ARTICLE IN PRESS

JOURNAL OF
SOUND AND
VIBRATION
Journal of Sound and Vibration 307 (2007) 527–544
www.elsevier.com/locate/jsvi

Nonlinear vibration characterization by signal decomposition

P. Frank Pai
Department of Mechanical and Aerospace Engineering, University of Missouri-Columbia, Columbia, MO 65211, USA
Received 2 June 2005; received in revised form 26 October 2006; accepted 13 June 2007
Available online 28 August 2007

Abstract

Methods for nonlinear vibration characterization by decomposing dynamic responses using the Hilbert–Huang
transform and a sliding-window fitting technique are presented. Numerical results show that Hilbert–Huang transform can
be used for decomposing nonlinear/non-stationary signals in order to reveal and estimate nonlinear effects. Major
nonlinear phenomena that can be extracted from transient and/or steady-state dynamic responses include different
nonlinearities, softening and hardening effects, intrawave amplitude- and phase-modulation, distorted harmonic responses
under a single-frequency harmonic excitation, interwave amplitude- and phase-modulation, and multiple-mode vibrations
caused by internal/external resonances. However, the discontinuity-induced Gibbs’ phenomenon makes Hilbert–Huang
transform analysis inaccurate around the two data ends. On the other hand, the sliding-window fitting analysis has no
Gibbs’ phenomenon at the two data ends, but it cannot extract accurate modulation frequencies due to the use of non-
orthogonal basis functions in the sliding-window least-squares curve fitting process.
r 2007 Elsevier Ltd. All rights reserved.

1. Introduction

In the last few decades structural engineers and researchers have been developing dynamics-based methods
for rapid damage inspection of large structures [1]. Damage detection is challenging because it is an inverse
engineering problem. Based on the complexity of sensor systems used, signal processing methods, and
accuracy of deduced damage indicators, dynamics-based damage detection methods can be separated into
three groups. Methods in the first group require a simultaneous full-field measurement tool (e.g., Moire
interferometry, digital shearography, or scanning laser vibrometers), and they process the measured
displacement, slope, or velocity field to compute strains and/or curvatures and then locate damage by
examining abnormality or sudden change of these spatially distributed data [2]. Methods in the second group
require simultaneous measurements of many points, and they use a well-calibrated structural model and a
modal expansion/update method to locate damage. Methods in the third group require simultaneous
measurements of only a few locations, and they use the measured time traces and the travelling sequence of
abnormality or sudden change in the time traces to locate damage. In-work dynamics-based damage detection
methods usually use the third approach. Methods in the first group process spatially distributed data to extract
physical variables (such as slopes, curvatures, and strains) to directly reveal damage locations, but the

Tel.: +1 573 884 1474; fax: +1 573 884 5090.

E-mail address: [email protected]

0022-460X/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jsv.2007.06.056
 ARTICLE IN PRESS
528 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

challenge is how to extract clear damage indicators from the large amount of data obtained from a full-field
measurement. Methods in the second group use the calibrated structural model to compute spatially
distributed physical variables to estimate damage locations, but the challenge is how to repeat the same setup
conditions used in the structural model and how to correlate the model with currently measured dynamic
responses that may be affected by unknown damage and/or changes of physical conditions (such as
temperature change, moisture absorption, and changes of boundary conditions). On the other hand, methods
in the third group process time-domain data to extract dynamic characteristics (i.e., natural frequencies,
damping ratios, and nonlinear effects) to reveal the existence of damage, but the challenge is how to extract
linear and nonlinear dynamic characteristics from dynamic responses (especially transient responses) and how
to correlate the damage indicators obtained at just a few physical locations to locate damage. Even under
small vibrations dynamic responses of damaged structures are often nonlinear because of damage, such as the
opening and closing of cracks. This paper presents methods for processing transient and/or steady-state time
signals to extract time-varying dynamic characteristics that can be used to reveal damage.
A linear dynamic system has constant natural frequencies, and it vibrates at the frequency of an externally
applied harmonic excitation. On the other hand, a nonlinear system has amplitude-dependent natural
frequencies, and it may vibrate at a frequency different from an externally applied harmonic excitation. Other
nonlinear phenomena include multiple-harmonic response under a single-frequency harmonic excitation,
intrawave amplitude- and phase-modulated motions, and multiple-mode vibrations caused by modal
interaction (i.e., interwave modulation) [3,4]. One major method for dynamics characterization of a nonlinear
system is to examine the harmonic components contained in the system’s response to a harmonic excitation.
To show the response of a nonlinear oscillator under a harmonic excitation we consider the following weakly
nonlinear system:
u€ þ mu_ þ o2 u þ m3 u_ 3 þ a2 u2 þ a3 u3 þ a4 u4 þ a5 u5 ¼ F cosðOt þ gÞ, (1a)
where u_  du=dt, t is the time, ai are constants, m and m3 are damping coefficients, F is the excitation
amplitude, o is the linear natural frequency, Oð oÞ is the excitation frequency, and g is the phase angle of the
excitation w.r.t. the response. The second-order asymptotic solution can be derived using the method of
multiple scales to be [5]
X
5
uðtÞ ¼ ai cosðiOtÞ þ b3 sinð3OtÞ,
i¼0
a2 a21 3a4 a41 a2 a21 a4 a41 a3 a31 5a5 a51
a0 ¼   ; a 2 ¼ þ ; a 3 ¼ þ ,
2O2 8O2 6O2 6O2 32O2 128O2
a4 a41 a5 a51 m3 Oa31
a4 ¼ ; a 5 ¼ ; b 3 ¼ , ð1bÞ
120O2 384O2 32
where a1 is a function of F , m, and m3 and needs to be obtained by solving the modulation equations derived
from perturbation analysis [5]. Eq. (1b) reveals that the nonlinear solution is a periodic function (period
T ¼ 2p=O) expanded into multiple harmonics. Because a5 appears in a3 and a4 appears in a2 and a0 in Eq.
(1b), it indicates that higher-order odd-power nonlinearities will behave like the cubic nonlinearity, and
higher-order even-power nonlinearities will behave like the quadratic nonlinearity. However, higher-order
nonlinearities will introduce more small-amplitude high-frequency harmonics.
To show an example we consider the following damped Duffing oscillator subjected to a harmonic
excitation having a frequency O close to the linear natural frequency o and its second-order asymptotic
perturbation solution:
u€ þ mu_ þ o2 u þ au3 ¼ F cos Ot; O  o,

aa3
uðtÞ ¼ a cosðOt  gÞ þ a3 cosð3Ot  3gÞ; a3  5a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 32O2
¼ a2 þ a23 þ 2aa3 cosð2Ot  2gÞ cosðOt  g þ YðtÞÞ
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 529

