Sunteți pe pagina 1din 13

Frequency slice algorithm for modal signal separation and damping identication

Zhonghong Yan
a,b,,1
, Ayaho Miyamoto
b,2
, Zhongwei Jiang
b,3
a
Biomedical Department, Chongqing University of Technology, Chongqing 400050, China
b
Yamaguchi University, 2-16-1 Tokiwadai, Ube, Yamaguchi 755-8611, Japan
a r t i c l e i n f o
Article history:
Received 13 March 2009
Accepted 30 July 2010
Keywords:
Timefrequency analysis
Random decrement technique
Signal process
Vibration signal
Modal parameter identication
a b s t r a c t
This paper focuses on modal signal separation and damping parameter identication by a new time
frequency analysis method. With the aid of the random decrement technique (RDT), an accurate
estimation method is rstly introduced both in time and frequency domains for single modal damping
identication. Next, the background of a new concept of frequency slice wavelet transform (FSWT) is
revealed clearly. Then, some new properties of the FSWT are briey discussed in contrast with the wave-
let transform (WT). Based on the analysis of RDT and FSWT, a frequency slice algorithm (FSA) is designed
for modal separation and parameter identication. The merits of FSWT and FSA with numerical simula-
tions and experiments are demonstrated in this paper. We nally apply the proposed methods to analyze
the free-decay responses (FDR) collected from a small laboratory bridge monitoring system (LBMS). Some
conclusions are drawn that the RDT being used directly in FSWT domain can bring a good damping esti-
mator. The FSA is not limited to FDR, and also can be used to random impacting response directly. FSWT
itself is a new kind of good lter, and has high performance against noise. It is signicant to get damping
parameter with higher accuracy through modal separation by FSWT, and FSWT can be controlled adap-
tively in modal separation by dynamic scale method.
Crown Copyright 2010 Published by Elsevier Ltd. All rights reserved.
1. Introduction
1.1. Problems and methods
In a vibration system, response frequency, mode shape and
damping, which are always relative with the system structure,
are often used in damage detection in structural health monitoring
system (SHMS) [1,2]. Meanwhile, modal analysis and parameter
identication are common-used methods for extracting these dy-
namic characteristics in SHMS. Quite a few researches [3,4] noted
that the identied natural frequencies and mode shapes by using
various system identication methods and different test data are
in excellent agreement, but the estimation uncertainty of damping
ratios is inherently larger than that of natural frequencies. Experi-
ments show that the damping features of a system, especially for
the high frequency components, are very signicant marks to dam-
age detection. However, in general, it is not easy to extract the
damping (ratio) exactly because the vibration signals always in-
clude many frequency components and noise. Therefore, in this
paper, we would like to pay more attention to modal signal sepa-
ration and damping identication, especially for the modal signals
with high damping and close frequency modes. Let us [5] still con-
sider a linear damped multi-degree-of-freedom (MDOF) system
with n real modes for system modal parameter identication,
and its free-decay response (FDR) is given as
xt

n
i1
A
i
e
2pf
i
f
i
t
cos2pf
di
t h
i
1
where A
i
is the amplitude, f
i
is the undamped natural frequency,
f
di
f
i

1 f
2
i
_
is the damped natural frequency, and f
i
is the damp-
ing ratio. Here a single modal signal is described as an exponentially
decayed sinusoid signal function: s(t) = e
at
cos(bt + h).
Generally, based on frequency domain decomposition (FDD),
modal identication techniques for vibration responses (for exam-
ple, the FDR in Eq. (1)) are widely recognized [6] as being simple.
However, it often leads to the loss of accuracy of the identied
modal parameters due to the spectrum of measured responses,
which cannot be estimated exactly, especially for high-damped
systems and systems with severe modal interference. Conversely,
many methods based on the time domain analysis have been
developed [79]. These approaches frequently provide accurate re-
sults if the measured responses are not severely contaminated by
noise. Nevertheless, the de-noising in time domain is not as conve-
nient as in frequency domain.
0045-7949/$ - see front matter Crown Copyright 2010 Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2010.07.011

Corresponding author at: Biomedical Department, Chongqing University of


Technology, Chongqing 400050, China. Tel./fax: +86 023 68660070.
E-mail addresses: yzh816@msn.com (Z. Yan), miya818@yamaguchi-u.ac.jp (A.
Miyamoto), jiang@yamaguchi-u.ac.jp (Z. Jiang).
1
Tel./fax: +81 836 85 9530.
2
Tel./fax: +81 836 85 9530.
3
Tel./fax: +81 836 85 91370.
Computers and Structures 89 (2011) 1426
Contents lists available at ScienceDirect
Computers and Structures
j our nal homepage: www. el sevi er . com/ l ocat e/ compst r uc
Combining the representation of the measured responses in
time and frequency domains simultaneously, some new tools such
as, WignerVille distribution (WVD), wavelet transform (WT) and
HilbertHuang transform (HHT), etc., are developed to construct
new frameworks for system identication and damage detection
(Refs. [1015]). Prominently, as one of important timefrequency
analysis tools, the continuous wavelet transform (CWT) and dis-
crete wavelet transform (DWT) have been fully developed in the
theoretical aspect over past two decades.
Yan et al. [13], noted that WT as well as the traditional schemes
can only supply a single estimation of the modal parameter, and its
accuracy depends on modal separation, end-effect, and the para-
metric selection of wavelet function, etc. In fact, even if for sin-
gle-degree-of-freedom (SDOF) response, by using the WT method
based on [10], most of estimations are only approximate. There-
fore, Tan et al. [14] discard the fussy selections of wavelet scale
and centre frequency in Yans method [13] that is based on mini-
mum Shannon entropy search to separate the close modals, and
they proposed a relatively simpler pattern search method to im-
prove the estimated results of [13].
Using a different idea in modal separation, Huang and Su [15]
have presented an eigenvalue model in wavelet domain based on
CWT method. Although the new method of Huang and Su could
overcome the problem of close modes decomposition by perform-
ing system identication in wavelet domain, their algorithm still
involves the expensive computation to rst solve a large-scale
overdeterminate system of linear algebraic equations, and then
nd the solutions of eigenvalue problem of a big size square ma-
trix. At the same time, the determination of both wavelet scale
and centre frequency still affects the accuracy of identication
problem. In fact, pursuing high timefrequency localization of
modal signal is always important to wavelet methods. Fortu-
nately, in this paper, we can easily overcome the difculty with
a new timefrequency transform method. We still base on the ba-
sic idea of modal separation by wavelet, the new wavelet trans-
form is to analyze their introduced questions. Notably, we can
simplify a lot of computation for modal separation and damping
identication.
In wavelet transforms, there is a common problem: how is it
possible to decide the center frequency and the time supporting
width of a mother wavelet? As we know, the characteristics of a
mother wavelet function always affect the performance of time
frequency analysis. For example, although the Gabor wavelet,
which is one of the most widely used analytic wavelets, has the
best timefrequency resolution, i.e. the smallest Heisenberg box,
the center frequency and the time supporting width of the mother
Gabor wavelet affect its timefrequency decomposition character-
istics. This means that, depending on the signals to be analyzed,
different Gabor wavelet shapes must be used. Since the character-
istics of signals are unknown in general, the determination of opti-
mal shape is usually difcult [16]. Based on the motivations, we
have developed a new timefrequency transform with better
properties than WT to improve the situations in application
[5,17]. The new transform is called the frequency slice wavelet
transform (FSWT). Actually, this paper will reveal the proposed
background of FSWT clearly. Frequency slicing processing is an
important idea in modal analysis and parameters identication
in this paper.
Many existing methods for modal identication are based on
FDR signals. Unfortunately, we usually cannot easily get the FDR
signals from a big system. The well-known random decrement
technique (RDT) (e.g. [18]) is usually used in computing random
decrement signatures from ambient random vibration data. How-
ever, this paper will study a new usage of RDT idea, and its trigger-
ing concept is employed to extract accurately the modal damping
in timefrequency domain for SDOF response. The method can also
be used to estimate the modal damping directly for MDOF signals
with an acceptable accuracy.
1.2. Main ideas
The main aim of this paper is to realize modal signals separation
and damping identication in Eq. (1) with a novel method. Firstly,
for SDOF signal, combined with logarithmic decrement method
(LDM) and RDT idea, a high accurate damping estimation in fre-
quency domain is introduced. Secondly, based on the estimation
method, FSWT as a new signal analysis tool is thus introduced in
this study, and the background of the FSWT analysis is rst re-
vealed clearly.
Since FSWT itself is a new kind of good lter [5,17], this paper
does not need any lter even if the obtained signal includes high
noise. We only focus on the application of FSWT in this paper.
Implementing the RDT directly in FSWT timefrequency domain
is the rst important application of FSWT in this paper, and modal
separation is another application of FSWT. FSWT has many better
properties [17] than WT. Such as, the center of FSWT timefre-
quency window is the observing center, and this property makes
it possible to construct an adaptive algorithm of modal parameter
identication in timefrequency domain. Dynamic scale of FSWT
as a new skill will be proved to be a powerful tool in modal
separation.
As a new result, a general estimator of modal damping is to ex-
press in FSWT domain. FSWT can very clearly represent the damp-
ing characteristics of multi-modal signal simultaneously in time
and frequency domain, a frequency slice algorithm (FSA) for modal
separation and parameter identication is therefore designed to
analyze the free-decay responses (FDR). Meanwhile, we introduce
timefrequency projective method of FSWT for modal separation.
A real FDR signal collected from a small laboratory bridge monitor-
ing system (LBMS) will be investigated. The merits of FSWT and
FSA with numerical simulations and experiments are demon-
strated in this paper. Some conclusions are drawn that the RDT
being used directly in FSWT domain can bring a good damping
estimator. FSA is not limited to FDR, and also can be used to ran-
dom impacting response directly. FSWT itself is an effectual lter
to noise. It is signicant to get damping parameter with higher
accuracy through modal separation by FSWT, and FSWT can be
controlled adaptively in modal separation by dynamic scale
method.
1.3. Notation
R denotes the set of real numbers. L
2
(R) denotes the vectors
space of measurable, square-integrable one-dimensional functions
f (x).
Fourier transform (FT) for function f (x) e L
2
(R).
Fff g :
^
f x
_
1
1
f se
ixs
ds 2
Fourier inverse transform:
F
1
f
^
f g : f t
1
2p
_
1
1
^
f xe
ixt
dx 3
The signal energy is recorded as:
kf k
2
2

