2

A signal decomposition theorem with Hilbert transform and its
application to narrowband time series with closely spaced

frequency components
Genda Chen
a,n
, Zuocai Wang
b
a
Department of Civil, Architectural, and Environmental Engineering, Missouri University of Science and Technology, 328 Butler-Carlton Hall, 1400 N. Pine Street,
Rolla, MO 65409-0030, USA
b
Department of Civil, Architectural, and Environmental Engineering, Missouri University of Science and Technology, 326 Butler-Carlton Hall, 1400 N. Pine Street,
Rolla, MO 65409-0030, USA
a r t i c l e i n f o
Article history:
Received 2 December 2010
Accepted 2 February 2011
Available online 18 February 2011
Keywords:
Empirical mode decomposition
Hilbert transform
Analytical mode decomposition
Closely spaced frequency components
System identication
a b s t r a c t
Empirical mode decomposition and Hilbert spectral analysis have been extensively
studied in recent years for the system identication of structures. It often encounters
three challenges: (1) unable to decompose a signal with closely spaced frequency
components such as wave groups in ocean engineering and beating responses in
structural and mechanical systems, (2) difcult to distinguish the frequency components
in a narrowband signal that is commonly seen in the free vibration of structures, and
(3) unable to separate small intermittent uctuations from a large wave. In this paper, a
new analytical mode decomposition theorem based on the Hilbert Transform of a
harmonics multiplicative time series is developed to address the challenges. The
theorem can be applied with two procedures based on the decomposition of the original
signal only or the previously decomposed (modied) signals in sequence. Numerical
examples for four representative engineering applications indicate that the new theorem
is superior to existing methods in decomposing a time series into many signals whose
Fourier spectra are non-vanishing over mutually exclusive frequency ranges separated
by bisecting frequencies. It is simple in concept, efcient in computation, consistent in
performance, and reliable in signal processing. The accuracy of the new theorem is
insensitive to noise. The discernable frequency spacing between the dominant frequen-
cies of decomposed signals is theoretically near zero but practically equal to twice the
frequency resolution of a nite length time series. Each bisecting frequency can be
selected as an average of its two nearby frequencies of interest and is insensitive to other
choices between 80% and 120% of the average value. The modied signal decomposition
procedure can be less accurate than the original signal decomposition procedure due to
potentially accumulated numerical errors in Hilbert transforms.
& 2011 Elsevier Ltd. All rights reserved.
1. Introduction
System identication is a process to estimate the characteristic properties of a structural system with dynamic
measurements. In the past fteen years, Hilbert Transform has received increasing attentions in the eld of system
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/ymssp
Mechanical Systems and Signal Processing
0888-3270/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2011.02.002
n
Corresponding author. Tel.: +1 573 341 4462; fax: +1 573 341 6215.
E-mail address: gchen@mst.edu (G. Chen).
Mechanical Systems and Signal Processing 28 (2012) 258279
identication and damage detection [15]. In order to admit a well-behaved Hilbert transform, Huang et al. [68]
developed the empirical mode decomposition (EMD) and Hilbert spectral analysis, called HilbertHuang Transform (HHT).
EMD separates a complicated data set into a nite number of intrinsic mode functions (IMFs), each having a well-behaved
Hilbert transform. In combination with the random decrement technique, Yang et al. [911] successfully applied the HHT
method into the identication of natural frequencies and damping ratios for linear structures. Yang et al. [12] further
proposed an EMD-based approach to determine the time and location of damage by identifying the damage-induced
pulses due to a sudden change of structural stiffness.
Although powerful in extracting the properties of nonlinear and non-stationary signals, HHT faces several challenges in
some engineering applications: (1) unable to decompose signals with closely spaced frequency components such as wave
groups in ocean engineering and beating responses in structural and mechanical systems, (2) difcult to distinguish
various components in narrowband signals that are commonly seen in free vibration of structures, and (3) unable to
separate small intermittent uctuations from a large ocean wave. Attempt has been made by a few investigators to address
these challenges. With the HHT method, Chen and Xu [13] identied the modal parameters of a linear system with closely
spaced modes of vibration. They established the minimum cutoff frequency criteria in the EMD sifting process to ensure
that the frequency bandwidths of IMFs be separated. Cautions must be exercised with necessary intermittency checks as
introducing the additional cutoff frequencies is a subjective intervention to the EMD process and may unintentionally
distort the IMFs. Peng et al. [14] used the wavelet packet transform to decompose a general signal into a set of narrowband
signals in order to derive single-component IMFs. Yang et al. [9] applied a band pass lter on the measured free vibration
time histories to identify the modal parameters to a multiple-degree-of-freedom system. Chen and Feng [15] took
advantage of the waves beating phenomena to facilitate the EMD process. Wang [16] decomposed a small intermittent
uctuation around a large ocean wave by injecting a known time function into the original signal for two purposes. The
rst purpose was to introduce a series of additional maxima/minima when the added time function is rich in high
frequency. The second purpose is to down shift the frequency components of the signal using so that the ratio of two close
frequencies in a wave group can be increased for a smooth EMD process. The wave-group method may not be applicable to
exible structures, such as long-span bridges, since their natural frequencies are already low. More recently, Zheng et al.
[17] proposed the singular value decomposition and band pass ltering technique to extract the parameters of closely
spaced modes. With this method, the challenge may remain as the space between two frequencies becomes smaller.
Overall, it is still quite a challenge to consistently and reliably identify the properties of structures with closely spaced
modes.
In this paper, a new signal decomposition theorem is developed to address the above three challenges. The essence of
this theorem is the exact separation of a general time series into two time functions whose Fourier spectra are non-
vanishing over two mutually exclusive frequency ranges separated by a bisecting frequency. With multiple steps of
bisecting, an original time series with multiple closely spaced frequency components can be decomposed into many
signals, each dominated by a single frequency component. Emphasis in this study is placed on understanding the
minimum frequency separation of two closely spaced frequency components, the effect of a bisecting frequency selection,
and the performance advantage of the proposed theorem over existing techniques in four representative engineering
applications.
2. A signal decomposition theorem based on Hilbert transform
2.1. Theorem
Let x(t) denote a real time series of n signicant frequency components o
1
,o
2
,. . .,o
n
; all positive in L
2
(N, +N) of
the real time variable t. It can be decomposed into n signals x
d
i
t (i =1,2,y,n) whose Fourier spectra are equal to
^
Xo over
n mutually exclusive frequency ranges (9o9oo
b1
), o
b1
o o j j oo
b2
,. . ., (o
b(n2)
o9o9oo
b(n1)
), and (o
b(n1)
o9o9).
That is,
xt
X
n
i 1
x
d
i
t 1
Here,
^
Xo is the Fourier transform of x(t), o represents a frequency variable, and o
bi
2 o
i
,o
i 1