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
a3
 a2 þ a23 1 þ cosð2Ot  2gÞ cosðOt  g þ YðtÞÞ,
a
a3 sinð2Ot  2gÞ a3
YðtÞ  tan1  sinð2Ot  2gÞ, ð2Þ
a þ a3 cosð2Ot  2gÞ a
where m, a, F , and g are constants. Because aba3 when a is small, the amplitude consists of a constant plus a
small amplitude varying at a frequency 2O, and the Y is a small angle varying at a frequency 2O. In other
words, it is a distorted harmonic with intrawave frequency- and amplitude-modulation caused by
nonlinearities. This phenomenon can be used to identify cubic nonlinearity. However, the challenge is how
to extract time-varying frequencies of amplitudes and phases from experimental time signals.
To show another example we consider the following nonlinear oscillator with quadratic nonlinearity and its
asymptotic perturbation solution:
u€ þ mu_ þ o2 u þ au2 ¼ F cos Ot; O  o,

aa2
uðtÞ ¼ a0 þ a cosðOt  gÞ þ a2 cosð2Ot  2gÞ; a2  5a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 6O2
¼ a0 þ a2 þ a22 þ 2aa2 cosðOt  gÞ cosðOt  g þ YðtÞÞ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
a2
 a0 þ a2 þ a22 1 þ cosðOt  gÞ cosðOt  g þ YðtÞÞ,
a
a 2 sinðOt  gÞ a2
YðtÞ  tan1  sinðOt  gÞ. ð3Þ
a þ a2 cosðOt  gÞ a
Because aba2 when a is small, the amplitude and phase vary at a frequency O. In other words, it is a distorted
harmonic with intrawave frequency- and amplitude-modulation caused by nonlinearities. This phenomenon
can be used to identify quadratic nonlinearity. Again, the problem is how to extract the time-varying
frequencies of amplitudes and phases from experimental time signals, especially if only transient data are
available.
For a continuous system with certain quadratic nonlinearities, if the dependent variable wðx; tÞ is discretized
using the system’s two linear mode shapes f1 ðxÞ and f2 ðxÞ and the corresponding modal coordinates u1 and u2
as
wðx; tÞ ¼ f1 ðxÞu1 ðtÞ þ f2 ðxÞu2 ðtÞ (4a)
and if the spatial distribution of the excitation force is orthogonal to f1 ðxÞ, one may obtain the following
modal equations:
u€ 1 þ 2m1 o1 u_ 1 þ o21 u1 ¼ u1 u2 ,
u€ 2 þ 2m2 o2 u_ 2 þ o22 u2 ¼ u21 þ F cos Ot, ð4bÞ
where oi are the linear natural frequencies, mi are the modal damping ratios, F is the amplitude of the external
excitation force, and O is the external excitation frequency. The first-order perturbation solution of Eq. (4b) is
given as [5]
u1 ¼ a1 cosð12Ot  12ðg1 þ g2 ÞÞ; u2 ¼ a2 cosðOt  g2 Þ, (4c)
where g1 and g2 are phase angles. It follows from Eq. (4a) that experimental measurement of wðx; tÞ at any
location will consist of u1 and u2 , which have different characteristic time scales. Then, the challenges are how
to separate u1 and u2 in order to know the participation of each mode and how to obtain the time-varying
frequencies of ui in order to estimate nonlinearities.
Hence the key for accurate vibration characterization is to decompose complex time signals into functions
of different characteristic time scales and extract their time-varying frequencies and amplitudes. This paper
investigates the use of Hilbert–Huang transform [6–10] and presents a sliding-window least-squares fitting
method for performing such signal decomposition.
 ARTICLE IN PRESS
530 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

2. Signal decomposition methods

Here we present and correlate three signal decomposition methods: the discrete Fourier transform, the
Hilbert–Huang transform, and a sliding-window fitting method.

2.1. Discrete Fourier transform

To extract harmonic components from a time signal uðtÞ the following discrete Fourier transform [11] is
usually used:
0 1
N=2 
X  N=2
X
2pitk 2pitk
uðkDtÞ  uk ¼ a0 þ 2 ai cos þ bi sin ¼ Real@ ðai  jbi Þejoi tk A, (5a)
i¼1
T T i¼N=2

1X N
1X N
1X N
a0 ¼ uk ; ai ¼ uk cos oi tk ; bi ¼ uk sin oi tk ; oi  2pi=T,
N k¼1 N k¼1 N k¼1

1X N
Uði=TÞ  ai  jbi ¼ uk ejoi tk , (5b)
N k¼1
pffiffiffiffiffiffiffi
where k ¼ 1; . . . ; N, j  1, tk  kDt, Dt is the sampling interval, 1=Dt is the sampling frequency, N is the
total number of samples, T ð¼ NDtÞ is the sampled period, Df ð¼ 1=TÞ is the frequency resolution, and the
maximum (Nyquist) frequency is 0:5=Dt ð¼ 0:5N=TÞ. U is the spectrum of uðtÞ, and ai and bi are called
spectral coefficients and they represent amplitudes of harmonic components. The expressions of ai and bi in
Eq. (5a) show that the orthogonality between uðtÞ and cos oi t and sin oi t is used to extract regular harmonics
from uðtÞ. Moreover, it is apparent that the time signal uðtÞ is presented as the summation of N=2 harmonics of
constant amplitudes and phases. Unfortunately, if uðtÞ is non-periodic and/or transient, uð0ÞauðTÞ and U will
include many high-frequency harmonics caused by Gibbs’ phenomenon, which makes it difficult to
understand uðtÞ from its spectrum.

2.2. Hilbert– Huang transform

Hilbert–Huang transform is a newly developed technique for processing nonlinear and non-stationary
signals [6–10]. The two major steps of Hilbert–Huang transform are (i) using the empirical mode
decomposition method to decompose a time-domain signal uðtÞ into n intrinsic mode functions ci
corresponding to different intrinsic time scales as
X
n
uðtÞ ¼ ci ðtÞ þ rn , (6)
i¼1

where rn is the residue, and (ii) performing Hilbert transform and computing the time-dependent frequency oi
and amplitude Ai of each ci [6,7]. Then, the time–frequency–energy (i.e., t  o  A) distribution is named the
Hilbert spectrum and is denoted by Hðo; tÞ.
An intrinsic mode function is a function that satisfies two conditions: (i) the number of extrema and the
number of zero crossing must either equal or differ at most by one in the whole data and (ii) the envelope
defined by the local maxima and minima is symmetric and hence the mean value of the maxima envelope and
the minima envelope is zero at any point. The empirical mode decomposition is based on the assumptions that
the signal has at least one maximum and one minimum, and the characteristic time scale is defined by the time
lapse between the extrema. Once the extrema are identified, all the local maxima are connected by a natural
cubic spline line as the upper envelope. Repeat the procedure for the local minima to produce the lower
envelope. The mean of the upper and lower envelopes is designated as m11 , and the first intrinsic function is
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 531

