Sie sind auf Seite 1von 23

50th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference<br>17th AIAA 2009-2651

4 - 7 May 2009, Palm Springs, California

An Improved Hilbert Transform for Nonlinear Vibration


Signal Analysis

Xingjie Fang* and Huageng Luo


GE Global Research Center, Niskayuna, NY 12309

The Hilbert transform (HT), as a powerful signal processing method, has been widely

applied in analyzing nonlinear and nonstationary signals. Currently, the implementation of

HT is realized by the discrete Fourier transform (FT), which, however, is based on the

assumption that the signal to be analyzed is periodic in nature and of infinite length.

Therefore, the conventional HT of nonlinear signals inherits the leakage and distortion

problems introduced by the discrete FT. Although the local maxima interpolation (LMI)

based approach demonstrated a significant improvement to the conventional FT based

approach, it still underestimates the Hilbert amplitude. In this paper, we propose an

improved LMI approach to address the underestimation issue. The HT obtained by the

improved approach is proven to be a very close approximation to the analytic HT.

Benchmark simulations as well as practical example will show that the new approach can

further improve the HT accuracy and consequently identify the dynamic characteristics of a

baseline blade successfully.

Nomenclature
a = Initial amplitude

f = Frequency (Hz)

h0 = Viscous damping characteristic

i = Imaginary unit

t = Time

v = Angular velocity function

*
Mechanical Engineer, Structural Mechanics and Dynamics Laboratory, KWD212B, Niskayuna, NY 12309, Member AIAA and
ASME.

Mechanical Engineer, Structural Mechanics and Dynamics Laboratory, KWD212A, Niskayuna, NY 12309, Member AIAA and
ASME.

American Institute of Aeronautics and Astronautics

Copyright 2009 by General Electric. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
x = Displacement response

~
x = Hilbert transform of x

x = Velocity response

x = Acceleration response

xa = Analytic signal of x

A = Hilbert amplitude (envelope) of the analytic signal xa

A = Approximate Hilbert amplitude

H[ ] = Hilbert transform

P = Cauchy principal value

X(f) = Fourier transform of x

Xa( f ) = Fourier transform of xa

= Damping ratio

= Phase of the analytic signal xa

= Initial phase

= Instantaneous frequency (rad/s)

0 = Undamped natural frequency (rad/s)

I. Introduction
Signal processing has tremendous applications in vibration analysis as it advances our knowledge of the

underlying physics that vibration signatures expose. Through signal processing, structural dynamic characteristics

(e.g., vibration frequency) or parameters (e.g., mass and stiffness) can be identified. The identified parameters can

then be used to construct a model representing the reality. Vibration signal collected from measurements or

obtained from simulations, usually is time-series data, which may have one or more of the following properties: (1)

the total time span is short; (2) the data are nonstationary; and (3) the data represent nonlinear process.1 Clearly,

each of these properties limits our options in selecting a proper signal processing method. Transient vibrations (e.g.,

impulse responses) that are widely used in modal analysis,2,3,4 most likely carry all these properties. Transient

responses can die out quickly due to damping, leaving short and nonstationary data. Also, many mechanical

American Institute of Aeronautics and Astronautics


systems indeed exhibit a certain level of nonlinearity, like stiffness hardening/softening5,6 and amplitude or strain

dependent damping.7,8,9 Hence, an appropriate signal processing method has to tackle these difficulties possessed by

transient responses.

Today, the most widely used technique in frequency domain analysis is Fourier spectral analysis. Although

Fourier analysis provides a valuable tool to examine the globally averaged energy-frequency distribution,10 it has

been proven that the Fourier transform (FT) is strictly restricted to linear system and periodic or stationary data. It is

well known that for periodic and band-limited signals the FT yields true Fourier coefficients when exactly one or

multiple periods are processed.11,12 A phenomenon termed leakage will occur if signal is non-periodic (i.e., not

exactly one or multiple periods). Leakage can result in corruption of the spectral density magnitude.13 In addition,

as the Fourier transform essentially defines the sine and cosine harmonics with time-invariant amplitudes and phases

over the whole time span, it may need many spurious harmonics to represent nonlinear or nonstationary signals.14

Therefore, the distorted Fourier spectral analysis for nonlinear transient response may make little physical sense.

