CEE 745: Recommended Reading

An Improved Method for Nonstationary

Spectral Matching
Linda Al Atika) and Norman Abrahamson,b) M.EERI

Seismic input to nonlinear dynamic analyses of structures is usually

defined in terms of acceleration time series whose response spectra are
compatible with a specified target response spectrum. Time domain spectral
matching used to generate realistic design acceleration time series is discussed
in this paper. A new and improved adjustment function to be used in modifying
existing accelerograms while preserving the nonstationary character of the
ground motion is presented herein. The application of the new adjustment
wavelet ensures stability, efficiency and speed of the numerical solution and
prevents drift in the resulting velocity and displacement time
series. 关DOI: 10.1193/1.3459159兴

INTRODUCTION
Seismic design of structures is generally based on a design response spectrum ob-
tained from hazard analysis for a specified return period. The disaggregation of the haz-
ard is then used to determine the controlling earthquake scenario in terms of magnitude
and distance. For many engineering applications, such as the design of critical facilities
or highly irregular buildings, a more complex dynamic nonlinear analysis is often con-
ducted. Such analysis requires input in the form of design time series with response
spectra that are consistent with the target design spectrum.
Design time series are developed by modifying initial time series that consist of em-
pirical recordings from past earthquakes representative of the design event or numerical
simulations of the ground motion for the design event. Two approaches exist for modi-
fying the time series to be consistent with the design response spectrum: scaling and
spectral matching. Scaling involves multiplying the initial time series by a constant fac-
tor so that the spectrum of the scaled time series is equal to or exceeds the design spec-
trum over a specified period range. Spectral matching involves modifying the frequency
content of the time series to match the design spectrum at all spectral periods.
Although spectral matching is commonly used in engineering practice, the concept
of using spectrum compatible time series in the seismic design of structures remains
controversial for two reasons. First, a time series that matches the entire design spectrum
represents more than one earthquake at a time since the design spectrum may be an en-
velope of multiple earthquakes. As a result, it is generally believed that such time series
overestimate the structural response. Second, spectrum compatible time series have

a)
Postdoctoral Scholar, Department of Civil & Environmental Engineering, University of California, Berkeley,
CA 94720
b)
Pacific Gas & Electricity Company, 245 Market Street, San Francisco, CA 94105

601
Earthquake Spectra, Volume 26, No. 3, pages 601–617, August 2010; © 2010, Earthquake Engineering Research Institute
602 L. AL ATIK AND N. ABRAHAMSON

smooth response spectra and are considered unrealistic when compared to typical earth-
quake response spectra that tend to have large peaks and troughs.
The advantage of using spectrum compatible time series is to reduce the number of
time series that need to be run in the engineering analysis. A gross rule of thumb is that
one spectrum compatible time series is worth three scaled time series in terms of the
variability of the mean of the nonlinear response of structures (Bazzurro and Luco
2006). For example, if it takes engineering analyses of nine scaled time series to get 20%
accuracy in the mean structural response, then it takes only analyses of three spectrum
compatible time series to get the same accuracy. If time series analyses are expensive,
then there can be a significant cost saving using the spectrum compatible approach. An
important additional benefit of the spectral matching approach is that the criteria for se-
lecting the initial time series are not as stringent as those for the scaling approach. For
example, if the design ground motion is for rock site conditions, initial time series on
deep soil sites can be used because the spectral matching process will correct for the
differences in frequency content on soil and rock sites. Adopting the spectral matching
approach, therefore, results in more choices for the initial time series selection compared
to the scaling approach.
Various methods have been developed to modify a reference time series so that its
response spectrum is compatible with a specified target spectrum. A review of spectral
matching methods is given by Preumont (1984). There are three basic approaches for
spectral matching: frequency domain method, frequency domain method with random
vibration theory (RVT), and time domain method. Early spectral matching approaches
used the frequency domain method. This method adjusts the Fourier amplitude spectrum
based on the ratio of the target response spectrum to the time series response spectrum
while keeping the Fourier phase of the reference time series fixed. While this approach is
straightforward, it has two drawbacks. First, it generally does not have good convergence
properties. Second, it often alters the nonstationary character of the time series to such a
large degree that it no longer looks like a time series from an earthquake. The second
approach uses random vibration theory to make initial large adjustments to the Fourier
amplitude spectrum, followed by the frequency domain method for small-scale adjust-
ments. This method generally works well in terms of the acceleration and velocity time
series, but often changes the character of the displacement time series.
The third approach for spectral matching adjusts the time series in the time domain
by adding wavelets to the initial time series. A formal optimization procedure for this
type of time domain spectral matching was first proposed by Kaul (1978) and was ex-
tended to simultaneously match spectra at multiple damping values by Lilhanand and
Tseng (1987, 1988). While the time domain spectral matching procedure is generally
more complicated than the frequency domain approach, it has good convergence prop-
erties and in most cases preserves the nonstationary character of the reference time se-
ries.
While the original Lilhanand and Tseng algorithm uses an adjustment wavelet that
ensures numerical stability of the time domain spectral matching method, their adjust-
ment function does not preserve the nonstationary character of the initial acceleration
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 603

