You are on page 1of 8

Soil Dynamics and Earthquake Engineering 129 (2020) 105918

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering


journal homepage: http://www.elsevier.com/locate/soildyn

Empirical decomposition of seismic response of soft soils


antara *
S.R. Garcia, L. Alc�
Instituto de Ingeniería, Departamento de Geotecnia, Universidad Nacional Aut�
onoma de M�exico, Mexico

A R T I C L E I N F O A B S T R A C T

Keywords: In this research the use of the Hilbert-Huang Transform HHT, for the decomposition and characterization of
Hilbert-Huang transform seismic ground response, is examined. The HHT, integrated by the Empirical Mode Decomposition EMD and the
Empirical mode decomposition Hilbert Transformation HT, enables engineers to analyze data from nonlinear and nonstationary processes and to
Time series analysis
obtain more detailed intensity descriptions on time-varying frequency diagrams. The usefulness of the method,
Ground motions
Fundamental frequency
particularly the EMD, to analyze recordings of accelerations of soft-soil deposits in Mexico City, is presented.
Peak ground acceleration Conclusions of the analysis of accelerograms indicate that this adaptive decomposition permits to extract motion
characteristics, which might not be exposed effectively by other conventional data processing techniques.

1. Introduction They are approximately orthogonal to each other with respect to the
inner product and consent for the definition of instantaneous fre­
The assessment of ground motions due to potential earthquakes is quencies through complexifying each component by the Hilbert trans­
perhaps the most challenging problem in geotechnical-earthquake en­ form. Even though the EMD method is developed on the basis of an
gineering. Analysis of available behavioral data collected during seismic algorithm, and hence, it is still subject of debate and contributions to
phenomena, is the natural first step to approach this demand. Faithful aspects of its theoretical framework, for the study made on the set of
extraction and representation of motion characteristics improve the time series of accelerations, it is considered that the main drawbacks do
knowledge of the underlying physical process the numbers expose. not affect the results obtained and do not have a negative impact on the
Clear non-stationary characteristics are manifested in the waves practical conclusions derived from them. However, in any case, care was
detected in earthquake records. It is recognized that during any seismic taken that the resulting frequencies of the components were
event there are variations of the intensity of the ground motion with non-negative and it was recognized that the amplitudes are interpreted
time (acceleration, velocity or displacement). After the arrival of the as envelopes.
first seismic wave the intensity accumulates quickly till a maximum In this paper, we formulate an optimization problem which takes
value that is sustained (for some seconds) to later diminish gradually into account important features desired of the resulting EMD. Specif­
until it vanishes. Additionally exists a variation of the frequency content ically, we propose a data-adapted iterative method which minimizes in
with time with a tendency to move to lower frequencies as time pro­ each iteration step a smoothness functional subject to inequality con­
gresses, what is known as “frequency dependent dispersive effect” [1]. straints involving the extrema. In this way, our method constructs a
The evolution of the amplitude/energy/power characteristics and its sparse data-adapted basis for the input function as well as a mathe­
relation with the frequency content is usually studied by spectral rep­ matically stringent envelope for the function. Moreover, we present an
resentations in two-dimensional time-frequency domain [2–4]. Have optimization based normalization to extract instantaneous frequencies
been recognized that transformed domain and response spectral analysis from the analytic function approach. We present corresponding algo­
be ill with shortcomings when are applied to accelerations time series rithms together with several examples.
[5]. This study proposes the use of the Hilbert-Huang Transform HHT [6, Local wave decomposition methods play an important role in seismic
7], particularly one of its components the Empirical Mode Decomposi­ signal processing for their superiority in significantly revealing the fre­
tion EMD, as an alternative for finding useful features into seismic re­ quency content of a signal changes with time variation [12]. used
cordings [8–11]. The EMD was developed as an iterative method to matching-pursuit decomposition for instantaneous spectral analysis to
decompose a nonlinear and nonstationary univariate function additively detect low-frequency shadows beneath hydrocarbon basins. Spectral
into multiscale components, the called intrinsic mode functions IMFs. decomposition added to coherency was exploited by Ref. [13] for

* Corresponding author.
E-mail addresses: sgab@pumas.iingen.unam.mx (S.R. Garcia), LAlcantaraN@iingen.unam.mx, leonardo@pumas.ii.unam.mx (L. Alc�
antara).

https://doi.org/10.1016/j.soildyn.2019.105918
Received 13 October 2017; Received in revised form 19 June 2019; Accepted 24 October 2019
Available online 2 November 2019
0267-7261/© 2019 Elsevier Ltd. All rights reserved.
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

