Asme 125 (Vibration & Acoustic) PDF

Transactions Journal of Vibration
of the ASME and Acoustics
Editor
LAWRENCE A. BERGMAN
Assistant to the Editor
LINDA CONWAY
Past Editors
F. EHRICH
T. CONRY
D. J. INMAN
Published Quarterly by The American Society of Mechanical Engineers
VOLUME 125 NUMBER 3 JULY 2003
TECHNICAL PAPERS
249
Spectrum of High-Frequency Acoustic Noise in Inviscid Liquid-Linear

Approximation for Spherical Waves
L. Likhterov and A. Berman
252
Theory on Pitch Noise and Its Application

Yukio Nakajima
257
Investigation of the Sound Transmission into an Advanced Grid-Stiffened

Structure
Jeffrey S. Vipperman, Deyu Li, Ilya Avdeev, and Steven A. Lane
267
Optimal Stiffener Design for Interior Sound Reduction Using a Topology

Optimization Based Approach
Jianhui Luo and Hae Chang Gea
274
New Evaluation Method on Gear Dynamics Using Continuous and

Discrete Wavelet Transforms
Yuji Ohue and Akira Yoshida
282
A Stochastic Model for Simulation and Diagnostics of Rolling Element

Bearings With Localized Faults
J. Antoni and R. B. Randall
290
Nonlinear Dynamic Analysis of a Rotor Shaft System With Viscoelastically

Supported Bearings
Nabeel Shabaneh and Jean W. Zu
299
Nonlinear Parameter Estimation in Rotor-Bearing System Using Volterra

Series and Method of Harmonic Probing
Animesh Chatterjee and Nalinaksh S. Vyas
307
Effect of Thrust Magnetic Bearing on Stability and Bifurcation of a

Flexible Rotor Active Magnetic Bearing System
Y. S. Ho, H. Liu, and L. Yu
317
Manager, Journals
JOAN MERANZE
Torsional Vibration Analysis of Complicated Multi-Branched Shafting

Systems by Modal Synthesis Method
Chun-Ping Zou, Duan-Shi Chen, and Hong-Xing Hua
324
Production Coordinator
RAY RAMONAS
Study on the Dynamics of a Rotor in a Maneuvering Aircraft

Fusheng Lin and Guang Meng
328
Control of Self-Excited Vibration of a Rotor System With Active Gas

Bearings
Jinhao Qiu, Junji Tani, and Taekyu Kwon
335
Vibration of Flex Circuits in Hard Disk Drives

J. A. Wickert
343
Vibration Control of a Traveling Suspended System Using Wave

Absorbing Control
M. Saigo, K. Tani, and H. Usui
351
Dynamic Stiffness Formulation and Its Application for a Combined Beam

and a Two Degree-of-Freedom System
J. R. Banerjee
359
Effect of Damping on Asymmetric Systems

Paolo Gallina
DESIGN ENGINEERING DIVISION

Associate Editors
J. CUSUMANO 2003
S. DYKE 2005
G. FLOWERS 2005
M. I. FRISWELL 2004
J. GINSBERG 2005
R. P. S. HAN 2003
J. MAIN 2005
D. QUINN 2005
R. PARKER 2005
C. H. TAN 2004
J. WICKERT 2003
NOISE CONTROL AND ACOUSTICS
DIVISION
R. KELTIE 2004
R. OHAYON 2003
BOARD ON COMMUNICATIONS
Chair and Vice-President
OZDEN OCHOA
OFFICERS OF THE ASME
President, REGINALD VACHON
Exec. Director
VIRGIL R. CARTER
Treasurer
R. E. NICKELL
PUBLISHING STAFF
Managing Director, Engineering
THOMAS G. LOUGHLIN
Director, Technical Publ.
PHILIP DI VIETRO
Production Assistant
MARISOL ANDINO
Transactions of the ASME, Journal of Vibration and
Acoustics (ISSN 1048-9002) is published quarterly
(Jan., April, July, Oct.) by The American Society of
Mechanical Engineers, Three Park Avenue, New York, NY
10016. Periodicals postage paid at New York, NY and
additional mailing offices. POSTMASTER: Send address
changes to Transactions of the ASME, Journal of
Vibration and Acoustics, c/o THE AMERICAN SOCIETY OF
MECHANICAL ENGINEERS, 22 Law Drive, Box 2300,
Fairfield, NJ 07007-2300. CHANGES OF ADDRESS must be
received at Society headquarters seven weeks before
they are to be effective. Please send
old label and new address.
STATEMENT from By-Laws. The Society shall not be
responsible for statements or opinions advanced in papers
or ... printed in its publications (B7.1, Par. 3).
COPYRIGHT 2003 by the American Society of Mechanical
Engineers. Authorization to photocopy material for
internal or personal use under circumstances not falling
within the fair use provisions of the Copyright Act is granted
by ASME to libraries and other users registered with
the Copyright Clearance Center (CCC) Transactional
Reporting Service provided that the base fee of $3.00 per
article is paid directly to CCC, 222 Rosewood Dr.,
Danvers, MA 01923. Request for special permission or bulk
copying should be addressed to Reprints/Permission
Department. INDEXED by Applied Mechanics Reviews and
Engineering Information, Inc. Canadian Goods and
Services Tax Registration #126148048
Contents continued on inside back cover
Contents continued
Journal of Vibration and Acoustics
Volume 125, Number 3
365
The Effect of Actuator and Sensor Placement on the Active Control of Rotor Unbalance
Marty E. Johnson, Luiz P. Nascimento, Mary Kasarda, and Chris R. Fuller
374
Robust Control of a 2D Acoustic Enclosure