i 1,2,. . .,n1 are n1
bisecting frequencies. Each signal has a narrow bandwidth in the frequency domain and can be determined by
x
d
i
t s
i
ts
i1
t,. . .,x
d
n
t xts
n1
t 2
s
i
t sino
bi
tHxtcoso
bi
tcoso
bi
tHxtsino
bi
t i 1,2,. . .,n1 3
where s
0
(t)=0, and H[.] represents the Hilbert transform of the function inside the square bracket.
G. Chen, Z. Wang / Mechanical Systems and Signal Processing 28 (2012) 258279 259
2.2. Proof
The signal decomposition theorem is proven in two steps. The rst step is to split a time series into two signals whose
Fourier spectra are non-vanishing over two mutually exclusive frequency ranges about a bisecting frequency. The second
step is to apply the bisecting process to derive the decomposed signals x
d
i
t (i =1,2,y,n) as stated in the theorem.
2.2.1. Bisection of a time series
A time series can generally be expressed into a summation of two signals:
xt s
1
t s
1
t 4
whose Fourier transforms
^
s
1
o and
^
s
1
o vanish for 9o94o
b
and 9o9oo
b
, respectively. Here, o
b
is an arbitrary positive
value referred to as the bisecting frequency. Due to the Parsevals theorem in Fourier transforms and the fact that ^ s
1
o and
^
s
1
o are equal to
^
Xo over two mutually exclusive frequency ranges of (N,o
b
) and (o
b
, +N), respectively, as
demonstrated in Appendix A:
Z
1
1
s
1
t
2
dt
Z
1
1
^ s
1
o
2
do=2pr
Z
1
1
^
Xo
2
do=2p
Z
1
1
xt
2
dt o1 5
Z
1
1
s
1
t
2
dt
Z
1
1
^
s
1
o
2
do=2pr
Z
1
1
^
Xo
2
do=2p
Z
1
1
xt
2
dt o1 6
Therefore, both s
1
t and s
1
t are real functions in L
2
(N, +N).
Let s
c
(t)=cos(o
b
t) and s
s
(t)=sin(o
b
t). The Hilbert transform of s
k
(t)x(t) (k=c, s) can be expressed into:
H s
k
t xt H s
k
ts
1
t Hs
k
ts
1
t 7
Since the Fourier transforms of s
c
(t) and s
s
(t) vanish at all frequencies except for 9o9=o
b
, they are non-vanishing over
mutually exclusive frequency ranges with the Fourier transforms s
1
(t) and s
1
t. The rst term on the right-hand side of
Eq. (7) has a low-pass function of s
1
(t) and a high-pass function of s
k
(t). The second term has a low-pass function of s
k
(t)
and a high-pass function of s
1
t. According to the Bedrosian theorem in Hilbert Transform [18], Eq. (7) can be rewritten as
H s
k
txt s
1
tHs
k
t s
k
tHs
1
t 8
When k is set to c and s, respectively, two algebraic equations for two unknowns, s
1
(t) and Hs
1
t, can be obtained,
leading to
s
1
t
s
s
tH s
c
txt s
c
tHs
s
txt
s
s
tH s
c
t s
c
tHs
s
t
9
H s
1
t
Hs
c
tH s
s
txt Hs
s
tHs
c
txt
s
s
tH s
c
t s
c
tHs
s
t
10
The Hilbert transforms of s
c
(t) and s
s
(t) can be expressed into:
Hs
c
t sino
b
t and Hs
s
t coso
b
t 11
Thus,
s
s
tHs
c
ts
c
tHs
s
t 1 12
and Eqs. (9) and (10) become:
s
1
t sino
b
tH xtcoso
b
t coso
b
tHxtsino
b
t 13
H s
1
t sino
b
tH xtsino
b
t coso
b
tHxt coso
b
t 14
The signal s
1
t and the Hilbert transform of s
1
(t) can be derived from Eq. (4) and expressed into:
s
1
t xts
1
t 15
H s
1
t H xt Hs
1
t 16
As demonstrated in Appendix A, by relating the Hilbert transform to Fourier transform, it can be veried that the
Fourier transform of s
1
(t) in Eq. (13) is equal to
^
Xo for ooo
b
and zero for o4o
b
. Similarly, the Fourier transform of s
1
t
in Eq. (15) is equal to
^
Xo for o4o
b
and zero for ooo
b
. Eq. (3) is the direct application of Eq. (13) with a bisecting
frequency of o
bi
i 1,2,. . .,n1.
2.2.2. Decomposition of a time series by repeating the bisecting process
By selecting o
b
o
b1
,o
b2
,. . .,o
bn1
, a time series can be bisected into two signals in various ways as follows:
xt s
1
t s
1
t s
2
ts
2
t s
n1
ts
n1
t 17
Eq. (2) can then be expressed into:
x
d
1
s
1
t, x
d
i
t s
i
ts
i1
t, i 2,3,. . .,n1, x
d
n
t s
n1
t 18
in which s
i
t i 1,2,. . .,n1 and s
n1
t can be evaluated by using Eqs. (13) and (15) with o
b
=o
bi
and o
b(n1)
,
respectively. This process is referred to as the original signal decomposition procedure.
Alternatively, by bisecting the previously decomposed signals with o
b
o
b1
,o
b2
,. . .,o
bn1
in sequence, a time series
can be decomposed into:
xt s
1
ts
1
t, s
1
t s
2
ts
2
t,. . .,s
n2
t s
n1
ts
n1
t 19
The decomposed signals in Eq. (1) can then be expressed into:
x
d
i
t s
i
t, i 1,2,. . .,n1 , x
d
n
t s
n1
t 20
Note that the s
i
(t) values in Eq. (20) are different from those in Eq. (18) except when i =1. This alternative process is
referred to as the modied signal decomposition procedure.
3. System identication
For the system identication of an n-degree-of-freedom (DOF) structure, the Fourier transform of an original measured
response, such as structural displacement or acceleration, can be plotted as illustrated in Fig. 1 for the case of n=3. The
frequencies o
1
,o
2
,. . .,o
n
correspond to the peaks of the Fourier transform and can be estimated from Fig. 1. The bisecting
frequencies o
b1
,o
b2
,. . .,o
bn1
can then be selected in between any two natural frequencies of the structure as shown in
Fig. 1.
Each decomposed signal from Eqs. (18) and (20) has a narrow frequency band corresponding to one mode of vibration.
For this reason, applications of the new theorem are hereafter referred to as an analytical mode decomposition (AMD)
method. Similar to the HHT method [9], the instantaneous amplitude A
pi
(t) and phase angle y
pi
(t) of the ith decomposed
mode x
d
pi
t at the pth DOF can be determined from:
A
pi
t e
jy
pi
t
x
d
pi
t jHx
d
pi
t 21
where j
1
p
is an imaginary unit. If the structure is subjected to an impulsive load, the instantaneous frequency and
damping ratio can be detected using the least-square t procedure in [9]:
o
i