time series and introduces drift to the resulting velocity and displacement time series.
Abrahamson (1992) developed the RspMatch program to implement the Lilhanand and
Tseng algorithm and proposed a new adjustment wavelet that preserves the nonstation-
ary character of the reference ground motion and ensures stability and efficiency of the
numerical solution. This adjustment wavelet, however, does not integrate to zero end ve-
locity and displacement and leads to drift in the resulting velocity and displacement time
series. The original RspMatch program, therefore, requires the application of a baseline
correction to the resulting acceleration time series to correct for the drift in the corre-
sponding velocity and displacement time series. Hancock et al. (2006) revised and
cleaned RspMatch to eliminate this drift by modifying the adjustment wavelets. The ad-
justment functions proposed by Hancock et al. (2006) and used in their updated version
of RspMatch (2005) consist of the original functions developed by Abrahamson (1992)
and Suarez and Montejo (2003, 2005) modified to include the baseline correction in
their functional form. As a result, analytical solution of the problem is no longer possible
and numerical speed and efficiency is compromised by the use of numerical integration.
Moreover, additional wavelets are sometimes needed in RspMatch2005 to prevent diver-
gence of the solution. These wavelets, nevertheless, sometimes have limited success in
ensuring solution convergence. Hancock et al. (2006) also allowed acceleration records
to be matched to a pseudo-acceleration target spectrum with different levels of damping
simultaneously.
In this paper, we propose a new adjustment function that allows the use of an ana-
lytical solution in the spectral matching algorithm and that readily integrates to zero ve-
locity and displacement without adding baseline correction to its functional form. The
new version of RspMatch, described herein, provides a stable and time-efficient solution
without introducing drift to the resulting velocity and displacement time series. It also
allows matching records to pseudo-acceleration response spectra and ensures conver-
gence and stability of the solution.

METHODOLOGY FOR TIME DOMAIN SPECTRAL MATCHING

The algorithm proposed by Lilhanand and Tseng (1987, 1988) and implemented in
RspMatch uses wavelet functions to modify the initial time series such that its response
spectrum is compatible with the design spectrum. A fundamental assumption of this
methodology is that the time of the peak response does not change as a result of the
wavelet adjustment. If a共t兲 is the initial acceleration time series, the goal is to modify
a共t兲 such that its computed response spectrum matches the target spectrum across the
whole frequency range while maintaining realistic velocity and displacement time series.
The difference between the target spectrum and the time series spectrum at a given
frequency 共␻i兲 and damping 共␤i兲, called the spectral misfit, is given by

⌬Ri = 共Qi − Ri兲Pi , 共1兲

where Qi is the target spectral value, Ri is the time series spectral value, and Pi is the
polarity of the peak response of the oscillator. Pi is equal to 1 if the maximum oscillator
response is positive and Pi is equal to −1 if the maximum oscillator response is negative.
604 L. AL ATIK AND N. ABRAHAMSON

Assuming that the time of the peak oscillator response, ti, will not be perturbed by add-
ing a small adjustment to a共t兲, the basic method is to determine an adjustment time se-
ries, ␦a共t兲, such that the oscillator response from ␦a共t兲 at time ti is equal to ⌬Ri for all
i. ␦a共t兲 can be written as
N
␦a共t兲 = 兺 bjfj共t兲, 共2兲
j=1

where fj共t兲 is a set of adjustment functions, bj is the set of amplitudes of the adjustment
functions (coefficients to be determined), and N is the total number of spectral points
(frequency and damping pairs) to match. The acceleration response of ␦a共t兲 for fre-
quency ␻i and damping ␤i at time ti is given by