Hemanshu R. Pota, Ian R. Petersen, and Atul G. Kelkar
384
Active Vibration Suppression With Time Delayed Feedback

Rifat Sipahi and Nejat Olgac
389
Bayesian Model Screening for the Identification of Nonlinear Mechanical Structures

Gaetan Kerschen, Jean-Claude Golinval, and Francois M. Hemez
JULY 2003
TECHNICAL BRIEFS
398
Closed-Form Exact Solution to H Optimization of Dynamic Vibration Absorbers Application to Different

Transfer Functions and Damping Systems
Toshihiko Asami and Osamu Nishihara
405
Determining Stability Boundaries Using Gyroscopic Eigenfunctions

Anthony A. Renshaw
407
Rotor Isolation for Vibration and Noise Reduction

Timothy A. Brungart and Eric T. Riggs
The ASME Journal of Vibration and Acoustics is abstracted and

indexed in the following:
Acoustics Abstracts, Aluminum Industry Abstracts, Applied Science & Technology
Index, AMR Abstracts Database, Aquatic Science and Fisheries Abstracts, Civil
Engineering Abstracts, Compendex (The electronic equivalent of Engineering Index),
Corrosion Abstracts, Current Contents, EEA (Earthquake Engineering Abstracts
Database), Electronics & Communications Abstracts, Engineered Materials Abstracts,
Engineering Index, Enviroline (The electronic equivalent of Environment Abstracts),
Environment Abstracts, Environmental Engineering Abstracts, Environmental Science
and Pollution Management, Excerpta Medica, Fluidex, Health & Safety Science
Abstracts, INSPEC, Materials Science Citation Index, Mechanical & Transportation
Engineering Abstracts, Mechanical Engineering Abstracts, METADEX (The electronic
equivalent of Metals Abstracts and Alloys Index), Pollution Abstracts, Referativnyi Zhurnal, Shock & Vibration Digest, Solid State and Superconductivity Abstracts, Steels Alert
L. Likhterov
A. Berman
Institutes for Applied Research
and Dept. of Biotechnology Engineering,
Ben-Gurion University of the Negev,
Beer-Sheva, 84105, Israel
Spectrum of High-Frequency
Acoustic Noise in Inviscid
Liquid-Linear Approximation for
Spherical Waves
The high-frequency asymptotics of the acoustic noise spectrum is considered for the case
of spherically symmetric waves propagating in an unbounded inviscid liquid. Using the
Kirkwood and Bethe hypothesis regarding kinetic enthalpy, the Euler equations, the equation of state in the Taits form and following linearization allow the kinetic enthalpy and
reduced pressure to be obtained. The Fourier transform yields the spectral density of
acoustic energy which proves to be inversely proportional to the square frequency and
decreases approximately by 6 decibels per octave with increase of a frequency.
DOI: 10.1115/1.1570446
Introduction
It is known that the acoustic noise spectrum has some slope in