_
1
1
jf tj
2
dt 4
||||
2
also denotes the classical norm in the space of square-integra-
ble functions.
We dene the following timefrequency localization features of
limited energy signals, which include wavelet functions and STFT
window functions etc.
Z. Yan et al. / Computers and Structures 89 (2011) 1426 15
The duration Dt
f
and bandwidth Dx
f
are dened as
Dt
f

1
kf k
2

_
1
1
t t
f

2
jf tj
2
dt

; Dx
f

1
kf k
2

_
1
1
xx
f

2
j
^
f xj
2
dx

; 5
where t
f
and x
f
are the centers of f(t) and
^
f x, respectively,
t
f

1
kf k
2
2
_
1
1
tjf tj
2
dt; x
f

1
kf k
2
2
_
1
1
xj
^
f xj
2
dx 6
2. Frequency slice expression of modal damping
2.1. RDT idea
Here we briey describe the RDT idea, and more details can be
found in [18]. Under a randomly exciting force, a structure has ini-
tial displacement a at time t, and we record the response as x(t) = a.
The most important idea of RDT is that the moving average func-
tion of response x(t) on a level crossing trigger condition is intro-
duced to get the free-decay response from random loads. The
following is a simple form, where X = {x(t
i
) = a} is called trigger
condition, which can be changed according to the needs, such as
X fa xt
i
bg,
xs
1
N

N
i1
xt
i
s X j ; 7
where N is the total number of triggering points.
The simplicity of the estimation process is obvious, since only
the detection of triggering points and averaging of the correspond-
ing time segments are performed. As the number of average in-
creases, the random part due to the random loads will be
eventually averaged out and be negligible. Furthermore, the sign
of the initial velocity is expected to vary randomly with time, so
the resulting initial velocity will be zero. Therefore, the free-decay
response from the initial displacement a will remain.
Considered a SDOF system response as the rst case, a new
usage of RDT concepts to extract the modal damping directly will
be introduced in the following section.
2.2. RDT expression of modal damping
2.2.1. Time domain expression
Let s(t) = Ae
at
cos(bt + h) be FDR signal in a simplest SDOF sys-
tem. If bT = kp, here k is a positive integer, then logarithmic decre-
ment method (LDM) is given as
a
1
T
lnjstj lnjst Tj : 8a
Unfortunately, the LDM in Eq. (8a) is always sensitive to noise
or when |s(t)| % 0. However, we can use the above RDT idea to se-
lect a suitable trigger condition to avoid the singularity of function
ln|s(t)| or increase the ability against random noise. For example,
the following is a discrete form under a triggering condition
a
1
NT

N
i1
lnjst
i
j

N
i1
ln st
i
T j j
_ _

f0 < a < st
i
j j < bg
8b
Nevertheless, Eq. (8b) still has not good capability to anti-noise
because when the signal is contaminated by noise, we can not be
sure |s(t
i
+ T)| 0 even if |s(t
i
)| satisfy the trigger condition
0 < a<|s(t
i
)|<b. Therefore, we attempt to translate Eq. (8b) into the
frequency domain by the following methods.
2.2.2. Frequency domain expression
The following T is called the width of the observing time win-
dow, t observing time, and x observing frequency.
Suppose that ^sx is Fourier transform of function s(t) e
L
2
(1, 1), and a window function in frequency domain such as
^ px
1e
ix
ix
, then we dene
Ft; x; T e
ixt
1
T
_
tT
t
sxe
ixx
dx; 9
where T > 0. Through using the Fourier transform, it is not difcult
to prove that F(t, x, T) can be further rewritten as
Ft; x; T
1
2p
_
1
1
^su^ p

u xTe
iut
du; 10
Where the star means the conjugate of the function.
Specially, if s(t) = Ae
at
cos(bt + h), then Eq. (10) becomes
1 Ft; x; T
1
2T
Ae
at
e
ihbt
1 e
aixbT

a x bi
_

e
ihbt
1 e
aixbT

a x bi
_
11
if bs = kp and s > 0, where k is a positive integer, then
2 a
1
s
lnjFt; x; Tj lnjFt s; x; Tj 12a
Eq. (11) is easy to verify directly. Because under the condition
bs = kp, Eq. (12a) is similar to Eq. (8a), the detailed proof is omitted
with the exception of a few comments.
Let us analyze the performance against noise in Eq. (12a). Note
that in a damping system, the damping ratio f ( 1, and a and b are
dened as
a 2pf f; b 2pf

1 f
2
_
; 13
thus, a ( b. According to Eq. (11), we know that F(t, x, T) attains its
approximate maximum at x= b. Therefore, for x% b and a ( b, it
is predicable that
jFt; b; Tj ! jFt; x; Tj > 0; andkFt s; xj > 0:
As the result, Eq. (12a) is always correct and more reasonable
than Eq. (2a) in time domain. The RDT formula of Eq. (12a) can
be established similarly as the following. For example, under a trig-
gering condition
a
1
Ns

N
i1
lnjFt
i
; x; Tj

N
i1
ln Ft
i
s; x; T j j
_ _

f0 < a < jFt


i
; x; Tj < bg; 12b
where bs = kp, N is the total number of triggering points.
2.2.3. Numerical demonstration
Notation. Noise level R of a signal s(t) is dened as s := s(t)(1+Rr(t))
called multiple noise (R%); or s := s(t)/max|s|+Rr(t) called addi-
tional noise (+R%); where R > 0 and r(t) is a normally distributed
random variable with zero mean and unit variance.
Suppose a general single modal FDR signal as
st; A; f ; f; h; t
0