In order to overcome the limitations of Fourier analysis, several time-frequency based signal processing

methods, like the short-time Fourier transform (STFT) and the wavelet transform, have been developed. The STFT

which can be considered as a derivative of FT, consists of pre-windowing a signal around a particular time instant,

calculating its FT, and repeating this process for each further time step.15 Since the window function (e.g., a

Gaussian function) effectively suppresses the signal outside the neighborhood around a time instant, the STFT can

provide a localized time-dependent spectrum. To localize an event precisely in time, the window function should be

narrow. However, a good frequency resolution needs a longer time span (i.e., a wide window). These conflicting

requirements lead to the fact that the STFT can not have simultaneously good resolutions in time and in frequency,

which makes it sometimes insufficient in practical applications. The wavelet transform analysis has been available

only in the last two decades. The essence of wavelet transform is the decomposition of a signal into several

components with different frequency scales by projecting the analyzed signal on a family of zero-mean basic

functions (wavelets). Readers may refer to Refs. 15 and 16 for a detailed mathematical description of the wavelet

transform. Compared with the STFT that only uses a constant window, the wavelet transform is more adaptive to

nonstationary data since it uses narrow windows at high frequencies and wide windows at low frequencies.15 Hence,

the wavelet analysis is very useful in analyzing vibration with gradual frequency change. However, the wavelet

transform is still a linear technique and inherits the same limitation the STFT has, that is, the time and frequency

American Institute of Aeronautics and Astronautics


resolutions can not be achieved at the same accuracy level. Moreover, the result of wavelet analysis is significantly

dependent on the selected basic wavelets and the leakage generated by the limited length of basic wavelets can not

be avoided.

Recently, a new time-frequency signal processing method based on the Hilbert transform (HT) technique, has

attracted extensive attention in the fields of spectral analysis,14,15 damage detection,16 system identification,9,17,18 etc.

The HT can be considered as a filter that simply shifts phase of a signal by 90 degrees. The original signal and its

HT then formulate a complex analytic signal. With this analytic signal, the time-dependent amplitude and phase

information embedded in the original signal can be easily extracted. Since the HT amplitude is an index of vibratory

energy and the phase is related to vibration frequency, such information is very desirable for characterizing dynamic

characteristics of a system. Also, the Hilbert spectrum (i.e., time-frequency-amplitude distribution) can provide the

information regarding the time dependency of a signal, which enables us to gain insights of a nonlinear system.

Although the HT is found to be very effective for nonlinear and nonstationary data, it has its own limitation that the

HT has physical meaning only by applying to mono-component frequency signal. However, most dynamical signals

obtained from engineering applications are not purely sinusoids. To resolve this, Huang et al.,1 proposed an

empirical mode decomposition (EMD) method, which can separate a complex signal with multi-component

frequencies into a series of intrinsic mode functions (IMFs). Each IMF represents a simple oscillatory mode and

admits a well-behaved HT. The integration of the EMD and the HT is now referred to as the Hilbert-Hung

transform (HHT).1,14 A major procedure of EMD is spline curve-fitting and imperfections, such as

overshooting/undershooting fitting1 and end swings,19 are associated with this procedure. Therefore, many

researches on HHT have been focused on the improvement of EMD.19,20,21 Nevertheless, the HHT requires the

conventional HT, which is strongly related to the FT as we will discuss later. Therefore, the HT also inherits

leakage and distortion issues from the FT. For example, in many previous HT studies,15,17,22 non-negligible errors

occur at the beginning and the end of the analyzed signal where leakage and distortion are most significant. Usually,

one can truncate the heavily distorted beginning and end portion of the data, and only analyze the less corrupted

middle portion. However, many signals, such as transient vibrations studies in this paper, may have short time span

(i.e., only few vibration cycles). The truncating method becomes infeasible since important information will be lost.

In addition, the HT analysis involves derivative operations, so the errors caused by leakage and distortion can be

American Institute of Aeronautics and Astronautics


easily amplified through the numerical differentiation. Analysis based on the distorted signal will lead to an

inaccurate interpretation of a system.

In this paper, an improved HT is proposed to address leakage and distortion in the conventional HT. In Ref. 23,

the authors have developed an envelope detection based method called local maxima interpolation (LMI) to

approximate the Hilbert amplitude of the original signal. This approximate Hilbert amplitude enables us to calculate

the HT easily, with much higher accuracy than that based on conventional FT based approach. In theory, the Hilbert

amplitude should always encompass all the data. However, due to the effect of damping, the LMI based approach

may not always encompass all the data. In this study, we develop an updating envelope detection method to obtain a

fully encompassing Hilbert amplitude. The updated Hilbert amplitude is proven to be a very close approximation to

the analytic Hilbert amplitude. The improved HT will show its high effectiveness in analyzing nonlinear vibration

with varying frequency or/and damping.

