You are on page 1of 12

Simulation of Nonstationary Stochastic Processes

by Spectral Representation
Jianwen Liang, M.ASCE1; Samit Ray Chaudhuri2; and Masanobu Shinozuka, Hon.M.ASCE3

Abstract: This paper presents a rigorous derivation of a previously known formula for simulation of one-dimensional, univariate,
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

nonstationary stochastic processes integrating Priestly’s evolutionary spectral representation theory. Applying this formula, sample func-
tions can be generated with great computational efficiency. The simulated stochastic process is asymptotically Gaussian as the number of
terms tends to infinity. This paper shows that 共1兲 these sample functions accurately reflect the prescribed probabilistic characteristics of the
stochastic process when the number of terms in the cosine series is large, i.e., the ensemble averaged evolutionary power spectral density
function 共PSDF兲 or autocorrelation function approaches the corresponding target function as the sample size increases, and 共2兲 the
simulation formula, under certain conditions, can be reduced to that for nonstationary white noise process or Shinozuka’s spectral
representation of stationary process. In addition to derivation of simulation formula, three methods are developed in this paper to estimate
the evolutionary PSDF of a given time-history data by means of the short-time Fourier transform 共STFT兲, the wavelet transform 共WT兲,
and the Hilbert-Huang transform 共HHT兲. A comparison of the PSDF of the well-known El Centro earthquake record estimated by these
methods shows that the STFT and the WT give similar results, whereas the HHT gives more concentrated energy at certain frequencies.
Effectiveness of the proposed simulation formula for nonstationary sample functions is demonstrated by simulating time histories from the
estimated evolutionary PSDFs. Mean acceleration spectrum obtained by averaging the spectra of generated time histories are then
presented and compared with the target spectrum to demonstrate the usefulness of this method.
DOI: 10.1061/共ASCE兲0733-9399共2007兲133:6共616兲
CE Database subject headings: Stationary processes; Simulation; Stochastic process; Fourier transform; Ground motion; Power
spectral density.

Introduction this purpose. Yang 共1972, 1973兲 showed that the fast Fourier
transform 共FFT兲 technique can be used to dramatically improve
Monte Carlo simulation methods play an important role in the the computational efficiency of the spectral representation algo-
response analysis of civil engineering structures and systems with rithm, and also proposed a formula to simulate random envelop
stochastic properties and/or subjected to stochastic loadings. For processes. Shinozuka 共1974兲 extended the application of the FFT
the purpose of the Monte Carlo analysis, generation of sample technique to multidimensional cases. In recent years, the spectral
functions of the stochastic processes involved in the problem representation method has been extended in various ways, e.g.,
must accurately provide the probabilistic characteristics of these to simulate stochastic waves 共Deodatis and Shinozuka 1989兲,
processes. Among the various methods that have been developed non-Gaussian stochastic fields 共Yamazaki and Shinozuka 1988兲,
to generate such sample functions, the spectral representation spatially incoherent multidimensional and multivariate random
method pioneered by Shinozuka 共Shinozuka and Jan 1972; processes and fields 共Ramadan and Novak 1993兲, multivariate
Shinozuka 1972兲 appears to be most versatile and widely used ergodic stochastic processes 共Deodatis 1996b兲, multivariate
today. Shinozuka 共1972兲 developed an analytical expression for nonstationary stochastic processes 共Deodatis 1996a兲, and non-
generating sample functions of the envelope of a stochastic pro- Gaussian multidimensional multivariate fields 共Popescu et al.
cess observing the usefulness of the Hilbert transformation for 1998兲, etc.
From the rich bibliography related to the various methods for
1 the simulation of nonstationary stochastic processes, representa-
Dept. of Civil Engineering, Tianjin Univ., Tianjin 300072, China.
E-mail: jwliang@tj.cnuninet.net
tive studies are the following: Shinozuka and Sato 共1967兲 by fil-
2
Postdoctoral Researcher, Dept. of Civil Engineering, Univ. of tered Gaussian white noise process models, Shinozuka and Jan
California, Irvine, CA 92697-2175. E-mail: samit@uci.edu 共1972兲 by spectral representation method with amplitude modu-
3
Distinguished Professor and Chair, Dept. of Civil and Environmental lation, Ohsaki 共1979兲 by phase difference method, Deodatis and
Engineering, Univ. of California, Irvine, CA 92697-2175. E-mail: Shinozuka 共1988兲 by autoregressive method, Li and Kareem
shino@uci.edu 共1991兲 by fast Fourier transform technique, Ramadan and Novak
Note. Associate Editor: George Deodatis. Discussion open until 共1993兲 by spectral representation method with amplitude modu-
November 1, 2007. Separate discussions must be submitted for individual lation, and Deodatis 共1996a兲 by spectral representation method
papers. To extend the closing date by one month, a written request must
with amplitude and frequency modulation. It should be noted that
be filed with the ASCE Managing Editor. The manuscript for this paper
was submitted for review and possible publication on March 14, 2002; the modulated envelope functions used for making stationary sto-
approved on December 9, 2006. This paper is part of the Journal of chastic process nonstationary in the previous references are usu-
Engineering Mechanics, Vol. 133, No. 6, June 1, 2007. ©ASCE, ISSN ally based on the analytically expedient postulation derived from
0733-9399/2007/6-616–627/$25.00. general observations of nonstationary time history data. However,

616 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


the estimation of evolutionary power spectral density function dH共␻兲 = S共␻兲d␻ 共4兲
共PSDF兲 from a particular nonstationary record 共e.g., an earth-
quake ground motion acceleration兲 is necessary to represent spec- where S共␻兲⫽PSDF of the associated stationary stochastic pro-
tral properties of the nonstationary process. cess, and
This paper presents a rigorous derivation of the computation- ␧关dZ*共␻兲dZ共␻⬘兲兴 = 0, ␻ ⫽ ␻⬘ 共5兲
ally efficient cosine series formula presented by Shinozuka and
Jan 共Shinozuka and Jan 1972兲 for generation of sample functions with an asterisk indicating the complex conjugate. Then, one can
of one-dimensional and univariate nonstationary stochastic pro- also show that


cesses. The spectral representation method, which is used to de- ⬁
rive this simulation formula, integrates the theory of evolutionary ␧关f 0共t兲兴 = A共t,␻兲ei␻t␧关dZ共␻兲兴 = 0 共6兲
process introduced and promoted by Priestly 共1965, 1967兲. The −⬁
simulated stochastic process is asymptotically Gaussian as the
and
number of terms tends to infinity. This paper shows that 共1兲 the
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

sample functions generated by the present method accurately re- R f 0 f 0共t,t + ␶兲


flect the prescribed probabilistic characteristics of the stochastic
process when the number of the terms in the cosine series is large, = ␧关f *0共t兲f 0共t + ␶兲兴

冕冕
i.e., the ensemble averaged evolutionary PSDF and autocorrela- ⬁ ⬁
tion function approach the corresponding target function as the = A*共t,␻兲A共t + ␶,␻⬘兲e−i␻tei␻⬘共t+␶兲␧关dZ*共␻兲dZ共␻⬘兲兴
sample size increases, and 共2兲 the simulation formula, under cer- −⬁ −⬁


tain conditions, can be reduced to that for nonstationary white ⬁
noise process or Shinozuka’s stationary process 共Shinozuka and = A*共t,␻兲A共t + ␶,␻兲ei␻␶␧关兩dZ共␻兲兩2兴 共7兲
Jan 1972; Shinozuka 1972兲. −⬁
In the later part, three methods are developed to estimate the
evolutionary PSDF of a given time-history data by using the For ␶ = 0, Eq. 共7兲 can be written as
short-time Fourier transform 共STFT兲, the wavelet transform