interpreting incised valleys [14]. operated windowed spectral analysis


to produce discrete-frequency energy cubes for applications in reservoir
characterization [15]. showed that an average frequency attribute,
produced from sine curve-fitting, correlates with shale volume in
particular regions [16]. described a data-adaptive method for
time-frequency broad mapping. The continuous-wavelet transform CWT
provides a different approach to time-frequency analysis. Instead of
producing a time-frequency spectrum, it produces a time-scale map
called a scalogram [17,18,19,and 20]. [21] proposed a time-frequency
analysis to identify Rayleigh and Love waves exploiting the Stockwell
Transform [22] coupled with filters based on the Normalized Inner
Product.
However, between the local wave decomposition methods (e.g. the
EMD, the Synchro-Squeezing Transform SST [23], and the Variational
Mode Decomposition VMD [24], the one that Huang et al. [7] for
meeting the requirements for the Hilbert transformation is the EMD.
The subject under analysis is a site in Mexico City, a megalopolis
where the lacustrine environment propitiated the deposition of large
volumes of fly ash and other pyroclastic materials that formed excep­
tional soft clays [25–27]. The seismic response of this location is Fig. 1. Hilbert-Huang Transform, the overall algorithm layout (modified
examined in order to achieve detailed information about behaviors that from Ref. [32]).
could not be discovered with restrictive tools. The well-defined
description of the time-frequency-amplitude content obtained using functions of time, the Hilbert spectrum, which displays the relative
the HHT, clarifies the nature of the ground motions producing beneficial amplitude or energy (square of amplitude) contributions for a certain
data for safer designs and improved practical earthquake engineering. frequency at a specific time, can be constructed as H(ω,t). Then, a
marginal spectrum can be calculated as in equation (5), where the
2. Hilbert-Huang transform spectrum is summed over the time domain of 0 and T.
Z T
The HHT was proposed by Huang and co-workers [6] and consists of hðωÞ ¼ Hðω; tÞdt (5)
two parts: i) Empirical Mode Decomposition EMD, and ii) Hilbert 0

Spectral Analysis. Signals to be analyzed are decomposed into a finite The marginal spectrum represents the sum of all amplitudes (en­
number of intrinsic mode functions or IMFs in what is called the EMD ergies) over the entire data span. This can be directly compared to the
process. An IMF is described as a function satisfying the following Fourier spectrum which was shown in equation (5) [7,8]. showed that
conditions: i) the number of extrema and zero-crossings must either not all functions give “good” Hilbert transforms, meaning those which
equal or differ at most by one; and ii) at any point, the mean value of the produce physical instantaneous frequencies. For example, functions
envelope defined by the local maxima and the envelope defined by the with non-zero means will give negative frequency contributions using
local minima is zero [6]. An IMF permits directly Hilbert transforms. the Hilbert transform [7,8]. Therefore, the signals which can be
analyzed using the Hilbert transform must be restricted so that their
calculated instantaneous frequency functions have physical meaning.
2.1. Hilbert spectral analysis
Next, the empirical mode decomposition will be described. It is essen­
tially an algorithm which decomposes nearly any signal into a finite set
The Hilbert transform permits to compute instantaneous frequencies
of functions which have “good” Hilbert transforms that produce physi­
and amplitudes and to describe the signal more locally. Equation (1)
cally meaningful instantaneous frequencies.
displays the Hilbert transform, byðtÞ which can be written for any
Fig. 1 shows the complete block diagram of the HHT. To the reader
function x (t) of Lp class [5]. The PV denotes Cauchy’s principle value
interested in deepening the subject it is recommended to read the works
integral.
of [7,29,30].
Z ∞
1 xðτÞ
H½xðtÞ� � b
y ðtÞ ¼ PV dτ (1)
π ∞ t τ
2.2. Empirical mode decomposition
[6,7] determined that an analytic function can be formed with the
Hilbert transform pair as shown in equation (2). At any given time, the data may involve more than one oscillatory
y ðtÞ ¼ AðtÞeiθðtÞ
zðtÞ ¼ xðtÞ þ ib (2) mode; that is why the Hilbert transform cannot provide the full
description of the frequency content for general data [31]. EMD can
where decomposed diachronic signals into a fitness of IMF component and,
together with HT, which provides meaningful time-varying frequency,
�b
y� pffiffiffiffiffiffi
y2
AðtÞ ¼ x2 þ b
�1=2
; θðtÞ ¼ tan 1
and i ¼ 1 (3) can deal with nonlinear and non-steady state signals. Since EMD is
x derived from original vibrations-diachronic signals, the following un­
A(t) and θðtÞ are the instantaneous amplitudes and phase functions, derlying assumptions exist when dealing with nonlinear and
respectively [28]. The instantaneous frequency can then be written as non-resident state data: i) there must be two extremes of the signal: a
the time derivative of the phase, as shown in equation (4). maximum and a minimum, ii) signal characteristic time scale is defined
as the time difference between the two extremes and iii) if the data
dθðtÞ contains no extreme but inflection points, the process does once or more
ω¼ (4)
dt times of differential with the signal until finding the extreme. The final
Note that the analytic function z (t) is the mathematical approxi­ results can be obtained from integral component differential. EMD is
mation to the original signal x (t). intuitive, direct, a posteriori and adaptive, with the basis of the
Because the amplitude and frequency functions are expressed as decomposition based on, and derived from, the data.

2
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

Table 1
Seismic events recorded in SCT and used in this investigation.
Number Date Focal Depth km Magnitude Epicentral Distance km

1 19/09/85 15 Ms ¼ 8.10 425


2 02/12/85 1 M¼3 17
3 04/01/86 9 Ms ¼ 5.4 928
4 05/01/86 <5 Mc ¼ 3.5 31
5 24/10/93 19 Ms ¼ 6.6 315
6 23/05/94 23 Mc ¼ 5.6 212
7 10/12/94 20 Mc ¼ 6.3 296
8 15/07/96 20 Ms ¼ 6.5 302
9 11/01/97 16 Ms ¼ 6.9 442
10 22/05/97 59 Ms ¼ 6.0 305
11 19/07/97 5 Ms ¼ 6.3 400
12 20/04/98 66 Mb ¼ 5.9 245
13 15/06/99 69 Me ¼ 7.0 219
14 21/06/99 54 Me ¼ 6.2 267
15 30/09/99 16 Me ¼ 7.6 442
16 29/12/99 82 Mb ¼ 6.1 307
17 21/07/00 48 Me ¼ 6 145
18 09/08/00 16 Me ¼ 7.0 402
19 22/01/03 9 Ms ¼ 7.6 540
20 26/04/11 7 M ¼ 5.5 302
21 05/05/11 11 M ¼ 5.5 309

