Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

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

net/publication/290591733

On the stopping criterion to apply the Hilbert Huang Transform


method to sea wave records

Article · January 2005


DOI: 10.1201/9781439833728.ch134

CITATIONS READS

6 69

3 authors, including:

Carlos Guedes Soares Albena Veltcheva


University of Lisbon Waseda University
2,709 PUBLICATIONS 53,339 CITATIONS 24 PUBLICATIONS 275 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Albena Veltcheva on 10 September 2021.

The user has requested enhancement of the downloaded file.


Maritime Transportation and Exploitation of Ocean and
Coastal Resources – Guedes Soares, Garbatov & Fonseca (eds)
© 2005 Taylor & Francis Group, London, ISBN 0 415 39036 2

On the stopping criterion to apply the Hilbert Huang Transform method


to sea wave records

R. Pascoal & C. Guedes Soares


Unit of Marine Technology and Engineering, Technical University of Lisbon, Instituto Superior Técnico,
Av. Rovisco Pais, Lisbon, Portugal

A.D. Veltcheva
IO-BAS, Bulgaria, Present address: ECOH Co. Yokohama, Japan

ABSTRACT: Some specific properties of the Hilbert Huang Transform (HHT) method are examined on the
basis of application of this new time series analysis method of nonlinear and non-stationary sea wave data. Field
data, collected at the Portuguese coast are used in this study. The performance and limitations of the Empirical
Mode Decomposition (EMD), a key part of HHT method, are discussed and examined in detail. The decomposi-
tion is based on the local characteristic time scale of the data and thus the wave data are decomposed into a finite
number of Intrinsic Mode Functions (IMF), a set that is unique to each record. An important step in HHT method
is to determine an adequate stopping criterion in order to obtain physically meaningful IMFs. In this study, several
stopping criteria are used and the Empirical Mode Decomposition is performed for the wave records. The confi-
dence limit for EMD is determined. The orthogonality of the sets of Intrinsic Mode Function is also checked. The
characteristics of the obtained sets of IMFs are compared in order to investigate the effect of stopping criterion on
the extracted modes. The contribution of different Intrinsic Mode Functions to the wave data is examined.

1 INTRODUCTION Actually there is no way to determine this from the


data because the surface elevation, even for freely
Nowadays it is still very common to use the assumption propagating waves, is already a response of a dynamic
of linearity, stationarity and homogeneity when analyz- system to some form of excitation, such as wind for
ing sea wave data. A number of numerical design pro- instance, which itself is best modeled as a random
cedures for ocean structures are also based on these process.
assumptions and need parameters determined under The development of methods that do not assume
these hypotheses. However a deeper insight of wave stationarity, and therefore do not assume ergodicity, is
data analysis will result not only in better understanding a topic of research. The wavelet transform has been
of the sea wave phenomenon, but it will also enlarge the applied for analysis of sea wave data, for instance by
horizon for improving dependent procedures. Massel (2001), and by Liu and Mori (2000). The resolu-
It is a fact that most sea data are nonlinear to a tion in the time-frequency domain of the wavelet
degree that should not be underestimated. For instance, spectrum is restricted due to the Heisenberg’s Uncer-
several design procedures are already using at least tainty Principle as the wavelet solves the resolution
second order theory to better model sensitive phenom- dilemma only to a certain extent. Some nonparamet-
ena such as wave impact on the bow of floating plat- ric methods still rely on the assumption of slowly
forms (Guedes Soares et al. 2004), or a green water varying processes and therefore apply “short span”
on the deck (Hellan et al. 2001). The wave theories version of methodologies meant to be applied to sta-
are based on perturbation methods together with fre- tionary processes. An example of such a method is
quency domain analysis of the nonlinear wave differ- the Short Time Fourier Transform, or the spectrogram
ential equations, such that the analysis is based on (Guedes Soares and Cherneva 2004) frequently used
the appearance of sum and difference frequency to produce time-frequency power spectral densities.
components. There are also adaptive parametric methods such as the
Each record of wave elevation data is usually under- Adaptive Auto-Regressive method. The effort has been
stood as a realization of the elevation random variable. to make no assumption whatsoever of linearity or