共WT兲, and the Hilbert-Huang transform 共HHT兲. Three evolution- ␧关f 20共t兲兴 = 兩A共t,␻兲兩2S共␻兲d␻ 共8兲
ary PSDFs thus developed maintain the same total energy pos- −⬁
sessed by the time history data. In fact, N-S component of the where
well-known 1940 El Centro earthquake record is used as an ex-
ample of such a time history and the evolutionary PSDFs derived S f 0 f 0共t,␻兲 = 兩A共t,␻兲兩2S共␻兲 共9兲
from the record by using these three methods are compared. Ef-
fectiveness of the simulation formula proposed here is demon- is the 共two-sided兲 evolutionary PSDF 共Priestley 1965, 1967兲. It
strated by generating time histories from the three evolutionary may be mentioned here that if f 0共t兲 is stationary, it admits the
PSDFs, while using the same random number sequences for them. following well-known spectral representation 共e.g., Cramer and
Further, 20 sample functions are generated for each of the three Leadbetter 1967兲:
evolutionary PSDFs and the mean of 5% damped acceleration

+⬁
response spectra of these simulated motions are computed and f 0共t兲 = ei␻tdZ共␻兲 共10兲
compared with the target spectrum. −⬁

As A共t , ␻兲 is complex, Eq. 共1兲 can be written as

冕 冕
+⬁ +⬁
Spectral Representation of Nonstationary
Stochastic Processes f *0共t兲 = A*共t,␻兲e−i␻tdZ*共␻兲 = A*共t,− ␻兲ei␻tdZ*共− ␻兲
−⬁ −⬁

Based on evolutionary spectral representation 共Priestley 1965, 共11兲


1967兲, a zero mean nonstationary stochastic process f 0共t兲 共com- But f 0共t兲 is real valued if and only if f 0共t兲 = f *0共t兲, i.e., if and only
plex valued in general and defined in − ⬁ ⬍ t ⬍ + ⬁兲 admits if both A共t , ␻兲 and dZ共␻兲 are Hermitian for all ␻
the representation
A共t,␻兲 = A*共t,− ␻兲 共12兲


+⬁
f 0共t兲 = A共t,␻兲ei␻tdZ共␻兲 共1兲
dZ共␻兲 = dZ*共− ␻兲 共13兲
−⬁

where A共t , ␻兲⫽deterministic modulating function of both t and Hence, introducing


␻ ; Z共␻兲⫽spectral process with orthogonal increments and has a dU0共␻兲 = Re关dZ共␻兲兴, dV0共␻兲 = − Im关dZ共␻兲兴 共14兲
distribution function in the interval 共− ⬁ , + ⬁ 兲. The properties of
Z共␻兲 can be expressed as it can be shown that
dU0共␻兲 = dU0共− ␻兲, dV0共␻兲 = − dV0共− ␻兲 for all ␻ 共15兲
␧关dZ共␻兲兴 = 0 共2兲
In Eq. 共14兲, Re is used to denote the real part and Im is used to
denote the imaginary part. Modulating function A共t , ␻兲 can be
␧关兩dZ共␻兲兩 兴 = dH共␻兲
2
共3兲 written in the polar form as follows:
where ␧ represents the ensemble average and H共␻兲⫽integrated
A共t,␻兲 = ␣共t,␻兲 + i␤共t,␻兲 = 兩A共t,␻兲兩ei␪共t,␻兲 共16兲
spectrum.
Now, if H共␻兲 is differentiable, one can write with

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 617

J. Eng. Mech. 2007.133:616-627.


兩A共t,␻兲兩 = 共␣2共t,␻兲 + ␤2共t,␻兲兲1/2 共17兲 ␧关dUt共␻兲兴 = ␣共t,␻兲␧关dU共␻兲兴 + ␤共t,␻兲␧关dV共␻兲兴 = 0
and
␧关dVt共␻兲兴 = 0, 共following the previous steps兲 共31兲
␪共t,␻兲 = tan−1共Im共A共t,␻兲兲/Re共A共t,␻兲兲兲 = tan−1共␤共t,␻兲/␣共t,␻兲兲
共18兲 ␧关dU2t 共␻兲兴 = ␣2共t,␻兲␧关dU2共␻兲兴 + ␤2共t,␻兲␧关dV2共␻兲兴

Then, with the aid of Eqs. 共13兲–共15兲, the nonstationary process + 2␣共t,␻兲␤共t,␻兲␧关dU共␻兲dV共␻兲兴
f 0共t兲 defined in Eq. 共1兲 can be written as = 共␣2共t,␻兲 + ␤2共t,␻兲兲2S共␻兲d␻



= 2兩A共t,␻兲兩2S共␻兲d␻
f 0共t兲 = 关␣共t,␻兲 + i␤共t,␻兲兴关cos ␻t + i sin ␻t兴
−⬁ = 2S f 0 f 0共t,␻兲d␻
⫻关dU0共␻兲 − idV0共␻兲兴
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.


⬁ ␧关dV2t 共␻兲兴 = 2兩A共t,␻兲兩2S共␻兲d␻
= 关cos ␻t + i sin ␻t兴兵关␣共t,␻兲dU0共␻兲 + ␤共t,␻兲dV0共␻兲兴 = 2S f 0 f 0共t,␻兲d␻, 共following the previous steps兲
−⬁
共32兲
+ i关␤共t,␻兲dU0共␻兲 − ␣共t,␻兲dV0共␻兲兴其



␧关dUt共␻兲dUt共␻⬘兲兴 = ␣共t,␻兲␣共t,␻⬘兲␧关dU共␻兲dU共␻⬘兲兴
= cos ␻t关␣共t,␻兲dU0共␻兲 + ␤共t,␻兲dV0共␻兲兴
−⬁ + ␤共t,␻兲␤共t,␻⬘兲␧关dV共␻兲dV共␻⬘兲兴
− sin ␻t关␤共t,␻兲dU0共␻兲 − ␣共t,␻兲dV0共␻兲兴 共19兲 + ␣共t,␻兲␤共t,␻⬘兲␧关dU共␻兲dV共␻⬘兲兴
Eq. 共19兲 can further be written as + ␤共t,␻兲␣共t,␻⬘兲␧关dV共␻兲dU共␻⬘兲兴 = 0



f 0共t兲 = cos ␻t关␣共t,␻兲dU共␻兲 + ␤共t,␻兲dV共␻兲兴 ␧关dVt共␻兲dVt共␻⬘兲兴 = 0, 共following the previous steps兲
0 共33兲
− sin ␻t关␤共t,␻兲dU共␻兲 − ␣共t,␻兲dV共␻兲兴 共20兲 and
where ␧关dUt共␻兲dVt共␻⬘兲兴 = ␣共t,␻兲␤共t,␻⬘兲␧关dU共␻兲dU共␻⬘兲兴
dU共␻兲 = dU0共␻兲 + dU0共− ␻兲 = 2dU0共␻兲 for ␻ ⬎ 0 共21兲 − ␤共t,␻兲␣共t,␻⬘兲␧关dV共␻兲dV共␻⬘兲兴
− ␣共t,␻兲␣共t,␻⬘兲␧关dU共␻兲dV共␻⬘兲兴
dU共0兲 = dU0共0兲 共22兲
+ ␤共t,␻兲␤共t,␻⬘兲␧关dV共␻兲dU共␻⬘兲兴 = 0
and
共34兲
dV共␻兲 = dV0共␻兲 − dV0共− ␻兲 = 2dV0共␻兲 for ␻ ⱖ 0 共23兲
Therefore, Ut共␻兲 and Vt共␻兲 are each orthogonal, and in addition,
On simplification, Eq. 共20兲 can further be expressed as they are mutually orthogonal, constituting the spectral processes