Once a signal has been fully decomposed, the signal D (t) can be
written as the finite sum of the IMFs and a final residue as shown in
equation (6).
X
n
DðtÞ ¼ Rn ðtÞ þ IMFj ðtÞ (6)
j¼1

Using equations (6) and (8), the analytic function can be formed as
shown in equation (2).
hX
n R i
DðtÞ Rn ðtÞ ¼ Re Aj ðtÞei ωj ðtÞdt
(7)
j¼1

Also, for reference, equation (7) shows the Fourier decomposition of


Fig. 2. Empirical Mode Decomposition flowchart, from the figure it can be a signal, x (t).
inferred that as the algorithm progresses, the IMFs generated are of lower
hX i
frequency than the previous ones (modified from Ref. [33]).
n
DðtÞ ¼ Re Aj eiωj t (8)
j¼1
The EMD algorithm starts to decompose almost any signal into a
Notice that the EMD decomposition can be considered a generalized
finite set of functions. The resulted functions are called intrinsic mode
Fourier decomposition, because it describes a signal in terms of ampli­
functions IMFs whose Hilbert transformation gives physical instanta­
tude and basis functions whose amplitudes and frequencies may fluc­
neous frequency values. The algorithm employs an iterative sifting
tuate with time [5,29]. The HHT will now be used on a number of
process which successively subtracts the local mean from a signal, as
different data sets to analyze its applicability.
follows (Fig. 2):

3. Intrinsic oscillations of seismic response


1. Settle on the local extrema (maxima, minima) of the signal.
2. Link the maxima with an interpolation function, producing an upper
3.1. Geotechnical database
envelope of the signal.
3. Link the minima with an interpolation function, producing a lower
Data recorded in the SCT site (Table 1), an accelerographic station
envelope of the signal.
placed in central Mexico City, are analyzed to illustrate the HHT ad­
4. Calculate the local mean as half the difference between the upper and
vantages for studying seismic response of soft soils. SCT is located within
lower envelopes.
a densely urbanized zone, where earthquake related damages have
5. Take from the local mean from the signal.
concentrated recurrently in the past. The soil formation is ordered in a
6. Iterate on the residual.
sequence of soft clay strata interspersed with layers of harder clayey silts
with sands (Fig. 3).
These six steps are repeated until the produced signal converges to
The following list of properties highlights the unique characteristics
the definition of an IMF, which is a signal with a zero-mean and whose
of Mexico City Soils, in the context of most other natural soils [25–27],:
number of extrema and zero-crossings differ by at most one, considered
very high specific surface (Ss ¼ 40–350 m2/g), void ratio (e ¼ 3–9),
monocomponent functions with no riding waves [5]. Then, the IMF is
plasticity (wL ¼ 140–380, wP ¼ 55–112), and friction angle (φ¼ 43–47� ).
subtracted from the original signal, and the sifting process is repeated on
The shear wave velocity is relatively constant with depth in the upper
the remainder. This is iterated until the final residue is a monotonic
40 m (Vs ¼ 70–90 m/s). SCT has almost constant shear modulus and
function. The ultimate extracted IMF is the lowest frequency component
damping less than 4% for shear strains as high as 1%; and no strength
of the signal, better known as the trend. The characterization of an IMF
loss even when the amplitude of the dynamic cycles is as high as 80% of
was created to guarantee that the signal give physical frequency values
the quasi-static strength.
when using the Hilbert transform.

3
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

Fig. 3. SCT site description: stratigraphic column and water content W%, volumetric weight γ, cone penetration resistance qc (Cone Penetration Test CPT), shear
wave velocities Vs and pore pressure u profiles.

Fig. 5. Integration of a signal using IMFs; adding modes, from the first to the
last, it is verified that the accelerogram is re-constructed (the example shown is
that of the September 30th, 1999 event).

Fig. 4. Intrinsic Mode Functions obtained from the accelerogram recorded


during the September 30th, 1999 event (the trend or residue is not shown). These changes have had effect on the dynamic properties; e.g., the
evolution on shear wave propagation velocities is field evidence of
process of stiffening of the clay strata over, more evident for greater
The clayey soils have suffered a gradual depletion of pore pressures
depths.
(from 0.002 to 0.014 kPa/year) at different depths [34]. The water
content, volumetric weight and undrained strength profiles at different
dates (Fig. 3) expose that. 3.2. Decomposition of accelerograms

(a) thicknesses of the relevant clay strata have decreased as a According with the procedure described above (EMD algorithm) the
consequence of regional subsidence; selected accelerograms were decomposed into their IMF components.
(b) water content reductions are especially significant below 15 m The results are like those shown, by way of example, in Fig. 4 (IMFs
approximately which is consistent with the fact that pumping obtained from the 30th September 1999 event). For this accelerogram
induced consolidation propagates upwards from the base of the the algorithm yielded 9 components plus a residue.
clayey soils to the surface; For the whole data set, the number of IMFs needed to decompose the
(c) changes in soil density (volumetric weight) confirm that soil signals fluctuates from 9 to 10 modes what is in agreement with
strata has densified as the clay masses consolidate. empirical recommendations [10]. To validate that the decomposition
was performed correctly and to discriminate between modes, proof the