dy
pi
t
dt
and B
i

d lnA
pi
t

o
i
dt
22
In general, the response of the pth DOF x
p
(t) can be expanded into a summation of modal contributions, each being the
product of the pth modal element f
pi
and the modal coordinate g
i
t i 1,2,. . .,n. That is
x
p
t
X
n
i 1
x
d
pi
t
X
n
i 1
j
pi
g
i
t 23
Considering the ith modal contributions from the pth and qth DOFs, j
pi
g
i
(t) and j
qi
g
i
(t), the instantaneous mode shape
can be described by
j
pi
j
qi
A
pi
t
A
qi
t
24
The difference between the phase angles of the pth and qth modal elements j
pi
and j
qi
can be expressed into
j
i
pq
y
pi
ty
qi
t, which is in theory equal to either 72mp or 7(2m+1)p (m is an integer). In frequency domain, the
A
m
p
l
i
t
u
d
e
Frequency
Fig. 1. Representative Fourier spectrum of a structural response.
modal elements can be directly evaluated by
j
pi
j
qi
Im^ x
d
pi
o
i
Im^ x
d
qi
o
i
25
for a time-invariant structure. In Eq. (25), ^ x
d
pi
o
i
and ^ x
d
qi
o
i
are the Fourier transforms of x
d
pi
t and x
d
qi
t i 1,2,. . .,n
and Im[.] represents the imaginary part of the complex argument inside the square bracket.
4. Parametric analysis
To understand the accuracy of the two proposed procedures and the effect of various bisecting frequency selections, a
time series with three close frequency components was considered: x(t)=x
1
(t)+x
2
(t)+x
3
(t). The three frequencies were set
to f
1
=1.1 Hz, f
2
=1.2 Hz, and f
3
=1.3 Hz with a frequency spacing of 0.1 Hz. A time step of 0.02 s was used in numerical
simulations.
4.1. Comparison between the two proposed procedures
The three-component time series used to test the two analysis procedures is a 20-s time function of the following
original signals: x
1
t cos2pf
1
t, x
2
t cos2pf
2
t p=6, and x
3
t cos2pf
3
t. The frequency resolution in Fourier
transforms is 0.05 Hz. The decomposed signals obtained following the original signal decomposition procedure and their
Fourier transforms (normalized by the number of data points in the time series) are presented in Figs. 2 and 3, respectively.
It can be seen from Figs. 2 and 3 that the decomposed signals are in excellent agreement with the exact signals both in
time and frequency domains. The AMD method can accurately separate the three components. Note that a slight difference
observed at the beginning and end of the time series is due to the end effect of Hilbert transforms as a result of a nite time
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
Decomposed signal Origianal signal x1
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
Decomposed signal Original signal x2
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
Fig. 2. Decomposed signals in time domain: original signal decomposition procedure.
series. In this study, the mirror image technique about the rst or the last maxima/minima of the original signal was used
to minimize the end effect in the numerical integration of Hilbert transforms [19].
The signals obtained following the modied signal decomposition procedure and their Fourier transforms are presented
in Figs. 4 and 5. Similar to the original signal decomposition procedure, the decomposed signals are in excellent agreement
with the exact signals. By comparing Figs. 4 and 5 with Figs. 2 and 3, no appreciable difference is observed, indicating the
similar accuracy that can be achieved with the two procedures. However, as the number of frequency components
increases, the numerical error associated with a series of Hilbert transforms on the previously decomposed signals may be
accumulated with the modied signal decomposition procedure. The original signal decomposition procedure only
involves one Hilbert transform and, in theory, is more accurate than the modied signal decomposition procedure.
4.2. Selection of bisecting frequencies
To investigate the effect of bisecting frequency selections on the accuracy of signal decompositions, two frequency
resolutions were considered: Df =0.05 and 0.025 Hz. They correspond to a time series of 20 and 40 s, respectively.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x3
Fig. 3. Fourier spectra of the decomposed signals: original signal decomposition procedure.
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
Fig. 4. Decomposed signals in time domain: modied signal decomposition procedure.
Furthermore, three relative intensities of the three frequency components were chosen to understand their potential
inuence on the selection of the bisecting frequency. Therefore, a total of six cases were analyzed with the following
original signals: (1) x
1
t cos2pf
1
t, x
2
t cos2pf
2
t p=6, and x
3
t cos2pf
3
t; (2) x
1
t 0:5cos2pf
1
t, x
2
t
cos2pf
2
t p=6, and x
3
t 0:5cos2pf
3
t; and (3) x
1
t cos2pf
1
t, x
2
t 0:5cos 2pf
2
t p=6

, and x
3
t cos 2pf
3
t .
The normalized Fourier spectra for the above six cases are presented in Fig. 6. The amplitudes of the pulses depend
upon the frequency resolution used in numerical integrations with the Fast Fourier Transform technique. As the frequency
resolution improves from 0.05 to 0.025 Hz, the three pulses in each case become narrower and higher and their relative
amplitudes are more accurate. This is because the amplitude of a pulse is expected to decrease with the frequency,
Fig. 6(b), as a result of the increasing aliasing effect and the increasing impact by the incomplete data set whose total
number of points is not equal to the multiplier of 2 in the Fast Fourier Transform. However, selection in the frequency
resolution has little effect on the Fourier spectrum comparison between the decomposed signal and the exact signal so
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x3
Fig. 5. Fourier spectra of decomposed signals: modied signal decomposition procedure.
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
Frequency (Hz)
A
m
p
l
i
t
u
d
e
A
m
p
l
i
t
u
d
e
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Fig. 6. Fourier spectra of time series in six cases: (a,b) for xt cos2pf
1
tcos2pf
2
t p=6 cos2pf
3
t, (c,d) for xt 0:5cos2pf
1
t
cos2pf
2
t p=6 0:5cos2pf
3
t, and (e,f) for xt cos2pf
1
t 0:5cos2pf
2
t p=6 cos2pf
3
t. (a) Df =0.05Hz, (b) Df =0.025Hz, (c) Df =0.05Hz,
(d) Df =0.025Hz, (e) Df =0.05Hz and (f) Df =0.025Hz.
long as they use the same resolution as done in Figs. 3 and 5. Therefore, all examples to illustrate the superiority of the
AMD theorem over existing technologies were analyzed with a minimum frequency resolution of 0.05 Hz. For system
identication and damage detection, a frequency resolution of 0.025 Hz is recommended to ensure the accurate
representation of various frequency components in a structural response. Note that a weak signal in between two strong
signals requires a higher frequency resolution to detect.
The amplitude ratio between the Fourier transforms of the decomposed signal ^ x
d
i
o
i
and the original signal ^ x
i
o
i
is
presented in Fig. 7 for i =1,2,3 in each of the six cases. The bisecting frequency was considered as a weighted average of the
beginning and end frequencies of a range: o
b1
/2p=af
1
+(1a)f
2
and o
b2
/2p=af
3
+(1a)f
2
, in which 1a represents the
weight factor on the center frequency f
2
. The extreme condition a=0 means o
bi
=2pf
2
(i =1, 2). It can be seen from Fig. 7
that the relative error in signal decomposition is nearly independent of the relative signal amplitudes and less than 3.3%
when the bisecting frequency takes 80120% of the frequency at a=0.5. Therefore, the bisecting frequency can be selected
to be the average of the beginning and end frequencies and the accuracy in mode decomposition is insensitive to this
selection.
4.3. Comparison with a frequency ltering technique
The signal decomposition theorem is developed to split a time series into n signals whose frequency contents are
mutually exclusive about n1 bisecting frequencies. This process appears somewhat similar to a band pass ltering
technique with multiple rectangular windows in the frequency domain. Therefore, a comparative study was conducted to
understand their difference. For fair comparisons, the cutoff frequencies in frequency band pass lters were set to equal
the bisecting frequencies used with the AMD method.
0.9
1.0
1.1
0.9
1.0
1.1
0.9
1.0
1.1
0.9
1.0
1.1
0.9
1.0
1.1
Decomposed signal x1
0.9
1.0
1.1
0