The outline of the paper is as follows. In Section II the improved HT is described in detail. Comparisons with

the conventional HT and LMI approach are made in Section III to demonstrate the necessity of the improvement.

Section IV shows the identification of dynamic characteristics of a baseline blade using the improved HT based

technique. Main conclusions are drawn in Section V.

II. The Conventional and Improved Hilbert Transform

A. The Conventional Hilbert Transform

Here we consider vibration response x(t ) = ae (t ) v( t )t cos[v(t )t + ] . Its HT ~


x (t ) is defined as1,14

x( )

~ 1
x (t ) = H [ x(t )] = P d (1)
t

Through the HT, the magnitude of ~


x (t ) is the same as that of x(t ) , while the vibration phase is shifted by 90

degrees. In theory, x(t ) and ~


x (t ) form the complex conjugate pair, so we can define the analytic signal xa (t ) as

x (t ) = A(t )e i (t )
x a (t ) = x (t ) + i~ (2)

where the Hilbert amplitude A(t ) = x 2 (t ) + ~


x 2 (t ) , the phase (t ) = tan 1 ( ~
x (t ) / x(t )) , and i = - 1 . Based on the

above definition, we can see that the analytic signal xa (t ) is the best local fit in time-domain of an amplitude- and

American Institute of Aeronautics and Astronautics


phase-varying trigonometric function to the signal x(t ) . The instantaneous frequency (t ) can be described as the

rate of change of the phase (t ) of the analytic signal xa (t ) ,

x (t ) x(t ) ~
x (t ) ~ x (t )
(t ) = (t ) = (3)
A 2 (t )

Unfortunately, the analytic form of the digital signal is not a priori. Furthermore, the neat closed-form analytic

solution to Eq. (1) may not be found. Thus, the HT ~


x (t ) defined by Eq. (1) is often obtained numerically by

utilizing the FT because in frequency domain the analytic signal xa (t ) is a one-sided FT of x(t ) where the negative

frequency contents are removed, the direct component remain unchanged, and the positive frequency contents are

doubled, 24 i.e.,

0 for f < 0

X a ( f ) = X ( f ) for f = 0 (4)
2 X ( f ) for f > 0

So the numerical procedure to calculate xa (t ) is proceeded as follows

FFT Eq.( 4) IFFT


x(t ) X ( f ) X a ( f ) x a (t ) (5)

As we know, the FT is based on the assumption that the signal to be analyzed is periodic in nature and of infinite

length. When the FT is used for nonlinear and nonstationary signal x(t ) , it yields distorted X ( f ) . Therefore, the

subsequent xa (t ) estimated from the distorted X ( f ) could lead to an inaccurate interpretation of the signal.

B. The Improved Hilbert Transform


The Bedrosian theorem25 states that if two functions g (t ) and h(t ) have non-overlapping Fourier spectra and the

frequency range of h(t ) is higher than that of g (t ) , the HT for the product of g (t ) and h(t ) is equal to

H [ g (t )h(t )] = g (t ) H [h(t )] . Therefore, the analytic signal xa (t ) can be written as

x (t ) = ae (t ) v (t ) t cos[v (t )t + ] + iae ( t ) v( t )t sin[v(t )t + ] = A(t )e i[ v (t ) t + ]


x a (t ) = x (t ) + i~ (6)

where A(t ) = x 2 (t ) = ae ( t ) v( t )t .
x 2 (t ) + ~

When x(t ) reaches its local extrema, it yields x (t ) = 0 , i.e.,

ae (t ) v( t ) t (vt vt v) cos[v(t )t + ] ae ( t ) v (t )t (v + vt ) sin[v(t )t + ] = 0

American Institute of Aeronautics and Astronautics


Assuming we are dealing with non-DC signal, i.e., (t ) = (t ) = v + vt is nonzero, we then obtain

vt + (v + vt )
tan[ v (t )t + ] = (7)
v + vt

First consider signals with constant damping (i.e.,  = 0 ). The local extrema appear at

v (t )t + = n tan 1 ( ), n = 0,1,  (8)

and,

2 1
sin 2 [ v(t )t + ] = , cos 2 [v (t )t + ] = (9)
1+ 2 1+ 2

In virtue of Eqs. (6) and (9), we can easily derive A(t ) = 1 + 2 | x (t ) | at the local extrema. If the damping is

small (i.e., << 1 ), it yields A(t ) | x(t ) | , that is, when the damping is small, the local extrema of x(t ) are a close

approximation to the Hilbert amplitude A(t ) . Our later study will confirm that even with varying damping (i.e.,

 0 ), this conclusion can still hold since usually  << v and it has cos 2 [v (t )t + ] 1 at local extrema of x(t ) .

