Probabilistic Engineering Mechanics 74 (2023) 103495

Contents lists available at ScienceDirect

Probabilistic Engineering Mechanics

journal homepage: www.elsevier.com/locate/probengmech

Wavelet packets-based simulation of non-stationary multivariate ground

motions
Pan Wen a ,∗, Renjie Liu a , Ruizhi Wen b
a College of Civil Engineering, Yantai University, No. 30 Qingquan Road Yantai, Shandong 264000, China
b Key Laboratory of Earthquake Engineering and Engineering Vibration, Institute of Engineering Mechanics, China Earthquake Administration, China

ARTICLE INFO ABSTRACT

Keywords: This study proposes a novel approach to simulate non-stationary multivariate earthquake accelerograms for
Multivariate non-stationary ground motions the seismic analysis of extended structures. This approach is based on an extension of the traditional
Wavelet packet stochastic decomposition technique applied to the wavelet packet theory using the spectral representation
Stochastic decomposition technique
method. Using the advantages of a wavelet packet to identify the time and frequency characteristics of ground
Multi-support excitation
motion effectively, the generated recordings successfully reflected the non-stationarity in both the time and
Coherency function
frequency domains. Accordingly, the spatial correlation trend was consistent with the selected coherence
function. Compared to the ground motions observed using the traditional method, the proposed approach was
exemplified by simulating ensembles of ground motions at three points spaced 200 m apart. The multivariate
ground motions simulated by different methods were applied as multi-support excitations for the dynamic
time–history analysis of a 190 m continuous rigid-frame box girder bridge. Notable variations in structural
response were observed when considering the non-stationarity of the frequency domain, contrasting with the
cases where it was not taken into account. This study provides an effective approach for generating non-
stationary multivariate earthquake accelerograms, which is of great significance for structural response analysis
in engineering applications.

1. Introduction characteristics in both the time and frequency domains are reflected
in the evolutionary spectral density (EPSD) matrix [29–31]. Thráinsson
The analysis of extended structures requires ground motion input et al. [32] proposed models for the probabilistic distribution of phase
that considers variations at different points, which influence the seismic differences with different Fourier amplitudes. Based on earthquake
response of structures [1–5]. Despite a substantial increase in the parameters, Sato et al. [33,34] proposed a model that described the
number of observed recordings in recent decades, limitations persist mean and variance of the phase derivative. By aligning the probabilistic
due to the random nature of the data; therefore, simulated recordings distribution of the phase difference with these models, instead of being
are important in engineering applications. Several researchers have uniformly distributed, their simulated results effectively capture the
attempted to generate non-stationary multivariate earthquake accelero- non-stationary characteristics.
grams for engineering applications. These methods include the spec- In most of the abovementioned simulation approaches, the spatial
tral representation method [6–11], covariance matrix decomposition non-stationary ground motions are generated by the deterministic en-
method [12], regressive model method [13–15], and conditional simu- velope function [35–37] or phase difference model [38–43]. However,
lation [16–22]. The results of the aforementioned simulation methods considering that the envelope function and phase difference of an earth-
reflect the statistical properties of the target field. Hao et al. [6] quake accelerogram are closely related [44,45], the simulation results
first proposed a method for obtaining spatial ground motions using can be affected by separating their relationships. However, it may be
a trigonometric series sum. Considering the simplicity of simulation unrealistic to establish a model that contains information on both the
schemes or sites with varying conditions, further studies based on this envelope function and phase difference. A direct resolution method
method have been conducted [23–28]. However, these methods ignore uses a recorded time history as a parent accelerogram containing
the non-stationarity in the frequency domain because they assume that the phase and Fourier spectra. Because the non-stationarity of earth-
phase angles are uniformly distributed. This non-stationarity in the quake accelerograms is related to their temporal variations, the time-
frequency domain also affects the dynamic responses of long and large frequency transformation can be used to capture the non-stationary
structures [4,5]. To overcome this shortcoming, the non-stationary characteristics. The wavelet transform technique performs significantly

∗ Corresponding author.
E-mail addresses: [email protected] (P. Wen), [email protected] (R. Liu), [email protected] (R. Wen).

https://doi.org/10.1016/j.probengmech.2023.103495
Received 21 March 2023; Received in revised form 18 June 2023; Accepted 17 July 2023
Available online 21 July 2023
0266-8920/© 2023 Published by Elsevier Ltd.
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