0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 7. Effect of bisecting frequency selection on amplitude ratio: (a,b) for x
1
t cos2pf
1
t, x
2
t cos2pf
2
t p=2, and x
3
t cos2pf
3
t;
(c,d) forx
1
t 0:5cos2pf
1
t, x
2
t cos2pf
2
t p=6, and x
3
t 0:5cos2pf
3
t; and (e,f) for x
1
t cos2pf
1
t, x
2
t 0:5cos2pf
2
t p=6, and
x
3
t cos2pf
3
t. (a) Df =0.05Hz, (b) Df =0.025Hz, (c) Df =0.05Hz, (d) Df =0.025Hz, (e) Df =0.05Hz and (f) Df =0.025Hz.
Fig. 8 compares the decomposed signals of the time series used in Figs. 2 and 4 by band pass ltering with Df =0.05 Hz.
It can be clearly seen from Fig. 8 that the decomposed signals are signicantly different from the original signals,
particularly at the beginning and end of the signals. In comparison with Figs. 2 and 4, the band pass ltering technique is
less accurate than the AMD method. The transient oscillations at the beginning and end of the signals due to ltering are
substantially more signicant than those by the AMD method.
With band pass ltering, a time series of nite length in engineering applications is rst transformed into the frequency
domain (Fourier transform). In the time domain, the nite length signal is equivalent to the application of a rectangular
window into the innite long time series. After having been multiplied by a rectangular window in the frequency domain,
the Fourier spectrum is transformed back to the time domain (inverse Fourier transform). Both transforms introduce
numerical errors in two ways: signal modication by windowing and end effect due to sharp edges or brick walls of the
windows. Applying a rectangular window in the frequency/time domain is equivalent to execute a convolution between a
sinc function and the time/frequency function over an innite range. This means that the time/frequency function is now a
distorted signal as indicated in Fig. 8. The brick wall with the rectangular window or the sudden change of frequency
corresponds to an innitely long oscillation in the time domain that cannot be represented accurately with the Fast Fourier
Transform and its inverse transform. An improper selection of the beginning and end of a signal could introduce an
articial oscillation as illustrated in Fig. 8. In the case of splitting closely spaced frequency components, there is no space to
soften the brick wall by designing two smooth edges of a window.
On the other hand, the AMD method only involves two Hilbert transforms that were evaluated with the inverse Fourier
transforms of one-sided spectra without rectangular windows. Therefore, the problems associated with frequency band
pass ltering do not exist with the AMD method. However, the issue with the nite length of a time series remains in
practical applications. Only for a time series (even function) with locally symmetric data points about the rst and last
maxima/minima such as cosine signals with a zero phase angle can the mirror image technique [19] smoothly extend the
nite length to innity and eliminate the end effect in Hilbert transforms. The end effects in Figs. 2 and 4 result from the
incomplete cycles in the time series.
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (Sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (Sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
3
Time (Sec)
Fig. 8. Decomposed signals by frequency ltering and exact signals.
5. Numerical examples in various engineering applications
To demonstrate the sensitivity to noise and the performance advantage over existing techniques in the literature, the
proposed method was applied to analyze ve examples representing various engineering applications. The results are
compared with those from the existing techniques.
5.1. Signal decomposition with noise effects
Consider a 20-s, two-frequency time series x(t)=x
1
(t)+x
2
(t)+rand(s) with the Gaussian white noise rand(s): x
1
t
sin2pf
1
t and x
2
t sin2pf
2
t where f
1
=1.0 Hz and f
2
=1.5 Hz. The ratio of the two frequencies is f
2
/f
1
=1.5. The standard
deviation of the mean-zero white noise is 10% of the signal strength. That is, s=0.1max9x
1
(t)+x
2
(t)9. A time step of 0.02 s
was used in simulations.
Based on the EMD, four IMFs were determined from the time series. The rst two IMFs had wideband frequency
components that were mainly contributed by the white noise. The 3rd and 4th IMFs, IMF3 and IMF4, and their Fourier
transforms are compared with their respective exact signals in Figs. 9 and 10, respectively. It can be clearly seen that both
IMFs have signicant errors and the two frequency components are still mixed.
To apply the proposed AMD method, o
b
=2p(f
1
+f
2
)/2=2.5p rad/s was selected. Figs. 11 and 12 present the decomposed
signals and their Fourier transforms, respectively. As one can see, the two frequency components were completely
separated and the decomposed signals are in excellent agreement with their respective exact signals. Therefore, the
proposed AMD is superior to the EMD.
5.2. Signal decomposition with close frequency components
A two frequency component signal was considered: x(t)=x
1
(t)+x
2
(t), x
1
(t)=cos(2pf
1
t+p/6) and x
2
(t)=cos(2pf
2
t). Both
long period ocean wave and short period mechanical wave were analyzed.
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
I
M
F
4

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
1
IMF4 Original signal x1
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
I
M
F
3