the high range of frequencies. The typical Knudsen sea-noise
spectral slope is about 5 decibels per octave 1, see p. 336.
Acoustic noise spectra generated by oscillating bubbles have been
studied most completely. Analysis of the acoustic energy generated versus frequency 2 suggests that the 5 decibels per octave
wind dependent ambient noise slopes of the Knudsen curves are
caused by the shorter lifetimes of high-frequency bubbles, rather
than significantly lower peak pressure. The average of several
acoustic spectra from a single energetic spill was shown a slope of
5 decibels per octave over the frequency range up to 8 kHz
according to the same authors. The experimental data cited by
Longuet-Higgins 3 show that deep water bubble noise spectra in
high-frequency range over 10 kHz have the slope of 57
decibels per octave. Pumphrey and Crum 4 have obtained acoustic power spectra for rain drops falling onto a lake in slightly
windy conditions which have the slope approximately 5 decibels per octave. In the spacious review by Prosperetti and Oguz
5, the comparison between measured and computed underwater
noise spectra demonstrates the decrease of sound levels about
5.5 decibels per octave. According to Urick 6, see p. 209, the
spectral slope of a spectrum of deep-sea noise is 5 6 dB/
octave in the frequency band from one up to hundred kilohertz,
and further, the thermal noise leads to the 6 dB/octave increase of
the spectrum level. However, such a rise gives the unlimited
acoustic energy (E ac 0 Sd , where S is the spectral density and
is the frequency. It is of interest to elucidate analytically the
asymptotic behavior of an acoustic spectrum of spherical waves
spreading in unbounded inviscid medium for following comparison with experimental data.
t
r
r
r
with addition of the Taits equation of state for water
pB
p 0 B
0
(1)
Contributed by the Technical Committee on Vibration and Sound for publication

in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received May 2002;
Revised February 2003. Associate Editor: R. F. Keltie.
(3)
2
c
r h
t
r
2
0,
(4)
the Euler equations 1 and 2 can be written according to Cole

10 as
Initial Equations
the motion equation
where B300 MPa3000 atm and n7.

As has been noted by Vogel, Bush and Parlitz 7, the equation
of state given by Rice and Walsh 8 may be used but this equation can not easily be incorporated into the Gilmore 9 model that
has been developed on the basis of the Kirkwood and Bethe hypothesis limited to ongoing waves described in the monograph
by Cole 10.
Related more rigorous and general results can be obtained for
the radiated pressure from a bubble taking into account the heat
conduction in the fluid outside the bubble and the effects of the
liquid compressibility 11. However, the Kirkwood and Bethe
theory gives good results and therefore found wide use for studying the generation and propagation of shock waves. Beside that, it
is assumed that the considered process will be isothermal. In the
analysis that follows, the above-mentioned hypothesis is employed. In terms of the enthalpy, dhdp/ , the sound speed, c 2
p/ , and above-mentioned hypothesis which postulates that
the kinetic enthalpy propagates in medium with the velocity c
:
The initial system of governing equations is the Navier-Stokes

equations which for inviscid medium are reduced to the Euler
equations the spherical case is considered and the radial velocity
is denoted r omitting the index:
1 p
t
r
r
the continuity equation (2)
,
t
r
r
(5)
2 1 h
2
.
r
r
c0 t
(6)
It should be noted that Eq. 4 is not an equivalent to / t(c

) / r (r )0, where is the flow potential, but its use allows the relatively simpler result to be obtained.
According to the Taits equation, the enthalpy, h, can be expressed in following form:
Copyright 2003 by ASME
JULY 2003, Vol. 125 249
c0
p n1/n 1 ,
n1 *
(7)
where c 20 n/ 0 (p 0 B) is the sound speed in undisturbed medium and p pB/p 0 B is the reduced pressure. The den*
sity, , will then be determined as
0 p 1/n .
*
1 h 2
,
r c 20 t
r
p
*
(11)
(12)
(21)
nI h
c 20
n/n1
(22)
(23)
and neglecting p 0 as compared with B, the acoustic pressure is

pB
(13)
Using Eq. 6 once more, Eq. 13 can be rewritten in the following form:
where all nonlinear terms are collected in the right-hand part of

this equation.
The evaluation of the nonlinear terms shows that the first term
contains a multiplier c 4
0 , the second and fourth terms are propor1
tional to c 1
M is the Mach number and the third term
0 M
corresponds to the spherical expansion which can be considered
as sufficiently small at large distances from the source of perturbations. Therefore, these terms may be neglected if the linear
approximation is considered, and the linear equation follows for
determination of the enthalpy:
(15)
Assuming hA(r)e t , where A is a function of r only, the

equation for determination of A(r) is obtained:
C2
sin kr t
e .
r
(24)
p 1 sin kr t
e .
k
r
(25)
Acoustic Spectrum
The Fourier transform of an aperiodic function has the form:
p t
L i e itd ,
(26)
(27)
where
L i
1
2
p t e i t dt
is the complex function of the frequency, .