In Ref. 23, a local maxima interpolation (LMI) approach was developed to approximate the Hilbert amplitude by

interpolating the local extrema. The basic procedures of LMI are listed as follows:

1. Local extrema detection. Find out the local extrema of a given signal and flip the local minima to the

positive side as shown in Fig. 1. The flip-over can give a more accurate interpolation in step 2.

2. Envelope interpolation. Approximate the Hilbert amplitude(envelope) through cubic spline interpolation of

the local extrema as shown in Fig. 1.

For signals with constant damping, Eq. (8) indicates when the damping is positive, the local extrema are lagging

behind the exact Hilbert amplitude. While, if the damping is negative, they are leading ahead of the exact solution.

This conclusion is easily seen in Fig. 2(a) for a positive damping ratio and in Fig. 2(b) for a negative damping. In

both figures, the blue solid lines are the signals; the red dot-dash lines are the true envelopes; the black dash lines are

the LMI envelopes; the black circles are the local maxima; and the red squares are the true envelopes.

In theory, the Hilbert amplitude should always encompass all data. However, the approximate Hilbert amplitude

may not encompass the data around local extrema as shown in Fig. 2. This phenomenon can be explained by Eq.

(9), which indicates that due to the effect of damping the approximate Hilbert amplitude is slightly smaller than the

American Institute of Aeronautics and Astronautics


analytic. Since the HT analysis involves derivative operation, it can be very sensitive to this difference though its

small. Therefore, in this paper we develop an updating Hilbert amplitude(envelope) detection method to amend the

difference. This updating Hilbert amplitude detection method is performed as follows:

(1) identify data points not encompassed by the interpolating envelope;

(2) calculate the difference between the identified data and the interpolation (i.e., the difference between the

black dash line and the blue solid line in Fig. (2));

(3) update the local extrema to the new data at which the interpolation and signal have the maximum difference

(i.e., moving the black circle point to the red square point in Fig. (2)); and

(4) interpolate a new envelope through the new local extrema. Clearly, this updated Hilbert amplitude would be

a very close approximation to the analytic Hilbert amplitude as shown in Fig. 3.

After the updating interpolation, the approximate Hilbert amplitude is A (t ) . So, ~


x (t ) can be estimated as

~
x (t ) A 2 (t ) x 2 (t ) or ~
x (t ) A 2 (t ) x 2 (t ) (10)

x (t ) at any time is exclusive, positive or negative. Since the signal considered here is continuous, ~
The sign of ~ x (t )

should change continuously from positive to zero, then to negative, and vice versa. In addition, the phase of ~
x (t )

always lags 90 degrees behind that of x(t ) . Therefore, it is easy to estimate the sign of ~
x (t ) with above properties.

With ~
x (t ) the instantaneous frequency (t ) would be available by calling Eq. (3). Based on the above analysis, we

can see that, in contrast to the conventional HT, the improved HT estimates the Hilbert amplitude first, and then

calculates the HT. Since it does not require the FT and the inverse FT, the improved HT is more efficient than the

conventional HT. The effectiveness of the improved HT will be examined in the following case studies.

American Institute of Aeronautics and Astronautics


Figure 1. Local extrema detection and envelope interpolation in LMI.

(a) (b)
Figure 2. LMI envelope: (a) positive damping; (b) negative damping.

(a) (b)
Figure 3. Improved LMI envelope: (a) positive damping; (b) negative damping.

American Institute of Aeronautics and Astronautics


III. Case Studies
In this section we perform a series of numerical analyses. The leakage and distortion associated with the FT are

illustrated by two benchmark examples. Comparison studies are carried out to demonstrate the significant

improvement of the new HT approach over both the traditional HT and the LMI approach. Also, a system

identification of a single degree-of-freedom (SDOF) nonlinear system is performed.