better in capturing temporal variations in the frequency domain of Conversely, the earthquake accelerogram can be reconstructed from
ground motions than other time-frequency techniques, such as the the wavelet-packet coefficients using the following reverse transform:
short-time Fourier transform. Wavelet transforms have already been 𝑗 𝑁−𝑗
∑
2 2∑
used to generate non-stationary ground motions [46–52] or analyze 𝑥(𝑡) = 𝑖
𝑐𝑗,𝑘 𝑖
𝜓𝑗,𝑘 (𝑡) (4)
their non-stationarity characteristics [53,54]. 𝑖=1 𝑘=1
However, the wavelets at adjacent time and frequency axes are not where 2𝑁 represents the data points in the earthquake accelerogram.
orthogonal for a continuous wavelet transform. Although the wavelet The time interval (𝑑𝑡𝑤 ) and frequency interval (𝑑𝑓 𝑤 ) between the
basis function of the discrete wavelet transform (DWT) is orthogonal, centers of the adjacent wavelet packet coefficients are calculated as
the resolution is insufficient for high-frequency bands. The wavelet follows:
packet transform (WPT) was developed based on a wavelet transform
2𝑁 𝑑𝑡
that can capture the high-frequency components of ground motions. 𝑑𝑡𝑤 = = 2𝑗 𝑑𝑡 (5)
2𝑁−𝑗
Amiri et al. [55] used WPT to generate non-stationary ground motions
compatible with the target response spectrum. However, the character- 𝑓𝑁 1 1
𝑑𝑓𝑤 = = (6)
istics of ground motions, such as the magnitude, epicentral distance, 2𝑗 2𝑑𝑡 2𝑗
fracture propagation, and site conditions, are difficult to describe us- where dt represents the interval of earthquake accelerogram and 𝑓𝑁
ing only the target response spectrum model. Yamamoto et al. [56] represents the Nyquist frequency.
proposed a model to describe complex time-varying ground motions In this study, the stochastic model proposed by Yamamoto [56] is
using WPT, which contained 13 parameters related to seismological used to generate a non-stationary ground motion, as shown in Fig. 1(a)
parameters, such as earthquake magnitude, epicentral distance, and (𝑀 = 7, 𝑅𝑟𝑢𝑝 = 100 km, 𝑉𝑎𝑝𝑝 = 400 m∕s); Fig. 1(b) shows the
site condition. Huang and Wang [57] developed a method to generate corresponding Fourier spectrum. In this example, the time interval of
energy- and spectrum-compatible ground motions based on WPT. dt = 0.01 s separates the entire time series containing 2𝑁 = 16384
In this study, considering the model proposed by Yamamoto as the (𝑁 = 14) data points, and parameter j is chosen as 𝑗 = 8 to standardize
parent accelerogram, a wavelet-packet-based approach is proposed to the process. Therefore, the time interval (𝑑𝑡𝑤 ) and frequency interval
(𝑑𝑓 𝑤 ) between the centers of adjacent wavelet packets coefficients are
generate the non-stationary multivariate ground motions that matched
2.56 s and 0.1953 Hz, respectively, from Eqs. (5) and (6). Fig. 1(c) illus-
the characteristics of the parent accelerograms and assumed coherence
trates the wavelet packet coefficient spectrum, showing the time- and
function. In this approach, the WPT is combined with a traditional
frequency-domain distributions of the absolute values of the wavelet
spectral representation method, as reported by Deodatis et al. [7].
packet coefficients |𝑐𝑗,𝑘
𝑖 | for the given earthquake accelerogram, which
The proposed approach is illustrated by generating three correlated
demonstrates the non-stationary characteristics in both the time and
ground motions for stations 200 – 400 m apart. The generated results
frequency domains. Fig. 2 shows the reconstructed time series of the
are validated by comparing them with those obtained using the tradi-
acceleration presented in Fig. 1(a) based on the inverse WPT (see (4)).
tional method. Finally, a continuous rigid-frame bridge is taken as an
To successfully verify the accuracy of the reconstructed result, the time
example to conduct a structural dynamic response analysis using the
domain between 50 – 80 s is captured and compared with the original
spatial ground motions generated by different methods as multi-support
ground motion. Considering that the wavelet packet spectrum contains
excitations.
information about the original ground motion, the reconstructed result
mostly coincides with the original ground motion (see Fig. 3).
2. Wavelet-packet transform
3. Simulation produce of spatial non-stationary ground motions
The WPT can decompose an earthquake accelerogram into a set
of wavelet packet coefficients localized in the time and frequency The most widely used spectral representation method for generating
domains. The WPT is defined as follows: non-stationary spatial ground motions was proposed by Deodatis [7].
∞ In this process, the evolutionary power spectral density (EPSD) function
𝑖 𝑖
𝑐𝑗,𝑘 = 𝑥(𝑡)𝜓𝑗,𝑘 (𝑡)𝑑𝑡 (1) matrix of multivariate ground motions was defined as follows, p, 𝑞 = 1,
∫−∞
2, . . . , n:
where x(t ) represents the earthquake accelerogram, 𝑐𝑗,𝑘 𝑖 represents
the 𝑖th set of wavelet packet coefficients at the 𝑗th decomposition ⎡𝑆11 (𝜔, 𝑡) 𝑆12 (𝜔, 𝑡) … 𝑆1𝑛 (𝜔, 𝑡)⎤
⎢ ⎥
level and 𝑘th translation parameter. The parameters j and k indicate ⎢𝑆21 (𝜔, 𝑡) 𝑆22 (𝜔, 𝑡) … 𝑆2𝑛 (𝜔, 𝑡)⎥
the frequency and time axis resolutions, respectively, the parameter i 𝐒(𝜔, 𝑡) = ⎢ ⎥ (7)
⎢ … … … ⎥
indicates the location of the wavelet packet in frequency resolution j. ⎢ ⎥
𝑖 represents the wavelet packet basis function, which is chosen as the
𝛹𝑗,𝑘 ⎢ ⎥
⎣𝑆𝑛1 (𝜔, 𝑡) 𝑆𝑛2 (𝜔, 𝑡) … 𝑆𝑛𝑛 (𝜔, 𝑡)⎦
Meyer wavelet function in this study owing to its orthogonal nature and
Here, the 𝑆𝑝𝑝 (𝜔, t ) in diagonal term denotes the auto EPSD of each
good localization in both the time and frequency axes [58]. Its explicit
point and 𝑆𝑝𝑞 (𝜔, t ) denotes the corresponding cross EPSD between
expression is stated as follows:
points p and q, which is defined as follows:
⎧ 𝜔 𝜋 3 2 4 √ √
⎪ exp(𝑖 2 ) sin[ 2 𝜈( 2𝜋 |𝜔| − 1)] 3
𝜋 ≤ |𝜔| ≤ 𝜋
3 | | 𝑑𝑝𝑞
⎪ 𝑆𝑝𝑞 (𝑖𝜔) = 𝑆𝑝𝑝 × 𝑆𝑞𝑞 × 𝛾𝑝𝑞 (𝑖𝜔) = 𝑆𝑝𝑝 × 𝑆𝑞𝑞 × |𝛾𝑝𝑞 (𝑖𝜔)| × exp(−𝑖𝜔 )
| | 𝜐𝑎𝑝𝑝
𝑖 ⎪ 𝜔 𝜋 3 4 8
𝜓𝑗,𝑘 (𝜔) = ⎨exp(𝑖 ) cos[ 𝜈( |𝜔| − 1)] 𝜋 ≤ |𝜔| ≤ 𝜋 (2)
⎪ 2 2 4𝜋 3 3 (8)
⎪
⎪ 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 where 𝛾(i𝜔) is the complex coherence function, 𝑑𝑝𝑞 represents the
⎩ surface distance between points p and q, and 𝑣𝑎𝑝𝑝 is the surface apparent
where 𝜈(x) is differentiable infinitely: wave velocity.
{ S(𝜔, t ) used in Eq. (7) denotes the positive definite Hermitian
0 𝑥≤0
𝜈(𝑥) = matrix, which can be processed by Choleskey decomposition into a
1 𝑥≥1 (3) lower triangular matrix L(𝜔, t ) and upper triangular matrix L𝐻 (𝜔, t ).
𝜈(𝑥) + 𝜈(1 − 𝑥) = 1 (−∞ < 𝑥 < ∞) 𝐒(𝜔, 𝑡) = 𝐋(𝜔, 𝑡) × 𝐋𝐻 (𝜔, 𝑡) (9)