Ae
2pf ft
cos2pf
d
t h t ! t
0
0 t < t
0
_
; 14
where A is the amplitude of this mode, f is the undamped natural
frequency, f
d
f

1 f
2
_
is the damped natural frequency, f is the
damping ratio. Record f
s
as the sampling ratio, and T
s
is the sample
time.
16 Z. Yan et al. / Computers and Structures 89 (2011) 1426
Table 1 shows the results obtained by applying Eq. (12b) to sig-
nal s(t, 1, f, f, 0, t
0
), where f = 1 Hz and f = 0.05. A random noise
with different levels is added into the signal in order to test the
effectiveness of RDT formula in Eq. (12b). Meanwhile the inuence
on sampling frequency f
s
and the width of the observing time win-
dow T
s
are also investigated. Note that T
s
is decreased to 5s.
Since Eqs. (10) and (12b) actually imply ltering, therefore, in
the computation of Table 1, we do not use any lter to de-noise.
Notably, by using Eq. (12b), we can give a high accurate estimator
of modal damping even if the obtained signal includes stronger
noise and shorter sample time than ones of Ref. [14]. At the same
time, the inuence of additional noises is larger than that of multi-
ple noises, and the high sampling rate is advantageous in test.
Eq. (12b) is an accurate estimator of damping a for SDOF re-
sponse, but it is always approximate for multi-degree-of-free-
dom (MDOF) responses, especially it is still difcult to
distinguish the closely spaced modes. Consequently, a simple
idea is that, rst, a way should be found to separate these
modes, and then Eq. (12b) is used to complete the damping
computation. Although the search methods in [13,14] can be
used to separate the close modes, they always involve hard opti-
mal computation.
Interestingly, the estimation method of damping ratio directly
carries out a new timefrequency transform. F(t, x, T) denoted in
Eq. (10) is called a frequency slice representation of signal s(t). A
more general transformation is to discuss in the following sections
for modal damping expression and modal separation.
3. Introduction of FSWT Tool
For any function f(t) e L
2
(R), the frequency slice wavelet trans-
form (FSWT) is dened directly in frequency domain as
W
f
t; x; r
1
2p
_
1
1
^
f u^ p

u x
r
e
iut
du 15
where the scale r is a constant or a function of x, t and u, and the
star means the conjugate of a function. Here we call x and t the
observed frequency and time, and u the assessed frequency. ^ px is
also called frequency slice function (FSF). In fact, the FSWT is a mod-
ied version of the traditional wavelet transform in frequency do-
main. By using Parseval equation, if r is not function of the
assessed frequency u, then Eq. (12b) can be translated into its time
domain.
W
f
t; x; r re
ixt
_
1
1
f se
ixs
p

rs tds: 16
Eqs. (15) and (16) can be found in [5]. FSWT fast discrete algorithm
with aid of fast Fourier transform (FFT), more comparisons based on
application with FFT, CWT, DWT, short time Fourier transform
(STFT), and WVD etc., can be found in [5]. An overall theoretical
description of FSWT can be seen in [17]. Therefore, this paper will
only pay more attention to the application of FSWT in modal sepa-
ration and damping identication. We then briey discuss some
new properties of FSWT in contrast with the wavelet transform
(WT).
Firstly, as a new timefrequency transform, FSWT has better
performance than CWT [17]. In [17], we have analyzed that
j^ pxj and |p(t)| are to select as even functions respectively. Both
of functions j^ pxj and |p(t)| have perfect symmetry, and so it is
possible that FSWT has better properties than the traditional WT.
For example, the center of timefrequency window is always the
observing center in contrast to the WT. Therefore, the timefre-
quency window is adaptive to the observing center of the analyzed
signal, and the scale r is a balance factor between the time resolu-
tion and frequency resolution.
Secondly, if one let r
x
j
, i.e. j
x
r
, then Eq. (15) naturally be-
comes into
W
f
t; x; j
1
2p
_
1
1
^
f u^ p

j
u x
x
e
iut
du 17
Note that the parameter j in Eq. (17) is the unique parameter that
should be chosen in application. Consider the bandwidth-to-fre-
quency ratio property of FSF. We dene frequency resolution ratio
of an FSF as
g
p

half width of frequency window
center frequency

rDx
p
x

Dx
p
x=r

Dx
p
j
18
where Dx
p
is computed by Eq. (5). Thus, in FSWT, g
p
may not be
constant.
The frequency resolution ratio g
s
of the measured signal is sim-
ilarly dened as
g
s

Dx
s
x
s
19
From [17], we can assume g
p
= g
s
, and then have a basic choice
about the scale parameter j
Table 1
Identication of modal damping ratio using Eq. (12b) for SDOF response with different sample time and noise levels.
Noise level(R) Modal parameter Sampling parameter Test of Eq. (12b) statistic times = 100
Modal parameter average Variance
f (Hz) f f
s
(Hz) T
s
(s) E(f) E(f) Var(f) Var(f)
25% 1 0.05 20 10 1.000 0.0501 0 1.3e06
1 0.05 100 10 1.000 0.0500 0 3.4e07
1 0.05 20 5 1.000 0.0499 0 9.7e06
1 0.05 100 5 1.000 0.0501 0 2.5e06
50% 1 0.05 20 10 1.000 0.0501 0 7.7e06
1 0.05 100 10 1.000 0.0500 0 1.4e06
1 0.05 20 5 1.000 0.0489 0 4.1e05
1 0.05 100 5 1.000 0.0498 0 7.6e06
+25% 1 0.05 20 10 1.000 0.0489 0 3.0e05
1 0.05 100 10 1.000 0.0496 0 6.9e06
1 0.05 20 5 0.9975 0.0506 9.8e05 6.7e05
1 0.05 100 5 1.000 0.0492 0 1.1e05
+50% 1 0.05 20 10 0.9972 0.0495 2.8e04 1.1e04
1 0.05 100 10 1.000 0.0496 0 3.1e05
1 0.05 20 5 0.9896 0.0492 4.6e04 2.5e04
1 0.05 100 5 0.9987 0.0497 5.1e05 5.1e05
Z. Yan et al. / Computers and Structures 89 (2011) 1426 17
j
Dx
p
g
s
20
As stated above in Eq. (20), FSWT have another important prop-
erty: FSWT can be controlled by the frequency resolution ratio g
s
of
the measured signal.
We can easily design an FSF. The following gives two simple
examples.
FSF1 : ^ px e

1
2
x
2
; pt
e

1
2
t
2
FSF2 : ^ px
1
1 x
2
; pt e
jtj
Note that the centers of timefrequency windows in FSF1 and
FSF2 are always at the origin of the timefrequency plane. Fig. 1
shows the principle of FSWT; and the window functions in FSF1
present both in time and frequency domains. Both functions ^ pu
and p(t) are symmetric around their center points respectively.
Their energy is concentrated at the origin of time and frequency
plane.
Inverse transform [17]: if ^ px satises ^ p0 1, then the origi-
nal signal f(t) can be reconstructed by
f t
1
2p
_
1
1
_
1
1
Ws; x; je
ixts
dsdx 21
As an important result of Eq. (21), the reconstruction proce-
dure is independent from the selected FSF. Reconstruction
independency is an important feature in FSWT, but this charac-
teristic is not allowable in traditional wavelet. Therefore, if the
condition ^ p0 1 remains unchanged in computation of the
FSWT, we can easily realize dynamic scale controlling. For exam-
ple, j can be adaptively controlled by the signal spectrum as the
following
j j
0
^
f u

=
^
f x

; 22
where j
0
> 0 is a constant that satises Eq. (20),
^
f x

represents
the energy of the signal at the observing frequency in Eq. (17) and
^
f u

for the assessing frequency. See more explanations in [17].