estimated as c11 given by

c11 ¼ uðtÞ  m11 . (7)
Ideally the upper and lower envelopes should cover all the data between them. In reality, overshoots and
undershoots may exist, and they generate new extrema and shift or exaggerate the existing ones after the
process shown in Eq. (7). New extrema generated in this way actually recover the proper modes lost in the
initial examination, and this process can recover low-amplitude riding waves with repeated sifting. This
process is like sifting because it uses the characteristic time scale to separate the finest local mode from the data
first. The sifting process is to eliminate riding waves and to make the wave-profiles more symmetric. Toward
this end, the sifting process has to be repeated more times. In the kth sifting process, c1k1 is treated as the
data, then
c1k ¼ c1k1  m1k ; k ¼ 2; . . . ; K. (8)
Keep repeating the process until all the local maxima are positive, all the local minima are negative, and waves
are almost symmetric. Then c1k is accepted as c1 . A systematic method of determining the end of iteration is to
limit the deviation Dv computed from the two consecutive sifting results as
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sP
N 2
i¼1 ðc1k ðti Þ  c1k1 ðti ÞÞ
Dv  PN 2 (9)
i¼1 c1k1 ðti Þ

to be a small number and/or to limit the maximum number of iterations [6,7]. Here ti ¼ iDt and NDt ¼ T is
the sampled period. After the c1 is obtained, define the residue r1 , treat r1 as the new data, and repeat the steps
shown in Eqs. (7) and (8) as
c21 ¼ r1  m21 ; r1  uðtÞ  c1 ,
c2k ¼ c2k1  m2k ; k ¼ 2; . . . ; K. ð10Þ
After the c2 is obtained, define the residue r2 , treat r2 as the new data, and repeat the steps as
c31 ¼ r2  m31 ; r2  uðtÞ  c1  c2 ,
c3k ¼ c3k1  m3k ; k ¼ 2; . . . ; K. ð11Þ
The whole sifting process can be stopped when the residue rn becomes a monotonic function from which no
more intrinsic mode function can be extracted. In other words, the last intrinsic mode function has no more
than two extrema. For data with a trend, rn should be the trend. After all ci ðtÞ are extracted, one can perform
Hilbert transform to obtain d i ðtÞ of each ci . Then one can define
zi ðtÞ  ci ðtÞ þ jd i ðtÞ ¼ Ai ejyi ,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ai ¼ c2i þ d 2i ; yi ¼ tan1 d i =ci ; oi ¼ dyi =dt. ð12Þ
Replacing the ci ðtÞ in Eq. (6) with the zi ðtÞ in Eq. (12) and neglecting rn yields
! Z t
Xn
jyi ðtÞ
uðtÞ ¼ Real Ai ðtÞe ; yi ðtÞ ¼ oi ðtÞ dt ¼ tan1 d i =ci . (13)
i¼1 0

Eqs. (5a) and (13) reveal that the Hilbert–Huang transform (i.e., Eq. (13)) is a generalized Fourier expansion
allowing the use of distorted harmonics.

2.2.1. Discussions
The Hilbert–Huang transform is an adaptive method based on the local characteristic time scales of the data
under processing. The intrinsic mode functions are usually physical because the characteristic scales are
physical. Because distorted harmonics with time-dependent frequencies and amplitudes are allowed in the data
decomposition, it does not need spurious harmonics to represent nonlinear/non-stationary signals. Hence, the
Hilbert–Huang transform is more appropriate than the Fourier transform [11] and the wavelet transform [12]
for signal decomposition and time–frequency–energy presentation of nonlinear/non-stationary signals [6,7].
 ARTICLE IN PRESS
532 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

As shown in Eq. (6), the ci represent a complete set of basis functions and they are local and adaptive, but
they may not be really orthogonal. Even pure harmonics of different frequencies are not exactly orthogonal,
and it is why the continuous wavelet in the most commonly used Morlet form suffers severe leakage [12].
However, the empirical mode decomposition used in the Hilbert–Huang transform does not use orthogonality
to extract ci . RT
The marginal spectrum hðoÞð 0 Hðo; tÞ dtÞ from Hilbert–Huang transform analysis offers a measure of
total amplitude contribution from each frequency value, and it represents the cumulated amplitude over the
entire time span in a probabilistic sense. However, the frequency in Hðo; tÞ or hðoÞ has a meaning different
from that in Fourier spectra. In Fourier analysis, the existence of energy at a frequency means a regular
harmonic wave persisted through the entire time span. In Hilbert–Huang transform analysis, the existence of
energy at a frequency only means that, in the entire time span,
R 1=Dt there is a higher likelihood for such a wave to
appear locally. The instantaneous energy density IEðtÞð 0 H 2 ðo; tÞ doÞ can be used to check the energy
fluctuation with time.

2.3. Sliding-window fitting

Several sliding-window fitting methods for time–frequency analysis have been proposed by researchers [13].
Here we use a special set of functions to perform sliding-window fitting in order to show the physical
implication of Hilbert–Huang transform, to confirm the accuracy of Hilbert–Huang transform, and to
understand the merits and limitations of Hilbert–Huang transform.
If a time signal uðtÞ is identified from its Fourier spectrum to have two major frequencies o1 and o2 ðoo1 Þ,
one can assume that
uðtÞ ¼ e1 cosðo1 tÞ þ e2 sinðo1 tÞ þ e3 cosðo2 tÞ þ e4 sinðo2 tÞ þ e5 þ e6 t
¼ C 1 cosðo1 t̄Þ þ C^ 1 sinðo1 t̄Þ þ C 2 cosðo2 t̄Þ þ C^ 2 sinðo2 t̄Þ þ C 3 þ C^ 3 t̄, ð14Þ
where ei are constants, t̄ð t  ts Þ is a moving time coordinate, ts is the observed instant, and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
e2
C 1  e21 þ e22 cosðo1 ts  f1 Þ; C^ 1   e21 þ e22 sinðo1 ts  f1 Þ; tan f1  ,
e1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
e4
C 2  e23 þ e24 cosðo2 ts  f2 Þ; C^ 2   e23 þ e24 sinðo2 ts  f2 Þ; tan f2  ,
e3
C 3 ¼ e5 þ e6 t s ; C^ 3 ¼ e6 . ð15Þ
To obtain the coefficients C j and C^ j for the data point at t̄ ¼ 0 we use the data points around t ¼ ts to
minimize the square error E rror , which is defined as
X
m
E rror  ai ðui  ūi Þ2 , (16)
i¼m