2
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 1. The information for the given ground motion obtained from the stochastic model (Yamamoto 2013): (a) waveform of earthquake accelerogram, (b) Fourier spectrum of
earthquake accelerogram, (c) wavelet packet spectrum for time series.

Fig. 2. The reconstructed and original ground motions obtained from the stochastic model (Yamamoto 2013).

To consider the non-stationary information of parent ground mo- follows:

tions, the power spectral density matrix S is suggested to be expressed 1
𝑆(𝑓𝑖 , 𝑡) = [𝑥𝑖 (𝑡)]2 (𝑖 = 1, 2, … , 2𝑗 ) (11)
by a series of instantaneous spectral density matrices called EPSD. It 𝛥𝑓𝑖 ⋅ 𝑇
can be estimated using a wavelet packet transform to leverage on its
where T represents the duration of ground motion, 𝛥𝑓𝑖 = f /2𝑗 .
function of capturing the non-stationarity characteristics of earthquake
The total energy of the earthquake time history x(t ) is defined as
accelerograms. Next, we discussed generating spatial non-stationary
the sum of the squared wavelet packet coefficients as follows:
ground motions, followed by a combination of the spectral representa-
𝑗 𝑁−𝑗
tion process with the wavelet packet transform. Therefore, confirming ∑
2 2∑
| 𝑖 |2
the relationship between the EPSD and wavelet packet coefficient is key 𝑥(𝑡)2 𝑑𝑡 = |𝑐𝑗,𝑘 | (12)
∫ | |
𝑖 𝑘
to generating non-stationary earthquake accelerograms.
According to Eq. (4), the earthquake accelerogram can be divided For each frequency 𝑓𝑖 , the corresponding energy of the sub-signal
into 2𝑗 adjacent and non-overlapping sub-frequency bands. The 𝑖th cor- can also be represented by the sum of the squared wavelet packet
responding signal is defined as follows, whose frequency band ranges coefficients as follows:
between ((i-1)/2𝑗 )f and (i/2𝑗 )f : 𝑁−𝑗
2∑
| 𝑖 |2
[𝑥𝑖 (𝑡)]2 𝑑𝑡 = |𝑐𝑗,𝑘 | (13)
𝑁−𝑗
2∑ ∫ | |
𝑘
𝑖 𝑖 𝑖 𝑗
𝑥 (𝑡) = 𝑐𝑗,𝑘 𝜓𝑗,𝑘 (𝑡) (𝑖 = 1, 2, … , 2 ) (10)
𝑘=1 Considering the similarity between the integrals and the sum,
| 𝑖 |2
Eq. (11) is equal to [𝑥𝑖 (𝑡)]2 = |𝑐𝑗,𝑘 | . This equation can be defined as
According to the definition of the power spectral density function, | |
| |2
the corresponding EPSD in each sub-frequency band is expressed as [𝑥(𝑡)] = |𝑐𝑗,𝑘 | for the total signal.
2
| |

3
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 3. Flowchart of proposed methodology.

