304 Int. J. Vehicle Noise and Vibration, Vol. 4, No.
4, 2008
Identification of noise modes for automobile
operational modal analysis
Zhang Shouyuan
CAE Department,
Nanjing Automobile Research and Development Center,
No 128, Hongshan Road, Xuanwu District,
Nanjing, Jiangsu 210028, P.R. China
E-mail: [email protected]
Zhang Yimin*, Li He and Wen Bangchun
School of Mechanical Engineering and Automation,
Northeastern University,
No 11, The Third Ally, Wenhua Road, Heping District,
Shenyang, Liaoning 10004, P.R. China
E-mail:
[email protected] E-mail:
[email protected] E-mail:
[email protected] *Corresponding author
Abstract: In rotating machinery, the input excitation does not meet the
requirement of operational modal analysis (OMA). In this article, an effective
and improved method of distinguishing forced harmonic response from
measured vibration signal is presented based on the differences of the statistical
properties of a stochastic signal and a harmonic signal. It makes the OMA
applicable to the situation of existing periodic excitation. In order to prove the
effectiveness of the proposed method, a modal experiment of a steel plate
excited by a motor is performed. Then, an OMA is performed to solve the
vehicle interior noise problem. Combining the projection channel technique
with the advanced method, most noise modes, especially harmonic modes are
successfully separated. Precise modal parameters of the car are obtained in
typical working conditions which provide guidance of the dynamic
modification and show that the harmonic identification method has great
practical value.
Keywords: EFDD; enhanced frequency domain decomposition; harmonic
mode; OMA; operational modal analysis; probability density function; vehicle
noise and vibration.
Reference to this paper should be made as follows: Shouyuan, Z., Yimin, Z.,
He, L. and Bangchun, W. (2008) ‘Identification of noise modes for automobile
operational modal analysis’, Int. J. Vehicle Noise and Vibration, Vol. 4, No. 4,
pp.304–316.
Biographical notes: Zhang Shouyuan works in the Nanjing Automotive
Research and Development Center with responsibilities in the area of NVH
experiment and simulation. He researches mainly on automobile NVH and
corresponding finite element analysis. He has worked on many projects and his
publications include many papers in the areas of vehicle interior noise, active
vibration control and rotor dynamics.
Copyright © 2008 Inderscience Enterprises Ltd.
Identification of noise modes for automobile OMA 305
Zhang Yimin is a Chang Jiang Scholar, specially appointed Professor in
Northeastern University and Dean of School of Mechanical Engineering and
Automation. He visited Hong Kong City University between the year 2001 and
2002. He went to the Ohio State University in 2006 as a visiting Scholar doing
research in the mechanical department. His research interest focuses mainly on
mechanical dynamic design and mechanical reliability. He has published more
than 140 papers and authored or co-authored four books.
Li He is an Associate Professor in the School of Mechanical Engineering and
Automation, Northeastern University. He has presented many High-Tech
Research and Development Programmes of China. His current major interests
are vehicle dynamics, rotor dynamics and tribology/contact mechanics. He has
over 30 publications in these areas.
Wen Bangchun is an Academician in the Chinese Academy of Sciences. He is
an expert in the field of vibration and has made a great contribution to the
development of mechanical dynamic design. He developed a new discipline
‘vibration utilisation engineering’ which researches on how to utilise vibration
to serve humanity.
1 Introduction
Compared with experimental modal analysis (EMA), operational modal analysis (OMA)
is a new way of obtaining modal parameters just from the output response signals. It does
not need the excitation signal to compute the frequency response function. It could be
used to conduct modal test when the equipment is working or in some particular
situations which is impossible to make an excitation. OMA is first used in the field of
civil engineering to get modal parameters of buildings, off-shore platforms, bridges, etc.
In recent years, it is gradually attached importance to in the field of mechanical
engineering and applied in automotive and aeronautic industries.
OMA has conspicuous advantages compared with EMA (Lingmi, Brincker and
Andersen, 2002). It does not need expensive excitation system and corresponding sensors
to get input signals. Boundary condition is no need to be simulated and the modal test is
simplified to response measurement which is more convenient to make. Some special
dynamic characteristics of the whole system which is impossible to get using EMA
method could be acquired in typical working conditions and real vibration circumstances.