␦Ri = 冕 ␦a共␶兲h 共t − ␶兲d␶ ,

⬁

0
i i 共3兲

where hi共t兲 is the acceleration impulse response function for a single-degree-of-freedom

oscillator with frequency ␻i and damping ␤i, and ␶ is the integration time parameter.
Substituting Equation 2 into Equation 3 gives

冕 f 共␶兲h 共t − ␶兲d␶ .
N ⬁
␦Ri = 兺 bj j i i 共4兲
j=1 0

The acceleration impulse response function is given by

− ␻i
冑1 − ␤2i cos共␻⬘i t兲兴,
hi共t兲 =
冑1 − ␤2i exp共− ␻i␤it兲关共2␤i − 1兲sin共␻⬘i t兲 − 2␤i 共5兲
2

where

␻⬘i = ␻i冑1 − ␤2i , 共6兲

and hi共t兲 = 0 for t ⬍ 0. Let cij be the response at time ti for the ith frequency and damping
resulting from the adjustment function fj共t兲, then

cij = 冕 f 共␶兲h 共t − ␶兲d␶ .

0
ti
j i i 共7兲

Substituting Equation 7 into Equation 4 gives

N
␦Ri = 兺 bjcij . 共8兲
j=1

If the response of the adjustment time series,␦Ri, is equal to the spectral misfit, ⌬Ri,
then
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 605

N
⌬Ri = 兺 bjcij . 共9兲
j=1

The amplitude of each wavelet used in the adjustment is determined by

b = C−1␦R 共10兲
where C is a square matrix with elements describing the amplitude of each single degree
of freedom response at the time the response needs to be adjusted, under the action of
each wavelet.
Given bj, the adjustment time series, ␦a共t兲, can be computed using Equation 2. The
new adjusted time series for the first iteration is given by

a1共t兲 = a共t兲 + ␥␦a共t兲, 共11兲

where ␥ is a relaxation parameter (between 0 and 1) to damp the adjustments. In the
second iteration, the algorithm is repeated using the adjusted time series, a1共t兲, in place
of a共t兲. The iterations are continued until the desired accuracy of the spectral match is
achieved.

SELECTION OF INITIAL TIME SERIES

Ideally, time series having the same source, path, and site properties as the design
earthquake should be selected. These properties include earthquake magnitude, distance,
style-of-faulting, directivity condition, and site condition. Typically, there are not enough
recordings in the empirical strong motion database to satisfy all of these conditions. The
alternatives are to use numerical simulations that have all of the desired features or to
relax some of these conditions to get enough candidate recordings.
Artificially generated spectrum compatible ground motions are the result of numeri-
cal simulations that can include detailed modeling of rupture propagation, path and site
effects. The disadvantage of using such ground motions is that there is less confidence
that they capture important features in the nonlinear analysis of structures. As a result,
the use of empirical recordings, whenever possible, is generally preferred over the use of
artificially generated spectrum compatible ground motions.
In selecting empirical recordings, the key parameters that affect the nonstationary
character of the wave form are magnitude, distance, and directivity direction (for sites
located close to large faults). Magnitude is the first parameter that should be included in
defining the search window. Whenever possible, recordings should be within 0.5 mag-
nitude units of the design earthquake. Depending on the size of the event, it may be pos-
sible to restrict the magnitude to a tighter range (such as within 0.25 magnitude units).
Source to site distance should be restricted to be within ±10 km in the near fault region
but can be extended to be within ±20 km at larger distances from the fault. Site condition
has a strong influence on the frequency content of ground motions. It is, therefore, an im-
portant parameter for the scaling approach but not as important for spectral matching as long
as extreme effects such as soft soil sites are excluded. Directivity effects are an important
606 L. AL ATIK AND N. ABRAHAMSON

parameter in the selection of initial time series. If near-source effects are a factor in the de-
sign, the selection of initial time series should identify records at short distances and in the
forward directivity zone. Finding a record with the same style-of-faulting as the design earth-
quake is not an important criterion in the initial time series selection.
If multiple recordings are used, they should be selected such that they have a range
of nonstationary characteristics sampling the range of ground motions. This can be ac-
complished by selecting recordings from multiple earthquakes and/or multiple azimuths
(e.g., directivity directions) for a single earthquake.