Multiplying both sides of the equation by 1∕(𝛥𝑓𝑖 ⋅𝑇 ), the relationship

between the EPSD and wavelet packet coefficients is expressed as
follows:
1 1 | |2
𝑆(𝑓𝑖 , 𝑡) = ⋅ [𝑥(𝑡)]2 = ⋅ |𝑐 | (14) Fig. 4. Configuration of each point on the ground surface.
𝛥𝑓𝑖 ⋅ 𝑇 𝛥𝑓𝑖 ⋅ 𝑇 | 𝑗,𝑘 |
Similarly, the cross-spectral density function and cross-EPSD func-
tion for different earthquake accelerograms are expressed as follows for of the instantaneous EPSD matrix. Then, these coefficients are inverse-
each sub-frequency. transformed to generate the corresponding earthquake accelerograms
2∑ 𝑁−𝑗 for each point (see Eq. (4)).
𝑖 1 Finally, the iterative process to adjust the wavelet packet coef-
𝑆𝑚𝑛 (𝑓𝑖 ) = (𝑐𝑚𝑖𝑗,𝑘 𝑐𝑛𝑖𝑗,𝑘 ) (15a)
𝛥𝑓𝑖 ⋅ 𝑇 𝑘=1 ficients of the earthquake accelerograms to ensure their consistency
𝑖 1 with the target spectrum of the parent ground motions is expressed as
𝑆𝑚𝑛 (𝑓𝑖 , 𝑡) = ⋅ (𝑐𝑚𝑖𝑗,𝑘 ⋅ 𝑐𝑛𝑖𝑗,𝑘 ) (15b) follows:
𝛥𝑓𝑖 ⋅ 𝑇
√
where the wavelet packet coefficients of points m and n are denoted as 𝑖,mod 𝑆𝑇 𝑖
𝑐𝑗,𝑘 = 𝑐 (20)
𝑐𝑚𝑗,𝑘 and 𝑐𝑛𝑗,𝑘 respectively, to distinguish different ground motions. 𝑆 𝑗,𝑘
The (m, n)th element of the EPSD matrix, according to Choleskey, where 𝑆 𝑇 is the target power spectrum and S is the power spectrum of
from Eq. (9) can be expressed as follows [59]: the simulated ground motion. The power spectrum of the generated
∑
𝑚 ground motion matches the target spectrum well after n iterations.
𝑺 𝑚𝑛 (𝑓 ) = 𝐿𝑚𝑟 (𝑓 ) ⋅ 𝐿𝐻
𝑛𝑟 (𝑓 ) (16) Typically, after two or three iterations, a good match is achieved.
𝑟=1 A flowchart of the process of generating non-stationary multivari-
Term-by-term extraction of the right-hand-side expressions of ate ground motions based on the wavelet packet transform is shown
Eqs. (16) and (15b) yields the following: in (3).

1 1 ∑ 𝑚
√ 𝑐𝑚𝑖𝑗,𝑘 ⋅ √ 𝑐𝑛𝑖𝑗,𝑘 = 𝐿𝑚𝑟 (𝑓 ) ⋅ 𝐿𝐻
𝑛𝑟 (𝑓 ) (17) 4. Numerical example of the proposed procedure
𝛥𝑓 ⋅ 𝑇 𝛥𝑓 ⋅ 𝑇 𝑟=1

Assuming that 𝑛 = 𝑚, Eq. (17) is expressed as follows: The proposed approach is illustrated by generating non-stationary
multivariate motions at three different stations with non-stationarity
1 1 ∑
𝑚
characteristics of the parent accelerogram (see Fig. 1(a)). The configu-
√ 𝑐𝑚𝑖𝑗,𝑘 ⋅ √ 𝑐𝑚𝑖𝑗,𝑘 = 𝐿𝑚𝑟 (𝑓 ) ⋅ 𝐿𝐻
𝑚𝑟 (𝑓 ) (18)
𝛥𝑓 ⋅ 𝑇 𝛥𝑓 ⋅ 𝑇 𝑟=1 ration of the three points is illustrated in Fig. 4, where point 1 belongs
to the parent accelerogram. The coordinate values are 0 m, 200 m, and
The wavelet packet coefficients of point m can be generated as
400 m. The Loh-lagged coherence model fitted by the recordings of the
follows:
SMART-1 array is selected for this example [60].
∑
𝑚
√
𝑐𝑚𝑖𝑗,𝑘 = 𝛥𝑓𝑖 ⋅ 𝑇 ⋅ 𝐿𝑚𝑟 (19) 𝜔𝑑
|𝛾(𝜔, 𝑑)| = exp(−0.125 ) (21)
𝑟=1 2𝜋𝜈𝑎𝑝𝑝
Observe from Eq. (19) that the process of generating non-stationary Here, d represents the distance between different points, and the ap-
multivariate earthquake accelerograms involves estimating the wavelet parent propagation velocity is set to 400 m/s, which is equal to that of
packet coefficients of different points from the Cholesky compositions the stochastic model (Yamamoto 2013). Using the parent accelerogram

4
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 5. Comparison between the target and generated power spectra after the iterative process for each point.