Most OMA parameter identification methods so far are advanced on the basis of
assumption that structures are subjected to zero-mean Gaussian white noise excitation
(Ventura and Horyna, 1997). But the demand is usually difficult to be met in the field of
mechanical engineering because of broad existence of rotating machinery. When the
rotating machinery works, they will be subjected to a combination of both periodic and
random excitations. If one of the harmonic components of the periodic excitation is
strong, the response at that frequency will become as big as or even bigger than those at
the real structure modal frequencies (Huang and Lin, 2001). Thus, in the identification
process, some harmonic responses will cause pseudo modes. This kind of pseudo mode is
called ‘harmonic mode’ which could not be separated by current OMA methods
(Brincker, Andersen, and Moller, 2002).
A straightforward way to distinguish harmonic components in response signal is
treating them as nominal modes which has zero damping (Shuwei and Huaiqiao, 2005).
306 Z. Shouyuan et al.
After the identification process, the mode with very small damping could be treated as
harmonic mode. But, the convergent property of this method is poor especially for the
structure with closely spaced modes. Numerical filter could also be used to eliminate
harmonic components from the response signal (Allemang, Brown and Fladung, 1994).
But in fact, the filter is not ideal which will do harm to response caused by stochastic
excitation. Shuwei and Huaiqiao (2005) improved the least-squares complex frequency
domain method to distinguish harmonic components in the response, but the computation
process is extremely complex. In this article, a simple and effective method is presented
based on the differences of the statistical properties of harmonic response and stochastic
structural response. Firstly, the method is validated by a plate trial in which a steel plate
is excited by a motor with eccentric mass. Then, an operational modal experiment is
made on an automobile in several typical working conditions and enhanced frequency
domain decomposition (EFDD) method is used to identify operational modal parameters.
After that the method presented in this article is applied on separating harmonic and
structural modes. Finally, three harmonic modes and eight structural modes are identified.
2 Theoretical background
2.1 Enhanced frequency domain decomposition
Frequency domain decomposition (FDD) method could be used to obtain modal
frequencies and mode shapes from response signal on condition that structure has small
damping ration (usually <5%) and is loaded by Gaussian white noise excitation (Brown
and Zimmerman, 1979). But FDD cannot estimate damping ratios. In order to overcome
the shortage of FDD, EFDD method is advanced. It transfers the singular value data with
corresponding singular vector having enough high modal assurance criteria (MAC) value
back to time domain via inverse Fast Fourier Transform (FFT), which is the process of
getting approximation of correlation function of the single degree of freedom (SDOF)
system (Brincker, Zhang and Andersen, 2000). Then, the modal frequency and damping
ratio can be calculated by the logarithmic decrement technique.
Since at a particular frequency, there is usually a limited number of modes (only one
or two modes) that contribute to most of the response while others can be ignored, the
response power spectral density (PSD) matrix can be written as the following final form
(Mehdi, 2002):
§ d k ȥ k ȥTk d k* ȥ *k ȥ Tk ·
[G yy ( jZ )] ¦ ¨¨
ksub(Z ) ©
jZ Ok
jZ Ok*
¸¸.
¹
(1)
With the kth pole
Ok V k jZd k ,
where sub(Z) is the set of modes that contribute most at the particular frequency, \k
being the kth mode shape, Vk being the kth mode decaying factor, Zdk being the kth
damped natural frequency, dk being a scalar constant.
After the computation of the response PSD matrices, EFDD technique performs the
singular value decomposition (SVD) for each of the response PSD matrices at each
Identification of noise modes for automobile OMA 307
frequency and for each measurement. The SVD of a complex PSD matrix [Gyy ( jZ )] is
the following factorisation:
[G yy ( jZ )] [) ][6][) ]H , (2)
with
[) ]H [) ] [ I ]
[) ] [{M1}{M2 }{M3 }!{Mr }],
where >6@ is a diagonal matrix that contains the real singular values. The decomposition
corresponds to a SDOF identification of the system for each singular value. The number
of non-zero elements in the diagonal of the singular matrix corresponds to the rank of
each PSD matrix. The singular vector ^Mi` stands for the estimation of the mode shape
corresponding to the ith singular value.
Based on the computation above, the mode shape identified by peak picking on the
SVD curves is taken as reference vector. Then, on both sides of the picked frequency a
MAC between the reference vector and the singular vectors corresponding to a certain
frequency is calculated. The MAC value indicates the level of correlation between two
modes (it takes a value between 0 and 1). MAC is defined as follows for two vectors
) and Ȍ :
2
{)}* ,{<}
MAC({)},{<}) . (3)
{)}2 ,{<}2
In terms of the experiment divided into several measurements, if the largest MAC value
of a singular vector is above a user specified MAC rejection level, the corresponding
singular value is included in the identification of the SDOF system. Then, the singular
vectors are averaged to obtain an improved estimation of mode shape. The average is
being weighted by multiplying the singular vectors with their corresponding singular
values which means that the closer the singular vector is to the peak the more weight, it
has on the mode shape estimation.