In Section 5, we will show that the dynamic scale j is efcient
for modal separation. It is possible that by using dynamic scale
method one does not have to sacrice the time resolution to in-
crease the frequency resolution, conversely, the same reason is also
possible to increase the time resolution [17]. However, this kind of
dynamical characteristic is also not allowable in traditional wave-
let because its reconstructed equation must depend on the selected
wavelet base and its scale.
4. FSWT modal analysis
4.1. FSWT expression of modal damping
In this paper, as the main application of FSWT, the damping
parameter in Eq. (1) can be analyzed by the following theorem.
Theorem 1. Let f t
Ae
at
cosbt h t ! t
0
0 t < t
0
_
; and the FSF p(t)
satises t<0, p(t) = 0.
Moreover, if t ! t
0
and bT = kp, hence
jWt T; x; rj e
aT
jWt; x; rj 23
a
1
T
lnjWt; x; rj lnjWt T; x; rj 24a
a
1
NT

N
i1
lnjWt
i
; x; rj

N
i1
lnjWt
i
T; x; rj
_ _
f0 < a < jWt
i
; x; rj < b; t
i
> t
0
j g 24b
Here, we call that if t < 0, p(t) = 0 p(t) is single side function (SSF) in
time domain.
Remark. It is necessary that p(t) should be an SSF to maintain the
accuracy of Eq. (24a) for SDOF response. However, in traditional
wavelet theory, we cannot suppose that p(t) should be a single side
function. For example, the Gaussian function is not an SSF.
Therefore, many estimators by means of the asymptotic techniques
and Taylors formula based on general wavelet or Morlet wavelet
transform (MWT) (e.g. [10]) are always approximate even for SDOF
signal.
In fact, Eq. (23) reveals that |W(t, x, r)| can be viewed as the
envelope of FDR f(t), latter we will show the FSWT characteristic.
Eq. (24b) then completes the logarithmic decrement method to
get the damping parameter. Usually, Eq. (24b) may be sensitive
to noise or when |W(t, x, r)| = 0. To avoid the singularity of func-
tion ln() or increase the ability against noise, we can also use the
RDT idea in FSWT domain, and then Eq. (24b) can be easily chan-
ged into Eq. (24b) similar with Eq. (12b). Eq. (24b) can be a good
approximate expression of damping for MDOF response due to
the localization of FSWT in timefrequency domain and another
fact that FSWT can be easily controlled by frequency resolution ra-
tio g
s
of the measured signal and dynamic scale method. Finally,
Eq. (24b) is used to compute the damping of random response di-
rectly. Latter the computational result will be shown in an
experiment.
Inverse
Fourier
Transform
F
F
T


S
p
e
c
t
r
u
m

u
u
) (

u f
) (

u
p
A

F
r
e
q
u
e
n
c
y

S
l
i
c
e

R
e
s
p
o
n
s
e

t
Multiple

Move Slice Window


S
l
i
c
e

F
u
n
c
t
i
o
n


) ( u p
u
) (t p
0 0
t
Fig. 1. Schematic diagram of FSWT.
18 Z. Yan et al. / Computers and Structures 89 (2011) 1426
4.2. Local characteristics of frequency slices
In this section, a simulated damping vibration signal is used to
observe the characteristics of frequency slices and compare the de-
tails of the FSWT with the CWT. We can further analyze the FDR
signal with FSWT based on these characteristics.
Example 1. the original signal in Fig. 2 is stated as s = s
1
+ s
2
+ s
3
,
where s
1
, s
2
and s
3
are simulated by Eq. (14) with the parameters
described in Table 2.
Let FSF be the Gaussian function ^ px e

1
2
x
2
, hence Dx
p

2
p
=2: According to Eq. (20), we can take j = 0.707/g, and set
g = 0.025, hence j = 28.28 and g
p
= 0.025. So the chosen parame-
ters for FSWT Eq. (17) are summarized as follows
^ px e

1
2
x
2
; r
x
j
; g
s
0:025; g
p
0:025; j 28:28
25
Fig. 2 reveals that FSWT has three clear separated peaks that
only indicate one modal. In fact, this signal is equivalent to a SDOF
response for multiple impacts. The Fourier spectrum Fig. 2b only
shows one peak since it cannot show the same frequency signals.
Fig. 3 shows two groups of three slices located near the maximum
response frequency of the modal signal s under two FSFs. Fig. 4
demonstrates these slices of the FSWT under high noise (+25%).
Fig. 5 compares with the MWT without noise or under high noise
(+25%); some drawbacks of the MWT method are revealed out,
where the complex Morlet wavelet function is taken as
wt e

1
a
t
2
e
ibt
26
The observed features of damping vibration signals are summarized
as below
(1) FSWT shows the details of time and frequency components
for each modal individually, such as its main frequency
and the response time.
(2) Each modal signal on the 2D map of FSWT coefcients is a
connected area. Note that this is an important feature for
modal signal separation.
(3) All frequency slices of a single modal signal demonstrate the
damping envelopes: Ae
2pf ft
(as a result of Theorem 1).
(4) Different FSFs maintain similar properties. There are some
differences in FSWT amplitudes, but this does not affect
the damping estimation, because the damping in Eq. (24b)
is only a ratio of the FSWT amplitudes.
(5) The FSWT shown in Fig. 4 presents high performance against
noise, but the MWT shown in Fig. 5 is more sensitive to
noise.
(6) FSWT can be controlled only by the frequency resolution
ratio g
s
of the signal, but the CWT must depend on the center
frequency and the bandwidth parameter simultaneously.
Fig. 5 points out that the MWT may have much lower fre-
quency resolution ([17]) even if under the same parameters
g
w
= 0.025 and g
p
= 0.025 with the FSWT. At the same time,
the MWT has the frequency-bands energy leakage obviously
([17]).
Based on the analysis of Eqs. (17), (19)(22) etc., the timefre-
quency localization of FSWT can be adaptively controlled by the
frequency resolution ratio of signal. Therefore, by using timefre-
quency localization of FSWT, the multi-modals signal can be sepa-
rated into single modal. Moreover, all separation errors can be
viewed as a certain level of random noise; and Eq. (24b) or Eq.
(12b) can further eliminate the noise in damping computation.
This is the main idea proposed in this paper. Consequently, com-
bining the characteristics of the FSWT timefrequency image, we
will introduce the modal separation method in the following.
4.3. Determining modal domains
Example 2. Fig. 6a shows a real FDR signal obtained from a small
laboratory bridge monitoring system (LBMS), which will be
introduced in latter Section 5.3. The Fourier transform spectrum
Fig. 6(b) shows that, the rst modal signal is a clear indication of
the greatest energy at 29.3 Hz and the other peaks are smaller.
Fig. 6ch shows the results of the FSWT method, where all of FSWT
parameters are assumed in Eq. (25). The more clearly damping
characteristics of the signal in timefrequency domain are revealed
in Fig. 6c and d. Note that the second modal not the rst has the
highest amplitude at about 112.3 Hz. We then use this example to
explain the modal separation ow as below
100
150
0
50
0
5
10
20
0
2
4
6
8
FSWT
=28.28
A
m
p
l
i
t
u
d
e

T
im
e

(
S
e
c
)
F
re
q
u
e
n
c
y
(H
z
)
T
i
m
e

(
S
e
c
.
)

5
25
0 5 10Hz
0
10
150 0
0 5 10Hz 0 5 10 15 20 Sec.
-1.5
1.5
0
3
9
0
6
F
o
u
r
i
e
r