and coherence model, the non-stationary multivariate ground motions using the traditional spectral representation method. The configuration
are generated in this section. of each point is shown in Fig. 4. The apparent propagation veloc-
The process of generating ground motions is made spectrum- ity and coherence model are also equivalent to those mentioned in
compatible after the iterative process to be consistent with the power the previous section. The simulated time series and corresponding
spectrum of the parent accelerogram (see Fig. 1(a)). The corresponding wavelet packet spectra for different points after the iterative pro-
iteration process at each point is shown in Fig. 5, where the solid gray cess are shown in Fig. 9. Observe that the simulated time series is
line represents the initial simulated power spectrum before the iteration similar to the parent accelerogram. However, the non-stationary char-
process, red solid line represents the simulated power spectrum after acteristics in the frequency domain are hardly evident in the wavelet
two iterations using Eq. (20), and the black dotted line represents the packet spectrum, which differs from the distribution shown in Fig. 4(a).
power spectrum of the parent accelerogram (see Fig. 1(a)). Observe
This phenomenon indicates that the method proposed in this study
that the power spectrum after iteration accurately matches the power
is more advantageous for simulating ground motions with obvious
spectrum of the parent accelerogram. Considering that the wavelet
defects in the frequency domain compared to the traditional spectral
packet coefficients of point 1 contain information regarding the par-
representation method.
ent accelerogram, its power spectrum matches well with the power
The number of zero crossings over time for the ground motions
spectrum of the parent accelerogram until the end of the iterative
process. simulated using different methods is illustrated in Fig. 10. Observe that
The accuracy of the power spectrum obtained through the wavelet the increasing trend in proposed approach (see from Fig. 7) behaves
packet transform is verified by comparing it with the results obtained nonlinearly, which is consistent with the non-stationarity character-
through a Fourier transform. Fig. 6(a) demonstrates the high similarity istics of frequency domain. However, the increasing trend obtained
of the power spectrum of the parent accelerogram in different ways. using the traditional approach (see Fig. 9) behaves linearly, which
Furthermore, the effects of cross-power spectrum are illustrated in indicates that the spectral non-stationary behavior is not well repre-
Fig. 6(b–d). This also demonstrates the high similarity obtained using sented [61]. A significant difference is also observed when comparing
the traditional method (see (6)) and wavelet packet transform (see the Arias intensities [62] obtained using different methods, as shown in
(12)a). The above comparison proves the reliability of signal analysis Fig. 11. The total cumulative energy of the ground motions generated
based on wavelet packet transform. using the traditional spectral representation method is lower than
The wavelet packet spectra and time series for different points that generated using the method proposed in this study. The above
are illustrated in Fig. 7. The non-stationary characteristics are evident comparisons demonstrate that the traditional spectral representation
in both the time and frequency domains from the distribution of method can change both the inherent non-stationary characteristics in
the wavelet packet coefficients. Meanwhile, observe that the simu- the frequency domain and total energy of the parent accelerograms.
lated time series is very similar to the parent accelerogram (see from
Fig. 1(a)). The coherence functions are also computed for different
6. Application case study
points. Fig. 8 shows a comparison of the coherence functions values
with the selected lagged coherence model (see Eq. (19)). Observe that
In this section, a continuous rigid frame bridge is considered as
the computed coherence functions for different points broadly agree
an example to illustrate the structural dynamic response analysis con-
with the selected Loh coherence model.
ducted using the abovementioned generated multivariate ground mo-
5. Comparison with traditional approach tions in different ways. As shown in Fig. 12, this structure has three
spans with a width of 90 m + 200 m + 90 m. The two piers are
For comparison, this section considers the parent accelerogram marked as No. 1 and No. 2, respectively, for convenience of description.
used above to simulate a set of earthquake accelerogram ensembles The height of both piers is 7.7 m, where the left and right piers are

5
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 6. Comparison between the wavelet packet transform and Fourier transform: (a) the comparison of auto-power spectrum, (b-d) comparison of cross-power spectrum.

Fig. 7. The wavelet packet spectrum and time series of generated ground motions for the three target points obtained using the proposed approach.

Fig. 8. Comparison between the spatial coherence functions between the different points and selected Loh coherence model.

6
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 9. The wavelet packet spectrum and time series of generated ground motions for the three target points obtained using the traditional approach.

Fig. 10. Comparison between the number of zero crossings over time for the simulated ground motions obtained by the proposed approach and traditional approach.

Fig. 11. Comparison between the Arias intensity for the simulated ground motions obtained by the proposed approach and traditional approach.

Table 1 points 1 and 2 by the approach proposed in this study (see Fig. 7)
The first ten orders of the bridge modal periods.
are first used as the inputs from the bottom of No. 1 and No. 2
Mode Frequency (Hz) Mode Frequency (Hz)
piers, respectively. Meanwhile, the ground motions of points 1 and 2
1 0.332 6 2.184 simulated by the traditional spectral representation method (see Fig. 9)
2 0.443 7 2.87
are also used as multi-supports at the bottom of No. 1 and No. 2 piers,
3 0.737 8 3.242
4 1.317 9 3.704 respectively. The former contains non-stationary characteristics in the
5 1.495 10 5.495 frequency domain, whereas the latter contains nothing. In addition to
comparing the time-response curves of the longitudinal displacement
at the top of each pier, the bending moment at the bottom of each pier
and middle-span vertical displacement are also compared for different
marked as No. 1 and No. 2, respectively. Considering the simplicity of
this model, the soil-to-foundation effect is ignored in this study. The cases. As shown in Table 2, significant differences are observed at
FEM software is used to model the type of bridge abutment and upper different parts of the bridge with respect to different excitation cases.
partition of the pier as solid units. The first ten orders of the modal For some parts, the peak values of the time–history curves using multi-
periods of the bridge are listed in Table 1. support excitations that are non-stationary in the frequency domain are
The dynamic response of the bridge under multi-support excitation larger than those in the frequency domain. The differences between
is computed for two different cases. The ground motions obtained at the longitudinal displacement values of both the pier and bottom

7
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

Fig. 12. The finite element model of continuous rigid frame bridge: (a) 3D finite element model, (b) longitudinal section of the model, (c) cross-section of the model.