Assuming that p(t)0 at t0, the integral in Eq. 25
can be calculated making use of the formula 3.351.3 from Gradstein and Ryzhik 13 containing there in a exponent of a power
coefficient n0)
L i K
(16)
n
c 20
The constant C 2 can be determined from the condition that p

p 1 at t0 and r0 taking into account that limr0 sin kr/r
k. Then, the expression for the pressure can be written as
2h 2 h 1 2h 1 h 2 4 h 8 2 2h
2 2
,
r 2 r r c 20 t 2 c 40 t
r c 20 t
r
c0 tr
(14)
d 2 A 2 dA
k 2 A0,
dr 2 r dr
pB
n
1 2 h . . .
p 0 B
c0
2h 2h
1 h 2 4 2 h
2 h
2
.
4
r r c0 tr r r
r
c0 t2
c0 t
2h 2 h 1 2h
0.
r 2 r r c 20 t 2
1 t
e .
r
therefore, in the first approximation,
and Eq. 11 becomes
(20)
Since the enthalpy must remain limited, the constant C 1 has to

be equated to 0. Using the expression for enthalpy 7, the reduced pressure is determined as
2 4 4 2 1 h 2
2 4
,
r
r r
r
c0 t
(19)
and, consequently, the solution of Eq. 15 is
(10)
h h 4 2 h 1 h
2
2 2 2.
r
r
c0 tr r r2
c0 t
2
1
1
cos krC 2 sin kr,
r
r
h C 1 cos krC 2 sin kr
Now, differentiating Eq. 10 with respect to r, differentiating the

continuity equation 6 with respect to t, and subtracting the second from the first, it follows
The square of the continuity equation is
AC 1
(9)
1 h 2
h
2
.
t
r
r
c0 t
whose general solution for k0) is
Therefore, the solution of Eq. 17 can be written in the form:
and this expression can be substituted into the motion equation 5

that gives
(18)
uC 1 cos krC 2 sin kr.
From Eq. 6, the radial derivative of the velocity is:
u k 2 u0,
(8)
Determination of Enthalpy
According to the formula 2.101 from the handbook of Kamke

12, it should be assumed u(r)rA, and Eq. 17 reduces to the
linear equation
e 1i t dtK
1
,
1i
(28)
where k /c 0 .
Multiplying this equation by r, this equation connected with
the Bessels equation becomes:
where the coefficient K1/2 p 1 /k sin kr/r is introduced for

brevity.
The spectral density of an acoustic energy is given by the
modulus square of L(i ):
rA 2A k 2 rA0.
S L i 2,
250 Vol. 125, JULY 2003
(17)
(29)
Transactions of the ASME
therefore, with the accuracy up to numerical multiplier, the spectral density changes inversely proportional to the frequency
square:
S 2 .
(30)
Conclusion
The result obtained shows that a linearization of initial equations gives the decrease of the spectral density with the frequency
growth approximately 10 log 226 decibels per octave for
spherical waves propagating in inviscid medium. This may be
found useful for acoustic signal processing on background of an
ambient noise and separation and classification of acoustic
sources.
References
1 Medwin, H., and Clay, C. S., 1998, Fundamentals of Acoustical Oceanography, Academic Press.
2 Updegraff, G. E., and Anderson, V. C., 1991, Bubble Noise and Wavelet
Spills Recorded 1 m Below the Ocean Surface, J. Acoust. Soc. Am., 86, pp.
2264 2279.
3 Longuet-Higgins, M. S., 1990, Bubble Noise Spectra, J. Acoust. Soc. Am.,