ADJUSTMENT FUNCTIONS
The Lilhanand and Tseng algorithm provides a good solution to the spectral match-
ing problem but suffers from two shortcomings. First, the adjustment wavelets corrupt
the velocity and displacement time series of the computed accelerograms by introducing
a long period drift. Baseline correction is therefore needed to correct this resulting drift
in the displacement time series. Second, the method is not always stable and diverges if
matching of closely spaced periods and multiple damping levels is attempted. The new
version of RspMatch presented in this paper addresses these shortcomings and improves
the stability and efficiency of the algorithm.
The key to the non-stationarity of the Lilhanand and Tseng method is the selection of
the adjustment function fj共t兲. The adjustment function should be selected to yield real-
istic spectrum compatible acceleration time series. The numerical stability and speed of
the algorithm must also be considered when selecting the form of fj共t兲. For the method
to work efficiently, the timing of fj共t兲 should be such that the response of fj共t兲 is in phase
with the peak response of a共t兲. For numerical speed, fj共t兲 should be chosen such that the
elements of C given by the integral in Equation 7 can be computed analytically. For nu-
merical stability, the off-diagonal terms of C should be as small as possible. Two alter-
native models for the adjustment function used in the original version of RspMatch are
discussed in this section. The improved adjustment wavelet proposed by the authors is
also presented.

WAVELET USED BY LILHANAND AND TSENG

The original wavelet used by Lilhanand and Tseng (1987, 1988) consists of the true
acceleration impulse response function in reverse time order given by

− ␻j
fj共t兲 = hj共tj − t兲 =
冑1 − ␤2j exp共− ␻j␤j共tj − t兲兲关共2␤j − 1兲sin共␻⬘i 共tj − t兲兲
2

− 2␤j冑1 − ␤2j cos共␻⬘j 共tj − t兲兲兴, 共12兲

where tj is the time of the peak response of the jth oscillator under the action of the jth
wavelet. An example of this wavelet form is shown in Figure 1.
The wavelet form used by Lilhanand and Tseng has several desirable numerical fea-
tures that help keep the solution algorithm stable. First, it leads to a symmetric C matrix.
Second, the abrupt stop insures that the response will peak at time tj and will not reso-
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 607

Figure 1. Acceleration, velocity, and displacement time series of (a) reverse acceleration im-
pulse response wavelet and (b) tapered cosine wavelet.

nate to larger values at greater values of t. The desirable numerical features of this model
lead to some undesirable features from a seismological point of view, particularly at long
periods. As shown in Figure 1, the adjustment is emergent and stops abruptly. This is
contrary to the behavior of strong motion time series that generally have a sharp initia-
tion and gradual decay. Moreover, the adjustment function is non-zero only before the
time of the peak response which can lead to unrealistic time series. For example, forcing
the long period adjustment early in the time series can lead to unrealistic ground motions
if the long period response peaks early in the record as can be the case for near-fault
recordings with distances less than 5 km. Finally, the reverse acceleration impulse response
wavelet introduces drift to the velocity and displacement time series as seen in Figure 1. This
608 L. AL ATIK AND N. ABRAHAMSON

drift requires the application of baseline correction to the resulting time series after each it-
eration which can partially undo the spectral match and is time-consuming.

TAPERED COSINE WAVELET

The tapered cosine wave shown in Figure 1 can be described by

fj共t兲 = cos关␻⬘j 共t − tj + ⌬tj兲兴exp关− 兩t − tj + ⌬tj兩␣j兴, 共13兲

where ⌬tj is the difference between the time of peak response tj and the reference origin
of the wavelet. ⌬tj is given by

⌬tj =
−1
tan 冋 冑1 − ␤2j
␤j
册 . 共14兲
␻⬘j
␣j is a frequency-dependent factor that should be selected such that the adjustment
wavelet and the reference time series have consistent durations at frequency ␻j. In other
words, if the reference time series has a short duration at a particular frequency, ␣j
should be selected such that the adjustment function at that frequency will also have a
short duration (Hancock et al. 2006). A tri-linear model for ␣共f兲 is given by

␣共f兲 = a1f for f ⬍ f1 共15兲

冉
␣共f兲 = a1 + 共a2 − a1兲
共f − f1兲
共f2 − f1兲
f冊 for f1 ⬍ f ⬍ f2

␣共f兲 = a2f for f ⬎ f2 .