⬁ of nonstationary stochastic process f 0共t兲. It may be noted that
f 0共t兲 = cos ␻t dUt共␻兲 − sin ␻t dVt共␻兲 共24兲 when Ut共␻兲 and Vt共␻兲 are time independent, i.e., dUt共␻兲
0 = dU共␻兲 and dVt共␻兲 = dV共␻兲, Eq. 共24兲 is then reduced to the spec-
tral representation of a stationary stochastic process 共Cramer and
where
Leadbetter 1967兲. Now, Eq. 共24兲 can be rewritten in the following
dUt共␻兲 = ␣共t,␻兲dU共␻兲 − ␤共t,␻兲dV共␻兲 共25兲 discrete form:

dVt共␻兲 = ␤共t,␻兲dU共␻兲 − ␣共t,␻兲dV共␻兲 共26兲 f 0共t兲 = 兺
k=0
关cos共␻kt兲dUt共␻k兲 − sin共␻kt兲dVt共␻k兲兴 共35兲
As U共␻兲 and V共␻兲⫽real and imaginary part, respectively, of the
spectral process Z共␻兲 defined for nonnegative ␻, they admit the where
following conditions 共Cramer and Leadbetter 1967兲:
␻k = k⌬␻ 共36兲
␧关dU共␻兲兴 = ␧关dV共␻兲兴 = 0, ␻艌0 共27兲
with sufficiently small but finite ⌬␻ such that Eq. 共35兲 can be
used for Eq. 共24兲. If dUt共␻k兲 and dVt共␻k兲 are defined as
␧关dU 共␻兲兴 = ␧关dV 共␻兲兴 = 2S共␻兲d␻,
2 2
␻艌0 共28兲
dUt共␻k兲 = 冑2Bt,k cos ⌽k 共37兲
␧关dU共␻兲dU共␻⬘兲兴 = ␧关dV共␻兲dV共␻⬘兲兴 = 0, ␻ ⫽ ␻ ⬘, ␻,␻⬘ 艌 0
共29兲 dVt共␻k兲 = 冑2Bt,k sin ⌽k 共38兲
and where

␧关dU共␻兲dV共␻⬘兲兴 = 0, ␻,␻⬘ 艌 0 共30兲 Bt,k = 关2S f 0 f 0共t,␻k兲⌬␻兴1/2 共39兲

Then, it can be easily verified from their definitions that for ␻, and ⌽k⫽independent random phase angles uniformly distributed
␻⬘ ⱖ 0, Ut共␻兲 and Vt共␻兲 satisfy following equations: in the range 关0 , 2␲兴, then it can be shown that Eqs. 共31兲–共34兲

618 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


are satisfied. Indeed, the following expressions can be written for ␧关cos ⌽k sin ⌽⬘k 兴 = ␧关cos ⌽k兴␧关sin ⌽⬘k 兴 = 0 for k ⫽ k⬘ 共47兲
Eq. 共31兲
When k = k⬘, the term ␧关cos ⌽k sin ⌽⬘k 兴 can be written as
␧关dUt共␻k兲兴 = ␧关冑2Bt,kcos ⌽k兴 = 冑2Bt,k␧关cos ⌽k兴

= 冑2Bt,k 冕
+⬁
cos ␾k p⌽k共␾k兲d␾k 共40兲
␧关cos ⌽k sin ⌽⬘k 兴 = ␧关cos ⌽k sin ⌽k兴 = ␧ 冋 1
2
sin 2⌽k 册

−⬁ 2␲
1
= sin 2␾k d␾k = 0 for k = k⬘ 共48兲
where p⌽k共␾k兲⫽probability density function of random phase 0 2␲
angle ⌽k given by


Combining Eqs. 共47兲 and 共48兲, the following result is established:
1/2␲, 0 ⱕ ␾k ⱕ 2␲
p⌽k共␾k兲 = 共41兲 ␧关dUt共␻k兲dVt共␻⬘k 兲兴 = 0 for k ⫽ k⬘ and k = k⬘ 共49兲
0 otherwise
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

Therefore, all the requirements 关Eqs. 共31兲–共34兲兴 imposed


Now, Eq. 共40兲 can be written as
on dUt共␻k兲 and dVt共␻k兲 are satisfied by the expressions given

冕 in Eqs. 共37兲 and 共38兲. Then, substituting Eqs. 共37兲 and 共38兲 into
2␲
␧关dUt共␻k兲兴 = 冑2Bt,k
1
cos ␾k d␾k = 0 共42兲 Eq. 共35兲, the following series representation is obtained:
0 2␲

and in exactly the same way it can be shown that ␧关dVt共␻k兲兴 = 0. f 0共t兲 = 兺 兵cos共␻kt兲冑2关2S f f 共t,␻k兲⌬␻兴1/2 cos ⌽k
The requirement described in Eq. 共32兲 can be expressed as k=0
0 0

␧关dU2t 共␻k兲兴 = ␧关2B2t,k cos2 ⌽k兴 − sin共␻kt兲冑2关2S f 0 f 0共t,␻k兲⌬␻兴1/2 sin ⌽k其


= 2B2t,k␧关cos2 ⌽k兴 ⬁

= 冑2 兺

+⬁ 关2S f f 共t,␻k兲⌬␻兴1/2 cos共␻kt + ⌽k兲
0 0
共50兲
= 2B2t,k cos ␾k p⌽k共␾k兲d␾k
2 k=0

−⬁ Eq. 共50兲 can also be written as


2␲
1 1 ⬁
= 2B2t,k 共1 + cos 2␾k兲 d␾k
0 2 2␲ f 0共t兲 = 冑2 兺 关2兩A共t,␻k兲兩2S共␻k兲⌬␻兴1/2 cos共␻kt + ⌽k兲 共51兲
k=0
= B2t,k
where S f 0 f 0共t , ␻k兲⌬␻ = 兩A共t , ␻k兲兩2S共␻k兲⌬␻, which represents the
= 2S f 0 f 0共t,␻k兲⌬␻ 共43兲 discretization of evolutionary PSDF introduced by Priestley
共1965, 1967兲 as shown in Eq. 共9兲.
and in exactly the same way it can be shown that ␧关dV2t 共␻k兲兴 It is important to note that the derivations of Eqs. 共50兲 and 共51兲
= 2S f 0 f 0共t , ␻k兲⌬␻. initially retained the terms involving ␪共t , ␻兲, but they did not
The condition described in Eq. 共33兲 can be written as appear in the final result. In the present study, each of dU0共␻兲 and
dV0共␻兲 关or, dU共␻兲 and dV共␻兲兴 is considered to represent a white
␧关dUt共␻k兲dUt共␻⬘k 兲兴 = ␧关2Bt,kB⬘t,k cos ⌽k cos ⌽⬘k 兴
noise with S共␻兲 = 1 with appropriate dimension. Therefore, if it is
= 2Bt,kB⬘t,k␧关cos ⌽k cos ⌽⬘k 兴 estimated for example, by the STFT 共Flanagan 1972兲, A共t , ␻兲 is
nothing but the evolutionary PSDF.
= Bt,kB⬘t,k␧关cos ⌽k兴␧关cos ⌽⬘k 兴 共44兲
and the last equality in Eq. 共44兲 is valid as ⌽k and ⌽⬘k are inde-
pendent random phase angles for k ⫽ k⬘. Eventually, Eq. 共44兲 is Simulation of Nonstationary Stochastic Processes
written as