S
p
e
c
t
r
u
m

A
m
p
l
i
t
u
d
e

15
20
10
Fig. 2. (a) Simulated signal by Table 2; (b) Fourier spectrum; (c) and (d) are 2D and 3D maps of the FSWT coefcients, where g
p
= 0.025.
Table 2
The simulated signal shown in Fig. 2(a).
S A F f h t
0
s
1
1 5 Hz 0.02 0 4.0s
s
2
1 5 Hz 0.02 0 10.0s
s
3
1 5 Hz 0.02 0 16.0s
s = s
1
+ s
2
+ s
3
T
s
= 25s, f
s
= 400 Hz.
Z. Yan et al. / Computers and Structures 89 (2011) 1426 19
(1) The FSWT translates the multi-modals signal into a time
frequency image, such as Fig. 6(c) and (d).
(2) By processing the image, we can search for the Interest
Regions (IRs) called the objective signals or modal domains
in which the main energy is concentrated. IR blocks of the
multi-modals signal are to segment into single modal. For
example, see the red triangles marked in Fig. 6c.
(3) Each objective signal can be reconstructed from one IR block
by Eq. (21). The reconstructed modal signals shown in
Fig. 6eh are also desirable, where there are three pathways
to direct them.
FSWT provides a new approach where both ltering and seg-
menting can be processed simultaneously in time and frequency
domain. Whether signal ltering or signal segmentation, IRs still
needs to be found. Many image segmentation methods can be
implemented to determine the IRs. This paper ignores the discus-
sion, but a robust real time object detection method [19] is
recommended.
As stated above, due to the localization of FDR with FSWT anal-
ysis, the modal domain can be found from the FSWT image. There-
fore, we here introduce the timefrequency projectors of the FSWT
image to determine the modal domains.
(1) Frequency projector of the FSWT coefcients can be stated
as
P
f
x
_
Ts
0
jWs; x; rjds; 27
where T
s
is the sample time.
For example, there is a frequency projector shown in Fig. 7a. It is
clear that each peak of this curve points to the maximum fre-
quency response. We can use those peaks to segment all the
frequencies into a number of slices, and denote the Frequency
Slices as
20 Sec.
40
0
0 10 0 10
120
80
0 10 20 Sec.
-1.5
1.5
0
T
i
m
e

(
S
e
c
.
)

20
0 5 20Hz
0
10
150
0
FSWT
Noise=25%
M
a
g
n
i
t
u
d
e

A
m
p
l
i
t
u
d
e

20 Sec.
40
0
120
80
M
a
g
n
i
t
u
d
e

Fig. 4. (a) A noise version (+25%) of the original signal shown in Fig. 2a. (b) 2D map of FSWT coefcients. Compare with Fig. 3a and b under high noise, (c) and (d) are a group
of slices of modal signal s at 5 Hz shown in (a) and (b) with slice functions ^ px e
1=2x
2
and ^ px 1=1 x
2
respectively.
0 5
40
0
20 Sec. 15 10
120
80
40
0
120
80
M
a
g
n
i
t
u
d
e

o
f

F
S
W
T

C
o
e
f
f
i
c
i
e
n
t
s


0 5 10 15 20 Sec.
Fig. 3. Compares two groups of frequency slices of FSWT representation. The original signal is shown in Fig. 2a. (a) The rst group is sliced by function ^ px e
1=2x
2
, red
color is the slice of frequency f = 5 Hz, green one is f + 0.4 Hz, and blue one is f 0.4 Hz. (b) is the same group of slices but with another function ^ px 1=1 x
2
. (For
interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
40
0
120
80
0 5 10 15 20 Sec.
40
0
120
80
T
i
m
e

(
S
e
c
.
)
T
i
m
e

(
S
e
c
.
)
0 5 10Hz
150
0
MWT
Noise=25%
0 5 10 15 20 Sec.
M
a
g
n
i
t
u
d
e
M
a
g
n
i
t
u
d
e
0 5 10Hz
20
0
10
20
0
10
Fig. 5. Compare the MWT with the FSWT for same parameters g
p
= 0.025 and g
w
= 0.025; (a) 2D map of the MWT of Fig. 2a; (b) 2D map of the MWT of Fig. 4a; (c) and (d) are
two groups of slices of modal signal s at 5 Hz shown in (a) and (b).
20 Z. Yan et al. / Computers and Structures 89 (2011) 1426
x
0
; x
1
; x
1
; x
2
; . . . x
n1
; x
n

Here the segments are not strict, but it is necessary that each slice
include the main energy of one modal signal. One of the simplest
choices is [x Dx, x+ Dx] and Dx= kx, where k is recom-
mended to select 0.10.2 [14].
(2) Time projector of FSWT can be stated as
P
t
t
_
x
i1
x
i
jWt; x; rjdx 28
Here [x
i
, x
i+1
] is a frequency slice and it may include many near
frequency signals.
For example, there is a time projector shown in Fig. 7b. It is clear
that a peak of this curve points to the start time of the signal. We
call these peaks the time trigger points of the signal. In general, a
real signal may have a number of time trigger points. We can di-
vide the sample time into many Time Slices, and denote them as
t
0
; t
1
; t
1
; t
2
; . . . t
m1
; t
m

Here the time segments are not strict too, but it is necessary
that each time slice includes the main energy of one trigger re-
sponse signal.
As stated above, we can summarize the modal analysis ow as
Fig. 8. According to the ow and Eq. (24b) or Eq. (12b), we can pro-
vide the following algorithm for damping parameter identication.
4.4. Algorithm for modal parameter identication
Algorithm 1 is called the frequency slice algorithm (FSA) for
modal parameter identication. After Step 1 in FSA, one can esti-
mate the modal damping directly by Eq. (24b). However, before
clear out other modal frequencies for MDOF signal, Eq. (24b) is
only an approximate estimator. The latter Example 3 will show
the difference between Eqs. (12b) and (12b). At the same time,
we will show a comparable example that Eq. (24b) can directly
give an acceptable estimation of damping for random response.
Therefore, from Step 2 to Step 4 in FSA is somewhat necessary to
separate MDOF response into each single modal signal for getting
higher accuracy damping ratio in application, especially for close
modes with high damping ratio.
1
0
2
b
150 0
100
150
0
50
0
2
4
6
0
50
100
150
d
200
T
i
m
e

(
S
e
c
.
)
1
3
5
0 40 80 120 160 HZ
c
a
0
2
4
0 1 2 3 4 Sec.
0
1
-1
0
0.5
-0.5
0
0.5
-0.5
0
0.5
-0.5
F
S
W
T

R
e
c
o
n
s
t
r
u
c
t
e
d

e
f
g
h
S
y
n
t
h
e
s
i
z
e
d

S
i
g
n
a
l
FSWT
=28.28
0 1 2 3 4 Sec.
0
0 40 80 120 160 Hz
MWT
1
F
F
T
S
p
e
c
t
r
u
m
-1
A
m
p
l
i
t
u
d
e
T
im
e