where ui denotes uðt̄i Þ from Eq. (14) and ūi denotes the experimental data at t̄i . The total number of points used
is 2m þ 1, and ai is a weighting factor, which can be chosen to be
1
ai ¼
1þ j 99i=m j
or others. The six equations to determine C i and C^ i for the point at t̄ ¼ 0 are given by
qE rror Xm
qui
¼ 2ai ðui  ūi Þ ¼ 0; C j ¼ C 1 ; C^ 1 ; C 2 ; C^ 2 ; C 3 ; C^ 3 , (17)
qC j i¼m
qC j

which implies that C j are extracted by using the orthogonality between the functions used in Eq. (14) and the
experimental data ūi . After C i and C^ i are determined, it follows from Eq. (14) that
uðts Þ ¼ C 1 þ C 2 þ C 3 . (18)
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 533

It shows that uðts Þ consists of the ‘‘amplitude’’ C 1 of the harmonic cos o1 t̄, the ‘‘amplitude’’ C 2 of the
harmonic cos o2 t̄, and the moving average C 3 . Moreover, it follows from Eq. (15) that
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
e2 C^ 1
A1  e21 þ e22 ¼ C 21 þ C^ 21 ; f1 ¼ tan1 ¼ o1 ts  tan1 ,
e1 C1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
e4 C^ 2
A2  e23 þ e24 ¼ C 22 þ C^ 22 ; f2 ¼ tan1 ¼ o2 ts  tan1 . ð19Þ
e3 C2
If f1 and f2 are constant, it follows from Eq. (19) that
P2
dy1 ½y1 ðts þ iDtÞ  y1 ðts þ ði  1ÞDtÞ
o1 ¼  i¼1 ,
dt 4Dt
P2
dy2 ½y2 ðts þ iDtÞ  y2 ðts þ ði  1ÞDtÞ C^ i
o2 ¼  i¼1 ; yi  tan1 . ð20Þ
dt 4Dt Ci
To reduce the influence of noise on calculated oi , each oi at t ¼ ts is obtained by averaging over 4Dt, as shown
in Eq. (20). The same approach is also used to calculate the oi in Eq. (12).
Because harmonic functions are not orthogonal to the polynomial C 3 þ C^ 3 t̄, one needs to choose an
appropriate window length to enforce the orthogonality in order to obtain unique values for each C i and C^ i .
Numerical results show that an appropriate choice is 2mDtX4p=o2 , i.e., two periods of the lowest harmonic.
This method can be used to extract as many harmonics as needed by adding harmonics identified from the
signal’s Fourier spectrum to Eq. (14).

3. Nonlinear vibration characterization

As indicated by Eqs. (2), (3) and (5a) that a distorted harmonic is a periodic time signal with intrawave
amplitude- and phase-modulation and it can be decomposed into several regular harmonics by discrete
Fourier transform analysis. However, perturbation solutions show that nonlinearities cause nonlinear
oscillators to have distorted harmonic vibrations under a harmonic excitation. Hence, a distorted harmonic is
better than multiple regular harmonics in presenting the response of a nonlinear system to a harmonic
excitation. However, the challenge is how to extract quantitative dynamic characteristics and nonlinear effects
from dynamic responses, especially if only transient responses are available. Next we demonstrate how to use
Hilbert–Huang transform and sliding-window fitting to extract such information from linear, nonlinear,
stationary, and/or non-stationary signals.

3.1. Intrawave frequency and amplitude modulation

A steady-state time signal uðtÞ having a major frequency O and an amplitude modulating at a frequency o
can be presented as
 
uðtÞ ¼ ð1 þ  cos otÞ cos Ot ¼ cos Ot þ cosðO þ oÞt þ cosðO  oÞt, (21)
2 2
where  is a small parameter. It reveals that the Fourier spectrum of uðtÞ from discrete Fourier transform
analysis (see Eq. (5b)) will consist a major harmonic (at O) with two small harmonics (at O  o) close to it. We
_ Moreover, a steady-state
note that, if o ¼ 0:5O, uðtÞ has a period-doubled trajectory on the phase plane u  u.
time signal uðtÞ having a major frequency O modulating at a frequency o can be presented as
uðtÞ ¼ cosðOt þ  cos otÞ  ð1  142 Þ cos Ot  ð12  16
1 3
 Þ½sinðO þ oÞt þ sinðO  oÞt
 182 ½cosðO þ 2oÞt þ cosðO  2oÞt þ 48
1 3
 ½sinðO þ 3oÞt þ sinðO  3oÞt,

^  dðOt þ  cos otÞ ¼ O  o sin ot,

O (22)
dt
 ARTICLE IN PRESS
534 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

where Taylor’s expansion was performed by treating  cos ot as a small variable. Eq. (22) shows that, if o5O,
the Fourier spectrum of uðtÞ consists of several small, uniformly spaced components around O. If o5O and
O=o is an irrational number, it is a quasi-periodic motion with a period of infinity. Then, the Fourier spectrum
of uðtÞ contains many extra harmonics due to uð0ÞauðTÞ, and one may erroneously identify uðtÞ to be chaotic.
Eqs. (21) and (22) clearly show that a distorted harmonic consists of many regular harmonics of small
amplitudes.
To show intrawave amplitude- and frequency-modulation we consider the following function:
uðtÞ ¼ ð1 þ 0:1 cos oa tÞ cosðOt þ a sin op tÞ þ snðtÞ, (23)
where O ¼ 2p (i.e., 1 Hz), oa ¼ 0:1O, op ¼ 0:3 O, and a ¼ 0:5. s ¼ 0:004 and Dt ¼ 0:02 are used. (Dt ¼ 0:02 is
used in all presented examples, except otherwise stated.) The noise nðiDtÞ, i ¼ 1; . . . ; 1500, are obtained using
the MATLAB command randn and are normally distributed random numbers with a mean zero, a variance
one and a standard deviation one. Fig. 1 shows the results from Hilbert–Huang transform and sliding-window
fitting analyses, where the thin broken lines in Figs. 1c–f are the original functions in Eq. (23). Figs. 1c–f show
that the Hilbert spectra from Hilbert–Huang transform and sliding-window fitting will capture the amplitude-
modulation frequency ðoa ¼ 0:1 HzÞ, the variation of amplitude ð¼ 0:1Þ, the phase-modulation frequency
ðop ¼ 0:3 HzÞ, and the variation of frequency ð¼ aop ¼ 0:15 HzÞ. We note that, for this case, the accuracy
of Hilbert spectrum from Hilbert–Huang transform is better than that from sliding-window fitting. It is
obvious that the error patterns are different because Hilbert–Huang transform extracts the highest-frequency
intrinsic mode function first and the lowest-frequency intrinsic mode function last, but the sliding-window
fitting extracts all intrinsic mode functions at the same time. Non-orthogonality of the functions used in Eq.
(14) causes leakage errors in sliding-window fitting analysis and the extracted functions may be different from
intrinsic mode functions. If op ¼ 0:2O, the errors in the sliding-window fitting become smaller because of
function orthogonality.