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
2
Fig. 9. IMF3 and IMF4 by EMD versus exact signals.
0 0.5 1 1.5 2 2.5 3
0
0.1
0.2
0.3
0.4
0.5
Frequency (Hz)
A
m
p
l
i
t
u
d
e
0 0.5 1 1.5 2 2.5 3
0
0.2
0.4
0.6
0.8
Frequency (Hz)
A
m
p
l
i
t
u
d
e
IMF4
IMF3
Fig. 10. Fourier spectra of IMF3 and IMF4 by EMD.
5.2.1. Long period ocean wave
The processing of an ocean wave data was considered [16]. A time step of 0.1 s was adopted in this example. The two
frequencies were taken to be f
1
=1/32 Hz and f
2
=1/30 Hz. Their ratio was f
2
/f
1
=1.067. Figs. 13 and 14 present the 1st and
2nd IMFs, IMF1 and IMF2, and their Fourier transforms based on the wave group method with a downshift frequency of
1/28 Hz [16]. Similarly, Figs. 15 and 16 show the decomposed signals and their Fourier transforms obtained from the
proposed AMD method. The results indicate that both the wave group method with EMD and the AMD method can
effectively separate the two frequency components in this case. However, the Fourier spectra in Fig. 16 are still cleaner
than those in Fig. 14 outside the two frequency components f
1
and f
2
.
5.2.2. Short period mechanical wave
A 20-s signal with f
1
=1.1 Hz and f
2
=1.2 Hz was considered. The ratio of the two frequencies was f
2
/f
1
=1.091. A time
step of 0.02 s was used. Figs. 17 and 18 present the rst two IMFs and their Fourier spectra based on the wave group
method with a downshift frequency of 1.3 Hz. Figs. 19 and 20 depict the decomposed signals and their Fourier spectra
based on the proposed AMD method. It can be clearly seen that the wave group method cannot completely separate the
two frequency components even though the frequency ratio of 1.091 is greater than 1.067 in Section 5.2.1. This indicates
that the wave group method with EMD cannot ensure that a reasonable solution be achieved for general cases due to the
inherent issue with EMD. On the other hand, the proposed AMD can consistently separate the two close frequency
components with high accuracy.
5.3. Signal decomposition with amplitude decays
Consider a two frequency component signal: x(t)=x
1
(t)+x
2
(t), x
1
t 10e
B2pf
1
t
cos2pf
1
t, and x
2
t 1:8e
B2pf
2
t
cos2pf
2
t. The two frequencies and the damping ratio are f
1
=2.0 Hz, f
2
=4.3 Hz, and B=0.02. A time step of 0.02 s was
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
Fig. 11. Decomposed signals by AMD versus exact signals.
0 0.5 1 1.5 2 2.5 3
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.5 1 1.5 2 2.5 3
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 12. Fourier spectra of the decomposed signals by AMD versus exact signals.
0 100 200 300 400 500 600 700 800 900 1000
-2
-1
0
1
2
Time (Sec)
I
M
F
2

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
1
0 100 200 300 400 500 600 700 800 900 1000
-2
-1
0
1
2
Time (Sec)
I
M
F
1

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
2

Fig. 13. IMF1 and IMF2 by the wave group method with EMD versus exact signals.
0 0.01 0.02 0.03 0.04 0.05 0.06
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
IMF2
Original signal x1
0 0.01 0.02 0.03 0.04 0.05 0.06
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
IMF1
Original signal x2
Fig. 14. Fourier spectra of IMF1 and IMF2 and exact signals.
0 100 200 300 400 500 600 700 800 900 1000
-2
-1
0
1
2
Time (sec)
0 100 200 300 400 500 600 700 800 900 1000
-2
-1
0
1
2
Time (sec)
used. Figs. 21 and 22 present the two decomposed signals and their Fourier spectra by introducing a temporary signal of
x
temp
t 20cos8pt [16]. Similarly, Figs. 23 and 24 show the decomposed signals and their Fourier spectra based on the
proposed AMD method. Figs. 21 and 22 indicate that the existing method [16] cannot completely separate the two
frequency components. Figs. 23 and 24 clearly demonstrate that the two close frequency components have been separated
successfully by the AMD theorem.
5.4. Signal decomposition with small intermittent uctuations around a large wave
The two frequency component signal [16] was considered: x(t)=x
1
(t)+x
2
(t), x
1
(t)=cos(2pt/640) (0oto3200 s), and,
x
2
(t)=0.02 sin(2pt/32) (1264rt r1296 s). The time step in this case was taken to be 0.1 s. Figs. 25 and 26 present the
0 0.01 0.02 0.03 0.04 0.05 0.06
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.01 0.02 0.03 0.04 0.05 0.06
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 16. Fourier spectra of the decomposed signals by AMD and exact signals.
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (Sec)
I
M
F
2

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
1
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (Sec)
I
M
F
1

a
n
d

o
r
i
g
i
n
a
l

s
i
g
n
a
l

x
2
Fig. 17. IMF1 and IMF2 by wave group method with EMD versus exact signals.
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
IMF2
Original signal x1
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
IMF1
Original signal x2
Fig. 18. Fourier spectra of IMF1 and IMF2 by wave group method with EMD and exact signals.
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (sec)
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 20. Fourier spectra of the decomposed signals by AMD and exact signals.
0 2 4 6 8 10 12 14 16 18 20
-10
-5
0
5
10
Time (Sec)
D
e
c
o
m
p
o
s
e
d

a
n
d
o
r
i
g
i
n
a
l

s
i
g
n
a
l
s
0 2 4 6 8 10 12 14 16 18 20
-4
-2
0
2
4
6
Time (Sec)
D
e
c
o
m
p
o
s
e
d

a
n
d
o
r
i
g
i
n
a
l

s
i
g
n
a
l
sDecomposed signal Original signal x2
Fig. 21. Decomposed signals by adding a temp-signal versus exact signals.
decomposed signals and their Fourier transforms that were determined by EMD with an added temporary signal of
x
temp
(t)=0.2 cos(2pt/60). Figs. 27 and 28 show the decomposed signals and Fourier spectra by the AMD theorem with three
bisecting frequencies of f
b
=o
b
/2p=0.003 Hz, 0.0075 Hz and 0.0164 Hz, respectively. The last bisecting frequency
represents the average of the frequencies of signals x
1
(t) and x
2
(t). For this example, both methods can decompose the
small intermittent uctuation from the large long period wave. However, the decomposed signal x
2
(t) by the EMD method
is greatly affected by the signal x
1
(t) as shown in Fig. 26. The proposed AMD theorem gives clean, well separate, and
signicantly more accurate results as illustrated in Fig. 28. Figs. 27 and 28 also illustrate that in this case, the optimum
bisecting frequency in the AMD should weigh more of the large wave frequency due mainly to the distributive nature of
the Fourier spectrum of a short time intermittent signal. As indicated in Fig. 28, more accurate results can be obtained if
the bisecting frequency is reduced to slightly above the frequency of x
1
(t), such as 0.003 Hz.
0 1 2 3 4 5
0
0.5
1
1.5
2
2.5
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 2 4 6 8 10
0
0.1
0.2
0.3
0.4
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 22. Fourier spectra of the decomposed signals by Wangs method and exact signals.
0 2 4 6 8 10 12 14 16 18 20
-10
-5
0
5
10
Time (Sec)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
2
Time (Sec)
0 1 2 3 4 5
0
0.5
1
1.5
2
2.5
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decompose signal
Original signal x1
0 2 4 6 8 10
0
0.05
0.1
0.15
0.2
0.25
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 24. Fourier spectra of the decomposed signals and exact signals.
0 500 1000 1500 2000 2500 3000
-2
-1
0
1
2
Time (Sec)
D
e
c
o
m
p
o
s
e
d