冕 Simulation Formula
2␲
1
␧关dUt共␻k兲dUt共␻⬘k 兲兴 = Bt,kB⬘t,k cos ␾k d␾k
2␲ Consider a stochastic process f 0共t兲 with mean value equal to zero,
0
autocorrelation function R f 0 f 0共t , t + ␶兲 and two-sided PSDF


2␲
1 S f 0 f 0共t , ␻兲. In the following, distinction will be made between the
⫻ cos ␾⬘k d␾⬘ = 0 for k ⫽ k⬘ stochastic process f 0共t兲 and its simulation f共t兲.
0 2␲ k
From the infinite series representation shown in Eq. 共50兲, it
共45兲 follows that the stochastic process f 0共t兲 can be simulated by the
and in exactly the same way it can be shown that following series as N → ⬁;
␧关dVt共␻k兲dVt共␻⬘k 兲兴 = 0, for k ⫽ k⬘. N−1
Finally, the requirement described in Eq. 共34兲 can be expressed f共t兲 = 冑2 兺 关2S f 0f0
共t,␻n兲⌬␻兴1/2 cos共␻nt + ⌽n兲 共52兲
as n=0

␧关dUt共␻k兲dVt共␻⬘k 兲兴 = ␧关2Bt,kB⬘t,k cos ⌽k sin ⌽⬘k 兴 where

= 2Bt,kB⬘t,k␧关cos ⌽k sin ⌽⬘k 兴 共46兲 ⌬␻ = ␻u/N 共53兲


When k ⫽ k⬘, the expected value appearing in the last term of
␻n = n⌬␻, n = 0,1,2, . . . ,N − 1 共54兲
Eq. 共46兲 can be written as follows: as ⌽k and ⌽⬘k are independent
random phase angles and it is assumed that

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 619

J. Eng. Mech. 2007.133:616-627.


S f 0 f 0共t,␻0 = 0兲 = 0 共55兲
␧关f共t兲兴 = ␧ 冋 N−1
冑2 兺 关2S f
n=0
0f0
共t,␻n兲⌬␻兴1/2 cos共␻nt + ⌽n兲 册
In Eq. 共53兲, ␻u represents an upper cutoff frequency beyond N−1

which the evolutionary PSDF S f 0 f 0共t , ␻兲 may be assumed to = 冑2 兺 关2S f f 共t,␻n兲⌬␻兴1/2␧关cos共␻nt + ⌽n兲兴
0 0
共56兲
be zero for either mathematical or physical reasons. As such, n=0

␻u⫽fixed value and hence ⌬␻ → 0 as N → ⬁ so that N⌬␻ = ␻u. where the expected value ␧关cos共␻nt + ⌽n兲兴 can be shown to be
The ⌽0, ⌽1, ⌽2 , . . . , ⌽N−1 in Eq. 共52兲 are independent random equal to zero in the following way:
phase angles distributed uniformly over the interval 关0 , 2␲兴. The

+⬁
simulated stochastic process f共t兲 is asymptotically Gaussian as
␧关cos共␻nt + ⌽n兲兴 = p⌽n共␾n兲cos共␻nt + ␾n兲d␾n
N → ⬁ because of the central limit theorem. It will be shown next −⬁
that the ensemble expected value ␧关f共t兲兴 and the ensemble auto-


Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

2␲
correlation function R f f 共t , t + ␶兲 of the simulated stochastic process 1
f共t兲 are identical to the corresponding targets ␧关f 0共t兲兴 and = cos共␻nt + ␾n兲d␾n = 0 共57兲
2␲ 0
R f 0 f 0共t , t + ␶兲, respectively.
1. Show that: ␧关f共t兲兴 = ␧关f 0共t兲兴 = 0. Combining Eqs. 共56兲, 共57兲, and 共6兲, it is easy to show that
Proof: Utilizing the property that the operations of mathemati-
␧关f共t兲兴 = 0 = ␧关f 0共t兲兴 共58兲
cal expectation and summation are commutative, the ensemble
expected value of the simulated stochastic process f共t兲 can be 2. Show that: R f f 共t , t + ␶兲 = R f 0 f 0共t , t + ␶兲.
written as Proof:

R f f 共t,t + ␶兲 = ␧关f共t兲f共t + ␶兲兴 = ␧ 冋 N−1


冑2 兺 共2S f
n=0
0f0
共t,␻n兲⌬␻兲 1/2
N−1

cos共␻n共t兲 + ⌽n兲 ⫻ 冑2 兺 共2S f


m=0
0f0
共t + ␶,␻m兲⌬␻兲1/2 cos共␻m共t + ␶兲 + ⌽m兲 册
N−1 N−1

=2 兺 兺 共2S f f 共t,␻n兲⌬␻兲1/2共2S f f 共t + ␶,␻m兲⌬␻兲1/2 ⫻ ␧关cos共␻nt + ⌽n兲cos共␻m共t + ␶兲 + ⌽m兲兴


n=0 m=0
0 0 0 0
共59兲

As random phase angles ⌽n 共n = 0 , 1 , 2 , . . . , N − 1兲 are indepen- N−1

dent, for n ⫽ m, the expected value shown in Eq. 共59兲 can be R f f 共t,t + ␶兲 = 2 兺 共2S f f 共t,␻n兲⌬␻兲1/2
0 0
written as n=0

1
⫻共2S f 0 f 0共t + ␶,␻n兲⌬␻兲1/2 cos共␻n␶兲 共63兲
␧关cos共␻nt + ⌽n兲cos共␻m共t + ␶兲 + ⌽m兲兴 2

= ␧关cos共␻nt + ⌽n兲兴␧关cos共␻m共t + ␶兲 + ⌽m兲兴 = 0 共60兲 Taking the limit as ⌬␻ → 0 and N → ⬁ and keeping in mind that
␻u = N⌬␻ is constant and that S f 0 f 0共t , ␻兲 = 0 for 兩␻ 兩 ⱖ ␻u, leads to



It may be noted that only the terms with n = m remain in Eq. 共59兲.
Hence, the ensemble autocorrelation function of the simulated R f f 共t,t + ␶兲 = 2 关S f 0 f 0共t,␻兲兴1/2关S f 0 f 0共t + ␶,␻兲兴1/2 cos共␻␶兲d␻
0
process f共t兲 can be written as
共64兲
N−1 From Eq. 共7兲, we have


R f f 共t,t + ␶兲 = 2 共2S f 0 f 0共t,␻n兲⌬␻兲1/2共2S f 0 f 0共t + ␶,␻n兲兲1/2 ⬁
n=0
R f 0 f 0共t,t + ␶兲 = A*共t,␻兲A共t + ␶,␻兲ei␻␶S共␻兲d␻
· ␧关cos共␻nt + ⌽n兲cos共␻n共t + ␶兲 + ⌽n兲兴 共61兲 −⬁



= 关S f 0 f 0共t,␻兲兴1/2关S f 0 f 0共t + ␶,␻兲兴1/2ei␻␶d␻
where −⬁

共65兲
␧关cos共␻nt + ⌽n兲cos共␻n共t + ␶兲 + ⌽n兲兴 Observing Eqs. 共64兲 and 共65兲, it is easy to show
1
= 2 ␧关cos共␻nt + ␻n共t + ␶兲 + 2⌽n兲 + cos共␻n共t + ␶兲 − ␻nt兲兴
R f f 共t,t + ␶兲 = R f 0 f 0共t,t + ␶兲 共66兲
1 1
= 2 ␧关cos共␻n␶兲兴 = 2 cos ␻n␶ 共62兲
then