The power spectrum density function of a SDOF system identified above is
transferred back to time domain via inverse FFT, which is approximation of free decay
function of the SDOF system. The damping ratio can be obtained by calculating
logarithmic decrement:
2 § r ·
Gk In ¨ 0 k ¸, (4)
p ¨ rpk ¸
© ¹
where r0k is the initial value of the free decay function. rpk is the pth extreme, then the
damping ratio is written as:
Gk
[k . (5)
4S 2 G k2
308 Z. Shouyuan et al.
The modal frequency is obtained by determining zero X-axis-crossing times of the free
decay function. If the correlation function decays to a sufficiently small level, the modal
frequency and damping ratio can be well estimated (Andersen et al., 1999; Brincker,
Zhang and Andersen, 2000).
2.2 Statistical properties of stochastic response and harmonic response
According to the central limit theorem, supposing that any independent random variables
X1, X2, }, Xn whose mathematical expectation is E(XK) = P and variance is D(Xk) = V2
have the same distribution. The normal variable of the sum of the stochastic variables is
Yn ¦ n
k 1
X k nP / nV . In terms of any x, the distribution function Fn(x) of Yn is as
follows:
f
1 2
lim Fn ( x) ³ e t /2
dt ) ( x ). (6)
n of
f
2S
Based on the formula above, if any structure is excited by a sufficiently large number of
independent random excitations, the response can be seen as the result of linear
superposition of many random excitations and the probability density function of the
response is close to Gaussian distribution. This is true independently of what is done to
the signal, such as filtering, decimation, etc.
As to the probability density function of a harmonic response, if the stochastic
variables X and Y with the probability density functions fX(x), fY(y) are related by x = h(y),
then the following relation exists:
° f X [h( y )] | h' ( y ) |, D y E
fY ( y ) ® , (7)
°̄0
where hc(y) is the derivative of h(y), D, E is the value range of y. Now considering the
harmonic response y g ( x) a sin(Z x) , the probability density function of x can be
treated as constant fX,. Thus, the probability density function of the harmonic response
could be calculated as
ª1 º
fY ( y ) fX g 1 « arcsin( y / a ) »
¬Z ¼
(8)
1 1
fX .
Z a 1 ( y / a) 2
The formula above shows that the density function of zero-mean harmonic response goes
to infinite as y o a or y o a . It means that there will be two distinct peaks in the
probability density curve. This is true for any frequency as long as the amplitude is
constant and the response remains harmonic.
The theoretical probability function curve of stochastic response and harmonic
response are shown in Figure 1.
Identification of noise modes for automobile OMA 309
Figure 1 Theoretical probability density functions
left: structural response; right: harmonic response.
According to empirical estimation of probability density function, we divide the range of
the observation samples ^xt; t = 1, 2 } N` into k equal intervals supposing it is 'x and
then count the number of data falling into (xi, xi + 'x) supposing it is Ni. so the
probability of a sample in (xi, xi + 'x) can be written as p{x ( xi , xi 'x)} lim Ni / N .
n of
The estimation of probability density can be gained as 'xo0. Because the number of
observation samples is limited in practice, 'x is impossible to approach zero infinitely
when dividing the range. So empirical formula is usually used to determine the number of
intervals.
3 Plate trial
In order to determine the effectiveness of the identification method proposed, a steel plate
with 0.56 m length by 4.42 m width by 0.02 m thick was used to do the operational
modal experiment. The plate was placed horizontally on a tire tube approximating the
free boundary condition. An electrical motor with eccentric mass was fixed in the middle
of the plate to excite the plate periodically. The speed of the motor was kept at
1,410 r min1. The responses were measured using Brüel and Kjær PULSE 3560D multi-
analysis system with 12 Brüel and Kjær 3508-B accelerometers placed symmetrically on
the plate. The sampling rate was 2,048 samples per second and the length of time was
100 sec corresponding to 1,000 cycles of the lowest natural frequency expected. In order
to get the precise modal parameters of the motor-plate system, a traditional modal
experiment was performed with a plastic-tip hammer after the motor was stopped. A
multi-input single-output method was used. The Nyquist frequency was set to 1,600 Hz.
Figure 2 shows the SVD plot of the spectral density matrix of the data. The measurement
had 12 channels of response data which resulted in 12 SVD curves. It can be seen in the
SVD curves that the first three peaks have similar intervals and the second and third
peaks are approximately two or three times the frequency for the first one. These might
be suspected to be harmonics, but cannot be affirmed from the identification result itself.