The use of the tapered cosine wave as an adjustment function has the advantage of
preserving the nonstationary character of the acceleration time series. However, this ad-
justment function introduces drift to the velocity and displacement time series as shown
in Figure 1. This drift is due to the functional form of the adjustment function and not to
numerical errors. It requires applying baseline correction to the adjusted acceleration
time series.

IMPROVED TAPERED COSINE WAVELET

In order to prevent drift in the adjusted velocity and displacement time series, an
adjustment wavelet was chosen such that its functional form integrates to zero velocity
and displacement. Moreover, to ensure speed and efficiency of the solution, the adjust-
ment function was chosen such that the elements of the C matrix can be computed ana-
lytically. The improved adjustment function therefore consists of a cosine function ta-
pered with a Gaussian function. The improved tapered cosine wavelet differs from the
corrected tapered cosine wavelet proposed in Hancock et al. (2006) in that the wavelet
presented herein has a functional form that readily integrates to zero velocity and dis-
placement without including any baseline correction in its functional form. Contrary to
the wavelets proposed in Hancock et al. (2006), the improved tapered cosine wavelet
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 609

Figure 2. Acceleration, velocity and displacement time series of the improved tapered cosine
wavelet.

allows the elements of the C matrix given in Equation 7 to be determined analytically,

ensuring speed, stability, and efficiency of the solution. The improved tapered cosine
wavelet, shown in Figure 2, can be described by

冋冉
fj共t兲 = cos关␻⬘j 共t − tj + ⌬tj兲兴exp −
t − tj + ⌬tj
␥j
冊册
2
, 共16兲

where ␥j is a frequency dependent coefficient used to adjust the duration of the adjust-
ment function. ␥共f兲 was developed to ensure a smooth taper that would include several
cycles in the adjustment function and would therefore lead to zero velocity and displace-
ment at all frequencies. ␥共f兲 is given by

␥共f兲 = 1.178f−0.93 共17兲

The advantage of using the Gaussian taper in the adjustment wavelet is that it is
smooth and continuous. As a result, the wavelet ends with zero velocity and displace-
ment as shown in Figure 2 and no drift appears in the velocity and displacement time
series of the adjusted ground motion.

IMPROVED SOLUTION PROCEDURE

The spectral matching program, RspMatch, was revised to implement the new ta-
pered cosine wavelet and improved to ensure the stability and efficiency of the solution.
Moreover, the improved spectral matching method has the advantage of preserving fling
in the acceleration time series. A flow chart presenting the spectral matching procedure
610 L. AL ATIK AND N. ABRAHAMSON

as employed in the original version of RspMatch is presented in Hancock et al. (2006).

The program’s main features and workflow are largely unchanged in the new version
presented herein except for what is mentioned below.
In the original version of RspMatch, the user needs to run the program different
times for the same record and target spectrum in order to match different frequency
ranges of the spectrum. Following each run, the user is required to apply baseline cor-
rection to the resulting acceleration time series to correct for the drift in the velocity and
displacement time series before using the resulting time series as input to the subsequent
run. This process of applying baseline correction after each run is cumbersome and
time-consuming when multiple runs are needed for the generation of one spectrum com-
patible time series. The application of the improved tapered cosine function results in
acceleration time series that have zero final velocity and displacement. As a result, the
new version of RspMatch requires the user to specify the number of runs and the fre-
quency ranges for each run in the input file. The program will automatically output the
results of all the runs, which makes the process of spectral matching significantly faster.
Other features that were added to RspMatch are discussed in the following sections.