1.5 1.3 1.3

1 1.2 1.2
ω1 ω1
frequency (Hz)

frequency (Hz)

0.5
1.1 1.1
u (t)

0
1 1
-0.5
-1 0.9 0.9

-1.5 0.8 0.8

0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

1.5 1.2 1.2

c1
1
1.1 1.1
amplitude

0.5
amplitude

A1
A1
ci (t)

0 1 1
-0.5
0.9 0.9
-1
-1.5 0.8 0.8
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 1. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (23) with oa ¼ 0:1O, op ¼ 0:3O, and a ¼ 0:5: (a) uðtÞ, (b) c1
and r1 , (c) frequency o1 of c1 , (d) amplitude A1 of c1 from Hilbert–Huang transform, (e) o1 of C 1 , and (f) A1 of C 1 from sliding-window
fitting.
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 535

The cubic spline fitting (affected by noise) of local extrema causes errors in the Hilbert–Huang transform,
and the most serious errors of Hilbert–Huang transform occur at the two ends and are mainly caused by
discontinuity (due to ci ð0Þaci ðTÞ) and hence Gibbs’ phenomenon induced during the Hilbert transform of
each intrinsic mode function. To reduce Gibbs’ phenomenon in the Hilbert transform of an intrinsic mode

1.5 1.5
c1
1 1
c2

0.5 0.5

ci (t)
u (t)

0 0

-0.5 -0.5

-1 -1

r2
-1.5 -1.5
0 10 20 30 0 10 20 30
time (s) time (s)

Fig. 2. Hilbert–Huang transform of Eq. (23) with itrmax ¼ 5: (a) the first set of upper and lower envelopes and (b) c1 , c2 , and r2 (dots).

15 2.5 2.5

10
frequency (Hz)

2 2
frequency (Hz)

ω1 ω1
5
u (t)

1.5 1.5
0
1 1
-5
ω2 ω2
-10 0.5 0.5
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

6 4 4
r2
4 3.5 3.5
3 3
amplitude

2 A1
amplitude

A1
2.5 2.5
ci (t)

0
2 2
-2
1.5 1.5
c1 A2 A2
-4 1 1
c2
-6 0.5 0.5
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 3. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (24) with oa ¼ 0:1O, op ¼ 0:2O, and a ¼ 0:5: (a) uðtÞ, (b) ci
and r2 , (c) frequencies oi of ci , (d) amplitudes Ai of ci from Hilbert–Huang transform, (e) oi of C i , and (f) Ai of C i from sliding-window
fitting.
 ARTICLE IN PRESS
536 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

function, one can add to the two ends characteristic waves to make the slightly enlarged data begin and end
with zero. However, because it is not a systematic approach, it is not exercised in any of the presented
examples. The c1 in Fig. 1b is obtained using itrmax ¼ 20 (the maximum allowable number of iterations) and
Dv ¼ 0:001. If itrmax ¼ 5 and/or a different set of nðiDtÞ are used, Fig. 2 shows that Hilbert–Huang transform
fails because the noise creates extra local extrema and messes up the cubic spline fitting, and the extracted
functions may be misunderstood as intermittent signals, as shown in Fig. 2b. We also note that, when itrmax
increases, the Gibbs’ phenomenon becomes more serious but more localized at the two ends. Fig. 2b also
shows that, if uðtÞ is really intermittent, one intrinsic mode function could contain two time scales and the
neighboring intrinsic mode functions might contain oscillations of the same scale. However, signals of the
same time scale would never occur at the same location in two different intrinsic mode functions.
Next we perform Hilbert–Huang transform and sliding-window fitting analyses of the following function:
uðtÞ ¼ ð1  0:1 cos oa tÞð3 sinð2OtÞ þ 5 þ 0:1t  0:015t2 Þ
þ 2ð1 þ 0:1 cos oa tÞe0:02t sinðOt þ a cos op tÞ þ snðtÞ, ð24Þ
where O ¼ 2p, oa ¼ 0:1O, op ¼ 0:2O, a ¼ 0:5, s ¼ 0:01, and Dt ¼ 0:02. Fig. 3 shows the results, where the
broken thin lines represent the original functions in Eq. (24). We note that the Hilbert spectra from
Hilbert–Huang transform and sliding-window fitting have about the same level of accuracy, and both well
capture the amplitude-modulation frequency ðoa ¼ 0:1 HzÞ, the variations of amplitudes ð¼ 0:1Ai Þ, the
phase-modulation frequency ðop ¼ 0:2 HzÞ, the variation of frequency ð¼ aop ¼ 0:1 HzÞ, and the low-
frequency transient part (i.e., r2 and C 3 (not shown)). Again, Gibbs’ phenomenon causes errors at the two
ends of Ai and oi from Hilbert–Huang transform analysis.
If uðtÞ ¼ cosðO þ Þt þ cosðO  Þt, it is an amplitude-modulated function with an average amplitude of zero
because uðtÞ ¼ 2 cosðtÞ cosðOtÞ. Unfortunately, because the Hilbert–Huang transform procedure implicitly
confines the amplitudes of ci to be positive, Hilbert–Huang transform will extract from uðtÞ only one

5 1.4 1.4

1.3 1.3
frequency (Hz)

frequency (Hz)

1.2 1.2
u (t)

0 1.1 ω1 1.1 ω1
1 1
0.9 0.9
-5 0.8 0.8
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

5 6 6

5 5
c1
4
amplitude

4
amplitude

A1
ci (t)

0 3 3 A1

2 2
r1 1 1
-5 0 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 4. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (25) with F ¼ 0, a ¼ 1, uð0Þ ¼ 5, and uð0Þ _ ¼ 0: (a) uðtÞ, (b) c1
and r1 , (c) frequency o1 of c1 , (d) amplitude A1 of c1 from Hilbert–Huang transform, (e) o1 of C 1 , and (f) A1 of C 1 from sliding-window
fitting.
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 537

amplitude-modulated intrinsic mode function with a varying frequency at locations where A1 ð40Þ is close to
zero. For such cases, Fourier transform produces the best results because the data are linear and stationary.