4
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

goodness of fit is necessary [35]. Simply adding modes, from the first to
the last, one by one, it is verified that the integration of IMFs
re-constructs the original signal. As can be seen in the example of Fig. 5,
the sum of the first modes is far from resembling the original signal but
when modes 3, 4 and 5 are added practically the recorded ground mo­
tion has been reassembled. Adding modes 6, 7, 8 and 9 does not sub­
stantially modify the signal obtained in the previous step. The
adjustment is verified throughout the data set in order to confirm for
each individual. It is supported that, with minimal differences, the
original signals are reconstructed summing the first five IMFs.

3.3. Characterizing the IMFs

The amplifications, caused mainly by lithological site effects, can be


significantly different for a strong event compared to small ones Fig. 6. Boxplot of selected IMFs, excluding the noise intrinsic oscillations
[36–43]. The non-linear soil behavior of the sedimentary materials in considered here as non-interpretable or unwanted component of signals
SCT affects the seismic site response because of a degradation of its shear recorded by seismometers.
modulus and an increase in its attenuation properties. This impact on
site response is experimented as a shift of the resonance frequency to­
wards lower values, together with a reduction in the associated ampli­
fications. Studies around the world have tried to interpret this effect, for
example [40], reported a decrease of the amplification from weak to
strong motions (factor between 1.4 and 7) while [44] showed a factor of
2 for site responses for all frequencies [45]. showed a factor 4 decrease
in the PGA amplification but [46–50] reported a factor of 2. It can be
inferred that the dispersion between the factors is mainly due to the
number of earthquakes used in these studies and the restriction of the
areas under analysis, but they give an idea of the effect on the amplifi­
cation response according to the intensity of the seismic input.
In the following, it is presented the EMD as an alternative technique
for study the ground motions without the limitations of conventional
filtering methods. The results on dominant frequency and ground ac­
celerations are analyzed when they are obtained directly from the
Fig. 7. Some Fourier spectra of the first intrinsic mode, determining funda­
intrinsic oscillations against those that are defined through conventional mental frequencies from flat spectra can become a complicated and subjective
domain transformation analyzes. task (observe that the spectral amplitudes seem to be not directly related with
the seismic intensity).
3.3.1. Fundamental frequency
Several past earthquake events have resulted in devastating failures deposit parameter, the predominant frequency of each IMF in data set, is
of built environment, which in most of the cases are attributed to strong analyzed. In what follows this predominant frequency will be called fIMF.
amplification of seismic waves. As earthquake energy propagates In Fig. 6 a boxplot of the fIMF corresponding to each mode is depicted.
through soil media, its amplification or attenuation is mainly related to The first (IMF1) and last modes (IMF6 to IMF10) are not contained in the
fundamental natural frequency fn of the soil deposit. This situation has graph. Opposing to the other IMFs behavior, the IMF1s have flat spectra
been explained by means of the concept of the mechanical resonance, where is not possible to determine a predominant frequency (see some
that is the tendency of a mechanical system to respond at greater examples in Fig. 7). The last modes (from IMF6 to IMF10) have very
amplitude when the frequency of its oscillations matches the system’s short box sections with exceptionally low frequencies values (near to
natural frequency of vibration than it does at other frequencies. zero) and small amplitudes compared with those registered in the other
The fn of a soil deposit is dependent on its thickness, low strain modes. In this investigation following the definitions of [58,65], rather
stiffness and density. Also empirical evidences for non-linear soil than assuming that a signal is a linear combination of predetermined
behavior reported a frequency shift towards lower values from weak to basis functions, the data are thought of as a superposition of fast oscil­
strong motions [39,41]. Assessment of fn has been the subject of several lations onto slow oscillations. The slow modes, those with extremely low
research studies, yet inconsistencies and uncertainties associated with fIMF values, are considered noise. The term is taken from signal pro­
their estimation have not been resolved conclusively [51,59]. Hence, cessing lexicon and is labeled here as the generic name for the relatively
many researchers tend to corroborate their calculations by using data persistent vibration of the ground, due to a multitude of causes that is a
recorded in surface during earthquake events. The best tactic to estab­ non-interpretable or unwanted component of signals recorded by seis­
lish fn is to compute it using an abstract method and then compare this mometers. The noise intrinsic oscillations are random (unpredictable)
value with that exposed by recorded data. and carry no useful information [66].
For soft soil deposits in Mexico City the more general and accepted Analyzing the modes that are relevant to study the movement of soil
approach to calculate fn is to interpret the zone as homogeneous soil strata, observe that the fIMF progressively decays as the mode number
deposits overlying a rigid bedrock, with constant shear wave velocity increases. One remarkable finding is that the fIMF of IMF5 is closely
profiles throughout its depth [53,54,60,61] Theoretically fundamental related to the fn of SCT site and despite of the wide distribution of seismic
frequency for SCT (clay deposit �40 m thick) calculated using average of inputs (magnitudes, epicentral depths, generating mechanisms, direc­
shear wave velocity, is fn ¼ 0.5 Hz. Using microtremors [55,56,62,63] tivity, and dates), a condensed boxplot is resulted. The dispersion is
and MDOF Multi Degree of Freedom System methods [57,64] the minimal and it should be noted that this graph includes minimum en­
computed values occur in the vicinity of the above, slightly lower, fn ergy events (M < 4) and the extreme earthquake of 1985 (M8.1).
~0.47 Hz. The median of the fIMF of IMF3 is comparable to the second vibrate
Attempting to relate the intrinsic oscillations with this important