1113
slow variation of the averages (all types – autocorre- The time series X(t) is first decomposed by EMD
lation and higher order included). The last result of into a finite number, n, IMF, Cj, j  1, n, which extract
research in this direction is probably the Hilbert- the energy associated with various intrinsic time scales
Huang Transform (HHT). HHT method uses the data and residual rn. The superposition of the IMF and the
itself to determine the adequate time scales but assumes residues reconstruct the data record:
no prior knowledge of the underlying dynamics, as
discussed by Veltcheva et al. (2005).
(1)
A major step in the HHT method is the identifica-
tion and extraction of the intrinsic time scales of the
data by Empirical Mode Decomposition (EMD) into a By definition, an IMF has to satisfy two conditions.
set of Intrinsic Mode Functions (IMF), orthogonal to a First, the number of extremes must be equal or differ
desired precision. This is a remarkable result, provid- at most by one from the number of zero-crossings,
ing a link to the much explored normal modes in linear which is similar to the traditional narrow band require-
systems. An important constraint during this step is to ments for the stationary Gaussian process. In practice
have a stopping criterion, one that leads to physically this condition corresponds to elimination of riding
meaningful IMFs, and this is discussed in detail here. waves in the time series. Second, at any point the
Interestingly, a formal presentation of the existence of mean value of the envelope, defined by the local max-
normal modes in nonlinear systems has been set up and ima and the envelope defined by the local minima is
methods used in the analysis of nonlinear dynamic sys- zero. These upper and lower envelopes are determined
tems have been introduced by Vakakis et al. (1996). by using cubic splines as suggested by Huang et al.
Several stopping criteria, like Cauchy convergence (1998). This second condition ensures a phase function
criterion (Huang et al. 1998), simplified stopping cri- without bias and consequently well-defined instanta-
terion (Huang et al. 1999), confidence limit of HHT neous frequency. The important condition for IMF is
(Huang 2003) and stopping criterion proposed by that only one maximum or minimum exists between
Rilling et al. (2003), are utilized in the decomposition successive zeros. The sifting process of the EMD is
of wave data in this work. repeated until some conditions, termed as a stopping
The characteristics of the obtained sets of IMFs are criterion, are satisfied.
compared in order to investigate the effect of the stop- The set of IMF obtained in this way is unique and
ping criterion on the extracted modes. The ability of specific for the particular time series, since it is based
the Empirical Mode Decomposition to extract the dif- on and derived from the local characteristics of these
ferent oscillation modes, embedded in the sea surface data. IMF could be considered as a more general case of
elevation records, is demonstrated. The contribution the simple harmonic functions, but it can be claimed
of different Intrinsic Mode Functions to the wave data that, due to their specific derivation, IMF have also a
is examined. physical meaning in addition to mathematical one.
In the second step of data analysis, the Hilbert
transform is applied to these IMF:
2 HILBERT HUANG TRANSFORM METHOD

The Hilbert Huang Transform method relies on data (2)


sifting according to an intrinsic time scale. A time
interval between successive extremes in the time series
is defined as a relevant time scale. The main idea of where P indicates the Cauchy principal value. The
EMD is first to identify the time scale that will reveal amplitude aj, the phase j and the instantaneous fre-
the physical characteristics of the studied process quency j are calculated by
recorded as a time series, and to extract them into
Intrinsic Mode Functions. The EMD is a data sifting
(3)
process to eliminate locally riding waves as well as to
eliminate locally the asymmetry of the time series
profile. A procedure for data sifting and several appli-
cations of HHT method are presented in Huang et al. (4)
(1998, 1999). On the basis of the results of EMD of
the field wave data, Veltcheva (2002) investigated wave
and group transformation in the coastal zone. The
energy hierarchy of Intrinsic Mode Functions in decom- (5)
position of wave records was examined by Veltcheva
and Guedes Soares (2004) to identify the components Using (3), (4) and (5) the original data X(t) can be
of wave spectra. expressed as a real part (Re) of the complex expansion