3.2. Different nonlinearities

3.2.1. Cubic nonlinearity

We consider the following oscillator with cubic nonlinearity:
u€ þ 0:2u_ þ ð2pÞ2 u þ au3 ¼ F cosðOtÞ. (25)
For a free damped vibration with F ¼ 0, a ¼ 1, uð0Þ ¼ 5, and uð0Þ _ ¼ 0, Fig. 4 shows the results of
Hilbert–Huang transform and sliding-window fitting analyses. The Hilbert spectrum from Hilbert–Huang
transform (i.e., Figs. 4c and d) reveals that the natural frequency o1 is about 1 Hz and the amplitude- and
frequency-modulation frequency op ¼ 2o1 , which indicates the nonlinearity is cubic, as shown by Eq. (2).
Moreover, because o1 increases when the amplitude A1 increases, it is a hardening nonlinearity (i.e., a40)
because o1 ¼ o0 þ 3aA21 =ð8o0 Þ (o0 ¼ 2p) from perturbation analysis [5]. The a can be reasonably estimated
using the perturbation solution as a ¼ 8o0 ðo1  o0 Þ=ð3A21 Þ. We note that the Hilbert spectrum from sliding-
window fitting also indicates op ¼ 2o1 but it is not clear because of leakage errors caused by function non-
orthogonality.
For a forced vibration with F ¼ 5, a ¼ 1, O ¼ p, uð0Þ ¼ 4, and uð0Þ_ ¼ 0, Fig. 5 shows the results from
Hilbert–Huang transform and sliding-window fitting analyses. It is obvious that c1 is the distorted, damped
natural harmonic, c2 is the distorted harmonic caused by the excitation at O ¼ 0:5 Hz and the cubic
nonlinearity, and r2 is the lowest-frequency harmonic representing the trend (i.e., the moving average) of the
signal and is about zero for this case. Figs. 5c and d clearly show that the frequency of c1 modulates at 2o1
when A1 is large, and the frequency of c2 modulates at 1 Hz ð¼ 2OÞ because of the cubic nonlinearity.

4 1.2
1
1
frequency (Hz)

2 ω1
frequency (Hz)

ω1 0.8
0.8
u (t)

0 0.6
0.6 0.4 ω2
-2
ω2 0.2
0.4
-4 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

4 4 4

2 c2 3 3
c1
amplitude

amplitude

A1
ci (t)

0 2 2
A1
-2 1 1
A2 A2
-4 0 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 5. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (25) with F ¼ 5, a ¼ 1, O ¼ p, uð0Þ ¼ 4, and uð0Þ _ ¼ 0: (a)
uðtÞ, (b) ci and r2 , (c) frequencies oi of ci , (d) amplitudes Ai of ci from Hilbert–Huang transform, (e) oi of C i , and (f) Ai of C i from sliding-
window fitting.
 ARTICLE IN PRESS
538 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

Moreover, because o1 decreases when the amplitude A1 increases, it is a softening cubic nonlinearity (i.e.,
ao0). We note that the Hilbert spectrum from sliding-window fitting (see Figs. 5e and f) cannot show the
correct modulation frequency because the fixed value assigned to the o1 in Eq. (14) is obtained from the
Fourier spectrum of discrete Fourier transform analysis and it cannot cope with the significant variation of o1
shown in Fig. 5c. Hence, the non-orthogonality causes leakage errors in sliding-window fitting analysis and the
spectrum is wrong, especially o2 and A2 . The spectra from the wavelet analysis or the discrete Fourier
transform analysis consist of 0.5, 1.0, and 1.5 Hz harmonics and many other higher harmonics caused by
transient effects, instead of just two intrinsic mode functions. It is obvious that Hilbert–Huang transform can
extract transient and steady-state harmonics (distorted or not) of different time scales and reveal nonlinearities
from non-stationary data without dealing with the higher harmonics caused by transient effects.

3.2.2. Quadratic nonlinearity

Next we consider the following oscillator with quadratic nonlinearity:
u€ þ 0:2u_ þ ð2pÞ2 u þ au2 ¼ F cosðOtÞ. (26)
For a free damped vibration with F ¼ 0, a ¼ 1, uð0Þ ¼ 5, and uð0Þ _ ¼ 0, Fig. 6 shows the results of
Hilbert–Huang transform and sliding-window fitting analyses. Fig. 6c reveals that the natural frequency o1 is
about 1 Hz and the frequency-modulation frequency op ¼ o1 , which indicates that the nonlinearity is
quadratic, as shown by Eq. (3). Moreover, it shows that o1 decreases when the amplitude A1 increases, as
predicted by the perturbation solution o1 ¼ o0  5a2 A21 =ð12o30 Þ (o0 ¼ 2p) [5]. The a can be reasonably
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
estimated using the perturbation solution as a ¼ 12o30 ðo0  o1 Þ=ð5A21 Þ. Moreover, the expression of a0 in
Eq. (1b) indicates that a40 because r1 ðtÞo0, as shown in Fig. 6b. We note that the Hilbert spectrum from
sliding-window fitting also shows the softening effect but it does not clearly indicate op ¼ o1 due to leakage
errors caused by non-orthogonality of the functions used in Eq. (14).

6 1.05 1.05
4
frequency (Hz)
frequency (Hz)

2 1 1
ω1
u (t)

0 ω1
-2 0.95 0.95
-4
-6 0.9 0.9
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

6 6 6
4 c1 5 5
amplitude

2 4 4
amplitude
ci (t)

0 3 3
-2 2 2
r1 A1 A1
-4 1 1
-6 0 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 6. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (26) with F ¼ 0, a ¼ 1, uð0Þ ¼ 5, and uð0Þ _ ¼ 0: (a) uðtÞ, (b) c1
and r1 , (c) o1 of c1 , (d) A1 of c1 from Hilbert–Huang transform, (e) o1 of C 1 , and (f) A1 of C 1 from sliding-window fitting.
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 539

5 1.2 1.2

1 1

frequency (Hz)
frequency (Hz)
ω1 ω1
0.8 0.8
u (t)

0
0.6 0.6
ω2 ω2
0.4 0.4
-5
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

5 5 5
r2
c1 4 4

amplitude
amplitude

3 3
ci (t)

0
2 2
A1 A1
c2 1 A2 1 A2
-5 0 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 7. Hilbert–Huang transform and sliding-window fitting analyses of Eq. (26) with F ¼ 10, a ¼ 1, O ¼ p, uð0Þ ¼ 5, and uð0Þ _ ¼ 0: (a)
uðtÞ, (b) ci and r2 , (c) oi of ci , (d) Ai of ci from Hilbert–Huang transform, (e) oi of C i , and (f) Ai of C i from sliding-window fitting.