5
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

Fig. 8. The depletion of pore pressures within the clayey soils in Mexico City
changes soil properties stiffening the material, shortens the dominant period of
the soil deposits, and diminishes the amplification potential, in the figure it is
shown how 1-D wave propagation model (with strain-dependent nonlinearity)
can be better adjusted using the behavioral trend of IMF5.

mode f2 (evaluated using the MDOF method [55], f2 ~1.4 Hz). Its box­
plot is taller than the IMF5 and contrary to what the fifth mode shows,
IMF3 has evident dispersion in the lower and upper whiskers.
Many have been the efforts to analyze the changes Mexico City’s
seismic response due to land subsidence as a consequence of ground
Fig. 9. MIAs boxplot according to the Intrinsic Mode Functions, it is clear that
water extraction [34,52], being the most significant aspect to be the amazing energy of the Michoac� an earthquake is well above the seismic
demonstrated the effect on the spectral ordinates (with respect to the history recorded on the site, it is necessary to eliminate this atypical case from
current values specified for design purposes) and the relative values of the graph in order to study the behavior of the other events (in b) are included
the structures and site periods. With specific geotechnical and geological those events with M � 7.5).
information [60], presented an incremental procedure to evaluate the
progressive evolution of fn over time, in order to evaluate the effects of approach of using a modification on the Rayleigh’s method can be
these changes on the earthquake ground response. The adjustments improved using these IMFs as the natural modes of the undamped elastic
expected in selected values of fn, from 1980 to 2020, can be appreciated system with a direct application for simple-shaped homogeneous
in Fig. 8. What can be anticipated, as the evolution of this phenomenon, alluvial-valleys, like that of the Mexico City. Using the fn calculated from
is that after the initial reference year the parameter decrease mono­ Ref. [62] method, the predicted fundamental frequency could be seen as
tonically as the times span increases [61]. an upper limit of the true value of fn. This is important to decide whether
Observe what occurred when the values of fn, from original data, are performing more sophisticated analyses of site effects and whether the
superimposed on the graph, It is seen that the frequencies obtained from use of standard code spectral shapes is appropriate.
the recorded accelerograms hardly follow the trend set by Ref. [60].
Most events do not fit the specific curve for SCT (1/fn ¼ 2) and what the 3.3.2. Peak ground acceleration PGA
expected drop patterns if input intensity is fixed, of fn cannot be verified. PGA, as the maximum ground acceleration that occurred during
However, the behavior shown by the IMF3 and IMF5 fits very well to two earthquake shaking at a location [63], is used as an intensity measure
decrement curves obtained with the iterative model, each belonging to and the design-basis-earthquake ground motion is often defined in terms
the natural frequency (according to code) and a value of fn slightly of PGA [64]. Even damage to buildings and infrastructure is related
displaced towards lower values. Recognized that surficial soil layers act more closely to ground motion, of which PGA is a measure, rather than
as a filter on propagating seismic waves that modifies the frequency the magnitude of the earthquake itself [65].
content of input motions, understanding these changes can help to Analyzing the largest absolute acceleration of each IMF is useful for
explain the physical processes behind the measured responses. The detecting those more related to the registered PGAs. In Fig. 9a the
non-linearity could be analyzed through separated intrinsic oscillations, boxplots of the maximum instant amplitudes, associated with each IMF,
those that stand out before smaller earthquakes and those that do it are shown. The outliers are two of the most intense seismic events
when the seismic intensity exceeds a certain threshold. registered in Mexico City (M > 7.0), being the extreme value, the
The IMF5 keeps on the curve of 1/fn ¼ 2.0 and helps to corroborate devastating 1985 Michoaca �n earthquake (M8.1).
the stiffening that is suffering the clay deposit. Actually, a deep analysis Removing the M8.1 event to properly observe each box (Fig. 9b).
of the path shown by fIMF could drive to two stages of the evolution, the Minor boxes for the last modes are observed, if this characteristic is
first one with a more dramatic decrement of the 1/fn since 1985 (or associated with their fundamental frequencies (very close to zero), the
before) and towards the year 2000, and a second one where the fre­ conclusion that they are irrelevant, is reinforced. Now examine the
quencies change more slowly. Even, the model’s tendency that lower relatively small box of the IMF1 maximum amplitudes. Only for the first
values of 1/fn change less dramatically over time is clearly illustrated mode the maximum amplitude during the extraordinary earthquake of
with the aligned values of fIMF - IMF3. 1985 is contained inside the rectangle, i.e., it does not appear as atyp­
The attractiveness of using IMF3 and IMF5 for studying particular ical. It seems that this first IMF does not respond to the input energy as
phenomena (e.g. the effect of soil stiffening) has been exposed but these the other modes do, so it is inferred that a mechanism, other than the soil
intrinsic oscillations naturally they could be used as the input signals in deposit vibrating, is the one that controls this oscillation. Based on this
more detailed analyzes on resonance frequencies. For example, the [62] observation and their flat Fourier spectra, the IMF1 is labeled as a mode