A. Benchmark Examples
Figure 4(a) shows the spectrum of a sinusoid x(t ) = sin(2 1.204t 7 / 8) by the discrete FT. The

conventional HT is compared with the analytic HT ~


x (t ) = sin(2 1.204t 7 / 8 / 2) in Fig. 4(b) (Note that

x(t ) has incomplete cycles). Due to leakage the spectrum of this simple sinusoid is expanded over a large

frequency range, though actually it only has one frequency line. Clearly, the corresponding Fourier series are

corrupted. The conventional HT based on these corrupted Fourier series has a significant discrepancy from the

analytic HT. Non-negligible errors occur over the whole time span, especially at the beginning and end.

(a) (b)
Figure 4. (a) Fourier spectrum of x(t ) = sin( 2 1.204t 7 / 8) ; (b) HT of x( t ) .

The second example studies a nonlinear transient vibration x(t ) = cos[ 2t + 0.6 sin(2t )]e 0.1t with a time-

dependent frequency 1 + 0.6 cos( 2t ) . The response has exact 8 vibration cycles as shown in Fig. 5(a). Although

the actual frequency ranges between 0.4 and 1.6Hz, the Fourier spectrum in Fig. 5(b) indicates that the frequency

content is spread out with the primary energy at 1, 2, and 3Hz. Such spurious harmonics arise from the fact that the

Fourier analysis always explains data in terms of a superposition of trigonometric functions with time-invariant

American Institute of Aeronautics and Astronautics


amplitudes and phases. Therefore, it usually needs fictitious harmonics to simulate nonsinusoidal signals. The

conventional HT and its corresponding Hilbert amplitude are shown in Figs. (6) and 7(a), respectively. For

comparison purpose, the analytic HT and the analytic Hilbert amplitude are also plotted. Obviously, the

conventional HT based on the distorted FT has a noticeable discrepancy from the analytic HT. Not only the

beginning and the end but also the inward data are corrupted. Moreover, high frequency fluctuations are observed in

the conventional HT as shown in Fig. 7(b) (note that the inlet figure is plotted in a zoom-in scale for a better

illustration). These fluctuations are due to the spurious high frequency harmonics generated by the FT. The mean

relative Hilbert amplitude error defined as the mean of the absolute error between the conventional and the analytic

divided by the analytic value is 4.02%. The error in the conventional HT can be easily magnified through numerical

differentiation, such as taking the derivative of phase to obtain the instantaneous frequency. Figure 8 shows that the

instantaneous frequency identified by the conventional HT ranges from 6.49Hz to 6.46Hz. Clearly, the distorted

HT results in a misinterpretation of the signal.

Figure 5. (a) Nonlinear transient response x(t ) ; (b) Fourier spectrum of x(t ) .

American Institute of Aeronautics and Astronautics


Figure 6. The analytic and conventional HT of x(t ) = cos[ 2t + 0.6 sin(2t )]e 0.1t .

0.005

-0.005

-0.01

-0.015

-0.02

-0.025

-0.03

-0.035

-0.04

7.65 7.7 7.75 7.8 7.85 7.9 7.95

(a) (b)
Figure 7. (a) Hilbert amplitude A(t ) ; (b) Error of the conventional Hilbert amplitude.

Figure 8. The instantaneous frequency identified by the conventional HT.

American Institute of Aeronautics and Astronautics


In the following, we utilize the improved HT to estimate the Hilbert amplitude and the instantaneous frequency.

In contrast to the conventional HT, the improved HT calculates the Hilbert amplitude first. Figure 9 shows the

approximate Hilbert amplitude obtained by the improved HT and its error with respect to the analytic solution.

Comparing Figs. (7) and (9), we can see that the improved HT successfully eliminates the distortion in the

conventional HT. Moreover, the estimated Hilbert amplitude is such a close approximation to the analytic

amplitude that the mean relative Hilbert amplitude error is 2.48 10 6 % only. The HT ~
x (t ) is then calculated by

Eq. (10) using the approximate Hilbert amplitude. Again, the calculated HT with the improved method is accurate

that it almost overlaps with the analytic HT in Fig. 10. With the original signal and the calculated HT, we can easily

compute the instantaneous frequency. The time-frequency distribution of the amplitude, namely the Hilbert

spectrum, is presented in Fig. 11 where the amplitude is contoured in the time-frequency plane. Compared with the

Fourier spectrum that only returns the average amplitude and frequency over the entire time, the Hilbert spectrum

provides the dependency of amplitude and frequency over time, which clearly is very important for analyzing

signals of nonlinear/nonstationary nature. In addition, it truthfully represents the time-dependent frequency

1 + 0.6 cos( 2t ) embedded in the signal. So we can conclude that the new HT method improves the analysis

accuracy significantly.