Then, Eq. 共61兲 takes a more compact expression S f f 共t,␻兲 = S f 0 f 0共t,␻兲 共67兲

620 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


The fact that the two-sided PSDF S f 0 f 0共t , ␻兲 is a real and even N−1

function of the frequency ␻ was used in deriving Eq. 共66兲. ␧关f共t兲兴 = ␧关冑2 兺 关2S f f 共t,␻n兲⌬␻兴1/2 cos共␻nt + ⌽n共t兲兲
0 0
n=0
N−1
Two Special Cases of the Simulation Formula
= 冑2 兺 关2S f f 共t,␻n兲⌬␻兴1/2␧关cos共␻nt + ⌽n共t兲兲兴 共76兲
It was shown earlier that sample functions generated by Eq. 共52兲 n=0
0 0

possess the properties represented by Eqs. 共58兲 and 共66兲. Keeping


the total energy a constant, Eq. 共66兲 may be relaxed to some where


extent as given by following equations: ⬁
␧关cos共␻nt + ⌽n共t兲兲兴 = p⌽n共t兲共␾共t兲 共t兲 共t兲
n 兲cos共␻nt + ␾n 兲d␾n
R f f 共t,t兲 = R f 0 f 0共t,t兲 共68兲 −⬁


2␲
or further 1
cos共␻nt + ␾共t兲 共t兲
n 兲d␾n = 0 共77兲
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

=
2␲


⬁ 0

R f f 共t,t + ␶兲 = R f 0 f 0共t,t + ␶兲 = ␦共␶兲 S f 0 f 0共t,␻兲d␻ 共69兲 Combining Eqs. 共77兲 and 共6兲, it is easy to show that Eq. 共58兲 is
−⬁ satisfied. Now, for Eq. 共68兲, one can write,
and

R f f 共␶兲 = R f 0 f 0共␶兲 共70兲


R f f 共t,t兲 = ␧关f 2共t兲兴 = ␧ 冋 N−1
冑2 兺 共2S f
0f0
共t,␻n兲⌬␻兲1/2 cos共␻nt + ⌽n共t兲兲


n=0
N−1
In the following two special cases, it will be shown that by relax-
ing Eq. 共66兲 to Eq. 共68兲, the energy is randomly distributed with ⫻ 冑2 兺 共2S f f 共t,␻m兲⌬␻兲1/2 cos共␻mt + ⌽m共t兲兲
0 0
m=0
frequency at any time instant of a sample function; and further by
Eq. 共69兲 the sample functions are nonstationary white noise pro- N−1 N−1

cesses. Also considering Eq. 共70兲, the energy is averaged with =2 兺 兺 共2S f f 共t,␻n兲⌬␻兲1/2共2S f f 共t,␻m兲⌬␻兲1/2
0 0 0 0
time, and thus, Eq. 共52兲 is reduced to the well-known Shinozuka’s n=0 m=0
simulation formula 共Shinozuka and Jan 1972; Shinozuka 1972兲 ⫻ ␧关cos共␻nt + ⌽n共t兲兲cos共␻mt + ⌽m共t兲兲兴 共78兲
for stationary stochastic processes.
As random phase angles ⌽0共t兲, ⌽1共t兲, ⌽2共t兲,…, ⌽N−1共t兲 are inde-
Special Case (I) pendent at any time instant, the expected value in Eq. 共78兲, for
If dUt共␻k兲 and dVt共␻k兲 in Eq. 共35兲 are defined as n ⫽ m, can be written as
␧关cos共␻nt + ⌽n共t兲兲cos共␻mt + ⌽m共t兲兲兴
dUt共␻k兲 = 冑2Bt,k cos ⌽k共t兲 共71兲
= ␧关cos共␻nt + ⌽n共t兲兲兴␧关cos共␻mt + ⌽m共t兲兲兴 = 0 共79兲
dVt共␻k兲 = 冑2Bt,k sin ⌽k共t兲 共72兲 Therefore, considering n = m, Eq. 共79兲 can be expressed as
N−1
where
R f f 共t,t兲 = 2 兺 共2S f f 共t,␻n兲⌬␻兲␧关cos2共␻nt + ⌽n共t兲兲兴
0 0
共80兲
Bt,k = 关2S f 0 f 0共t,␻兲⌬␻兴 1/2
共73兲 n=0

where
and ⌽k共t兲⫽independent random phase angles uniformly distrib-
uted in the range 关0, 2␲兴, and ⌽k共t1兲 ⫽ ⌽k共t2兲 for t1 ⫽ t2. It can be 1
␧关cos2共␻nt + ⌽n共t兲兲兴 = 2 ␧关cos共2␻nt + 2⌽n共t兲兲 + 1兴 =
1
共81兲
2
shown that Eqs. 共71兲 and 共72兲 satisfy Eqs. 共31兲–共34兲.
Substituting Eqs. 共71兲 and 共72兲 into Eq. 共35兲, the following Further
series representation is obtained N−1


R f f 共t,t兲 = ␧关f 2共t兲兴 = 兺 2S f f 共t,␻n兲⌬␻
0 0
共82兲

兺 兵cos共␻kt兲冑2关2S f f 共t,␻k兲⌬␻兴1/2 cos ⌽k共t兲


n=0
f 0共t兲 =
k=0
0 0
Taking the limit as ⌬␻ → 0 and N → ⬁ and keeping in mind that
␻u = N⌬␻ is constant and that S f 0 f 0共t , ␻兲 = 0 for 兩␻ 兩 ⱖ ␻u, lead to
− sin共␻kt兲冑2关2S f 0 f 0共t,␻k兲⌬␻兴1/2 sin ⌽k共t兲其

冕 冕
⬁ ⬁

R f f 共t,t兲 = 2 S f 0 f 0共t,␻兲d␻ = S f 0 f 0共t,␻兲d␻ 共83兲
= 冑2 兺 关2S f f 共t,␻k兲⌬␻兴1/2 cos共␻kt + ⌽k共t兲兲
0 0
共74兲 0 −⬁
k=0
Note that
Now, stochastic process f 0共t兲 can be simulated by finite series


representation
R f 0 f 0共t,t兲 = S f 0 f 0共t,␻兲d␻ 共84兲
N−1 −⬁

f共t兲 = 冑2 兺 关2S f f 共t,␻n兲⌬␻兴1/2 cos共␻nt + ⌽n共t兲兲


0 0
共75兲 and it is easy to show that
n=0
R f f 共t,t兲 = R f 0 f 0共t,t兲 共85兲
where all the variables have the same meaning as those in
Eq. 共52兲 except for ⌽k共t兲. Next, it will be shown that the sample Further, if independent random phase angles ⌽n共t兲 are distrib-
functions by Eq. 共75兲 satisfy Eqs. 共58兲 and 共68兲. Now uted uniformly over 关0 , 2␲兴 for any n and any t, i.e., ⌽n共t兲 is a

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 621

J. Eng. Mech. 2007.133:616-627.


random field, it can be proven that Eqs. 共71兲 and 共72兲 satisfy
冕冕
⬁ ⬁

Eqs. 共31兲–共34兲, and Eq. 共75兲 satisfies Eqs. 共58兲 and 共69兲. Thus, the 兩F共t,␻兲兩2dtd␻
sample functions generated by Eq. 共75兲 are nothing but nonsta- −⬁ −⬁

冕冕冕冕
tionary white noise processes. The detail will not be given here ⬁ ⬁ ⬁ ⬁
for saving space. = f共␶1兲f共␶2兲h共t − ␶1兲h共t − ␶2兲
−⬁ −⬁ −⬁ −⬁
Special Case (II) ⫻e−i␻共␶1−␶2兲d␶1d␶2dtd␻
If one relaxes the energy distribution along with time, e.g., if one