a
n
d
o
r
i
g
i
n
a
l

s
i
g
n
a
l
s
1200 1220 1240 1260 1280 1300 1320 1340 1360 1380 1400
-0.02
-0.01
0
0.01
0.02
Time (Sec)
D
e
c
o
m
p
o
s
e
d

a
n
d
o
r
i
g
i
n
a
l

s
i
g
n
a
l
s
Fig. 25. Decomposed signals by adding a temp-signal versus exact signals.
0 0.01 0.02 0.03 0.04
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x1
0 0.01 0.02 0.03 0.04 0.05 0.06
0
1
2
3
4
x 10
-4
Frequency (Hz)
A
m
p
l
i
t
u
d
e
Decomposed signal
Original signal x2
Fig. 26. Fourier spectra of the decomposed signals by Wangs method and original signals.
0 500 1000 1500 2000 2500 3000
-2
-1
0
1
2
Time (Sec)
1200 1220 1240 1260 1280 1300 1320 1340 1360 1380 1400
-0.02
-0.01
0
0.01
0.02
Time (Sec)
Fig. 27. Decomposed signals by AMD and exact signals.
5.5. System identication with closely spaced modes
Consider the 3-DOF massspringdashpot representation of a classically damped mechanical system with closely
spaced modes. As shown in Fig. 29, the system has three masses of m
1
=m
2
=m
3
=1000 kg, four springs with stiffness of
k
1
=k
4
=40 kN/m and k
2
=k
3
=15 kN/m, and four dashpots with damping coefcients of c
1
=c
4
=120 N s/m and c
2
=c
3
=45 N s/m.
The system has three natural frequencies of f
1
=0.673 Hz, f
2
=1.180 Hz, and f
3
=1.304 Hz. Three damping ratios of the
classically damped system are B
1
=0.63%, B
2
=1.11%, and B
3
=1.23%. The 2nd and 3rd natural frequencies are closely spaced. An
impulsive force of 10 kN over 0.02 s was applied on the mass m
3
at t=0 s. The displacement time histories at three masses
were analytically evaluated and used for system identication. A time step of 0.02 s was used in numerical integrations.
The EMD of the displacement at mass m
3
separated the 1st mode from the 2nd and 3rd, but failed to further separate
the 2nd and 3rd modes. With the proposed AMD method, all three modes can be separated as indicated in Figs. 30 and 31
in time and frequency domains, respectively. Fig. 32 presents the amplitude and phase angle of each mode from which the
natural frequencies and damping ratios are identied as given in Table 1. It is seen from Table 1 that the modal parameters
0.01
0
0
0.2
0.4
0.6
0.8
1
A
m
p
l
i
t
u
d
e
0.02
Frequency (Hz)
0.03 0.04 0
0
1
2
3
4
x10
-4

A
m
p
l
i
t
u
d
e
0.01 0.02 0.03 0.04 0.05 0.06
Frequency (Hz)
Fig. 28. Fourier spectra of the decomposed signals by AMD and original signals.
k
1
c
1
k
2
c
2
m
1
k
3
c
3
m
2
k
4
c
4
m
3
Fig. 29. 3-DOF representation of a mechanical system.
0 10 20 30 40 50 60 70 80 90 100
-5
0
5
x 10
-3
Time (sec)
F
i
r
s
t

m
o
d
e

(
m
)
0 10 20 30 40 50 60 70 80 90 100
-0.01
0
0.01
Time (sec)
S
e
c
o
n
d

m
o
d
e

(
m
)
0 10 20 30 40 50 60 70 80 90 100
-5
0
5
x 10
-3
Time (sec)
T
h
i
r
d

m
o
d
e

(
m
)
Mode 2
Mode 1
Mode 3
Fig. 30. Three modes of vibration decomposed by AMD.
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
1.2
x 10
Frequency (Hz)
A
m
p
l
i
t
u
d
e

(
m
-
S
e
c
)
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
x 10
Frequency (Hz)
A
m
p
l
i
t
u
d
e

(
m
-
S
e
c
)
0 0.5 1 1.5 2
0
1
2
3
4
5
6
x 10
Frequency (Hz)
A
m
p
l
i
t
u
d
e

(
m
-
S
e
c
)
Mode 3 Mode 2
Mode 1
Fig. 31. Fourier spectra of the decomposed modes of vibration by AMD.
0 10 20 30 40 50 60 70 80 90
-9
-8
-7
-6
-5
Time (Sec)
l
n
(
A
m
p
l
i
t
u
d
e
)

(
m
)
0 10 20 30 40 50 60 70 80 90
-200
0
200
400
Time (Sec)
P
h
a
s
e

a
n
g
l
e

(
r
a
d
)
Amplitude
Linear least-square fit
Phase angle
0 10 20 30 40 50
-10
-8
-6
-4
Time (Sec)
l
n
(
A
m
p
l
i
t
u
d
e
)

(
m
)
0 10 20 30 40 50
-200
0
200
400
Time (Sec)
P
h
a
s
e

a
n
g
l
e

(
r
a
d
)
Amplitude
Phase angle
0 10 20 30 40
-10
-8
-6
-4
Time (Sec)
l
n
(
A
m
p
l
i
t
u
d
e
)