PSEUDOSPECTRAL ACCELERATION
For direct displacement-based design as well as the design of long-period structures,
spectral displacements rather than spectral accelerations are required. Since spectral dis-
placements are directly related to the pseudospectral accelerations, matching time series
to a target pseudo-acceleration spectrum should be performed. The new version of Rsp-
Match was, therefore, revised to allow pseudospectral matching. The original algorithm
proposed by Lilhanand and Tseng (1987, 1988) was modified to use the pseudo-
acceleration impulse response of a single-degree-of-freedom oscillator instead of the
true acceleration impulse response presented in Equation 5. The pseudo-acceleration im-
pulse response function used to determine the acceleration response of the adjustment
time series ␦a共t兲 is given by

− ␻i
hi共t兲 =
冑1 − ␤2i exp共− ␻i␤it兲sin共␻⬘i t兲 共18兲

DYNAMIC PADDING OF ACCELERATION TIME SERIES

The improved tapered cosine wavelet used in generating spectrum compatible accel-
eration time series results in no drift in the velocity and displacement time series when
the duration of the wavelet is long enough such that it starts with zeroes. In such cases,
the areas under the positive and negative cycles of the acceleration time series cancel out
and drift is not observed in the resulting velocity and displacement records. This, how-
ever, depends on the time, tj, at which the response of the acceleration time series peaks
for a certain frequency and damping pair. The frequency-dependent coefficient of the
improved tapered cosine function, ␥j, was selected to ensure that drift in the velocity and
displacement time series of the wavelet does not occur for most tj values at different
frequencies. The minimum tj values required to prevent drift at different frequencies are
shown in Figure 3. When tj is lower than the minimum value at a certain frequency, the
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 611

Figure 3. Relation between frequency and minimum tj values.

improved tapered cosine wavelet does not start with zero and therefore does not inte-
grate to zero velocity and displacement. For a frequency of 0.1 Hz, Figure 3 shows that
drift occurs when tj is less than 27.5 sec. The minimum tj value required to prevent drift
decreases as the frequency increases. For frequencies greater than 1 Hz, drift does not occur
for tj greater than about 2 sec.
In some cases, the response of the acceleration time series at low frequency peaks at
tj smaller than the minimum value defined in Figure 3. For such cases, the new version
of RspMatch allows zero-padding the beginning of the acceleration time series such that
tj becomes equal to the minimum value at the specific frequency. The new version of
RspMatch automatically checks that tj value at a certain frequency is greater than the
corresponding minimum value and zero-pads the beginning of the acceleration time se-
ries if necessary. This precludes any drift in the resulting velocity and displacement time
series.
When the response of the acceleration time series peaks at high tj values relative to
the duration of the record, the improved tapered cosine wavelet does not end with zero.
Although the spectrum compatible acceleration, velocity and displacement time series
might end with nonzero values, drift does not occur for such cases. A taper can be ap-
plied outside RspMatch to bring the end of the resulting acceleration, velocity and dis-
placement time series to zero.

EXAMPLE OF STEP-BY-STEP MATCHING PROCESS

Figure 4 shows the initial acceleration, velocity, and displacement of a recorded time
series. The goal is to match the response spectrum of this initial time series, shown as
the black line in Figure 5, to the target response spectrum, shown as the grey line in the
same figure, while maintaining realistic and similar acceleration, velocity, and displace-
ment profiles.
612 L. AL ATIK AND N. ABRAHAMSON

Figure 4. Initial acceleration, velocity, and displacement of a recorded time series.

Since the short period spectral acceleration is influenced by long period wavelets,
only the short period range of the response spectrum is matched in the first pass. Match-
ing the long period range of the response spectrum occurs in the next passes. In the first

Figure 5. Initial response spectrum plotted against target response spectrum.

AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 613

Figure 6. Response spectrum of the adjusted time series after the first pass plotted against tar-
get spectrum.

pass, the response spectrum is matched only up to a period of 1 sec. The initial spectral
acceleration is scaled to match the target peak ground acceleration. The scaling flag is turned
off for subsequent passes. As shown in Figure 6, the first pass brings the initial spectrum very
close to target in the specified frequency range for matching (1 to 35 Hz for the first pass).
The spectrum of the new adjusted time series is shown as the solid black line. The adjusted
time series, which the new spectrum represents, is shown in the top frame of Figure 7. Note
that the displacement profile of the adjusted time series, shown in the lower frame of the
same figure, maintains the same general characteristics of the original recording shown in
grey. No drift is observed in the velocity and displacement time series. Since the new spec-
trum is approaching the target spectrum and the characteristics of the new time series are
consistent with the initial time series, we proceed to the second pass.
In the second pass, the spectrum is matched out to a period of 2 sec. Note that the
period range for matching is extended out to longer periods progressively. The resulting spec-
trum from the second pass is a very good match to the target spectrum. The resulting velocity
and displacement time series maintain the same general characteristics of the initial time se-
ries but have different amplitudes. This iterative process is carried out until the spectrum has
been matched out to the desired period, in this case 10 sec. In Figure 8, we can see that this
occurred on the fourth pass, where the new computed spectrum matches the target spectrum
across the whole period range of interest. Checking the acceleration, velocity, and displace-
ment profiles of the computed spectrum, Figure 9 shows that the nonstationary characteris-
tics of the initial time series have been maintained. Though the amplitude and timing may be
slightly off, the general characteristics of the initial time series have been preserved through-
out the spectral matching process. We also note in Figure 9 that the adjusted time series are
shifted to the right as a result of the zero-padding at the beginning of the acceleration time
series.
As a final check, we compare the Fourier amplitude and the normalized Arias inten-
sity of the initial time series and the final computed time series. In Figure 10, we see that
the Fourier amplitude of the adjusted time series maintains the same general character-
614 L. AL ATIK AND N. ABRAHAMSON