1114
in this direction, determining a confidence limit for
(6) the EMD. Several sets of IMF are obtained by EMD,
performed by different stopping criterion S and their
mean and standard deviation are calculated. Next, the
which is considered as a generalized form of the Fourier variation of the deviation of the individual cases from
expansion. Here both amplitude aj(t) and instantaneous the mean is examined in order to find a range of stop-
frequency j(t) are function of time t in contrast with ping criteria S, that produce relatively stable results in
the constant amplitude and frequency in the Fourier either IMFs or the Hilbert spectrum. As especially
expansion. stressed by Huang et al. (2003) the confidence limit
The frequency-time distribution of the squared can be applied to individual IMF, if the number of
amplitude is designated as Hilbert spectrum H(, t). IMF in the ensemble is the same. Whenever the num-
ber of IMF is not same in different IMF sets then, the
3 STOPPING CRITERION OF THE EMD confidence limit has to be applied to the Hilbert spec-
tra. The last is the situation observed in this work and it
The important step in the HHT method is the deter- is the most general, so the confidence limit is applied
mination of the stopping criterion in order to obtain to Hilbert spectra here.
physically meaningful IMFs. The sifting process is ter- The sifting is performed for several values of param-
minated when some conditions are fulfilled. Huang eter S and the mean value of the Hilbert spectrum is
et al. (1998) introduced the Cauchy convergence cri- determined as
terion for stoppage of the sifting process by limiting
the size of the standard deviation , computed from (8)
two consecutive sifting results

where H(Sj)  H(, t) ddt is the integration over


(7) time and frequency of the Hilbert spectrum H(, t),
when the EMD is performed by parameter S  Sj.
The squared deviation of each individual case from
this mean is
where hj,k1 and hj,k are candidates to be the “j” IMF Cj,
obtained on the k  1 and k step, respectively of the sift- (9)
ing procedure, and T is the duration of the data record.
Huang et al. (1998) discussed in detail the specific The parameter S is then determined from the min-
peculiarities of the sifting process and proposed some imal sd(Sj) range.
values for the limit . The choice of the value of  is Additionally the orthogonality index (OI)
not clear from a physical point of view. This stoppage
criterion is not connected with specific peculiarities
of given data series and seems to be subjective one. (10)
Later Huang et al. (1999) stressed that the Cauchy con-
vergence criterion “should be used with great care,
because the deviations between successive sifting are is checked as a function of stopping criterion S. The OI
controlled primarily by the appearance of new extremes provides a criterion to reject those IMF sets that are
from their previously hidden state”. non-orthogonal to the desired degree and it should be
The criterion (7) was revised and a simplified form as close to zero as possible because it corresponds to
for stopping proposed by Huang et al. (1999): the sifting a scaled sum of the absolute value of covariance matrix
process has to stop when the number of extremes equals off diagonal terms.
the number of zero-crossings for S successive sifting Other stopping criteria have been proposed, by
steps. This new simplified stoppage criterion relied on Rilling et al. (2003) who introduced thresholds that
the peculiarities of the specific data series. Although it force the final IMF to have globally low fluctuations
has the advantage of having a physical motivation, the of the mean value but that allow locally large signal
choice of value of S still remains an open question. excursions. These criteria have been put in the form of
The number of successive steps, S, of the simpli- three parameters, two thresholds, herein designated
fied criterion or limit  of Cauchy stopping criterion by a and b, and one that controls the portions (in per-
is a free parameter of sifting. Different sets of IMFs cent) of signal for which the thresholds apply, called
will be produced by EMD for different values of S or . . The threshold, a, that prescribes how much the
There is no generally accepted criterion about how to mean, m(t), can fluctuate with respect to the excursion
choose these parameters in order to obtain meaningful extremes, emax(t) and emin(t), is such that q(t)  m(t)/
results. Huang et al. (2003) provided good guidelines (emax(t)  emin(t))  a for 1  . The other allows for