(
m
)
0 10 20 30 40
-200
0
200
400
Time (Sec)
P
h
a
s
e

a
n
g
l
e

(
r
a
d
)
Amplitude
Phase angle
Fig. 32. Amplitude and phase angle of Hilbert transforms. (a) Mode 1, (b) Mode 2 and (c) Mode 3.
Table 1
Identied natural frequencies and damping ratios by AMD.
Mode Displacement on m
1
Displacement on m
2
Displacement on m
3
Exact
Damping ratio
(%)
Frequency
(Hz)
Damping ratio
(%)
Frequency
(Hz)
Damping ratio
(%)
Frequency
(Hz)
Damping ratio
(%)
Frequency
(Hz)
1 0.63 0.673 0.64 0.673 0.64 0.673 0.63 0.673
2 1.11 1.180 1.11 1.181 1.11 1.180
3 1.24 1.304 1.24 1.304 1.23 1.304 1.23 1.304
of the 2nd mode cannot be identied from the displacement at mass m
2
since m
2
happens to be the node of the 2nd mode.
In comparison with the exact natural frequencies and damping ratios given in Table 1, the identied results are accurate to
1.6% for frequency and less than 1% for damping ratio.
The instantaneous mode shape identied with Eq. (24) is presented in Fig. 33 over different periods of time. This can
be explained by the modal responses of vibration as shown in Fig. 30. The vibration of the 1st mode lasts signicantly
longer than that of the 2nd mode, which is in turn longer than that of the 3rd mode. The 1st mode shape is more stable
after the 2nd and 3rd modes decay for 30 s. This indicates that, though not visible in Figs. 30 and 31, the 1st mode of
vibration decomposed from the system responses was slightly disturbed by the presence of the 2nd and 3rd modes. The
accuracy and sign of the identied modal elements were veried by the difference in phase angles between any two
modes. The results presented in Fig. 33 indicated that the phase difference between any two modes is approximately
either 01 or 1801.
For the time-invariant system as shown in Fig. 29, the modal elements can be directly determined with Eq. (25). The
real and imaginary parts of the Fourier transforms of all DOF displacements are presented in Fig. 34 for each decomposed
mode of vibration. Based on the ratio of the peak value of the imaginary part, the mode matrix [F] of the 3-DOF system
0 10 20 30 40 50 60 70 80 90
0
1
2
3
4
Time (Sec)
1st DOF
2nd DOF
3rd DOF
0 10 20 30 40 50 60 70 80 90
-1
-0.5
0
0.5
1
Time (Sec)
1st DOF
2nd DOF
3rd DOF
0 10 20 30 40 50
0
1
2
3
Time (Sec)
1st DOF
3rd DOF
0 10 20 30 40 50
-1
0
1
2
Time (Sec)
1st DOF
3rd DOF
0 5 10 15 20 25 30
0
1
2
3
Time (Sec)
1st DOF
2nd DOF
3rd DOF
0 5 10 15 20 25 30
-1
0
1
2
Time (Sec)
1st DOF
2nd DOF
3rd DOF
M
o
d
a
l

e
l
e
m
e
n
t
P
h
a
s
e

a
n
g
l
e

d
i
f
f
e
r
e
n
c
e

(
)
M
o
d
a
l

e
l
e
m
e
n
t
P
h
a
s
e

a
n
g
l
e

d
i
f
f
e
r
e
n
c
e

(
)
M
o
d
a
l

e
l
e
m
e
n
t
P
h
a
s
e

a
n
g
l
e

d
i
f
f
e
r
e
n
c
e

(
)
Fig. 33. Instantaneous mode shapes identied from the amplitude of Hilbert transforms. (a) Mode 1, (b) Mode 2 and (c) Mode 3.
was identied and compared with its theoretical values below:
F
identified

1:0 1:0 1:0
2:48 0 0:80
1:0 1:0 1:0
2
6
4
3
7
5 and F
theoretical

1:0 1:0 1:0
2:47 0 0:81
1:0 1:0 1:0
2
6
4
3
7
5:
It is clearly seen that the identied mode matrix is accurate with a maximum error of 1.25%. The 2nd mode is
symmetric about mass m
2
due to the symmetry of the 3-DOF system. The 1st and 3rd modes have masses m
1
and m
3
moving in the same amount but mass m
2
in the same or opposite direction as m
1
and m
3
.
0.5 0.6 0.7 0.8 0.9 1
-15
-10
-5
0
5
x 10
Frequency (Hz)
I
m
a
g
i
n
a
r
y

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
0.5 0.6 0.7 0.8 0.9 1
-1
-0.5
0
0.5
1
x 10
Frequency (Hz)
R
e
a
l

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
1 1.1 1.2 1.3 1.4 1.5
-6
-4
-2
0
2
4
6
x 10
Frequency (Hz)
I
m
a
g
i
n
a
r
y

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
1 1.1 1.2 1.3 1.4 1.5
-3
-2
-1
0
1
2
3
4
x 10
Frequency (Hz)
R
e
a
l

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
1 1.1 1.2 1.3 1.4 1.5
-3
-2
-1
0
1
2
3
x 10
Frequency (Hz)
I
m
a
g
i
n
a
r
y

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
1 1.1 1.2 1.3 1.4 1.5
-2
-1
0
1
2
x 10
Frequency(Hz)
R
e
a
l

p
a
r
t

o
f

F
o
u
r
i
e
r

t
r
a
n
s
f
o
r
m

(
m
-
s
e
c
)
displacement at m1
displacement at m2
displacement at m3
Fig. 34. Fourier transform of the three modes of vibration decomposed by AMD. (a) Mode 1, (b) Mode 2 and (c) Mode 3.
6. Conclusions
This paper presents a new analytical mode decomposition or AMD theorem for signal processing of a time series. This
mathematically sound theorem can address a number of issues associated with the empirical mode decomposition or EMD
method that has recently received increasing applications in civil and mechanical engineering. It potentially provides a
unied signal processing technique for many engineering applications. Based on the mathematical derivation and
numerical analysis, the following conclusions can be drawn:
1. A general time series can be decomposed into two signals that do not vanish in frequency domain over two mutually
exclusive frequency ranges about a bisecting frequency based on two Hilbert transforms of the harmonics multi-
plicative time series. The Hilbert transforms of the decomposed signals can be directly evaluated from those of the
harmonics multiplicative time series.
2. The proposed AMD theorem can directly separate closely spaced frequency components of both stationary and non-
stationary signals indicated by the steady-state and transient responses in mechanical systems and engineering
structures. It is insensitive to noise up to 10% of the signal strength.
3. The proposed AMD theorem can separate a small intermittent uctuation from a large stationary time history,
indicating a potential advantage in processing nonlinear system responses in civil and mechanical engineering.
4. The proposed AMD theorem is superior to the ltering technique, wave-group method, and tempered signal approach
that are currently available in the literature to deal with the decomposition of closely spaced modes, beating effects,
and small intermittent uctuations. It is simple in concept, efcient in computation, consistent in performance, and
reliable in signal processing.
5. The bisecting frequency in the AMD theorem can theoretically be selected as any value between the two frequencies of
interest but practically be recommended to be the average of the two frequencies due to the nite length of signals in
applications. It is insensitive to other choices in 80120% of the average value so long as the frequency resolution in the
Fast Fourier Transform does not exceed 50% of the spacing of the two frequencies. This translates into a discernable
frequency spacing of twice the frequency resolution of a time series or the required duration of the time series that is
equal to two divided by the minimum frequency spacing.
Acknowledgements
Financial support to complete this study was provided in part by Missouri Department of Transportation and the Center
for Transportation Infrastructure and Safety through a collaborative structure research initiative, and by the National
Science Foundation under Award No. CMMI0409420. The second author wishes to thank China Scholarship Council for
providing additional nancial support under Award No. NSCIS-[2007]-3020. The results and opinions expressed in this
paper are those of the authors only and they do not necessarily represent those of the sponsors.
Appendix A
The Fourier transforms of harmonic functions s
c
(t)=cos(o
b
t) and s
s
(t)=sin(o
b
t) can be expressed into:
^
s
c
o doo
b
doo
b