87, pp. 652 661.
4 Pumphrey, H. C., and Crum, L. A., 1990, Bubble Noise Spectra, J. Acoust.
Soc. Am., 87, pp. 142148.
5 Prosperetti, A., and Oguz, H. M., 1993, The Impact of Drops on Liquid
Surface and the Underwater Noise of Rain, Ann. Rev. Fluid Mech, 25, pp.
577 602.
6 Urick, R. J., 1983, Principles of Underwater Sound, 3rd edition, McGraw-Hill
Book Company.
7 Vogel, A., Bush, S., and Parlitz, U., 1996, Shock Wave Emission and Cavitation Bubble Generation by Picosecond and Nanosecond Optical Breakdown
in Water, J. Acoust. Soc. Am., 1001, July, 148 165.
8 Rice, M. H., and Walsh, H. M., 1957, Equation of State of Water to 250
Kilobars, J. Chem. Phys., 26, pp. 824 830.
9 Gilmore, R. F., 1952, The Growth and Collapse of a Spherical Bubble in a
Viscous Compressible Fluid, Calif. Inst. Tech. Rep., 26-4.
10 Cole, R. H., 1948, Underwater Explosions, Princeton U.P., Princeton, NJ.
11 Temkin, S., 1999, Radial Pulsation of a Fluid Sphere in a Sound Wave, J.
Fluid Mech., 380, pp. 138.
12 Kamke, E., 1959, Differentialgleichungen, Losungsmethoden und Losungen,
Vol. 1, Akad. Ver., Leipzig.
13 Gradstein, I. S., and Ryzhik, I. M., 1965, Tables of Integrals, Series and Products, Academic Press, New York.
JULY 2003, Vol. 125 251
Theory on Pitch Noise and Its

Application
Yukio Nakajima
Bridgestone Corporation
3-1-1, Ogawahigashi-Cho, Kodaira-Shi, Tokyo
187-8531, Japan
e-mail: nakaj1-y@bridgestone.co.jp
A theory on pitch noise of a tire has been developed based on the assumption that noise
is approximated by a series of Dirac delta-function. The proposed theory predicts the
existence of special lug shapes where tire noise becomes almost zero. In addition a tire
noise simulation system is developed based on this theory to predict tire noise from CAD
data. The prediction of overall noise of various tires with practical patterns is verified to
be in good agreement with the measurement. DOI: 10.1115/1.1569513
Assume the lug shape can be described by
Introduction
Noise nuisance problems have increased due to the increase of

traffic volume and have become a community noise environment
problem. In order to solve the noise problem, regulations have
become strict and manufacturers have invested in noise reduction
research. Vehicle manufacturers have been particularly successful
for significant noise reduction in engine, exhaustion and transmission systems. Tire manufacturers have also studied a tire noise
reduction 15. A tire, however, is becoming a dominant noise
source in a vehicle system and a tire causes 80 percent of total
vehicle noise in steady rolling and 23 percent in acceleration 6.
Research on tire noise reduction has dealt with tread pattern,
sidewall shape, construction, and material. This paper is related to
the research on tire pitch noise generated from tread pattern. In
various design elements of tread pattern as shown in Fig. 1, the
lateral groove called the lug groove has the biggest effect on tire
pitch noise. Pitch refers to the distance between one groove on a
tires tread and the corresponding point of the groove adjacent to
it along the tires circumference. Pitch noise is the noise generated
by the sequence of grooves. In this paper a theory on pitch noise
is developed to predict the effect of the lug angle on tire pitch
noise and a pattern to decrease tire pitch noise is proposed. Furthermore a tire noise simulation system is developed based on this
theory to predict tire noise from CAD data. The prediction of
overall noise of various tires with a practical pattern is verified to
be in good agreement with the measurement. The proposed theory
and the prediction system can be applicable to not only a tire but
also a timing belt or gear where gear or tooth pitch causes noise.
y i x g i x i .
g i (x) is the function of lug shape and i are coordinates determined by pitch variation. The peak amplitude of tire noise spectrum is reduced by a pitch variation treatment 7. The spectrum
with pitch variation treatment is widely distributed in a frequency
range like the white noise which has the same spectrum amplitude
in every frequency. The tonal noise, hence, can be improved by a
pitch variation treatment.
Substituting Eq. 3 for Eq. 1 and integrating it with respect to
the meridian direction, tire noise F(t) is obtained. By applying a
Fourier transformation, F(t) can be described by noise amplitude
a n at nth harmonics as follows:
F t
f t,x
w x tt x
i1
y i x
t i x
V
(1)
h t tt i dth t i .
(2)