6
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

Fig. 10. Examples of Response Spectra estimated using selected IMFs, events displayed better represented for the IMF5 (Fig. 10a) and those better represented for the
IMF3 (Fig. 10b); two cases are shown but the total of the records contained in the database were analyzed and all fell into one of the two classes.

that “corrupts” the registered motion of the soil deposit. them significantly. Once the signal is filtered, the remaining IMFs
It can be seen that the energy is concentrated mainly in 4th and 5th contain observable, physical information inherent to the ground motion.
oscillations. These modes most reacting to the extreme earthquake of Because the intrinsic oscillations can be related to the phenomena
1985, as can be seen in Fig. 9b, the dispersion is minimal but the atypical without restrictive hypothesis, the conclusions obtained from the
event is certainly distant which does not happen with the other oscil­ behavior of these modes are more natural than those achieved from
lations. It appears that the recognized amplification potential of this theoretical positions faraway the system, which generates the signal.
clayey deposit [52] is evident only for the central modes. Despite of soil properties in the SCT site have changed because of
For examining the effect of removing the irrelevant modes (IMF6 to regional subsidence, the recorded accelerations do not seem to exhibit
IMF10) on the interpreted behavior, the response spectra of accelera­ the expected stiffening effect. What is shown in this research is that the
tions from both sets of signals, the original and the “cleaned”, are time series, which serve as a basis for showing the soils hardening, are
calculated. Some selected responses are depicted in Fig. 10. These are corrupted by noise oscillations that mask actual responses. It was recog­
examples of behaviors that, only with the IMF3 or the IMF5, can be nized that IMF5 and IMF3 actually show the shift the decrement rate of
reproduced (with adequate approximation). The energy-frequencies the frequencies postulated in theoretical models. By detecting the os­
content of the original response spectra is represented with one cillations that do mainly correspond to the real ground motion, the EMD
intrinsic mode, the one with a MIA near to the PGA registered. It has is presented as a tool that allows to find directly the frequencies related
been found that the most intense events from the Pacific subduction with the highest energy concentrations.
zone tend to be related mainly with the 5th mode while for minor and The usefulness of these findings points to the need for further
intraplate events, the ground motion can be represented using only the investigation in qualification of the aptitude of tools for analyzing the
3rd mode. Deep examinations are necessary to clarify this finding but recordings and for obtaining conclusions, since they are the basis of the
the advantages of EMD, when analyzing directivity effect, is clear. process of calibration of all the theoretical geotechnical and seismo­
It should be emphasized that the “cleaned” signals represent simpler logical models.
and easier to interpret time series, with desirable density for minimizing
computational costs (when running complex ground-motion models). Acknowledgments

4. Conclusions The accelerogram records for this research were provided by the
Unidad de Instrumentacio
�n Sísmica, Instituto de Ingeniería, UNAM.
Many seismological data present strong periodic or quasi-periodic
signals masking the observations of the phenomena of interest. This Appendix A. Supplementary data
work has presented a conceptually different approach to filter compo­
nents in acceleration time-series and advantageous practical deductions Supplementary data to this article can be found online at https://doi.
were derived from the findings. The EMD method adopted here has org/10.1016/j.soildyn.2019.105918.
shown be useful for separating oscillation modes and to permit
discriminate between undesired signals to those more correlated to the
References
phenomena being modeled.
The frequency filtering approach consents to detect distorting signals [1] Jousset P, Neuberg J, Jolly A. Modelling low-frequency volcanic earthquakes in a
(those modes considered irrelevant) and, although is not able to viscoelastic medium with topography. Geophys J Int 2004;159:776–802.
completely eliminate these oscillations in the time-series, could reduce [2] Valerio J. Spectral analysis of earthquake accelerations as realizations of a non-
stationary stochastic process. Proc. World Conf. Earthq. Eng. 1992;2:901–7.