Fig. 13. Comparison between the time curves of structural response containing the frequency non-stationarity characteristic.

Table 2 7. Conclusions
The peak values of time–history curves under different cases.
Part of bridge Multivariate excitation This study proposes a new procedure to obtain spatial non-
Stationary in Non-stationary in stationary ground motions for a given parent accelerogram. The ac-
frequency domain frequency domain celerograms generated at different stations are consistent with the
Mid-span (mm) 3.95 −4.67 response spectrum of the parent accelerogram and incorporate the non-
No. 1 top pier (mm) −2.81 2.86
stationarity characteristics in both the time and frequency domains.
No. 2 top pier (mm) −2.95 −2.82
No. 1 bottom pier (103 KN m) 23.97 39.52
Meanwhile, the calculated coherence functions at any two stations
No. 2 bottom pier (103 KN m) 47.32 48.33 agree with the target coherence model. It is assumed that the power
spectrum of the parent accelerogram can represent each ground motion
generated at a uniform site.
The method proposed in this study is based on the traditional
moments of No. 2 pier are not as obvious as the structural dynamic spectral representation process, and the non-stationary multivariate
responses observed at other positions. The peak values of the longitu- ground motions are performed through a wavelet packet transform.
dinal displacement for No. 2 top pier are slightly smaller than those The wavelet packet transform can capture the amplitude and phase
obtained using multi-support excitations containing stationarity in the information of an earthquake accelerogram, and this signal can be de-
frequency domain. This phenomenon may have been caused by the composed into component-time histories containing energy character-
cumulative energy difference, as shown in Fig. 11. Observe that the istics depicted in non-overlapping frequency bands. Consequently, an
total cumulative energy of the earthquake accelerograms generated by expression for the EPSD is derived using the wavelet packet coefficients.
the proposed approach is larger than that generated by the traditional To demonstrate the proposed approach, multivariate ground mo-
approach. The time history curves of the structural response under tions are generated at three stations spaced 200 meters apart. These
different cases are illustrated in Fig. 13. Therefore, it is necessary to ground motions align with the power spectrum and non-stationary
consider the non-stationarity characteristics of the frequency domain characteristics of a specific parent accelerogram from the Yamamoto
when using the method proposed in this study to generate multivariate stochastic model. The simulated ground motions at different stations
earthquake accelerograms and apply them to large-span engineering exhibit the non-stationarity characteristics present in both the time and
input. frequency domains of the parent accelerogram and the specified coher-
Note that these differences may not be observed in all types of ence function model. Additionally, the generated results are compared
bridges; however, this reminds us that multiple supports containing with those of the traditional method to underline the advantages of the
non-stationarity in the frequency domain can influence the dynamic method proposed in this study. Considering a continuous rigid frame
structural responses by disturbing the cumulative energy of the gener- bridge as an example, the structural dynamic responses are compared
ated earthquake accelerograms. using the above-generated multivariate ground motions obtained from