1115
the locally large fluctuations of the mean q(t)  b and 5 RESULTS OF FIELD DATA ANALYSIS
they are chosen such that a  b and  is small. Rilling
et al. (2003) have proposed as defaults a  0.05, The choice of the stoppage criterion is important for
b  10a and   5%. the sifting process as mentioned before. In this work an
These proposed values are not connected with attempt is made to investigate the effect of the stop-
peculiarities of the given data series and seem to be page criterion on the proper extraction of the embed-
subjective ones. It is thought they should be determined ded oscillations in the data.
in a way analogous to parameter S of confidence limit The EMD is performed by selecting different stop-
of simplified criterion. Several combinations of (a, b) ping criteria.
have been used, the EMD was performed for each pair
and Hilbert spectrum H(, t) was calculated. The dif- 5.1 Stopping criterion
ferences between individual cases of Hilbert spec- The limit  of Cauchy convergence criterion can be
trum, integrated over time and frequency, and the overall considered as a size of the sieve, by which the sifting
mean was computed. The optimal values of the pair process of EMD is performed. The number n of IMF
(a, b) are determined from the range, where the dif- is examined as a function of  and results are pre-
ferences reach a local minimum. This opens the possi- sented in Figure 2 as the data of records 11, 13 and 17
bility of using some automated optimization procedure are used together. The number n of IMF increases
to find the optimal value of the pair (a, b). with reduction of the limit .
The influence of the stopping criterion on the spec-
tral characteristics of IMF is investigated. The zero-th
4 DATA moment mC0 j is proposed as a measure of integrally
determined energy of Cj IMF. The spectral peak fre-
The field data of sea surface elevation used in this quency of IMF fpCi is utilized as the representative fre-
work was collected at Figueira da Foz, Portugal, dur- quency. The peak frequency fp is determined here by
ing the period from March to June of 1994 (Paillard the simplest method, as a selection of the frequency
et al. 2000). The measurements were performed by a associated with the maximum spectral ordinate.
wave rider buoy at a depth of 90 m. The sea surface The uncertainty in the estimation of the spectral
elevation records are 20 min long at a sampling data peak frequency can be significant. However, it decreases
rate of 0.78 sec. with increasing degrees of freedom of the spectral
The peculiarities of the EMD and the importance estimates and with decreasing values of the normal-
of the stopping criterion is demonstrated on the basis ized frequency resolution df/fp, where df is the fre-
of the application of EMD to three records, represent- quency resolution (Young 1995). Here the normalized
ing different types of sea states, and it is discussed in frequency resolution is less than 0.05 and thus the
detail here. The energy spectra of these selected records spectral peak frequency fp is determined with com-
are shown in Figure 1. The spectra of record 13 and paratively little variability.
17, shown by dotted and thin line, are one peaked The energy mC0 i and peak frequency fpCi of the Ci
and correspond to swell with peak period Tp  16.7 s IMF from record 11 are presented as a function of 
and wind waves with Tp  7.7 s respectively. The data of in Figure 3(a) and (b), respectively. For simplicity
record 11 represent a mixture of two wave systems and
has a two peaked spectrum, thick solid line, with first 20
peak period Tp1  9.5 s and second one Tp2  16.0 s.

1.0E+01
Record 17
15
Number of IMF n

Record 13
Record 11
1.0E+00
0 0.05 0.1 0.15 0.2 0.25 0.3
Fourier Spectrum [m2.s]

1.0E-01
10

1.0E-02

1.0E-03 5
1.0E-08 1.0E-06 1.0E-04 1.0E-02 1.0E+00

1.0E-04
ε
frequency [Hz]

Figure 2. Number of IMF as a function of limit  of Cauchy


Figure 1. Fourier spectra of the wave records. convergence criterion.

