/I-18, pp. 167– Material Systems and Structures, Vol. 1,
175 (in Japanese). No. 4, pp. 447–475.

View publication stats