冕冕
⬁ ⬁
takes its average, the evolutionary PSDF S f 0 f 0共t , ␻兲 of a nonsta-
tionary process is then reduced to PSDF S f 0 f 0共␻兲 of a stationary = f 2共␶兲h2共t − ␶兲d␶dt 共90兲
−⬁ −⬁
process as follows:
In deriving Eq. 共90兲, the following equation is used:

冕 冕

Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

1 ⬁
S f 0 f 0共␻兲 = S f 0 f 0共t,␻兲dt 共86兲
T −⬁
e−i␻共␶1−␶2兲d␻ = ␦共␶1 − ␶2兲 共91兲
−⬁
where T⫽effective duration of evolutionary PSDF S f 0 f 0共t , ␻兲. If h2共t兲 = ␦共t兲, the total energy in Eq. 共90兲 is
Then, Eq. 共52兲 is simplified into

冕冕 冕
⬁ ⬁ ⬁
N−1 兩F共t,␻兲兩2dtd␻ = f 2共t兲dt 共92兲
f共t兲 = 冑2 兺 关2S f 0f0
共␻n兲⌬␻兴 1/2
cos共␻nt + ⌽n兲 共87兲 −⬁ −⬁ −⬁

n=0
This means that if the time window is chosen such that


which is nothing but the well-known Shinozuka’s simulation for- ⬁
mula 共Shinozuka and Jan 1972; Shinozuka 1972兲 for stationary h2共t兲dt = 1 共93兲
stochastic processes. −⬁

the total energy can be kept identical 共Parseval’s identity兲 in es-


timating the evolutionary PSDFs. For example, the selected time
Generation of Evolutionary Power Spectral Density window-squared functions may be
Function
1 −t2/2␴2
h2共t兲 = 共␴ = 0.25兲 共94兲
There are potentially a number of ways in which evolutionary 冑2␲␴ e
PSDF is estimated from a ground motion time-history record.
This study develops and compares simple methods of estimating or
evolutionary PSDF by means of the STFT 共Flanagan 1972兲, the
WT 共Daubechies 1992兲, and the HHT 共Huang et al. 1998兲. It may h2共t兲 = 1 共0 ⱕ t ⱕ 1兲 共95兲
be noted here that the STFT and the WT represent energy-time- which are nothing but the Gaussian time window squared with
frequency distribution of nonstationary processes with a chosen standard deviation ␴ = 0.25 s, and the rectangular time window
windowing function or wavelet, while the HHT represents squared with height and duration both equal to 1, respectively.
energy-time-frequency distribution by using a simple and adap- One might observe that both time windows satisfy the condition
tive technique of obtaining instantaneous frequencies. Following in Eq. 共93兲.
subsections present the methods to estimate evolutionary PSDFs
from a ground motion record using the three abovementioned
transforms. It may also be mentioned that the evolutionary PSDFs Wavelet Transform
estimated from these three methods obey Parseval’s identity of The WT of f 苸 L2共R兲 共finite energy function 兰兩f共t兲兩2dt ⬍ + ⬁兲 at
total energy. time u and scale s, and the corresponding inverse relationship are
given by Daubechies 共1992兲
Short-Time Fourier Transform
The STFT F共t , ␻兲 of a function f共t兲 is defined by the following W␺ f共u,s兲 =
1
冑s 冕

−⬁
f共t兲␺* 冉 冊 共t − u兲
s
dt, s⬎0 共96兲
convolution integral:
and

冕 冉 冊

冕冕
⬁ ⬁
F共t,␻兲 = f共␶兲h共t − ␶兲e−i␻␶d␶ 共88兲 1 1 共t − u兲
−⬁
f共t兲 = 2 W␺ f共u,s兲f共t兲␺ duds, s⬎0
2␲C␺ −⬁ −⬁ s s
where h共t兲 is an appropriate time window. The evolutionary 共97兲
PSDF S f 0 f 0共t , ␻兲 can be written as
where

冕冕 冕
⬁ ⬁

兩F共t,␻兲兩2 = f共␶1兲f 0共␶2兲h共t − ␶1兲h共t − ␶2兲e−i␻␶1ei␻␶2d␶1d␶2 兩␺ˆ 共␻兲兩2
C␺ = d␻ ⬍+⬁ 共admissibility condition兲 共98兲
−⬁ −⬁
−⬁ 兩␻兩
共89兲
In Eqs. 共96兲–共98兲, the wavelet function ␺ 苸 L2共R兲 known as
The total energy of f共t兲 can be estimated by “basic” or “mother” wavelet with zero average

622 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


冕 冕
⬁ ⬁
1 X共t⬘兲
␺共t兲dt = 0 共99兲 Y共t兲 = P dt⬘ 共106兲
−⬁ ␲ −⬁ t − t⬘

is centered in the neighborhood of t = 0, and as normalized where P indicates the Cauchy principle value. Here, X共t兲 and Y共t兲
兩兩␺ 兩 兩 = 1. ␺ˆ 共␻兲 denotes the Fourier transform of ␺共t兲 and is form the complex conjugate pair and thus an analytic function
given by Z共t兲 can be constructed as


⬁ Z共t兲 = X共t兲 + iY共t兲 = a共t兲ei␪共t兲 共107兲
1
␺ˆ 共␻兲 = ␺共t兲e−i␻tdt 共100兲
冑2␲ −⬁ in which
It may be noted that the WT decomposes signal f共t兲 over dilated
a共t兲 = 冑X2共t兲 + Y 2共t兲 共108兲
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

and translated wavelets. As W␺ f共u , s兲 is the convolution of f共t兲


with 共1 / 冑s兲␺*共−t / s兲, W␺ f共u , s兲 represents the contribution of the
function f共t兲 in the neighborhood of t = u and in the frequency ␪共t兲 = tan−1Y共t兲/X共t兲 共109兲
band corresponding to scale s. It can also be shown that 共see
Daubechies 1992兲 One can observe in Eq. 共107兲 that X共t兲 is the real part of
Z共t兲 关=共a共t兲ei␪共t兲兲兴 expressed in the polar coordinate to clarify the

冕 冕冕
⬁ ⬁ ⬁
1 1 local nature of X共t兲 by representing it in an amplitude- and phase-
兩f共t兲兩2dt = 兩W␺ f共u,s兲兩2duds 共101兲 varying trigonometric function.
−⬁ 2␲C␺ −⬁ −⬁ s2
Based on the definition of Hilbert transform, the instantaneous
Now, if any wavelet function satisfies the condition frequency of ␻共t兲 can be expressed as 共Huang et al. 1998兲


⬁ d␪共t兲
兩␺ˆ u,s共␻兲兩2d␻ = 1 共102兲 ␻共t兲 = 共110兲
dt
−⬁

However, as discussed by Huang et al. 共1998兲, for this definition


one can write Eq. 共101兲 as
to be meaningful, the data must be locally symmetric with respect

冕 冕 冋 冕冕 册
⬁ ⬁ ⬁ ⬁
to the zero mean. Unfortunately, the majority of data encountered
1 1 do not satisfy this condition. Empirical mode decomposition,
兩f共t兲兩2dt = 兩W␺ f共u,s兲兩2duds
−⬁ −⬁ 2␲C␺ −⬁ −⬁ s2 which is the first part of the HHT, decomposes the data into a
number of intrinsic mode functions 共IMF兲 to satisfy local symme-
⫻兩␺ˆ u,s共␻兲兩2d␻ 共103兲 try condition by using an adaptive numerical technique 共see
Huang et al. 1998兲. Empirical modes are obtained through a shift-
In Eqs. 共102兲 and 共103兲, ␺ˆ u,s共␻兲 represents the Fourier transform ing process, the result of which decomposes data into n empirical
of ␺关共t − u兲 / s兴 and can be expressed as ␺ˆ u,s共␻兲 = 冑s␺ˆ 共s␻兲ei␻u. modes and a residue rn, which can either be the mean trend of the
Then, using Parseval’s identity, one can write data or a constant and often negligible compared to the other IMF
components. After that, the instantaneous frequencies are ob-