7
S.R. Garcia and L. Alc�
antara Soil Dynamics and Earthquake Engineering 129 (2020) 105918

[3] Cohen L. Time- frequency analysis, Prentice- Hall PTR, vol. 299. Technology & [34] Ovando-Shelley E, Romo MP, Contreras N, Giralt A. Effects on soil properties of
Engineering; 1995. future settlements in downtown Mexico City due to ground water extraction.
[4] Shumway R, Stoffer D. Time series analysis and its applications. Springer Texts in Geofis Int 2003;42(2):185–204.
Statistics; 2011. p. 202. [35] Guang Y, Sun X, Zhang M, Li X, Liu X. Study on ways to restrain end effect of
[5] Huang, N. E., Shen, Z. Interdisciplinary mathematical sciences, volume 16 Hilbert-Huang transform. J Comput 2014;25.
Hilbert–Huang transform and its applications 2015; Norden E Huang, USA, pp 400. [36] Bard P, Bouchon M. The two-dimensional resonance of sediment-filled valleys. Bull
[6] Huang NE, Shen Z, Long SR. The empirical mode decomposition and the Hilbert Seismol Soc Am 1985;75:519–41.
spectrum for nonlinear and non-stationary time series analysis. Proc R Soc Lond A [37] Duval A-M, Bard P-Y, M� eneroud J-P, Vidal S. Mapping site effect with
1998:903–95. microtremors. In: International conference on seismic zonation; 1996. p. 1522–9.
[7] Huang NE, Attoh-Okine N. The Hilbert-Huang transform in engineering. CRC [38] LeBrun B, Hatzfeld D, Bard PY. Site effect study in urban area: experimental results
Taylor & Francis; 2005, ISBN 978-0849334221. in Grenoble (France). Pure Appl Geophys 2001;158:2543–57.
[8] Zhang R. The Role of Hilbert-Huang transform in earthquake engineering. In: [39] Frankel AD, Carver DL, Williams RA. Nonlinear and linear site response and basin
Proceedings of the world multiconference systemics, cybernetics informatics, vol. effects in Seattle for the M 6.8 Nisqually, Washington, earthquake. Bull Seismol Soc
XVII. USA: Orlando Florida; 2001. Am 2002;92:2090–109.
[9] Huang, et al. A new spectral representation of earthquake data: Hilbert spectral [40] Beresnev I, Wen K. Nonlinear soil response, A reality? Bull Seismol Soc Am 1996;
analysis of Station TCU129, Chi-Chi, Taiwan, 21 September 1999. Bull Seismol Soc 86:1964–78.
Am 2001;91(5):1310–38. [41] R�egnier J, Cadet H, Bonilla L, Bertand E, Semblat JF. Assessing nonlinear behavior
[10] Loh CH, Wu TC, Huang NE. Application of the Empirical mode decomposition- of soil in seismic site response: statistical analysis on KiK-net strong motion data.
Hilbert spectrum method to identify near-fault ground-motion characteristics and Bull Seismol Soc Am 2013;103:1750–70.
structural responses. Bull Seismol Soc Am 2001;91(5):1339–57. [42] Assimaki D, Li W. Site and ground motion-dependent nonlinear effects in
[11] Yang Jann N. System identification of linear structures based on Hilbert-Huang seismological model predictions. Soil Dyn Earthq Eng 2012:143–51.
spectral analysis. Part 2: complex modes. Earthq Eng Struct Dyn 2003;32(10): [43] R�egnier J, Bonilla LF, Bertrand E, Semblat J-F. Influence of the VS profiles beyond
1533–54. 30 m depth on linear site effects: assessment from the KiK-net data. Bull Seismol
[12] Castagna JP, Sun S, Seigfried RW. Instantaneous spectral analysis: detection of low- Soc Am 2014;104:2337–48. 2014.
frequency shadows associated with hydrocarbons. Lead Edge 2003;22:120–7. [44] Field E, Johnson PA, Beresnev IA, Zeng Y. Nonlinear ground-motion amplification
[13] Peyton L, Bottjer R, Partyka G. Interpretation of incised valleys using new 3-D by sediments during the 1994 Northridge earthquake. Lett Nat 1997:599–602.
seismic techniques: a case history using spectral decomposition and coherency. [45] Aguirre J, Irikura K. Nonlinearity, liquefaction, and velocity variation of soft soil
Lead Edge 1998;17:1294–8. layers in port island, Kobe, during the Hyogo-ken Nanbu earthquake. Bull Seismol
[14] Partyka G, Gridley J, Lopez J. Interpretational applications of spectral Soc Am 1997;87:1244–58.
decomposition in reservoir characterization. Lead Edge 1999;18:353–60. [46] Noguchi S, Sasatani T. Quantification of degree of nonlinear site response. In: The
[15] Hardy HH, Richard AB, Gaston JD. Frequency estimates of seismic traces. 14th world conference on earthquake engineering, Beijing, China; 2008.
Geophysics 2003;68:370–80. [47] Bonilla LF, Tsuda K, Pulido N, Regnier J, Laurendeau A. Nonlinear site response
[16] Jones DL, Baraniuk RG. An adaptive optimal-kernel time-frequency representation. evidence of K-net and KiK-net records from the Mw 9 Tohoku earthquake. Earth
IEEE Trans Signal Process 1995;43:2361–71. Planets Space 2011;58.
[17] Rioul O, Vetterli M. Wavelets and signal processing. IEEE Signal Process 1991;8(4): [48] Satoh T, Sato T, Kawase H. Nonlinear behavior of soil sediments identified by using
14–38. borehole records observed at the Ashigara valley, Japan. Bull Seismol Soc Am
[18] Hlawatsch F, Boudreaux-Bartels GF. Linear and quadratic time-frequency signal 1995;85:1821–34.
representations. IEEE Signal Process 1992;9(2):21–67. [49] Wen K-L, Huang J-Y, Chen C-T, Cheng Y-W. Nonlinear site response of the 2010
[19] Abry P, Goncalves P, Flandrin P. Wavelet-based spectral analysis of 1/f processes: Darfield, New Zealand earthquake sequence. In: 4th IASPEI/IAEE international