Compared with the results of the traditional modal experiment, peaks exist in all natural
modal frequencies in the SVD curves. In order to make further identification, the data of
the responses were imported to MATLAB. A Chebyshev band-pass filter was designed
310 Z. Shouyuan et al.
using MATLAB. Compared with other filters, Chebyshev filter has bigger slope in the
transition band which is quite useful in narrow-band situation and less ripples in the stop
band. The filter was designed with central frequency equalling to frequency of each peak
and 3 dB bandwidth as the pass-band width.
The empirical estimation of probability density function for the band-pass filtered
signals was then estimated. A small program written with MATLAB language was used
to do this. Some of the empirical density functions are shown in Figure 3.
Figure 2 Singular value decomposition plot of spectral density matrix of the data in plate trial
(see online version for colours)
Figure 3 Empirical probability density functions for eight identified modes
0.03
0.02
0.01
0
-0.15 -0.1 -0.05 0 0.05 0.1 0.15
23.47Hz
0.03
0.02
0.01
0
-0.01 -0.005 0 0.005 0.01
46.93Hz
Identification of noise modes for automobile OMA 311
Figure 3 Empirical probability density functions for eight identified modes (continued)
0.03
0.02
0.01
0
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
70.4Hz
0.03
0.02
0.01
0
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
93.87Hz
0.03
0.02
0.01
0
-0.1 -0.05 0 0.05 0.1
280.9Hz
0.03
0.02
0.01
0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
600.6Hz
0.03
0.02
0.01
0
-1.5 -1 -0.5 0 0.5 1 1.5
662.8Hz
0.03
0.02
0.01
0
-1 -0.5 0 0.5 1
828.3Hz
312 Z. Shouyuan et al.
It can be seen that the empirical density function for the first four modes has two-peak
distribution. So the first four modes could be identified as harmonic modes caused by
periodic excitation. For the rest modes, the density function is a one-peak distribution and
there are corresponding modes according to the result of traditional modal experiment. So
the rest modes could be identified as real structural modes obtained by OMA. This
experiment indicates that the method proposed in this article is effective and simple in
identifying harmonic mode.
4 Application in practice
Big vehicle interior noise influences the comfort property seriously. In order to prevent
the booming noise caused by modal coupling of car body and air in the cavity, an
operational modal experiment was performed. The modal parameters of the whole car in
real working conditions were obtained. As the engine is a complicated source of
excitation, the vibration caused by explosion pressure in cylinder rotation of crank shaft
movement of valve systems and so on could be both stochastic and periodic. This does
not meet the requirement of performing OMA and may produce harmonic modes in the
process of parameter identification. In this experiment, Brüel and Kjær PULSE 3560-D
multi-analysis system was used. The model of the car and the placement of
accelerometers were shown in Figure 3. In order to cover all the working conditions of
the engine, measurements were made separately from the engine speed of 800 r min1 in
idle condition to 4,500 r min1 at an interval of 500 r min1. The length of time for each
measurement was 250 sec. Fifteen accelerometers were used for measuring the 75 points
in six measurements by roving 13 accelerometers and having two accelerometers fixed as
reference degree of freedom (DOFs) (a DOF means a point and associated direction).
Because the booming noise is usually caused by low-frequency vibration, the
responses within 100 Hz were taken to analyse. The SVD of the spectral density matrix
of the responses is shown in Figure 4.
In this experiment, a large number of response DOFs were measured simultaneously.
The analysis process suffers from estimation of many noise modes compared with the
number of real structural modes. The main reason for this is that many measurement
channels contain the same physical information, but different random errors. A way to
reduce the amount of noise modes is to reduce the number of channels in the actual
estimation process. So the channel projection technique was used to reduce noise modes
caused by random errors. The first step is to calculate the correlation coefficients among
different channels and find the channel that correlates most with all the other channels.
This channel is most likely to contain maximum physical information. Then, the
remaining number of requested projection channels are found by similar search of the
correlation coefficient matrix. In this experiment, seven channels were used and therefore
seven SVD curves were obtained as shown in Figure 5.
After the channel projection, operational modal parameters could be gotten by the
EFDD method. Some modes were suspected as harmonic modes from the SVD curves
and the unsymmetrical modal shapes. In order to make further identification, the
measured data were imported to MATLAB. Using the program mentioned above, the first
eight empirical density functions are shown in Figure 6.