8
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

different methods. Incorporating non-stationarity characteristics in the [17] E.H. Vanmarcke, E. Heredia-Zavoni, G.A. Fenton, Conditional simulation of
frequency domain results in larger peak values in certain sections of spatially correlated earthquake ground motion, J. Eng. Mech. 119 (11) (1993)
2333–2352, ASCE.
the bridge as observed in the time-response curves.
[18] S. Jin, L.D. Lutes, S. Sarkani, Efficient simulation of multidimensional random
It is important to use the proposed approach to generate spatial fields, J. Eng. Mech. 123 (10) (1997) 1082–1089, ASCE.
non-stationary earthquake accelerograms containing information on [19] E. Heredia-Zavoni, S. Santa-Cruz, Conditional simulation of a class of non-
the target parent accelerograms, which can be utilized in site-specific stationary space–time random fields, J. Eng. Mech. 126 (4) (2000) 398–404,
seismic hazard analyses and engineering applications. ASCE.
[20] R. Jankowski, K. Wilde, A simple method of conditional random field simulation
of ground motions for long structures, Eng. Struct. 22 (5) (2000) 552–561.
Declaration of competing interest [21] S. Liao, Zerva A. Physically-compliant, Conditionally simulated spatially variable
seismic ground motions for performance-based design, Earthq. Eng. Struct. Dyn.
The authors declare that they have no known competing finan- 35 (7) (2006) 891–919.
[22] A. Tamhidi, N. Kuehn, S.F. Ghahari, Conditioned simulation of ground-motion
cial interests or personal relationships that could have appeared to
time series at uninstrumented sites using Gaussian process regression, Bullet.
influence the work reported in this paper. Seismol. Soc. Am. 112 (1) (2022) 331–347.
[23] A.A. Shama, Simplified procedure for simulating spatially correlated earthquake
Data availability ground motions, Eng. Struct. 29 (2) (2007) 248–258.
[24] K. Bi, H. Hao, Modelling and simulation of spatially varying earthquake ground
motions at sites with varying conditions, Probabilistic Eng. Mech. 29 (2012)
No data was used for the research described in the article.
92–104.
[25] P. Cacciola, I. Zentner, Generation of response-spectrum-compatible artifi-
Acknowledgments cial earthquake accelerograms with random joint time–frequency distributions,
Probab. Eng. Mech. 28 (2012) 52–58.
Professor Jack Baker is acknowledged for sharing the source code [26] M.D. Shields, Simulation of spatially correlated nonstationary response spectrum-
compatible ground motion time histories, J. Eng. Mech. 141 (6) (2015)
for WPT-based stochastic ground motion simulation at http://www. 04014161.
stanford.edu/~bakerjw/gm_simulation.html. [27] Y.X. Wu, N.Gao.Y.F. Zhang, D.Y. Li, Simulation of spatially varying ground
motions in V-shaped symmetric canyons, J. Earthq. Eng. 20 (6) (2016) 992–1010.
Funding [28] Y.X. Wu, Y.F. Gao, D.Y. Li, Simulation of spatially correlated earthquake ground
motions for engineering purposes, Earthq. Eng. Eng. Vibr. 10 (2011) 163–173.
[29] J.P. Conte, B.F. Peng, Fully nonstationary analytical earthquake ground-motion
This work was partly supported by the Chinese National Natural model, J. Eng. Mech. 123 (1) (1997) 15–24.
Science Fund (grant number: 51908518). [30] C. Vlachos, G. Konstantinos, Deodatis G. Papakonstantinou, A multi-modal
analytical non-stationary spectral model for characterization and stochastic
Ethics approval and consent to participate: This study was approved simulation of earthquake ground motions, Soil Dyn. Earthq. Eng. 30 (2016)
by Yantai University and Institute of Engineering Mechanics, China 177–191.
Earthquake Administration. All authors permit the researches in this [31] D. Wang, Z.L. Fan, S.W. Hao, D.H. Zhao, An evolutionary power spectrum model
paper. of fully nonstationary seismic ground motion, Soil Dyn. Earthq. Eng. 105 (2018)
1–10.
[32] H. Thráinsson, A.S. Kiremidjian, Simulation of digital earthquake accelerograms
References using the inverse discrete Fourier transform, Earthq. Eng. Struct. Dyn. 31 (12)
(2002) 2023–2048.
[1] K. Soyluk, Comparison of random vibration methods for multi-support seismic [33] T. Sato, Y. Murono, H.B. Wang, Design spectra and phase spectrum modeling to
excitation analysis of long-span bridges, Eng. Struct. 26 (11) (2004) 1573–1583. simulate design earthquake motions: a case study through design standards of
[2] Y. Bilici, A. Bayraktar, K. Soyluk, Stochastic dynamic response of dam–reservoir– railway facilities in Japan, J. Nat. Disaster Sci. 23 (2) (2001) 89–100.
foundation systems to spatially varying earthquake ground motions, Soil Dyn. [34] T. Sato, Y. Murono, A. Nishimura, Phase spectrum modeling to simulate design
Earthq. Eng. 29 (3) (2009) 444–458. earthquake motion, J. Nat. Disaster Sci. 24 (2) (2002) 91–100.
[3] Y.H. Zhang, Q.S. Li, J.H. Lin, Random vibration analysis of long-span structures [35] M. Shinozuka, Y. Sato, Simulation of nonstationary random process, J. Eng.
subjected to spatially varying ground motions, Soil Dyn. Earthq. Eng. 29 (4) Mech. Div. 93 (1) (1967) 11–40.
(2009) 620–629. [36] M. Amin, A. Ang, Nonstationary stochastic models of earthquake motions, J. Eng.
[4] Y. Li, J.P. Conte, M. Barbbato, Influence of time-varying frequency content in Mech. Div. 94 (2) (1968) 559–584.
earthquake ground motions on seismic response of linear elastic systems, Earthq. [37] P. Jennings, G. Housner, C. Tsai, Simulated earthquake motions for design
Eng. Struct. Dyn. 45 (8) (2016) 1271–1291. purpose, in: Proceeding 4th International Conference on Earthquake Engineering,
[5] F. Basone, L. Cavaleri, G. Muscolino, Incremental dynamic based fragility as- Santiago, 1969, A-1.
sessment of reinforced concrete structures: Stationary vs non-stationary artificial [38] V. Montaldo, A.S. Kiremidjian, H. Thrainsson, Simulation of the Fourier phase
ground motions, Soil Dyn. Earthq Eng. 103 (2017) 105–117. spectrum for the generation of synthetic accelerograms, J. Earthq. Eng 7 (3)
[6] H. Hao, C.S. Oliveira, J. Penzien, Multiple-station ground motion processing and (2003) 427–445.
simulation based on smart-1 array data, Nucl. Eng. Des. 111 (3) (1989) 293–310. [39] D. Boore, Phase derivatives and simulation of strong ground motions, Bullet.
[7] G. Deodatis, Simulation of ergodic multivariate stochastic processes, J. Eng. Seismol. Soc. Am. 93 (3) (2003) 1132–1143.
Mech. 122 (8) (1996) 778–787, ASCE. [40] M. Shrikhande, V.K. Gupta, On the characterization of the phase spectrum for
[8] A. Zerva, Response of multi-span beams to spatially incoherent seismic ground strong motion synthesis, J. Earthq. Eng. 5 (4) (2012) 465–482.
motions, Earthq. Eng. Struct. Dyn. 19 (6) (1990) 819–832. [41] D.Y. Zhang, W. Liu, W.C. Xie, Modeling of spatially correlated site-reflected,
[9] A. Zerva, Seismic ground motion simulations from a class of spatial variability and nonstationary ground motions compatible with response spectrum, Soil Dyn.
models, Earthq Eng. Struct. Dyn. 21 (4) (1992) 351–361. Earthq. Eng. 55 (2013) 21–32.
[10] A. Zerva, On the spatial variation of seismic ground motions and its effects on [42] L. Wang, M. Mccullough, Modeling and simulation of nonstationary processes
lifelines, Eng. Struct. 16 (7) (1994) 534–546. utilizing wavelet and Hilbert transforms, J. Eng. Mech. 140 (2) (2014) 345–360.
[11] M. Shrikhande, V.K. Gupta, Synthesizing ensembles of spatially correlated [43] K. Nagao, J. Kanda, Study of a ground-motion simulation method using a
accelerograms, J. Eng. Mech. 124 (11) (1998) 1185–1192, ASCE. causality relationship, J. Earthq. Eng. 18 (6) (2014) 891–907.
[12] A. Zerva, L.S. Katafygiotis, Selection of simulation scheme for the nonlinear [44] Y. Ohsaki, On the significance of phase content in earthquake ground motions,
seismic response of spatial structures, in: Proceedings of the fourth international Earthq. Eng. Struct. Dyn. 7 (5) (1979) 427–439.
colloquium on computation of shell and spatial structures, Chania (Greece), 2000, [45] N.C. Nigam, Phase properties of a class of random processes, Earthq. Eng. Struct.
IASS-IACM 2000. Dyn. 10 (5) (1982) 711–717.
[13] G. Deodatis, M. Shinozuka, Auto-regressive model for nonstationary stochastic [46] S. Mukherjee, V.K. Gupta, Wavelet-base generation of spectrum-compatible
processes, J. Eng. Mech. 114 (11) (1988) 1995–2012. time-histories, Soil Dyn. Earthq. Eng. 22 (9) (2002) 799–804.
[14] J.P. Conte, K.S. Pister, Nonstationary ARMA modeling of seismic motions, Soil [47] S. Mukherjee, V.K. Gupta, Wavelet-base characterization of design ground
Dyn. Earthq. Eng. 11 (7) (1992) 411–426. motions, Earthq. Eng. Struct. Dyn. 31 (5) (2002) 1173–1190.
[15] T. Sato, H. Imabayashi, Real time conditional simulation of earthquake ground [48] J. Iyama, H. Kuwamura, Application of wavelets to analysis and simulation of
motion, Earthq. Eng. Eng. Seismol. 1 (1) (1999) 27–38. earthquake motions, Earthq. Eng. Struct. Dyn. 28 (3) (1999) 255–272.
[16] H. Kameda, H. Morikawa, An interpolating stochastic process for simulation of [49] K. Gurley, A. Kareem, Applications of wavelet transforms in earthquake wind
conditional random fields, Probab. Eng. Mech. 7 (4) (1992) 243–254. and ocean engineering, Eng. Struct. 21 (2) (1999) 149–167.