冕冕
⬁ ⬁ tained through Hilbert transform.
1 1
兩F共␻兲兩2 = 兩W␺ f共u,s兲兩2兩␺ˆ u,s共␻兲兩2duds Let C j共t兲 be the jth empirical mode of X共t兲, then X共t兲 can be
2␲C␺ −⬁ −⬁ s2 expressed as
共104兲
n

where F共␻兲⫽Fourier transform of f共t兲. As the wavelet coeffi- X共t兲 = 兺 C j共t兲 + rn 共111兲
cients W␺ f共u , s兲 provides the localized information of signal f共t兲 j=1
at t = u, from Eq. 共104兲 the evolutionary PSDF S f 0 f 0共t , ␻兲 can be
expressed as Now, for a negligible residue, the original data can be approxi-
mated as the real part of the sum of the transform of each empiri-


⬁ cal modes
1 1
兩F共␻,t兲兩2 = 兩W␺ f共t,s兲兩2兩␺ˆ t,s共␻兲兩2ds 共105兲
2␲C␺ −⬁ s2 n

It may be noted that the previous expression of evolutionary


X共t兲 = RP 兺
j=1
a j共t兲ei兰␻ 共t兲dt
j 共112兲
PSDF obeys total energy equilibrium. Therefore, one can use any
wavelet basis, which satisfies Eq. 共102兲, to generate evolutionary where a j共t兲 = 共C2j 共t兲 + Y 2j 共t兲兲1/2, 兰␻ j共t兲dt = ␪ j共t兲 = arctan共Y j共t兲 / C j共t兲兲,
PSDF 关e.g., modified Littlewood-Paley basis as proposed by Basu and Y j共t兲 being the Hilbert transform of C j共t兲. Both the frequency
and Gupta 共1998兲兴 that maintains total energy. and the amplitude of each component are functions of time, thus,
the final representation is a curve in the three-dimensional space
of frequency-amplitude-time. This frequency-time distribution of
Hilbert-Huang Transform
the amplitude is termed as the Hilbert amplitude spectrum, or
The Hilbert transform Y共t兲 of any function X共t兲 of class L p is simply, the Hilbert spectrum. By expressing X2共t兲 in terms of
expressed by the following convolution integral 共see Bendat and C j共t兲 from Eq. 共111兲 and integrating with respect to t, one can
Piersol 1986兲 write

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 623

J. Eng. Mech. 2007.133:616-627.


instantaneous frequency is a function of time t, the evolutionary
PSDF S f 0 f 0共t , ␻兲 of X共t兲 can be expressed as
n
1
S f 0 f 0共t,␻兲 = 兩F共␻,t兲兩2 = 兺
j=1 2
␦关␻ − ␻ j共t兲兴a2j 共t兲 共114兲

It may be noted that the previous expression of evolutionary


PSDF obey energy equilibrium when IMF components are or-
thogonal to each other and the residue rn is zero. In all other
cases, we need to introduce a correction term cr for energy equi-
librium in the right-hand side of Eq. 共114兲, which can be obtained
by
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

兰⬁⬁X2共t兲dt
Fig. 1. Acceleration time history: N-S component of 1940 El Centro cr = 共115兲
兺 j=1 C2j 共t兲dt
n
earthquake record 兰⬁⬁

Now, if filtering and smoothing functions are applied to give a

冕 冕 兺兺
⬁ ⬁ n n better look of the estimated evolutionary PSDF, the energy
X2共t兲dt = C j共t兲Ck共t兲dt spreads over time and frequency. Thus, further correction terms
⬁ ⬁ j=1 k=1 must be introduced to preserve total energy of the time history in

冕兺 冕兺
⬁ n ⬁ n n generated evolutionary PSDFs.
=
⬁ j=1
C2j 共t兲dt + 2 兺
⬁ j=1 k=1;j⫽k
C j共t兲Ck共t兲dt

Results and Discussion


共113兲
Now, if the IMF components are orthogonal with respect to In order to generate the evolutionary PSDF from a given
each other, then the cross terms given in the right-hand side of earthquake record, N-S component of the well–known 1940 El
Eq. 共113兲 are zero. Although, in most of the cases, these cross Centro earthquake acceleration time history record is considered
terms are not zero, by virtue of the decomposition, each IMF 共see Fig. 1兲. The evolutionary PSDF S f 0 f 0共t , ␻兲 of the El Centro
component represents narrow banded data with minimal fre- earthquake acceleration record is estimated on the basis of the
quency content overlaps. Therefore, the cross terms are negligible methods developed by using the STFT, the WT, and the HHT. For
compared to the first term of Eq. 共113兲. Further, using this argu- the STFT, a Gaussian window squared defined by Eq. 共93兲 with
ment along with Parseval identity, and keeping in mind that the ␴ = 0.25 is used, and for the WT, a Morlet mother wavelet with

Fig. 2. Evolutionary PSDF of El Centro earthquake record in g2 – s using the method based on the STFT 共Gaussian window兲

624 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

Fig. 3. Evolutionary PSDF of El Centro earthquake record in g2 – s using the method based on the WT 共Morlet兲

␻0 = 6 is used. It was found that by using the proposed methods mated using the methods based on the STFT 共Gaussian window兲,
for generating evolutionary PSDFs, total energy of the time his- the WT 共Morlet basis兲, and the HHT, respectively. By comparing
tory is preserved in evolutionary PSDF with numerical errors Figs. 2–4, it may be observed that the methods based on the
of 0.01% for the method based on the STFT, 0.23% for the STFT and the WT give similar results in time-frequency repre-
method based on the WT. The numerical error and the error in- sentation. One can notice though that the method based on the
troduced by nonorthogonality of the extracted empirical modes in WT gives better resolution in both high and low frequencies,
case of the method based on the HHT is about 2.68%. Further, whereas the method based on the STFT gives not-so-good reso-
for the PSDFs estimated using this method, filtering and smooth- lution at high frequencies. On the other hand, the method based
ing functions are applied and energy balance is ensured by per- on the HHT shows that the energy is concentrated at some distinct
forming double integral of evolutionary PSDF and using the frequencies only. Further, in this example, one can notice from the
Perseval’s identity. vertical axes of Figs. 2–4 that the energy concentration at certain
Figs. 2–4 show zoomed view of the evolutionary PSDFs esti- frequency and time is approximately ten times higher in the case

Fig. 4. Evolutionary PSDF of El Centro earthquake record in g2 – s using the method based on the HHT

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 625

J. Eng. Mech. 2007.133:616-627.


Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

Fig. 6. 5% damped ground motion acceleration response spectrum


共GM兲 and the mean of five percent damped acceleration spectra
considering twenty sample functions generated from each of the three
evolutionary PSDFs using the STFT, WT, and HHT

spectra estimated by all the three methods capture the trend of the
target ground motion spectrum well with slight over estimation at
higher frequencies. The discrepancies observed in spectral ordi-
nates may be partly due to discretization scheme with some nu-
merical error, and partly due to the variation of sample functions
providing different levels of coefficient of variation in the spectral
values as a function of natural period 共see Fig. 7兲. Perhaps, a large
number of sample functions may give a better estimate. Also, one
may notice that out of the three methods, the method based on the
HHT predicts higher spectral ordinates, though it captures the
trend of the target spectrum well. This anomaly in spectral ordi-
nates may be due to a larger energy concentration at some dis-
crete frequencies.