1116
only characteristics of the first five IMF of the difficult to choose in advance the proper value of the
decomposition are drawn. The characteristics of the limit  of the Cauchy convergence criterion. The sim-
IMF, determined by the utilization of the simplified plified stoppage criterion relies on the specific pecu-
stoppage criterion are also listed in Table 1. liarities of the data and it is physically well founded
There is a tendency of increasing of the peak fre- and thus less subjective.
quency of IMF with decreasing of the limit . Utilizing The confidence limit for the EMD is first deter-
a finer size of the sieve, i.e., a small value of , shifts mined as a sifting process performed for 10 selected
the peak frequency f PCi of the IMF Ci into a higher fre- numbers of S: 3, 4, 5, 6, 7, 8, 9, 10, 15 and 20. The
quency range. squared deviation sd of the individual cases from an
The two dominant wave oscillations of record 11 ensemble mean for the 10 Hilbert spectra of records
are well separated by EMD with the simplified stop- 11, 13 and 17 are presented in Figure 4. The variation
page criterion and by criterion with limit   104. of orthogonal index OI as a function of stoppage num-
Similar estimations of energy and peak frequency are ber S is also shown in the same figures. The range of
provided by both stoppage criteria for some value of stopping number S, where the difference from the
the limit . Thus, with a special choice of limit , the computed ensemble mean has a minimum, is deter-
Cauchy convergence criterion and the simplified cri- mined for each record. For record 11 and 17 the range
terion can be considered identical. However, it is of S, where sd has a minimum is 5–10, while for
record 13 it is 4–9. The orthogonality index, OI, has
also small values for these values of S and thus the
obtained sets of IMF could be considered orthogonal
too for these values of S. All wave records are ana-
lyzed in this way to determine firstly the confidence
limit for the Empirical Mode Decomposition.

x102
2 0.2
sd (a)
OI
1.5 0.15

OI
sd

1 0.1

0.5 0.05

0 0
0 3 6 9 12 15 18 21
S

x103
8 0.2
sd
OI (b)
6 0.15

0.1
sd

4
OI

2 0.05

0 0
0 3 6 9 12 15 18 21
S
Figure 3. Spectral characteristics of IMF as a function of x103
limit , (a) energy (b) spectral peak frequency. 3
sd
0.2
(c)
2.5 OI
0.15
2
Table 1. Characteristics using the
OI
sd

1.5 0.1
simplified stopping criterion.
1
0.05
IMF m0 fp 0.5

0 0
C1 0.0070 0.3400 0 3 6 9 12 15 18 21
S
C2 0.0228 0.1125
C3 0.0206 0.0600
C4 0.0046 0.0500 Figure 4. The orthogonal index OI and squared deviation
C5 0.0003 0.0275 sd as a function of stopping number S for (a) record 11 (b)
record 13 and (c) record 17.

1117
100 Wave Data
C1
C2
C3
(b) C4
C5
C6
-1 C7
10
C8
C9
Fourier Spectrum

Figure 5. Level plot for the value (H  Hm)2, record 13. 10


-2

Next the Confidence limit for Rilling et al. (2003)


EMD algorithm is determined. Figure 5 represents the -3
10
surface plot for the values squared deviation, [H  Hm]2
for record 13. It is seen that independently of the rela-
tion a  b, that should be fulfilled due to physical con-
10-4
sideration, allowing strong local oscillations of the 0 0.05 0.1 0.15 0.2 0.25

envelope mean, it is further required that the values of Frequency [Hz]

a and b should be above a certain limit or else there is


over sifting. This limit is data dependent. Figure 6. Spectrum of record 11 and its IMF, determined
It has been observed that the minimization problem by (a) simplified stopping criterion (b) Rilling’s criterion.
is extremely sensitive. In Figure 5 the values greater
50 have been flattened to 50 so that the fine ridges of
low values can be distinguished. The minimum may parameters a  0.045, b  0.21. These values of sift-
sometimes be found surrounded by much larger val- ing are determined from the confidence limit of HHT
ues, as practically happens in Figure 5. This indicates method, discussed in details above.
that, to find the best EMD, the procedure should The Fourier spectrum is estimated as an average of
always be carried out for useful combinations of (a, b) the raw spectra, calculated in the overlapped segments,
to find the minimum to a desired precision. Maybe by which the wave record is divided. The spectral esti-
global search routines, such as genetic algorithms, can mations are also smoothed with a Hamming window.
be used to this end. Under a sampling rate of these data, the highest fre-
quency of Fourier spectrum is 0.64 Hz, but for clear-
5.2 Comparison of sets of IMF, determined by ness only until 0.25 Hz is shown.
different criteria A mixture of two wave systems is registered in the
record 11, whose Fourier spectrum is correspond-
The effect of stopping criterion on the decomposition ingly two-peaked, shown in thick solid line in Figure
is examined. The spectrum of wave record and its 6(a). These two dominant wave oscillations are well
IMF are calculated. The set of IMF are determined by separated by the EMD method into two different IMF-
using the simplified stopping criterion and the criteria C2 and C3 by both stopping criteria. The spectrum of
proposed by Rilling et al. (2003). The spectra of IMF of the first IMF C1, shown by a line with stars, covers
record 11, determined by sifting for S  7 successive very well the tail of the wave spectrum. The peaks of
times are shown in Figure 6(a), while Figure 6(b) pres- the spectrum of second C2 (open circles line) and third
ents the results of EMD by Rilling’s criterion with C3 (open triangles line) components well coincide