9
P. Wen, R. Liu and R. Wen Probabilistic Engineering Mechanics 74 (2023) 103495

[50] D. Cecini, A. Palmeri, Spectrum-compatible accelerograms with harmonic [56] Y. Yamamoto, W. Baker, Stochastic model for earthquake ground motion using
wavelets, Comput. Struct. 147 (2015) 26–35. wavelet packets, Bull. Seismol. Soc. Am. 103 (6) (2013) 3044–3056.
[51] A. Giaralis, P.D. Spanos, Wavelet-based response spectrum compatible synthesis [57] D. Huang, G. Wang, Energy-compatible and spectrum-compatible (ECSC) ground
of accelerograms-eurocode application (EC8), Soil Dyn. Earthq. Eng. 29 (1) motion simulation using wavelet packets, Earthq. Eng. Struct. Dyn. 46 (11)
(2009) 219–235. (2017) 1855–1873.
[52] K. Sarkar, K. Vinay, C. Riya, Wavelet-based generation of spatially correlated [58] Y. Meyer, Principle d’incertitude bases hilbertiennes et algebres d’operateurs,
accelerograms, Soil Dyn. Earthq. Eng. 87 (2016) 116–124. Séminaire Bourbaki 662 (1986) 209–223.
[53] P.D. Spanos, G. Failla, Evolutionary spectra estimation using wavelets, J. Eng. [59] M. Shrikhande, V.K. Gupta, Synthesizing ensembles of spatially correlated
Mech. 130 (8) (2004) 952–960. accelerograms, J. Eng. Mech. 124 (11) (1988) 1185–1192, ASCE.
[54] G. Huang, X. Chen, Wavelets-based estimation of multivariate evolutionary [60] C.H. Loh, Y.T. Yet, Spatial variation and stochastic modeling of seismic
spectra and its application to nonstationary downburst winds, Eng. Struct. 31 differential ground movement, Earthq. Eng. Struct. Dyn. 16 (4) (1988) 583–596.
(4) (2009) 976–989. [61] G.R. Saragoni, G.C. Hart, Simulation of artificial earthquakes, Earthq. Eng. Struct.
[55] G.G. Amiri, A. Shahjouei, S. Saadat, Hybrid evolutionary-neural network ap- Dyn. 2 (3) (1973) 249–267.
proach in generation of artificial accelerograms using principal component [62] A. Arias, A measure of earthquake intensity in Seismic Design for Nuclear Power
analysis and wavelet-packet transform, J. Earthq. Eng. 15 (1) (2011) 50–76. Plants, The MIT Press, Cambridge, MA, 1970, pp. 438–484,

10