Figure 9. (a) Hilbert amplitude A(t ) ; (b) Error of the improved Hilbert amplitude.

American Institute of Aeronautics and Astronautics


Figure 10. The analytic and improved HT of x(t ) = cos[ 2t + 0.6 sin(2t )]e 0.1t .

Figure 11. Contour plot of the Hilbert spectrum of x(t ) = cos[ 2t + 0.6 sin(2t )]e 0.1t .

(a) (b)
Figure 12. (a) Updated local extrema for envelope interpolation; (b) Instantaneous frequency identified by
LMI.

American Institute of Aeronautics and Astronautics


In the LMI approach, the local extrema of x(t ) are assumed to be equal to the true envelope. However, due to

the effect of damping, the true envelope is always larger than the local extrema. For example, in Fig. 12(a) the

response of x(t ) at t = 0.999s is a local maximum, which, however, is smaller than the true envelope at t = 0.999s .

The interpolation through the local maximum does not enclose all data. At t = 1s , it has the analytic HT

x (t ) = sin[2t + 0.6 sin(2t )]e 0.1t = 0 , so the response of x(t ) at t = 1s is equal to the true envelope as
~

A(t ) = x 2 (t ) + ~
x 2 (t ) . The improved updating envelope detection successfully corrects the error in LMI approach.

Without this correction, the identified instantaneous frequency shown in Fig. 12(b) has large jumps at the local

extrema as well as the beginning and the end. This example clearly illustrates the necessity of the improvement.

B. System Identification of SDOF Vibration System


In this sub-section, we will show the application of the improved HT for system identification. As we know,

many mechanical systems exhibit amplitude dependent stiffness and damping.5,9 Since the amplitude is varying

over time in transient vibrations, the stiffness and damping are indeed explicit functions of time. Hence, here we

consider free vibration of a SDOF nonlinear system whose equation of motion is governed by

x(t ) + (0.3 + 0.01t 0.29t 2 + 0.08t 3 ) x (t ) + (800 + 7t 10t 2 ) x (t ) = 0 , with x(0) = 2 and x (0) = 0 . Figure 13 shows

the numerical solution of x(t ) obtained by solving the governing equation by 4-5th order Runga-Kutta method

x (t ) , respectively. Clearly, ~
(ode45 in MATLAB), the approximate Hilbert amplitude, and the derived HT ~ x (t )

only has a phase shift with respect to x(t ) . It is of great interest to find the time- and amplitude-dependent

frequency and damping from response x(t ) . The instantaneous undamped natural frequency 0 (t ) and the

instantaneous damping characteristic h0 (t ) are estimated according to formulas:

 / A and h (t ) = A / A  /(2 ) .17 The (t ) and h (t ) identified by the


0 (t ) = 2 + A  /( A ) + 2( A / A) 2 A 0 0 0

conventional HT are compared with the analytic solutions 0 (t ) = 800 + 7t 10t 2 and

h0 (t ) = (0.3 + 0.01t 0.29t 2 + 0.08t 3 ) / 2 in Fig. 14. Apparently, the result has a poor accuracy. If these parameters

are used to construct a model, it would not truthfully represent the reality. Figures 15 and 16 show the identification

by the improved HT with respect to time and amplitude, respectively. Both the frequency and damping

characteristics have a complete agreement with the analytic solutions. Therefore, this improved HT based approach

American Institute of Aeronautics and Astronautics


can enable us to identify the modal parameters (natural frequency, damping, and their dependencies on amplitude)

accurately, hence providing us with physical insights of a complex system.

Figure 13. Free response, its improved HT and the approximate Hilbert amplitude.

Figure 14. Identified undamped natural frequency (a) and damping characteristic (b) by the conventional
HT.

American Institute of Aeronautics and Astronautics


Figure 15. Identified instantaneous undamped natural frequency (a) and damping characteristic (b) by the

improved HT.

Figure 16. Dependency of the instantaneous undamped natural frequency (a) and the instantaneous damping

characteristic (b) on amplitude.