=2 A:1
^ s
s
o doo
b
doo
b

=2j A:2
where d[.] is a Dirac Delta function of the argument in the square bracket. The Fourier transforms of y
c
(t)=x(t)s
c
(t) and
y
s
(t)=x(t)s
s
(t) can thus be derived by the following convolution integrations:
^ y
c
o
Z
1
1
^
Xou^ s
c
udu
^
Xoo
b
^
Xoo
b
=2 A:3
^ y
s
o
Z
1
1
^
Xou^ s
s
udu
^
X oo
b

^
Xoo
b
h i
=2j A:4
Based on the relation between the Hilbert transform and Fourier transform, the Fourier transforms of the Hilbert
transforms of y
c
(t) andy
s
(t) are equal to
^
H
yc t
o j sgno
^
y
c
o and
^
H
ys t
o j sgno
^
y
s
o A:5
in which
^
H
:
o denotes the Fourier transform of the Hilbert transform of the function in the square bracket and sgn(o) is a
signum function. The Fourier transform of s
1
(t) in Eq. (13) can then be determined by
^
s
1
o
Z
1
1
^
s
s
ouj sgnu
^
y
c
u du
Z
1
1
^
s
c
ouj sgnu
^
y
s
u du
^
Xo sgn oo
b
sgnoo
b
=2 A:6
As a result,
^ s
1
o
0,9o94o
b
^
Xo,9o9oo
b
^
Xo=2,9o9 o
b
8
>
>
<
>
>
:
A:7
The Fourier transform of s
1
t can be evaluated from Eq. (15):
^
s
1
o
^
Xo
^
s
1
o
^
Xo,9o94o
b
0,9o9oo
b
^
Xo=2,9o9 o
b
8
>
>
<
>
>
:
A:8
As indicated by Eqs. (A.7) and (A.8), ^ s
1
o and
^
s
1
o vanishes for 9o94o
b
and 9o9oo
b
, respectively. Both of them take
^
Xo=2 at the boundary 9o9=o
b
of two mutually exclusive frequency ranges.
References
[1] B. Liu, S. Riemenschneider, Y. Xu, Gearbox fault diagnosis using empirical mode decomposition and Hilbert spectrum, Mech. Sys. Signal Process. 20
(3) (2006) 718734.
[2] H.G. Chen, Y.J. Yan, J.S. Jiang, Vibration-based damage detection in composite wingbox structures by HHT, Mech. Sys. Signal Process. 21 (1) (2007)
307321.
[3] M. Feldman, Investigation of the natural vibrations of machine elements using the Hilbert transform, Soviet Mach. Sci. 2 (1985) 4447.
[4] M. Feldman, Non-linear system vibration analysis using Hilbert transform-I: free vibration analysis method FREEVIB, Mech. Sys. Signal Process. 8 (2)
(1994) 119127.
[5] M. Feldman, Non-linear free-vibration identication via the Hilbert transform, J. Sound Vib. 208 (3) (1997) 475489.
[6] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.C. Yen, C.C. Tung, H.H. Liu, The empirical mode decomposition and Hilbert spectrum
for nonlinear and nonstationary time series analysis, Proc. Royal Soc. London-Ser. A 454 (1998) 903995.
[7] N.E. Huang, Z. Shen, S.R. Long, A new view of nonlinear water waves: the Hilbert spectrum, Ann. Rev. Fluid Mech. 31 (1999) 417457.
[8] N.E. Huang, M.C. Wu, S.R. Long, S.S.P. Shen, W. Qu, P. Gloersen, K.L. Fan, A condence limit for the empirical mode decomposition and Hilbert spectral
analysis, Proc. Royal Soc. London-Ser. A 459 (2003) 23172345.
[9] J.N. Yang, Y. Lei, S. Pan, N. Huang, System identication of linear structures based on HilbertHuang spectral analysis, Part I: normal modes,
Earthquake Eng. Struct. Dyn. 32 (9) (2003) 14431467.
[10] J.N. Yang, Y. Lei, S. Lin, N. Huang, Identication of natural frequencies and damping of in situ tall buildings using ambient wind vibration data, J. Eng.
Mech. 130 (5) (2004) 570577.
[11] J.N. Yang, Y. Lei, S. Pan, N. Huang, System identication of linear structures based on HilbertHuang spectral analysis, Part II: complex modes,
Earthquake Eng. Struct. Dyn. 32 (10) (2003) 14431467.
[12] J.N. Yang, Y. Lei, S. Lin, N. Huang, HilbertHuang based approach for structural damage detection, J. Eng. Mech. 130 (1) (2004) 8595.
[13] J. Chen, Y.L. Xu, Identication of modal damping ratios of structures with closely spaced modal frequencies, Struct. Eng. Mech. 14 (4) (2002)
417434.
[14] Z.K. Peng, P.W. Tse, F.L. Chu, An improved HilbertHuang transform and its application in vibration signal analysis, J. Sound Vib. 286 (1-2) (2005)
187205.
[15] Y. Chen, M.Q. Feng, A technique to improve the empirical mode decomposition in the HilbertHuang transform, Earthquake Eng. Eng. Vib. 2 (1)
(2003) 7585.
[16] W. Wang, Decomposition of wave groups with EMD method, in: N. Huang, N.O. Attoh-Okine (Eds.), The Hilbert Transform in Engineering, Taylor &
Francis Group, LLC, FL2005, pp. 267280.
[17] M. Zheng, F. Shen, Y. Dou, X. Yan, Modal identication based on Hilbert-Huang transform of structural response with SVD preprocessing, Acta Mech.
Sin. 25 (6) (2009) 883888.
[18] E. Bedrosian, A product theorem for Hilbert transforms, Memorandum RM-3439-PR, U.S. Air Force Project RAND 1962.
[19] D. Huang, J. Zhao, J. Sun, Practical implementation of the Hilbert transform algorithm, Acta Oceanol. Sin. 25 (1) (2003) 111.

2

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2

Uploaded by

Copyright:

Available Formats

A signal decomposition theorem with Hilbert transform and its

application to narrowband time series with closely spaced

You might also like