Identification of noise modes for automobile OMA 313
Figure 4 Operational modal analysis model and transducer set-up (see online version for colours)
Figure 5 Singular value decomposition plot of spectral density matrix in idle condition
(see online version for colours)
314 Z. Shouyuan et al.
Figure 6 Empirical probability density functions for the first eight modes
0.01
0
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
17.01Hz
0.05
0
-0.1 -0.05 0 0.05 0.1
23.43Hz
0.02
0
-0.1 -0.05 0 0.05 0.1
26.52Hz
0.01
0.005
0
-0.03 -0.02 -0.01 0 0.01 0.02 0.03
33.72Hz
0.1
0.05
0
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
44.21Hz
0.02
0
-0.025 -0.02 -0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02 0.025
52.9Hz
Identification of noise modes for automobile OMA 315
Figure 6 Empirical probability density functions for the first eight modes (continued)
0.05
0
-6 -4 -2 0 2 4 6
58.86Hz x 10
-3
0.01
0
-4 -3 -2 -1 0 1 2 3 4 5
70.4Hz -3
x 10
Table 1 Structural modes identified for the whole car
Number of mode Natural frequency (Hz) Damping ratio (%)
1 23.43 0.3644
2 33.72 0.6065
3 44.21 2.40
4 58.86 2.422
5 70.40 1.048
6 78.8 1.97
7 90.6 1.88
8 99.3 1.248
It could be seen that three of them have two-peak distribution and can be confirmed as
harmonic modes. The rest of them have a Gaussian distribution and can be confirmed as
real structural modes. The final modal parameters of the whole car identified by EFDD
are shown in Table 1.
5 Conclusions
In this article, the EFDD method for identifying operational modal parameters are briefly
introduced and then the different statistical properties of harmonic response and
stochastic response are presented. Based on the theory above, a new method of
identifying noise mode from the OMA result obtained by EFDD method is advanced. In
order to show the effectiveness of the new method, both operational modal experiment
and traditional modal experiment are performed using a plate-motor system. Then, the
harmonic modes and real structural modes are successfully separated by the proposed
method.
After that the method is put into practical application. In order to control the vehicle
interior noise, OMA is made when a car is excited only by engine in several working
conditions. The modal parameters are identified by EFDD method and the empirical
density functions at each mode are obtained by proposed method. Three harmonic modes
and eight structural modes are differentiated successfully, which have great significance
316 Z. Shouyuan et al.
in dynamic modification of the car and reduction of the interior noise. The successful
application of the presented method indicates that not only could the mode shape and
MAC be used to identify pseudo mode, but also the statistical property of responses could
be used effectively.
Acknowledgements
We would like to express our appreciation to Program for Changjiang Scholars in
University, to High-tech Research and Development Program of China (2007AA04Z442)
and to Liaoning Innovative Research Team in University.
References
Allemang, R.J., Brown, D.L. and Fladung, W. (1994) ‘Modal parameter estimation: a unified
matrix polynomial approach’, Paper presented in the Proceedings of the IMAC, pp.501–514.
Andersen, P., Brincker, R., Peeters, B., De Roeck, G., Hermans, L. and Krämer, C. (1999)
‘Comparison of system identification methods using ambient bridge test data’, Paper presented
in the Proceedings of the 17th International Modal Analysis Conference, Kissimee, Florida.
Brincker, R., Zhang, L. and Andersen, P. (2000) ‘Modal identification form ambient responses
using frequency domain decomposition’, Paper presented in the Proceedings of the 18th
International Modal analysis Conference, San Antonio, Texas, 2000–2002, pp.257–262, 7–10
February.
Brincker, R., Andersen, P. and Moller, N. (2002) ‘An indicator for separation of structural and
harmonic modes in output-only modal testing’, B&K Technical Paper, pp.211–225.
Brown, D.L. and Zimmerman R. (1979) ‘Parameter estimation techniques for modal analysis’, SAE
Paper No.790221.
Huang, C.S. and Lin, H.L. (2001) ‘Modal identification of structures from ambient vibration, free
vibration and seismic response data via a subspace approach’, Earthquake Engineering and
Structural Dynamics, Vol. 30, pp.1857–1878.
Lingmi, Z., Brincker, R. and Andersen, P. (2002) ‘An overview of operational modal analysis:
major development and issues’, B&K Technical Paper, pp.152–161.
Mehdi, B. (2002) ‘Operational modal analysis-another way of doing modal testing’, B&K
Technical Paper, pp.235–324.
Shuwei, D. and Huaiqiao, Y. (2005) Improvement to the Least-Squares Complex Frequency-
Domain Method in Natural Loads Condition, Oriental Institute for Vibration Technology.
Ventura, C.E. and Horyna, T. (1997) ‘Structural assessment by modal analysis in Western Canada’,
Paper presented in the Proceedings of the 15th International Modal Analysis Conference,
Orllando, Florida, Oriando, pp.234–239.