in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Dec. 2001;
Revised Nov. 2002. Associate Editor: R. Keltie.
ae
n
i1
i n t
(4)
2n
.
T
(5)
T is a cycle of rotation. Then a n can be obtained by the following

inverse Fourier transformation:
a n
F t e i n t dt
f t,x e i n t dtdx.
(6)
Substituting Eqs. 1 and 3 for Eq. 6,

a n
e
j1
j /V
w x e i n g j x /V dx
(7)
is obtained. In Eq. 7 the term Nj1 e j /V is related to pitch variai n g j ( x ) /V

dx is related to lug shape.
tion and the term W
0 w(x)e
Assume that w(x) is given by
w x
h(t) is an arbitrary function and t i is included in the range of

integral.
252 Vol. 125, JULY 2003
f t,x dx
where W is half of tire contact width and n is given by
i1 . . . N
where w(x) is noise amplitude in the meridian direction represented by x, y i (x) is the ith lug position in the circumferential
direction measured by a reference point, and V is the velocity. is
Dirac delta-function defined by
Tire Pitch Noise
2.1 Theory on Tire Noise. Tire pitch noise is approximated

by the sequences of the following Dirac delta-function:
(3)
1
Wx x
W
(8)
where and are noise amplitude at the tire center (x0) and
noise amplitude at the edge of tire footprint (xW), respectively.
2.2 Theory on Magic Angle 12. Simplified tire pattern
is modeled by assuming that the lug is a straight line with the
same lug angle . The shape of the lug is described by
g i x tan x
i1 . . . N .
(9)
Furthermore, assume that noise amplitude is the same in the meridian direction, w(x)1. Substituting Eq. 9 for Eq. 7 and
integrating it, the following equation is obtained:
a n 2
2V
nW
tan d n
sin
n tan
2V
(10)
Fig. 1 Tire noise and design elements of tread pattern
where 2 is the Euclidian norm and d n is the term related to the

pitch variation described by
d n
e j /V
j1
(11)
where L(TV) is the circumference length of a tire. This equation expresses the effect of the lug angle on pitch noise. P n ( )
possesses multiple zeros where noise is never generated. We call
this lug angle the magic angle described by
The ratio of noise with lug angle and noise with zero lug angle
is derived by using Eq. 10 as follows:
P n
a n 2
2V
nW
tan
sin
a n 0 2
n W tan
2V
nW
tan
L
nW
tan
L
sin
(12)
iM tan1

iL
nW
i1 .
(13)
2.3 Theory on Magic Shape. If is not equal to in Eq.

8, the optimized lug shape, which minimizes pitch noise, is not
straight anymore. Hence, the optimized lug shape called the magic
shape must be determined by the optimization technique 8. Assume that all lug shapes g j (x) are the same in all pitches and
g j (x) is tangent to 2M second magic angle at tire center. We
approximate g j (x) by
Fig. 2 Magic shape
JULY 2003, Vol. 125 253
Fig. 3 Prediction of pitch noise of a tire with practical pattern
g j x g x x tan 2M
bx
i2
(14)
where P is the order of polynomials.

The design variables in the optimization are b i which are the
parameters to express lug shape in Eq. 14. The objective function is to minimize tire pitch noise. Lets consider the example
where W, L, n, P, and are 150 mm, 2000 mm, 50, 5, 1.0 and
0.5, respectively. In this example noise amplitude at the center
area is two times larger than that at the shoulder area. After applying various optimization techniques, the sequential quadratic
programming generated the best solution in this example.
The optimized lug shape is shown by the dotted line in Fig. 2
and the magic angle is shown by solid line as a reference. The
optimized lug shape is different from the magic angle ( 2M
46.8 degrees) and the tangential slope of the optimized lug becomes smaller at the shoulder area than at the center area. This is
because the noise amplitude at the shoulder area is small and the
amplitude needs to be increased by decreasing the lug angle at the
shoulder to match the amplitude at the center area. Hence, the
magic shape depends on the value of , and the footprint shape.
The amplitude of noise with the magic shape is one hundredth of
the noise with the magic angle.
Noise Prediction System of Practical Pattern
The noise prediction system of a tire with the practical pattern