Figure 7. Acceleration, velocity, and displacement of the adjusted time series after the first pass
plotted against the original time series.

Figure 8. Response spectrum of the adjusted time series after the last pass plotted against target
spectrum.
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 615

Figure 9. Acceleration, velocity, and displacement of the adjusted time series after the last pass
plotted against the original time series.

Figure 10. Comparison of the Fourier amplitudes of the initial and adjusted time series.
616 L. AL ATIK AND N. ABRAHAMSON

Figure 11. Comparison of the normalized Arias intensities of the initial and adjusted time
series.

istics of that of the initial time series. Figure 11 shows a very small difference between
the normalized Arias intensities of the original and the computed acceleration time se-
ries.

CONCLUSION
An improved method for the generation of spectrum compatible acceleration time
series is presented in this paper. An improved tapered cosine wavelet has been developed
for the adjustment of recorded ground motions resulting in acceleration time series that
have no drift in the corresponding velocity and displacement profiles. As a result, the
new method does not require baseline correction of the adjusted record after each pass.
The application of the new wavelet ensures stability and convergence of the spectral
matching solution. The spectral matching program, RspMatch, has been updated to
implement the new wavelet solution. The new version of RspMatch enables records to
be adjusted to target pseudospectral accelerations and allows zero-padding the beginning
and/or the end of the initial record, if needed, to ensure zero final velocity and displace-
ment. The new version of the program is available upon request from the corresponding
author.

REFERENCES
Abrahamson, N. A., 1992. Non-stationary spectral matching, Seismol. Res. Lett. 63, 30.
Bazzurro, P., and Luco, N., 2006. Do scaled and spectrum-matched near-source records pro-
duce biased nonlinear structural responses?, in Proc. of the 8th U. S. National Conf. on
Earthquake Engin., San Francisco, California.
Hancock, J., Watson-Lamprey, J., Abrahamson, N. A., Bommer, J. J., Markatis, A., McCoy, E.,
and Mendis, R., 2006. An improved method of matching response spectra of recorded earth-
quake ground motion using wavelets, J. Earthquake Eng. 10, 67–89.
AN IMPROVED METHOD FOR NONSTATIONARY SPECTRAL MATCHING 617

Kaul, M. K., 1978. Spectrum-consistent time-history generation, ASCE J. Eng. Mech. EM4,
781–788.
Lilhanand, K., and Tseng, W. S., 1987. Generation of synthetic time histories compatible with
multiple-damping response spectra, SMiRT-9, Lausanne, K2/10.
Lilhanand, K., and Tseng, W. S., 1988. Development and application of realistic earthquake
time histories compatible with multiple damping response spectra, in Ninth World Conf.
Earth. Engin., Tokyo, Japan, Vol 2, 819–824.
Preumont, A., 1984. The generation of spectrum compatible accelerograms for the design of
nuclear power plants, Earthquake Eng. Struct. Dyn. 12, 481–497.
Suarez, L. E., and Montejo, L. A., 2003. Generacion de registros artificiales compatibles con un
espectro de respuesta mediante la transformada wavelet, in Proc. of II Congreso Nacional de
Ingenieria Sismica, Medellin.
Suarez, L. E., and Montejo, L. A., 2005. Generation of artificial earthquakes via the wavelet
transform, Int. J. Solids Struct. 42, 5905–5919.
(Received 16 April 2009; accepted 10 January 2010兲