IV. Industrial Application


In practice, many mechanical systems exhibit a certain level of nonlinearity. In order to construct a model

representing the reality accurately, the nonlinear characteristics have to be identified first. Nonlinear system

American Institute of Aeronautics and Astronautics


identification is a challenging task as it is not uncommon that many system dynamic characteristics (e.g., frequency

and damping) are dependent on vibration level (amplitude, acceleration, etc). In this section, we present a practical

example to demonstrate the effectiveness of the improved HT for system identification of a baseline blade vibration.

In this example, the blade is clamp-mounted. The blade is excited at its first resonance by a noncontact electro-

magnetic shaker, and then the excitation is stopped by shutting down the power supply to the shaker. A laser

vibrometer is used to collect the blade displacement, while an accelerometer is attached to collect acceleration.

Figure 17 shows the collected displacement and acceleration of the free vibration, respectively. Here the damping

effect on the first mode is investigated. The identified undamped frequency and damping ratio

( (t ) = h0 (t ) / 0 (t ) ) as functions of time are plotted in Fig. 18. The dependency of vibration and damping on

vibration amplitude as shown in Fig. 19. We can see that the variation of vibration frequency is very small and the

mean of frequency is 75.22Hz, while the mean of damping ratio is 0.047% only. The reason is that the energy loss

mechanism here is mainly due to blade material damping, which usually is very small. However, since the blade

acceleration level is very high, we still can see that damping has a weak nonlinear dependency on acceleration as

shown in Fig. 20. After obtaining the amplitude-dependent natural frequency and damping ratio, we may then use

the equation of motion x(t ) + 2 (t ) 0 (t ) x (t ) + 02 (t ) x(t ) = 0 to predict the blade response with the given initial

conditions. The evolution of response is integrated in small time steps. With an initial amplitude x0 , we can find

the corresponding natural frequency and damping from Fig. 19. Then we can either use an analytical solution of the

equation of motion or apply a standard numerical integration scheme to calculate x in a time step. The natural

frequency and damping ratio for the next time step integration will be updated from Fig. 18. This procedure is

repeated until the total integration time is over. In Fig. 21, the predicted response using the identified characteristics

is compared with the test data. A very good agreement between the prediction and experiment is observed, which

indicates that the improved HT can extract blade dynamic characteristics accurately.

American Institute of Aeronautics and Astronautics


(a) (b)

Figure 17. Blade free response: (a) displacement; (b) acceleration.

Figure 18. Blade undamped frequency (a) and damping ratio (b) over time.

American Institute of Aeronautics and Astronautics


Figure 19. Blade undamped frequency (a) and damping ratio (b) vs. amplitude.

Figure 20. Dependency of blade damping ratio on acceleration level.

American Institute of Aeronautics and Astronautics


Figure 21. Comparison between predicted and test blade response.

V. Conclusion
The Hilbert transform is a powerful signal processing method for nonlinear and nonstationary data.

Nevertheless, the conventional HT implemented via the discrete Fourier transform, inherits the deleterious leakage

and distortion. Since the HT based technique involves derivative operations, the errors caused by leakage and

distortion can be easily amplified. The LMI approach demonstrated a much higher accuracy than the conventional

HT, but it still underestimates the Hilbert amplitude due to the damping effect. In this study, we proposed an

updating envelope detection method to amend the difference between the LMI based Hilbert amplitude and the

analytic Hilbert amplitude. The improved HT is mathematically proven to be a close approximation to the analytic

solution. Numerical benchmark studies illustrate the necessity of the improvement and the effectiveness as well.

The improved HT can be a very valuable approach in system identification as exemplified by successfully

identifying the dynamic characteristics of a baseline blade.

Acknowledgment
The authors would like to acknowledge and thank Robert Naumiec and Walter Smith for the experimental setup

and blade data acquisition.

American Institute of Aeronautics and Astronautics


References
1
Huang, N. E., Zheng, S., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., Yen, N.-C., Tung, C. C., and Liu, H. H., The

empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis, Proc. R. Soc.

London, Ser. A, V. 454, pp. 903-995. 1998.


2
Matsuda, H., Sakiyama, T., Morita, C, and Kawakami, M., Longitudinal impulsive response analysis of variable cross-section

bars, Journal of Sound and Vibration, Vol. 181, pp. 541-551, 1995.
3
Reynolds, R., and Dowell, E., The transient response of structures using asymptotic modal analysis, ASME Journal of

Applied Mechanics, Vol. 65, pp. 258-265, 1998.


4
Se, J., Weui, B., and Wan, S., Improvement of impulse response spectrum and its application, Journal of Sound and

Vibration, Vol. 288, pp. 1223-1239, 2005.