international conference on acoustic, speech and signal processing. IEEE Proc symposium: effects of surface geology on seismic motion; 2011.
1993;vol. 3:237–40. [50] Yu G, Anderson JG, Siddharthan R. On the characteristics of nonlinear soil
[20] Sinha S. Time-frequency localization with wavelet transform and its application in response. Bull Seismol Soc Am 1993;83:218–44.
seismic data analysis. M.S. thesis. University of Oklahoma; 2002. [51] Pando M, Cano L, Su� arez L, Ritta R, Montejo L. Comparison of site fundamental
[21] Meza-Fajardo KC, Papageorgiou AS, Semblat JF. Identification and extraction of period estimates using weak-motion earthquakes and microtremors. In: The 14th
surface waves from three-component seismograms based on the normalized inner world conference on earthquake engineering; 2008.
product. In: 6th international conference on earthquake geotechnical engineering; [52] Demetriu S, Trandafir R. Time-frequency representations of earthquake motion
2015 [New Zealand]. records. Analele Stiintifice ale Universitatii Ovidius Constanta 2003;11(2):57–68.
[22] Stockwell R, Mansinha L, Lowe R. Localization of the complex spectrum: the S [53] Kramer Steven L. Geotechnical earthquake engineering. Prentice Hall; 1996.
transform. IEEE Trans Signal Process 1996;44(4):998–1001. p. 653.
[23] Wang S, Chen X, Cai G, Chen B, Li X, He Z. Matching demodulation transform and [54] P�erez-Rocha LE, Schroeder M Ordaz, Angulo E Reinoso. Interpolaci� on espacial de
synchro squeezing in time-frequency analysis. IEEE Trans Signal Process 2014;62 las amplificaciones din� amicas del terreno en el Valle de M� exico, 80. Revista de
(1):69–84. Ingeniería Sísmica; 2009. p. 85–112.
[24] Zosso D, Dragomiretskiy K. Variational mode decomposition. IEEE Trans Signal [55] Lermo J, Ch� avez-García F. Are microtremors useful in site response evaluation?
Process 2013;62(3). Bull Seismol Soc Am 1994;84(5):1350–64.
[25] Díaz-Rodríguez JA, Leroueil S, Alem� an JD. Yielding of Mexico City clay and other [56] Ordaz M, Miranda E, Avil�es J. Proposal for seismic design spectra for Mexico’s
natural clays. J Geotech Eng. 1992;118(7):981–95. Division, ASCE. Federal District. Ingeniería de Estructuras 2003;8(2):189–207.
[26] Díaz-Rodríguez JA, Lozano-Santa Cruz R, D� avila Alcocer VM, Vallejo E, Gir�
on Py. [57] Bozorgnia Y, Bertero V. Earthquake engineering: from engineering seismology to
Physical, chemical, and mineralogical properties of Mexico City: a geotechnical performance-based engineering, vol. 450. B/W Illustrations; 2004. p. 976.
perspective. Can Geotech J 1998;35(4):600–10. [58] Flandrin P, Goncalves P, Rilling G. Empirical mode decomposition as a filter bank.
[27] Díaz-Rodríguez A. Los suelos lacustres de la Ciudad de M�exico. Rev. Int. de IEEE Signal Process Lett 2004;11(2).
Desastres Naturales, Accidentes e Infraestructura Civil 2006;6(2):111. [59] Ovando-Shelley Efraín, Ossa Alexandra, Miguel P, Romo. The sinking of Mexico
[28] Chen Yangkang, Zhou Chao, Yuan Jiang, Jin Zhaoyu. Application of empirical City: its effects on soil properties and seismic response. Soil Dyn Earthq Eng 2007;
mode decomposition in random noise attenuation of seismic data. J Seismic Explor 27:333–43.
2014;23:481–95. [60] Avil�es J, P�erez-Rocha L. Regional subsidence of Mexico City and its effects on
[29] Huang, N. E.; Wu Z. A review on Hilbert-Huang transform: method and its seismic response. Soil Dyn Earthq Eng 2010;30:981–9.
applications to geophysical studies 2008, Rev Geophys vol. 46. [61] Yin JH, Graham J. Elastic visco-plastic modelling of one dimensional
[30] Huang NE, Shen SSP. Hilbert-Huang transform and its applications. London: World consolidation. Geotechnique 1996;46(3):515–27.
Scientific; 2005, ISBN 978-9812563767. [62] Paolucci R. Shear resonance frequencies of alluvial valleys by Rayleigh’s method.
[31] Long SR, Huang NE, Tung CC, Wu ML, Lin RQ, Mollo-Christensen E, Yuan Y. The Earthq Spectra 1999;15(3):503–21.
Hilbert techniques: an alternate approach for non-steady time series analysis. IEEE [63] Douglas J. An investigation of analysis of variance as a tool for exploring regional
Geosci Remote Sens Soc Lett 1995;3:6–11. differences in strong ground motions. J Seismol 2004;8(4):485–96.
[32] Lee M, Shyu K, Lee P. Hardware implementation of EMD using DSP and FPGA for [64] Noeggerath J, Geller RJ, Gusiakov VK Fukushima. The myth of safety, the reality of
online signal processing. IEEE Trans Ind Electron 2011;58(6):2473–81. geoscience. Bull At Sci 2011. 67: 37.
[33] Hillenbrand Benedikt, Kulia Geir, Molinas Marta. Impact of time varying angular [65] Campbell KW, Bozorgnia Y. Updated near-source groundmotion (attenuation)
frequency on the separation of instantaneous power components instand-alone relations for the horizontal and vertical components of peak ground acceleration
power systems. In: 6th IEEE international conference on clean electrical power; and acceleration response spectra. Bull Seismol Soc Am 2003;93:314–31.
2017. p. 347–53. [66] Kaslovsky D, Meyer F. Noise corruption of empirical mode decomposition and its
effect on instantaneous frequency. Advances in Adaptive Data Analysis, vol. 2.
World Scientific Publishing Company; 2010. p. 373–96. 3.

You might also like