(
S
e
c
)
F
re
q
u
e
n
c
y
(H
z
)
Fig. 6. (a) A test signal in LBMS; (b) Fourier spectrum; (c) and (d) are the 2D and 3D maps of FSWT coefcients respectively; (e)(g) are the reconstructed signal by the inverse
FSWT; (h) The synthesized signal.
0 50 100 150 200
50
0
150
100
40
60
100
80
20
0
0 1 2 3 4 5
Fig. 7. (a) and (b) are the frequency and time projectors of the FSWT coefcients in Fig. 6c, respectively.
RDT Estimation of
Damping directly in
FSWT Domain
RDT Method
of damping for
SDOF Signal
FSWT
Time-Frequency
Transform
Determine Modal
Domain and
Modal Separation
Inverse FSWT
Transform
Time Domain
Data of
MDOF Signal
Projectors of Time-
Frequency Domain
Fig. 8. Modal separation and damping identication with FSWT.
Z. Yan et al. / Computers and Structures 89 (2011) 1426 21
Algorithm 1. Frequency slice algorithm (FSA) for modal
parameter identication.
Step 1. Input the signal data, and compute the FSWT
coefcients by Eq. (17)
Step 2. Determine frequency slices [x
i
,x
i+1
] of each modal
signal with the frequency projector of the FSWT coefcients
in Eq. (27)
Step 3. Determine time slices [t
j
,t
j+1
] of each modal signal with
the time projector of the FSWT coefcients in Eq. (28)
Step 4. Reconstruct the modal signal on each IR ([t
j
, t
j+1
][x
i
,
x
i+1
]) by FSWT Eq. (21) and compute the modal damping
parameter by Eq. (12b)
5. Application
5.1. Modal separation of close modal signal based on dynamic scale
In this section, an extreme example is to show the ability of the
FSWT. We will compare the FSWT with the CWT-based method for
modal separation in [13], where they have given a signicant re-
search about the selections of the wavelet scale and the centre fre-
quency based on the minimum Shannon entropy search by using
the MWT formula described by Eq. (26).Example 3 We consider a
simulated signal
f t
e
0:5pt
sin9:987pt e
0:55pt
sin10:986pt t ! 2
0 t < 2
_
;
where there are a couple of signals f
1
= 5 Hz, f
1
= 0.05 and f
2
=
5.5 Hz, f
2
= 0.05. The signal f(t) is shown on Fig. 9a where we set
f
s
= 100 Hz and T
s
= 6 s.
Although Example 3 is somewhat similar with the example in
[5], it is much more difcult than [5] to separate them because
their resonant frequencies are very close and the signal has high
damping ratio that cause the valid duration time of measured sig-
nal to be very short. The Fourier spectrum of the signal without
noise is shown in Fig. 9b. Fig. 9c and d show the time response
and spectrum of the original signal by superposition of a mild noise
(+15%). It is evident that two resonant peaks are difcult to recog-
nize from the Fourier spectrum.
All FSWT parameters are also determined as the same as Eq.
(25). Note that j = 28.28 is a constant. Fig. 9e represents the
timefrequency image of the FSWT coefcients. However, it is still
not easy to observe the different modes from that image. The fre-
quency projective curve of Fig. 9e is shown in Fig. 9f, where it
seems to be clear that there are two frequency signals at
f
1
= 5 Hz and f
2
= 5.5 Hz. Meanwhile, we cannot nd out another
suitable scale j to distinguish them even if we completely sacrice
the time resolution in Fig. 9e.
0
6
T
i
m
e

(
S
e
c
.
)
2
4
4.0 4.5 5.0 5.5 6.0 Hz
0
6
T
i
m
e

(
S
e
c
.
)
2
4
4.0 4.5 5.0 5.5 6.0 Hz
FSWT =28.28/2 ~28.28*2
MWT 35 . 8 , 15 . 1 = =
0 1 2 3 4 5 Sec.
0
1
-1
0 1 2 3 4 5 Sec.
0
1
-1
150
0
0 1 2 3 4 5 Sec.
0
2
-2 S
i
m
u
l
a
t
e
d

S
i
g
n
a
l
0 2 4 6 8 Hz
20
40
0
F
F
T
S
p
e
c
t
r
u
m
0
6
T
i
m
e

(
S
e
c
.
)
2
4
4.0 4.5 5.0 5.5 6.0 Hz
4.0 4.5 5.0 5.5 6.0 Hz
150
250
350
50
FSWT =28.28
0 2 4 6 8 Hz
20
40
0
F
T

S
p
e
c
t
r
u
m
N
o
i
s
e

V
e
r
s
i
o
n
0
2
-2
0 1 2 3 4 5 Sec.
60
Noise (+15%)
80
a b
c d
e f
g
h
i
j
Fig. 9. (a) Simulated signal; (b) Fourier spectrum of (a); (c) A noise version (+15%) of (a) and the followings (d)(j) are shown to analyze the noise version signal (c). (d) Fourier
spectrum; (e) 2D map of FSWT with scale j = 28.28; (f) frequency projector of (e); (g) 2D map of the MWT after the time-consuming iterations for minimum Shannon entropy
search; (h) 2D maps of FSWT with dynamic scale j j0
^
f u

=
^
f x

, where j
0
= 28.28/2; (i) and (j) are the reconstructed signals from the IRs shown in (h) respectively.
22 Z. Yan et al. / Computers and Structures 89 (2011) 1426
Similarly, the CWT-based method of [13] has the same problem.
Fig. 9g shows the hard effort by using the MWT in Eq. (26), where
the bandwidth parameter a = 4.5 and the center frequency b = 6
are the estimated parameters after the time-consuming iterations
for rst computing the MWT coefcients and then getting and
comparing the Shannon entropy. Unfortunately, it cannot distin-
guish the two close frequency components correctly. Moreover,
both of the time and frequency resolutions or localizations of
Fig. 9g are much lower than that of Fig. 9e. Meanwhile the result
of the MWT is very sensitive to noise. We have no better decision
for a wavelet scale that is able to separate them. Naturally, it is
very difcult to obtain high accuracy estimations for the damping
ratios.
Fortunately, with the same FSF as Eq. (25), scale j can be chan-
ged dynamically in terms of Eq. (22) where we not only still use
j
0
= 28.28 as Eq. (25), but also should take the frequency projector
Fig. 9(f) instead of the Fourier spectrum Fig. 9d since the real spec-
trum Fig. 9b is drown in the noise. Consequently, Fig. 9h indicates
that there are a couple of signals aligned separately, and it is obvi-
ous that there are two maximum response frequencies nearby
f
1
= 5 Hz and f
2
= 5.5 Hz. Therefore, we can easily separate them
into the single modal shown in Fig. 8i and j, where some errors
especially in the initial stages are due to the noise and the fact that
the segmentations of signal IRs cannot be rened. However, the
reconstructed signals are acceptable. Actually, in order to obtain
the similar result to Fig. 9h, we have many selections resembling
to Eq. (22), but here we omit the further discussion.
Finally, the damping ratios estimated directly by Eq. (24b) are
f
1
= 0.0453 and f
2
= 0.0558, and the errors are not over 13%. How-
ever, after the segmentation, the most accurate damping ratios ob-
tained by Eq. (12b) are f
1
= 0.0483 and f
2
= 0.0532, and the errors
are not over 6.5%. The difference between Eqs. (24b) and (12b)
are mainly due to that the signal f(t) is a MDOF response before
it is separated, but Eq. (12b) is accurate when f(t) is segmented into
SDOF signal.
Remark. This example provides a simple approach to overcome
the problem of close modes decomposition by using dynamic scale
technique, especially for high damping ratio signal. At the same
time, usually Eq. (24b) can directly give enough accuracy for
application. Nevertheless, if one want to get higher accurate
estimation of damping for MDOF response, in the rst step, it is
somewhat necessary that separate the MDOF signals into single
modal signal, and Eq. (12b) further gives a higher accurate
estimation.
5.2. Digital simulation and comparison
In this section, all of the FSWT parameters are still assumed in
Eq. (25). To compare with Refs. [13,14], we assume that the simu-
lated signal is under the same conditions with Ref. [14], where
f
1
= 1 Hz, f
1
= 0.03, f
2
= 1.1 Hz, f
2
= 0.02, f
3
= 3 Hz, f
3
= 0.01; all of
phase angles are 0, and f
s
= 20 Hz. Table 3 shows the identied re-
sults of modal parameters; note that the sample time T
s
is de-
creased to 9 s, which is shorter than T
s
= 12 s of reference [14].
From Table 3, the FSA algorithm can maintain the high accurate
estimation for modal parameters even though with stronger noise
and shorter sample time compared with Refs. [13,14]. Therefore,
FSA can be implemented to identify the modal parameters of a sys-
tem with high damping and close modal frequencies.
5.3. Experimental verication
Fig. 10 shows a small laboratory bridge monitoring system
(LBMS). There are 11 sensors of piezoresisitive ARF-10A accelera-
tion (at frequency response: 050 Hz) installed under the main
beams. DC-104R is applied to collect the free-decay responses by
an impact with a light hammer, where f
s
= 1000 Hz and T
s
= 5 s.
The measured free-decay signals are ready for identifying the mod-
al parameters. The FSA is applied to compute the damping. Two
examples are given to test the FSA method, and the single impulse
response and random excitation are presented as below.
5.3.1. Single impulse response
We collect the acceleration signals from LBMS by single impulse
with a hammer. Three responses at locations B1, C1 and D1 and
their FSWTs are shown in Fig. 11, where all of the FSWT parame-
ters are still assumed in Eq. (25). The rst three main modes are:
f
1
= 29.3 Hz, f
2
= 112.3 Hz, and f
3
= 164.0 Hz, which are almost the
same to all observing sensors.
Table 3
Identication of modal damping ratio using FSA for MDOF response with different noise levels.
Noise level (R) Modal parameter Sample time (s)
T
s
= 12 T
s
= 12 T
s
= 9
CWT-based [13] CWT-based & search [14] Proposed FSA, statistic times = 100, modal parameter average and variance
E() Var() E() Var()
5% f
1
0.989 1.0 0.9804 1.2e32 0.9804 1.2e32
f
1
0.0131 0.0302 0.0300 7.5e08 0.0338 1.3e07
f
2
1.111 1.1 1.1176 7.2e30 1.1176 7.2e30
f
2
0.0117 0.0198 0.0188 2.6e08 0.0200 6.5e08
f
3
3.00 3.00 3.00 0 3.00 0
f
3
0.0096 0.0099 0.0099 1.1e08 0.0099 1.7e08
20% f
1
0.989 1.001 0.9804 1.2e32 0.9802 3.8e6
f
1
0.0132 0.0296 0.0301 9.2e07 0.0339 2.2e06
f
2
1.11 1.1 1.1176 7.2e30 1.120 5.4e30
f
2
0.0118 0.0209 0.0187 4.3e07 0.0202 1.0e06
f
3
3.002 3.00 3.00 0 3.00 0
f
3
0.009 0.01 0.01 1.7e07 0.01 3e07
+50% f
1
0.9825 1.5e05 0.9712 1.0e4
f
1
0.0302 3.1e05 0.0330 4.0e05
f
2
1.1089 5.2e05 1.1178 3.8e06
f
2
0.0191 1.3e05 0.0200 1.5e05
f
3
3.001 4.2e05 3.002 4.7e05
f
3
0.0102 3.4e06 0.0101 3.7e06
Z. Yan et al. / Computers and Structures 89 (2011) 1426 23
The rst three computed modals (i.e. the high-energy responses
pointed by red arrows in Fig. 11) for nine observing positions are
presented in Table 4. From Table 4, note that the rst mode damp-
ing ratios and frequencies are almost same for the observing posi-
Fig. 10. LBMS: (a) Sensor location; (b) experimental model.
Single pulse responses
A
m
p
l
i
t
u
d
e
100
150
0
50
0
2
6
0
50
100
150
200 4
150
0
FSWT
=28.28
A
m
p
l
i
t
u
d
e
100
150
0
50
0
2
6
0
50
100
150
200 4
FSWT
=28.28
A
c
c
e
l
e
r
a
t
i
o
n