5
Touze, C., Thomas, O., and Chaigne, A., Hardening/softening behaviour in non-linear oscillations of structural systems using

non-linear normal modes, Journal of Sound and Vibration, Vol. 273, pp. 77-101, 2004.
6
DeMartini, B., Rhoads, J., Turner, K., Shaw, S., and Moehlis, J, Linear and nonlinear tuning of parametrically excited MEMS

oscillators, Journal of Microelectromechanical Systems, Vol. 16, pp. 310-318, 2007.


7
Gandhi, F., Influence of nonlinear viscoelastic material characterization on performance of constrained layer damping

treatment, AIAA Journal, Vol. 39, pp. 924-931, 2001.


8
Patsias, S., Saxton, C., and Shipton, M., Hard damping coatings: an experimental procedure for extraction of damping

characteristics and modulus of elasticity, Materials Science and Engineering A, Vol. 370, pp. 412-416, 2004.
9
Fang, X., Luo, H., and Tang, J., Investigation of granular damping in transient vibrations using Hilbert transform based

technique, ASME Journal of Vibration and Acoustics, Vol. 130, 031006, 2008.
10
Titchmarsh, E. C., Introduction to the theory of Fourier integrals, Oxford University Press, Oxford, UK, 1948.
11
Arruda, J., Analysis of non-equally spaced data using a regressive discrete Fourier series, Journal of Sound and Vibration,

Vol. 156, pp. 571-574, 1992.


12
Vanherzeele, J., Guillaume, P., Vanlanduit, S., and Verboven, P., Data reduction using a generalized regressive discrete

Fourier series, Journal of Sound and Vibration, Vol. 298, pp. 1-11, 2006.
13
Alvin, K., Robertson, A., Reich, G., and Park, K., Structural system identification: from reality to models, Computers and

Structures, Vol. 81, pp. 1149-1176, 2003.


14
Zhang, R., Ma, S., Safak, E., and Hartzell, S., Hilbert-Huang transform analysis of dynamic and earthquake motion

recordings, Journal of Engineering Mechanics, Vol. 129, pp. 861-875, 2003.


15
Schlurmann, T., Spectral analysis of nonlinear water waves based on the Hilbert-Huang transformation, ASME Journal of

Offshore Mechanics and Arctic Engineering, Vol. 124, pp. 22-27, 2002.

American Institute of Aeronautics and Astronautics


16
Pines, D., and Salvino, L., Structural health monitoring using empirical mode decomposition and the Hilbert phase, Journal

of Sound and Vibration, Vol. 294, pp. 97-124, 2006.


17
Feldman, M., Non-linear system vibration analysis using Hilbert transform I. free vibration analysis method FREEVIB,

Mechanical Systems and Signal Processing, Vol. 8, pp. 119-127, 1994.


18
Feldman, M., Considering high harmonics for identification of non-linear systems by Hilbert transform, Mechanical Systems

and Signal Processing, Vol. 21, pp. 943-958, 2007.


19
Xun, J., and Yan, S., A revised Hilbert-Huang transformation based on the neural networks and its application in vibration

signal analysis of a deployable structure, Mechanical Systems and Signal Processing, Vol. 22, pp. 1705-1723, 2008.
20
Yang, W., Interpretation of mechanical signals using an improved Hilbert-Huang transform, Mechanical Systems and Signal

Processing, Vol. 22, pp. 1061-1071, 2008.


21
Rato, R., Ortigueira, M., and Batisata, A., On the HHT, its problems, and some solutions, Mechanical Systems and Signal

Processing, Vol. 22, pp. 1374-1394, 2008.


22
Shi, Z. Y., and Law, S. S., Identification of linear time-varying dynamical systems using Hilbert transform and empirical

mode decomposition method, ASME Journal of Applied Mechanics, Vol. 74, pp. 223-230, 2007.
23
Luo, H., Fang, X., and Ertas, B., Hilbert transform and its engineering applications, AIAA Journal, Vol. 47, pp. 923-932,

2009.
24
Marple, S. L., Computing the discrete-time analytical signal via FFT, IEEE Transactions on Signal Processing, Vol. 47,

pp. 2600-2603, 1999.


25
Bedrosian, E., A product theorem for Hilbert transforms, Proc. IEEE, Vol. 51, pp. 868-869, 1963.

American Institute of Aeronautics and Astronautics

Das könnte Ihnen auch gefallen