1118
Wave Data
Wave Data C1
C1 0 (b) C2
C2
10 C3
C3 C4
0 C5
10 (b) C4
C5 C6
C6 C7
C7 C8
-1
C8 10 C9
C9
-1
10
Fourier Spectrum
Fourier Spectrum

-2
10
-2
10

-3
10
10-3

-4
10
-4 0 0.05 0.1 0.15 0.2 0.25
10
0 0.05 0.1 0.15 0.2 0.25 frequency [Hz]
frequency [Hz]

Figure 8. Spectrum of record 17 and its IMF, determined


Figure 7. Spectrum of record 13 and its IMF, determined by (a) simplified stopping criterion (b) Rilling’s criterion.
by (a) simplified stopping criterion (b) Rilling’s criterion.
simplified criterion estimates that energy is at a much
with the major two peaks of spectrum of sea surface lower frequency, 0.0925 Hz against 0.040 Hz, respec-
elevation in both Figure 6(a) and 6(b). tively. There is a small hump on the signal spectrum
The swell type of waves prevails in the data of record for a frequency between 0.095–0.0975 Hz that appar-
13 – Figure 7. The sifting is performed with S  6 for ently is what has been picked up by C3 using Rilling’s
simplified criterion and a  0.035 and b  0.14 for criteria but the origin of the hump is not directly
Rilling’s criterion. For this wave system, the second C2 understood from simple wave-wave interaction because
IMF, whose spectrum is presented by open circle line in around the peak frequency and EMD considers it to
Figure 7(a) and 7(b), carries the largest part of the be a unique mode.
embedded energy in the record of sea surface elevation. The energy and time characteristics of individual
Record 17 contains mainly young short wind waves. IMF are estimated and compared with wave charac-
The wave spectrum of record 17 and its IMF are pre- teristics in order to estimate quantitatively the contri-
sented in Figure 8, as S  9 and a  0.045, b  0.08, bution of different IMF to the wave data.
respectively. The largest part of wave energy is extracted The spectral characteristics of records 11, 13 and
into the first C1 IMF by both simplified and Rilling’s 17 and their IMF, as determined by the simplified
criteria, whose spectrum is shown by the line with stopping criterion, are presented on the two leftmost
asterisks in Figure 8(a) and 8(b). columns of Table 2 to Table 4. The energy of the IMF
Qualitatively and quantitatively the results for Record with index higher than 4 is significantly lower than
11 and 13 are very similar, but for Record 17 it is seen the energy of the first four IMF. This tendency is con-
the distribution of energy density of C3 is very differ- firmed by the analysis of all other records, considered
ent, with Rilling’s criteria estimating that most energy in this study. From energetic point of view, the first
is close to the peak frequency of the signal while the four IMF contribute mainly to the energy contents of