Fig. 5. Synthetic accelerogram using the same random number Conclusions


sequence and the methods based on 共a兲 the STFT; 共b兲 WT; and
共c兲 HHT This paper, integrating Priestly’s evolutionary spectral representa-
tion theory, presents a rigorous derivation of the computationally
efficient cosine series formula introduced by Shinozuka and Jan
of the HHT when compared with the one associated with the in 1972 for simulation of one-dimensional, univariate, nonstation-
STFT or the WT. ary stochastic processes. The stochastic process thus simulated is
In order to study the three estimated evolutionary PSDF in a asymptotically Gaussian as the number of terms tends to infinity.
different way, and to illustrate the methodology of simulating This paper shows that 共1兲 the sample functions accurately reflect
nonstationary ground motions, one sample function for each of
the three estimated evolutionary PSDFs 共see Figs. 2–4兲 is gener-
ated. Three corresponding acceleration time histories generated
by the methods based on the STFT, WT, and HHT are shown in
Fig. 5. Since the generated sample functions depend on selected
random number sequence, the same sequence of the random num-
bers is used for all the three cases in order to compare the results.
From Fig. 5, one can observe that the method based on the HHT
gives higher estimate of the peak ground acceleration when com-
pared with those based on the STFT and the WT. This may be due
to a higher energy concentration at some high frequencies for
HHT as observed in Fig. 4. Comparing Fig. 1 with Fig. 5, one can
notice that all simulated time histories reproduce the arrivals of
earthquake waves very well.
Figs. 6 and 7, respectively, show the plot of the mean and the
coefficient of variation of response spectra 共with 5% damping兲
of 20 simulated sample functions obtained from each of the three Fig. 7. Coefficient of variation of 5% damped acceleration spectra
estimated evolutionary PSDF of El Centro record 共Figs. 2–4兲. The as a function of natural period considering twenty sample functions
same twenty sequences of random numbers are used for the three generated from each of the three evolutionary PSDFs using the STFT,
simulations. From Fig. 6, it is found that, in general, the response WT, and HHT

626 / JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007

J. Eng. Mech. 2007.133:616-627.


the prescribed probabilistic characteristics of the stochastic pro- Daubechies, I. 共1992兲. Ten lectures on wavelets, Soc. Indust. Appl. Math.,
cess when the number of terms in the cosine series is large, and Philadelphia.
共2兲 the simulation formula can be reduced to that for nonstation- Deodatis, G. 共1996a兲. “Non-stationary stochastic vector processes:
ary white noise process or Shinozuka’s stationary process under Seismic ground motion applications.” Probab. Eng. Mech., 11共3兲,
certain conditions. 149–168.
Deodatis, G. 共1996b兲. “Simulation of ergodic multivariate stochastic pro-
Three methods are developed to estimate the evolutionary
cesses.” J. Eng. Mech., 122共8兲, 778–787.
PSDF of a given time-history data by the STFT, WT, and HHT. A Deodatis, G., and Shinozuka, M. 共1988兲. “Autoregressive model for non-
comparison of the generated PSDF using the well-known El Cen- stationary stochastic processes.” J. Eng. Mech., 114共11兲, 1995–2012.
tro earthquake record shows that the method based the STFT and Deodatis, G., and Shinozuka, M. 共1989兲. “Simulation of seismic ground
the WT give similar results, while the method based on the HHT motion using stochastic waves.” J. Eng. Mech., 115共12兲, 2723–2737.
gives concentrated energy at certain frequencies. Using the same Flanagan, J. L. 共1972兲. Speech analysis: Synthesis and perception, 2nd
random number sequence, it is found that the method based on Ed., Springer, New York.
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 12/23/12. Copyright ASCE. For personal use only; all rights reserved.

the HHT gives higher estimate of the peak ground acceleration Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q.,
when compared with that of the methods based on the STFT and Yen, N., Tung, C. C., and Liu, H. H. 共1998兲, “The empirical mode
the WT. More importantly, it is found that the sample functions decomposition and the Hilbert spectrum for nonlinear and nonstation-
generated by these three methods all preserve the history of ar- ary time series analysis.” Proc. R. Soc. London, Ser. A, 454, 903–995.
rivals of earthquake waves very well. Further, using 20 sample Li, Y., and Kareem, A. 共1991兲. “Simulation of multivariate nonstationary
random processes by FFT.” J. Eng. Mech., 117共5兲, 1037–1058.
functions, it is observed that, in general, the mean of the 5%
Ohsaki, Y. 共1979兲. “On the significance of phase content in earthquake
damped acceleration response spectra obtained from each of the ground motions.” Earthquake Eng. Struct. Dyn., 7, 427–439.
three estimated PSDFs captures the trend of the target spectrum Popescu, R., Deodatis, G., and Prevost, J. H. 共1997兲. “Simulation of
well with some discrepancies at high frequencies. Perhaps, a bet- homogeneous non-Gaussian stochastic vector fields.” Probab. Eng.
ter discretization scheme and a larger number of sample functions Mech., 13共1兲, 1–13.
will provide better results. It is also concluded that the evolution- Priestley, M. B. 共1965兲. “Evolutionary spectra and nonstationary pro-
ary PSDF estimated by the method based on the HHT overpre- cesses.” J. R. Stat. Soc. Ser. B (Methodol.), Series B, 27, 204–237.
dicts mean spectral ordinates. Priestley, M. B. 共1967兲. “Power spectral analysis of non-stationary ran-
dom processes.” J. Sound Vib., 6, 86–97.
Ramadan, O., and Novak, M. 共1993兲. “Simulation of spatially incoherent
random ground motions.” J. Eng. Mech., 119共5兲, 997–1016.
Acknowledgments Shinozuka, M. 共1972兲. “Monte Carlo solution of structural dynamics.”
Comput. Struct., 2共5,6兲, 855–874.
This research was supported by National Science Foundation Shinozuka, M. 共1974兲. “Digital simulation of random processes in engi-
共NSF兲 through Award Nos. 0509018 and 0352014. neering mechanics with the aid of FFT technique.” Stochastic prob-
lems in mechanics, S. T. Ariaratnam and H. H. E. Leipholz, eds.,
University of Waterloo Press, Waterloo, Ont., Canada, 277–286.
Shinozuka, M., and Jan, C. M. 共1972兲. “Digital simulation of random
References processes and its application.” J. Sound Vib., 25共1兲, 111–128.
Shinozuka, M., and Sato, Y. 共1967兲. “Simulation of nonstationary random
Basu, B., and Gupta, V. K. 共1998兲. “Seismic response of SDOF systems process.” J. Engrg. Mech. Div., 93共1兲, 11–40.
by wavelet modeling of nonstationary processes.” J. Eng. Mech., Yamazaki, F., and Shinozuka, M. 共1988兲. “Digital generation of non-
124共10兲, 1142–1150. Gaussian stochastic fields.” J. Eng. Mech., 114共7兲, 1183–1197.
Bendat, J. S., and Piersol, A. G. 共1986兲. Random data: Analysis and Yang, J. N. 共1972兲. “Simulation of random envelop processes.” J. Sound
measurement procedures, Wiley, New York. Vib., 21共1兲, 73–85.
Cramer, H., and Leadbetter, M. R. 共1967兲. Stationary and related sto- Yang, J. N. 共1973兲. “On the normality and accuracy of simulated random
chastic processes, Wiley, New York. processes.” J. Sound Vib., 26共3兲, 417–428.

JOURNAL OF ENGINEERING MECHANICS © ASCE / JUNE 2007 / 627

J. Eng. Mech. 2007.133:616-627.

You might also like