For a forced vibration with F ¼ 10, a ¼ 1, O ¼ p, uð0Þ ¼ 5, and uð0Þ _ ¼ 0, Fig. 7 shows the results of
Hilbert–Huang transform and sliding-window fitting analyses. Fig. 7c shows that c1 is the distorted, damped
natural harmonic, c2 is the distorted harmonic caused by the excitation at O ¼ 0:5 Hz and the quadratic
nonlinearity, and r2 is the lowest-frequency harmonic representing the trend (i.e., the moving average) of the
signal. Fig. 7c clearly shows that, because of the quadratic nonlinearity, the frequency of c1 modulates at o1
when A1 is large. Moreover, because r2 40, it follows from the a0 shown in Eq. (1b) that ao0. We note that
the Hilbert spectrum from sliding-window fitting does not clearly reveal the modulation frequencies due to
leakage errors caused by non-orthogonality.

3.2.3. Higher-order nonlinearities

To show the effects of higher-order nonlinearities, we consider the following oscillator:

u€ þ 0:2u_ þ ð2pÞ2 u þ au5 ¼ 0. (27)

For a free damped vibration with a ¼ 1, uð0Þ ¼ 5, and uð0Þ_ ¼ 0, Fig. 8 shows the results of Hilbert–Huang
transform analysis. The Hilbert spectrum reveals that the natural frequency o1 significantly changes with the
amplitude A1 and the amplitude- and frequency-modulation frequency is 2o1 ðtÞ, like cubic nonlinearity. The
a3 in Eq. (1b) being a function of a5 implies that higher-order odd-power nonlinearities behave like cubic
nonlinearity. Moreover, r1 ðtÞ ¼ 0 and the modulation waves of o1 ðtÞ and A1 ðtÞ are found to be distorted
harmonics (see Figs. 8e and f), which imply that the power of nonlinearity is odd and higher than 3. Because
o1 increases when the amplitude A1 increases, it is a hardening nonlinearity (i.e., a40) because o1 ¼
o0 þ 5aA41 =ð16o0 Þ (o0 ¼ 2p) from perturbation analysis [5]. The a can be estimated using the perturbation
solution as a ¼ 16o0 ðo1  o0 Þ=ð5A41 Þ. We note that the sliding-window fitting analysis cannot work for this
case because it has a wide frequency band, as shown by Fig. 8c.
 ARTICLE IN PRESS
540 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

5 3.5 3

3 2.8 ω1

frequency (Hz)

frequency (Hz)
2.5 2.6
u (t)

0
2 2.4
ω1
1.5 2.2

-5 1 2
0 10 20 30 0 10 20 30 2 2.5 3 3.5 4
time (s) time (s) time (s)

5 5 4.5

c1 4
amplitude

amplitude
3
ci (t)

0 A1 4
2
A1
1

-5 0 3.5
0 10 20 30 0 10 20 30 2 2.5 3 3.5 4
time (s) time (s) time (s)

_ ¼ 0: (a) uðtÞ, (b) c1 and r1 , (c) o1 of c1 , (d) A1 of c1 , (e)

Fig. 8. Hilbert–Huang transform analysis of Eq. (27) with a ¼ 1, uð0Þ ¼ 5, and uð0Þ
o1 of c1 , and (f) A1 of c1 .

3.3. Interwave frequency and amplitude modulation

Next we consider the following quadratically coupled nonlinear oscillator:

u€ 1 þ 0:05u_ 1 þ ð2pÞ2 u1 ¼ 20u1 u2 ,
u€ 2 þ 0:05u_ 2 þ ð4pÞ2 u2 ¼ 20u21 . ð28Þ
For a free damped vibration with u1 ð0Þ ¼ 1, u_ 1 ð0Þ ¼ 0, u2 ð0Þ ¼ 1, and u_ 2 ð0Þ ¼ 0, Figs. 9 and 10 show the
Hilbert–Huang transform analysis of u1 ðtÞ and u2 ðtÞ, respectively. Fig. 9b shows that u1 behaves like
u1 ¼ A1 ðtÞ cos o1 t. Because u21 ¼ 0:5A21 ð1 þ cos 2o1 tÞ, the drift term r1 ðtÞð40Þ of u2 in Fig. 10b is due to this
transient excitation 0:5A21 ðtÞ from u21 . Because Fig. 10b shows that u2 behaves like u21 and hence 20u1 u2 is like a
cubic nonlinearity to u1 (see Eq. (28)), it causes the 2 Hz ð¼ 2o1 Þ modulation of o1 and A1 in Figs. 9c and d.
The 0.2 Hz modulation in Figs. 9d and 10d is caused by the coupling of u1 and u2 and is determined by the
coupling nonlinearities and damping coefficients. If only u1 þ u2 can be measured in the actual situation (see
Eq. (4a)), Fig. 11 shows the results of Hilbert–Huang transform and sliding-window fitting analyses. Fig. 11c
shows that the interwave modulation between u1 ð¼ c2 Þ and u2 ð¼ c1 Þ messes up the distributions of extrema
(see Figs. 11a and 2a) and hence the Hilbert spectrum from Hilbert–Huang transform is less accurate than that
from sliding-window fitting analysis. The sliding-window fitting analysis is more appropriate for extracting
intrinsic mode functions with constant frequencies.

3.4. Discussions

The presented methods of extracting nonlinear phenomena can be used for structural damage detection
because damaged structures often behave nonlinearly even under small vibrations. For example, the opening
and closing of cracks in a structure may cause frequency- and amplitude-modulations and/or localized
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 541

2 1.5

1.4
1
1.3

frequency (Hz)
1.2
u (t)

0 ω1
1.1

1
-1

0.9

-2 0.8
0 10 20 30 0 10 20 30
time (s) time (s)

2 1.6

c1 1.4

1 A1
1.2
amplitude

1
ci (t)

0
0.8

0.6
-1

0.4

-2 0.2
0 10 20 30 0 10 20 30
time (s) time (s)

Fig. 9. Hilbert–Huang transform analysis of uðtÞ ð¼ u1 ðtÞÞ of Eq. (28): (a) u1 ðtÞ, (b) c1 and r1 , (c) o1 of c1 , and (d) A1 of c1 .