1119
Table 2. Spectral characteristics for record 11. wave data, measured at 90 m depth. Shallow water
records of sea surface elevation could be dominated
Record 11, m0  0.055, fp1  0.0600, fp2  0.1125 by large low frequency waves and their decomposi-
tion picture data will be quite different. For complete
Rilling et al. (2003)
description of the analyzed data, the complete set of
Simplified criterion criterion
all IMF is necessary.
IMF fp [Hz] m0 [m2] fp [Hz] m0 [m2] In the two rightmost columns of Table 2 to Table 4
are presented the spectral characteristics of the IMF
C1 0.3400 0.0070 0.3400 0.0070 determined from Rilling et al. (2003) criteria. In com-
C2 0.1125 0.0228 0.1050 0.0228 parison with the results obtained from the simplified
C3 0.0600 0.0206 0.0600 0.0201 stopping criteria it is possible to find that both lead to
C4 0.0500 0.0046 0.0500 0.0048 EMD with most energy concentrated in the first four
C5 0.0275 0.0003 0.0300 0.0006 IMF. There are however differences in the spectral
C6 0.0100 0.0003 0.0100 0.0003 characteristics. For record 11, energies are within 5%
C7 0.0075 0.0002 0.0075 0.0001
C8 0.0050 9.3E5 0.0025 2.0E5
relative error reached at C4 (assuming the smaller value
C9 0.0050 1.1E4 0.0025 1.0E6 as correct) and spectral peak is within 7% reached at
C10 0.0050 1.0E6 0.0000 0.0E0 C2. It is relevant to note that the peak periods are
exactly the same for both criteria when comparing C1,
C3 and C4 from both criteria. For record 13, spectral
peak for C1 is very different as determined from both
Table 3. Spectral characteristics for record 13. criteria and this happens because of some slight shifts
of energy, other three are exactly equal. Record 17 has
Record 13, m0  0.1329, fp  0.0575 some large differences starting at C3 especially due to
a shift in energy to C2 as determined with Rilling’s
Rilling et al. (2003) criteria.
Simplified criterion criterion In any case it is interesting to note the following, as
IMF fp [Hz] m0 [m2] fp [Hz] m0 [m2]
per conventional frequency domain analysis; for record
11, C3 has the frequency of the signal’s first spectral
C1 0.3550 0.0041 0.0550 0.0059 peak, C2 has approximately the same frequency of the
C2 0.0675 0.0382 0.0675 0.0391 signal’s second peak and C4 corresponds to the differ-
C3 0.0575 0.0746 0.0575 0.0698 ence frequency between the two peaks. Thus some of
C4 0.0325 0.0056 0.0325 0.0035 the dominant IMF may correspond to modes of wave-
C5 0.0200 0.0010 0.0200 0.0009 wave interaction as conventionally described through
C6 0.0100 0.0002 0.0125 0.0002 sum and difference frequencies with the important
C7 0.0050 0.0002 0.0050 0.0002 particularity that the amplitude is not constant.
C8 0.0050 3.2E5 0.0025 1.1E5
C9 0.0050 3.0E6 0.0000 3.0E6
C10 0.0050 3.0E6 0.0000 0
6 CONCLUSIONS

Table 4. Spectral characteristics for record 17. The EMD methodology provides a new look at wave
data in a manner that can be considered a generaliza-
Record 17, m0  0.1113, fp  0.1275 tion of linear normal modes. There is however a bulk of
the understanding that still relies on frequency domain
Rilling et al. (2003) analysis in a traditional sense, which is based on lin-
Simplified criterion criterion
earity, but in this case the analysis is performed on the
IMF fp [Hz] m0 [m2] fp [Hz] m0 [m2] IMFs and not directly on the signal.
The EMD depends on adequate criteria to attain a
C1 0.1275 0.0871 0.1300 0.0620 reliable estimate of the IMFs. It has been shown that
C2 0.1125 0.0299 0.1150 0.0437 different criteria may lead to slightly different decom-
C3 0.0400 0.0022 0.0925 0.0043 positions, but the fact that there is a search for the
C4 0.0200 0.0004 0.0300 0.0006 minimum deviation of the Hilbert spectrum from the
C5 0.0100 0.0002 0.0175 0.0003 average based on several parameter combinations,
C6 0.0050 0.0004 0.0150 0.0002 appears to be a strong indication that the IMFs are
C7 0.0050 0.0001 0.0075 0.0003 adequate representation of the nonlinear modes that
C8 0.0025 3.5E5 0.0025 3.0E6
C9 0.0025 9.0E6 0.0000 0
form the original signal. It is extremely important that
C10 0.0000 1.0E6 0.0000 0 such determinism exists, because otherwise there would
be no basis to using one or other decomposition.