(
g
)
0
-1
1
0 1 2 3 4 Sec.
At B1
A
c
c
e
l
e
r
a
t
i
o
n

(
g
)
0
-1
1
0 1 2 3 4 Sec.
At C1
A
c
c
e
l
e
r
a
t
i
o
n

(
g
)
0
-1
1
0 1 2 3 4 Sec.
At D1
0
A
m
p
l
i
t
u
d
e
100
150
0
50
0
2
6
50
100
150
200 4
FSWT
=28.28
a
b
c
d
e f
T
im
e
(
S
e
c
) Frequency (H
z)
T
im
e
(
S
e
c
) Frequency (H
z)
T
im
e
(
S
e
c
) Frequency (H
z)
Fig. 11. (a), (b) . . .(f) are the obtained signals at positions B1, C1 and D1 and their FSWT 3D maps, red arrows point to the rst three modals computed in Table 4. (For
interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
Table 4
Identication of modal parameter using FSA algorithm for free-decay responses of the LBMS.
State Modal parameter T
s
= 5s, f
s
= 1000 Hz, proposed FSWT, statistic times = 10
Modal parameter average
B1 B2 B3 C1 C2 C3 D1 D2 D3
Single impulse test f
1
29.27 29.27 29.27 29.27 29.27 29.27 29.27 29.27 29.27
f
1
0.0065 0.0065 0.0065 0.0068 0.0067 0.0067 0.0065 0.0066 0.0065
f
2
112.36 112.36 112.36 112.36 112.36 112.36 112.38 112.36 112.36
f
2
0.0093 0.0094 0.0095 0.0097 0.0095 0.0096 0.0099 0.0094 0.0095
f
3
164 164 164 164.03 164.05 164.04 164 164.05 164
f
3
0.0028 0.0030 0.0030 0.0037 0.0036 0.0036 0.0033 0.0030 0.0031
24 Z. Yan et al. / Computers and Structures 89 (2011) 1426
tions B1, B2, and B3, C1,C2 and C3, D1, D2 and D3 in the simple test.
From the Fig. 11 and Table 4, it can be concluded that FSWT and
FSA methods are steady.
5.3.2. Random impacting response
Fig. 12 shows an acceleration response of LBMS under random
impacts to compare the single impulse described as the above.
The measured response is equivalent to the combining of multiple
FDRs. Usually, for most of the existing approaches to computing
damping of random response, as the rst step, it is necessary to
compute the FDR by RDT, and then one can compute the modal
parameters (e.g.[1]). However, by using frequency slice projector
of FSWT shown in Fig. 12b, we can choose each maximumresponse
frequency slice (see the dot bars in Fig. 12b) as each modal signal,
then the RDT formula Eq. (24b) can be used to directly compute the
damping of random response. In fact, FSWT or Eq. (10) translates
the response in time domain into the distribution in timefre-
quency domain. It is important that FSWT divides the signal from
the noise spectrum automatically. By using Eq. (24b) directly for
random response, Table 5 gives the computational damping of
the chosen maximum response slices, and their frequencies are
very similar to Table 4. The maximum frequency error is not over
5% comparing with Table 4, and the maximum damping error is
not over 17%. Nevertheless, note that we only use 5 s time data
for comparable computation with Table 4. At the same time, many
tests show that the higher frequency modal signals have higher
damping errors by Eq. (24b). The main reason is probably that
the duration time of those modal signals is very short.
From the comparison between the single impulse and the ran-
dom excitation, FSWT and FSA can estimate the modal damping di-
rectly for MDOF signals with an acceptable accuracy. Moreover,
more applications of FSWT and FSA especially for damage detec-
tion will be carried on in the near future.
5.3.2.1. Comparing and remarks. The main theoretical advantage of
the proposed approach over other existing approaches based on
WT is that its measured responses processed are not limited to
free-decay responses, it is also unnecessary to use a lter to de-
noise in practical application because FSWT itself is a new kind
of good lter [5,17]. At the same time, applying the proposed ap-
proach to determine the modal parameters of a system from its
ambient measurements does not require the technique such as
the random decrement technique to convert the random responses
into free-decay responses. Although there are many damping iden-
tifying methods [715] such as NExT, ERA, SSI and Wavelet etc., as
the rst step, usually, it is necessary to compute the FDR by RDT.
However, in this paper, a new usage of the RDT idea can be imple-
mented directly in FSWT domain, Eq. (24b) or Eq. (12b) can be used
for random responses for getting damping parameter, and the
FSWT is therefore very simple and direct.
Notably, in Section 5.2, we have given a comparable result with
CWT in [13,14]. FSA can also give higher accurate estimation even
with higher damping and shorter sampling time since FSWT can
provide a modal separation method by dynamic scale controlling.
On the other hand, unlike the existing approaches based on CWT,
the wavelet function and the chosen scale parameter can always
affect the accuracy of the identied modal parameters, FSWT can
provide an adaptive method for the center frequency and an adap-
tive window scale for different frequency responses. Although this
work only demonstrates the feasibility of the proposed approach in
processing the responses of laboratory system by hammer impact-
ing excitation, the proposed methods are certainly suitable for
dealing with free-decay responses or the MDOF responses from
the random loads. After segmentation and reconstruction of FSWT,
Eq. (12b) can therefore give higher accuracy than Eq. (24b) directly
even when the responses and input included great noise with 50%.
Nevertheless, in this case, usually, Eq. (12b) should be based on the
FDR signal similar with many existing approaches.
6. Conclusion
(1) The background of the powerful FSWT method is rst intro-
duced clearly. By using RDT, a good damping estimation in
FSWT domain is obtained in this paper. Combining RDT
0 40 80 120 160 Hz
A
c
c
e
l
e
r
a
t
i
o
n