transient vibration. In Hilbert–Huang transform analysis, because mik is the average of the upper and lower
_ ik and m
cubic spline envelopes, mik , m € ik are continuous time functions. Consequently, the discontinuities of u,
_ and u€ remain in c1 ðtÞ (see Eq. (7)) and the discontinuities of third-order and other higher-order time
u,
derivatives of u are shared by all extracted ci and rn . Hence, one can process c1 ðtÞ, instead of the original data
uðtÞ, in using structural damage detection methods based on the use of discontinuities of u, u, _ and u.
€ Hence
Hilbert–Huang transform can be used for processing any transient and/or steady-state time signals to extract
time-varying dynamic characteristics to reveal damage in structures. The use of Hilbert–Huang transform for
structural damage detection will be reported separately [14].
The presented signal decomposition method is different from the conventional frequency tracking method
in the literature. The frequency tracking method is usually for on-line estimation of the major frequency of the
incoming signal for some applications that may not have access to blocked data and thus require continuously
updating frequency estimates. The applications include decoding digital information from a frequency-shift
keyed bit stream, demodulation of an FM radio signal, tracking the engine speed of a maneuvering vessel via
acoustic data, pitch-tracking in speech and music, and modeling of vibrato (frequency modulation) in music
performance. The presented method is more for system identification by decomposing a set of recorded block
 ARTICLE IN PRESS
542 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

2 2.5

2.4

1
2.3
ω1

frequency (Hz)
2.2
u (t)

0
2.1

2
-1

1.9

-2 1.8
0 10 20 30 0 10 20 30
time (s) time (s)

2 1.6

1.4

1
c1 1.2
amplitude

1
ci (t)

0
0.8
A1
0.6
-1

0.4

-2 0.2
0 10 20 30 0 10 20 30
time (s) time (s)

Fig. 10. Hilbert–Huang transform analysis of uðtÞ ð¼ u2 ðtÞÞ of Eq. (28): (a) u2 ðtÞ, (b) c1 and r1 , (c) o1 of c1 , and (d) A1 of c1 .

data into components of different time scales and extracting detailed time-varying frequencies and amplitudes
to reveal nonlinear characteristics of the data. Numerically a frequency tracking problem is somewhat more
complicated than a system identification problem because less data are available for accurate estimation.

4. Concluding remarks

Methods of using the Hilbert–Huang transform and a sliding-window fitting method to extract nonlinear
effects from nonlinear/non-stationary time signals are presented. Intrawave amplitude- and frequency-
modulation explains the distortion of harmonic waves by nonlinear effects, and interwave amplitude- and
frequency-modulation explains modal coupling of modes with different time scales. Major nonlinear
phenomena that can be extracted from transient and/or steady-state dynamic responses include quadratic,
cubic, and higher-order nonlinearities, softening and hardening effects, intrawave amplitude- and phase-
modulated motions, distorted harmonic responses under a single-frequency harmonic excitation, interwave
amplitude- and phase-modulated motions, and multiple-mode vibrations caused by internal/external
 ARTICLE IN PRESS
P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544 543

2 6 2.5
5 1

frequency (Hz)
frequency (Hz)
1
4 2
u (t)

0 3
ω1
1.5
2
-1 ω2
1
ω2 1
-2 0
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

2 1.6 1.6
r2
c2 1.4 1.4
1 1.2 A2
1.2 A2

amplitude
amplitude

1 1
ci (t)

0
0.8 0.8

-1 0.6 0.6
c1 0.4 A1 0.4 A1
-2 0.2 0.2
0 10 20 30 0 10 20 30 0 10 20 30
time (s) time (s) time (s)

Fig. 11. Hilbert–Huang transform and sliding-window fitting analyses of uðtÞ ð¼ u1 ðtÞ þ u2 ðtÞÞ of Eq. (28): (a) u1 þ u2 , (b) ci and r2 , (c) oi
of ci , and (d) Ai of ci from Hilbert–Huang transform, (e) oi of ci , and (f) Ai of C i from sliding-window fitting.

resonances. The discontinuity-induced Gibbs’ phenomenon at data ends in Hilbert–Huang transform analysis
needs further study in order to improve the accuracy and robustness.

Acknowledgements

This work is supported by the National Science Foundation through Grants CMS-0120798 and CMS-
0319853 and the NASA Langley Research Center through Grant NAG-1-01037. The support is gratefully
acknowledged.

References

[1] S.W. Doebling, C.R. Farrar, M.B. Prime, D.W. Shevitz, Damage identification and health monitoring of structural and mechanical
systems from changes in their vibration characteristics: a literature review, Report No. LA-13070-MS, Los Alamos National
Laboratory, 1996.
[2] P.F. Pai, B.S. Kim, J.H. Chung, Dynamics-based damage inspection of an aircraft wing panel, Journal of Intelligent Material Systems
and Structures 15 (2004) 803–821.
[3] A.H. Nayfeh, D.T. Mook, Nonlinear Oscillations, Wiley, New York, 1979.
[4] A.H. Nayfeh, P.F. Pai, Linear and Nonlinear Structural Mechanics, Wiley-Interscience, New York, 2004.
[5] A.H. Nayfeh, Introduction to Perturbation Techniques, Wiley-Interscience, New York, 1981.
[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 the Hilbert spectrum for nonlinear and non-stationary time series analysis, Proceedings of the Royal Society of
London Series A—Mathematical Physical and Engineering Sciences 454 (1998) 903–995.
[7] N.E. Huang, Z. Shen, S.R. Long, A new view of nonlinear water waves: the Hilbert spectrum, Annual Reviews of Fluid Mechanics 31
(1999) 417–457.
[8] N.E. Huang, M.C. Wu, S.R. Long, S.S.P. Shen, W. Qu, P. Gloersen, K.L. Fan, A confidence limit for the empirical mode
decomposition and Hilbert spectral analysis, Proceedings of the Royal Society of London Series A—Mathematical Physical and
Engineering Sciences 459 (2003) 2317–2345.
 ARTICLE IN PRESS
544 P. Frank Pai / Journal of Sound and Vibration 307 (2007) 527–544

[9] Z. Wu, N.E. Huang, A study of the characteristics of white noise using the empirical mode decomposition method, Proceedings of the
Royal Society of London Series A—Mathematical Physical and Engineering Sciences 460 (2004) 1597–1611.
[10] N.E. Huang, N.O. Attoh-Okine (Eds.), The Hilbert– Huang Transform in Engineering, CRC Press, Boca Raton, FL, 2005.
[11] E.O. Brigham, The Fast Fourier Transform, Prentice-Hall, Englewood Cliffs, NJ, 1974.
[12] G. Strang, T. Nguyen, Wavelets and Filter Banks, Wellesley-Cambridge Press, Wellesley, MA, 1997.
[13] P. Flandrin, Time– Frequency/Time-Scale Analysis, Academic Press, New York, 1999.
[14] P.F. Pai, L. Huang, Hilbert–Huang decomposition of time signals for structural damage detection, SPIE 11th Annual Symposium on
NDE for Health Monitoring and Diagnostics, San Diego, California, February 26–March 2, 2006.