1120
It has been shown that in double peaked spectra, nonlinear and non-stationary time series analysis”, Proc.
IMFs will contain information that is directly related R. Soc. Lond., Vol. 454, pp. 903–995
to wave-wave interaction as per conventional frequency Huang, N. E., Wu, M. L., Long, S. R., Shen, S. P., Per, W. Q.,
domain analysis, but richer because the amplitude is Gloersen, P. and Fan, K. L., 2003, “A confidence limit for
the empirical mode decomposition and Hilbert spectral
variable. analysis”, Proc. R. Soc. Lond. Vol. 459, pp. 2317–2345.
The Intrinsic Mode Functions are data dependent Huang, N. E., Shen, Z. and Long, S. R., 1999, “A new view
and finite number, about 10 or less are sufficient for of nonlinear water waves: The Hilbert spectrum”, Ann.
full reconstruction of ocean wave records. Rev. Fluid Mech. 31, pp. 417–457.
Liu, P. and Mori, N., 2000, “Wavelet spectrum of freak
waves in the ocean”, Proc. of ICCE 2000, p. 1092–1098.
Massel, S. R., 2001, “Wavelet analysis for processing of ocean
ACKNOWLEDGMENTS surface wave records”, Ocean Eng. 28, pp. 957–987.
Paillard, M., Prevosto, M., Barstow, S. and Guedes Soares, C.,
Work of the first author has been financed by the grant 2000, “Field Measurements of Coastal Waves and Currents
SFRH/BD/10527/2002 from the Portuguese Foundation in Portugal and Greece”, Coastal Engineering, Vol. 40,
for Science and Technology (Fundação para a Ciência pp. 285–296.
e Tecnologia). Rilling, G., Flandrin, P. and Goncalves, P., 2003, “On Empirical
Mode Decomposition and its Algorithms”, Proc. IEEE-
EURASIP Workshop on Nonlinear Signal and Image
Processing NSIP-03.
REFERENCES Vakakis, A. F., Manevitch, L. I., Mikhlin, Y. V., Pilipchuk, V. N.
and Zevin, A. A., 1996, Normal Modes and Localization
Guedes Soares, C. and Cherneva, Z., 2004, “Spectrogram Modes and Localization in Nonlinear Systems, J Wiley.
Analysis of the Frequency Characteristics of Ocean Wind Veltcheva, A. D., 2002, “Wave and group transformation by
Waves”, Ocean Eng. (accepted for publication). a Hilbert spectrum”, Coastal Engineering, Vol. 44, No. 4,
Guedes Soares, C., Pascoal, R., Antão, E., Voogt, A. and pp. 283–300.
Buchner, B., 2004, “An Approach to Calculate the Prob- Veltcheva, A. D. and Guedes Soares, C., 2004, “Identi-
ability of Impact on an FPSO Bow”, Proc. of 23rd Conf- fication of the components of wave spectra by the Hilbert
erence on Offshore Mechanics and Arctic Engineering, Huang transform method”, Applied Ocean Research,
ASME Paper, Canada, OMAE2004-51575. Vol. 26, pp. 1–12.
Hellan, Ø., Hermundstad, O. A. and Stansberg, C. T., 2001, Veltcheva, A. D., Pascoal, R. and Guedes Soares, C., 2005,
“Design Tool for Green Sea, Wave Impact, and Structural “Insight on the Nonlinearity of Ocean Waves by Hilbert
Response on Bow and Deck Structures”, Proc. Of Offshore Huang Transform Method”, Proc. of the 5th Inter-
Technology Conference, Paper OTC 13213, Houston, national Symposium on Ocean Wave Measurements and
Texas, U.S.A. Analysis (WAVES 2005), Madrid, Spain.
Huang, N. E., Shen, Z., Long, S.R., Wu, M.C., Shin, H. S., Young, I. R., 1995, “The determination of confidence limits
Zheng, Q., Yuen, Y., Tung, C. C. and Liu, H.H., 1998, “The associated with estimates of the spectral peak frequency”,
empirical mode decomposition and Hilbert spectrum for Ocean Engineering, Vol. 22, No. 7, 669–686.

1121
View publication stats

You might also like