(
g
)
0
-1
1
0 1 2 3 4 Sec.
100
150
0
50
0
2
4
0
50
100
150
6
FSWT
=28.28
80
0
120
40
200
M
a
g
n
i
t
u
d
e
T
im
e

(
S
e
c
)
F
re
q
u
e
n
c
y
(H
z
)
a b
c
Fig. 12. (a) Random impacting response, (b) FSWT 3D map, (c) frequency projective curve of FSWT and maximum response frequencies.
Table 5
Identication of modal parameter by using Eq. (24b) directly for random responses of LBMS.
State Modal parameter T
s
= 5s, f
s
= 1000 Hz, Proposed FSWT, Statistic Times = 10
Modal Parameter Average
B1 B2 B3 C1 C2 C3 D1 D2 D3
Random impacting test f
1
29.72 29.61 29.72 29.62 29.74 29.72 29.75 29.72 29.71
f
1
0.0071 0.0069 0.0057 0.0065 0.0062 0.0070 0.0070 0.0068 0.0061
f
2
113.25 113.12 113.25 113.23 113.22 113.22 113.34 113.32 113.25
f
2
0.0087 0.0103 0.0106 0.0101 0.0103 0.0105 0.0108 0.0107 0.0103
f
3
165.05 165.04 164.64 164.56 164.60 164.85 164.85 164.85 165.05
f
3
0.034 0.0035 0.0033 0.0031 0.0032 0.0035 0.0034 0.0036 0.0034
Z. Yan et al. / Computers and Structures 89 (2011) 1426 25
and FSWT modal separation method, a frequency slice algo-
rithm (FSA) is designed to get high accurate estimation of
damping parameter. Both numerical and experimental tests
demonstrate that FSA is effective to identify modal parame-
ters of a system response even with large damping and close
modal interference, and strong noise. The computational
results of FSA for modal parameters are accurate and steady.
(2) FSA is not limited to FDR, and also can be directly used to
random impacting response. It is usually a good estimator
for general applications. Notably, FSWT itself is a new kind
of good lter and has high performance against noise. It is
necessary to get damping parameter with higher accuracy
through modal separation by FSWT, and FSWT can be con-
trolled adaptively in modal separation by dynamic scale
method. FSWT is adaptive to analyze the damping vibration
signals in time and frequency domains simultaneously.
(3) Compared with the results obtained from the traditional FFT
and CWT etc., the FSWT method can show the damping
characteristics of modal signal in timefrequency domain
more clearly and easily. Only one group of FSWT parameters
in Eq. (25) is used to t all examples in this paper, therefore
FSWT presents strong robustness in transformation of vibra-
tion signal in an easy way. Frequency slicing processing is an
important idea in this application.
Appendix A. Proofs
A.1. The proof of Theorem 1
Since Eq. (6) can be changed into
Wt; x; r r
_
1
1
f s te
ixs
p

rsds;
and from the condition s < 0, p(s) = 0, it can be rewritten as
Wt; x; r r
_
1
0
f s te
ixs
p

rsds
According to the denition of f(t), the additional condition t ! t
0
and bT = kp, it is easy to know that f(s + t + T) = e
aT
f(s + t).
Moreover
jWt T; x; rj e
aT
jWt; x; rj
a
1
T
lnjWt; x; rj lnjWt T; x; rj:
Thus, we have proven Eq. (23) and (Eq. (24a). Finally, using RDT
method, we can obtain (24b).
References
[1] Sohan H, Farrar CR, Hemez FM, Shunk DD, Stinemates DW, Nadler BR. A review
of structural health monitoring literature: 19962001. Los Alamos National
Laboratory report, LA-13976-MS; 2003.
[2] Doebling, SW, Farrar CR. The state of the art in structural identication of
constructed facilities. A report by the ASCE Committee on Structural
Identication of the Constructed Facilities; 1999.
[3] Ndambi JM, Peeters B, Visscher JDe, Wahab MA, Vantomme J, Roeck GDe, et al.
Comparison of techniques for modal analysis of concrete structures. Eng Struct
2000;22:115966.
[4] He X, Moaveni B, Conte JP, Elgamal A. Comparative study of system
identication techniques applied to New Carquinez Bridge, <http://
www.health.monitoring.ucsd.edu/documentation/public>; 2006.
[5] Yan Z et al. Frequency slice wavelet transform for transient vibration response.
Mech Syst Signal Process 2009;23(5):147489.
[6] Bendat JS, Piersol AG. Engineering applications of correlation and spectral
analysis. second ed. New York: John Wiley; 1993.
[7] Bodeux JB, Golinval JC. Application of ARMAV models to the identication and
damage detection of mechanical and civil engineering structures. Smart Mater
Struct 2001;10:47989.
[8] James GH, Garne TG, Lauffer JP. The Natural excitation technique (NExT) for
modal parameter extraction from operating wind turbines. Report SAND92-
1666, UC-261. Albuquerque, New Mexico: Sandia National Laboratories; 1993.
[9] Qin Q, Li HB, Qian LZ. Modal identication of Tsing Ma bridge by using
improved Eigensystem realization algorithm. J Sound Vib 2001;247:32541.
[10] Lardies J, Gouttebroze S. Identication of modal parameters using the wavelet
transform. Int J Mech Sci 2002;44:226383.
[11] Kijewski T, Kareem A. Wavelet transforms for system identication in civil
engineering. Comput-aided Civil Infrastruct Eng 2003;18:33955.
[12] Neild SA, McFadden PD, Williams MS. A review of timefrequency methods for
structural vibration analysis. Eng Struct 2003;25:71328.
[13] Yan BF, Miyamoto A, et al. Wavelet transform-based modal parameter
identication considering uncertainty. J Sound Vib 2006;291:285301.
[14] Tan Jiu-Bin, Liu Yan, Wang Lei, Yang Wen-Guo. Identication of modal
parameters of a system with high damping and closely spaced modes by
combining continuous wavelet transform with pattern search. Mech Syst
Signal Process 2008;22:105560.
[15] Huang CS, Su WC. Identication of modal parameters of a time invariant linear
system by continuous wavelet transformation. Mech Syst Signal Process
2007;21:164264.
[16] Hong J-C, Kim YY. The determination of the optimal Gabor wavelet shape for
the best timefrequency localization using the entropy concept. Exp Mech
2004;44:38795.
[17] Yan Z et al. An overall theoretical description of frequency slice wavelet
transform. Mech Syst Signal Process 2010;24(2):491507.
[18] Asmussen JC. Modal analysis based on the random decrement technique
application to civil engineering structures. Ph.D thesis. Denmark: University of
Aalborg; 1997.
[19] Viola P, Jones M. Robust real time object detection. In: IEEE international
conference on computer vision workshop on statistical and computational
theories of vision. Vancouver, Canada; July 13, 2001. p. 90510.
26 Z. Yan et al. / Computers and Structures 89 (2011) 1426

S-ar putea să vă placă și