has been developed by calculating f (t,x) in Eq. 1 from CAD
data 2,9. Considering that tire noise with the Dirac deltafunction is generated just when the lug rolls in and out of ground,
f (t,x) can be easily calculated from CAD drawing as shown in
Fig. 3. In order to consider the relationship between tire pattern
and the edge of footprint in a rolling tire, it is convenient to move
the edge of the footprint in the circumferential direction on the tire
pattern. Assume that we observe a rolling tire at each time interval
of t. When tire velocity is V, the edge of the footprint moves on
the tire pattern by the length Vt. If noise amplitude is the same
in the meridian direction, noise amplitude is proportional to x
which is the lateral component of the length of intersection points
generated by the lug edge and the footprint edges at t and t
t. If noise amplitude has the distribution in the meridian di254 Vol. 125, JULY 2003
rection described by w(x), noise amplitude is proportional to

w(x)x. By summing up every pattern edges, tire noise in time
domain can be described by the right graph of Fig. 3. In the case
of a small lug angle noise becomes large, since noise amplitude
becomes large in the time domain due to large x, and vice versa.
Experiment and Discussion
Tire noise is measured in an anechoic room by a 3-m indoor

drum the surface of which is paved by a safety-walk. A microphone is placed on the tire centerline and the distance from a tire
to the microphone is one meter and 0.25 meter height from drum
surface. The tire is a passenger radial tire 175R/70R13.
In order to verify Eq. 12, tires with lug angles of 0, 25 and 40
degrees measured form the meridian direction are prepared by
hand cut. Lugs are grooved at the center area to eliminate the
effect of contact shape. The number of pitches is 60 and there is
no pitch variation treatment. Measured data are the average of
overall noise at 40, 60, 80 and 100 km/h and indicated by black
circles. The prediction of the proposed theory shows very good
agreement with measurement as shown in Fig. 4. Furthermore, in
order to verify the existence of the magic angle defined by Eq.
13, a few patterns with different lug angles near the first magic
angle 1M 56.3 degree are grooved on a tire. Tire noise is measured near the leading edge and the average of overall noise at 40,
60, 80 and 100 km/h is shown in Fig. 5 by a black circle. The
theory indicated by solid line is in good agreement with measurement and the propose theory is still valid near the magic angle.
Then, the predictability of the tire noise prediction system is
verified by tires with a practical pattern. Tire noise overall with
A-weighted sound level is measured in an anechoic room by a
3-m indoor drum with a safety-walk and tire speed is 100 km/h. A
microphone is placed on the tire centerline and the distance from
a tire to the microphone is one meter and 0.25 meter height from
the drum surface. Tires are passenger radial tires with various
sizes from 165R13 to 225/55R15. The following equation is used
for the prediction:
200
prediction
a A
n1
n 2
(15)
Fig. 4 Prediction of lug angle vs tire pitch noise
Fig. 5 Verification of magic angle measurement
n is the number of harmonics, A n is the parameter for A-weighted

sound level and a n 2 is the amplitude at nth harmonics. In this
prediction w(x) are assumed to be the same in the meridian direction and both the leading and trailing edges. Predicted overall
tire noise is compared with the measurement as shown in Fig. 6.
The prediction is in good agreement with the measurement and
the tire noise prediction system is verified to be effective for pattern design for noise reduction.
In order to improve predictability, a database needs to be established including such data as distribution of noise amplitude in the
meridian direction, noise amplitude at the leading and trailing
edges, relationship between noise and material characteristics, and
relationship between noise and block stiffness, etc.
Fig. 6 Comparison between prediction and measurement
Summary
The theory on pitch noise of a tire has been developed based on

the assumption that noise is approximated by a series of Dirac
delta-function. Proposed theory predicts the effect of lug angle on
tire noise and the existence of special lug angles named magic
angles where tire noise becomes zero. The proposed theory was
verified to quantitatively predict the effect of lug angle on tire
JULY 2003, Vol. 125 255
noise level and the existence of magic angles, where tire noise is
theoretically zero, by the noise measurement of a grooved tire.
The theory was extended to pitch noise where noise amplitude
has distribution in the meridian direction. In this case, the shape of
the lug in minimizing tire noise was no longer a straight line and
became a complicated shape. The optimized lug shape, called the
magic shape, was obtained by using optimization technique. Furthermore the tire noise prediction system was developed in order
to apply this idea to the practical pattern. The prediction was
verified to be in good agreement with the measurement.
The proposed theory and the prediction system can be applicable not only to a tire but also to a timing belt or gear where gear
or tooth pitch causes noise.
Nomenclature
a n 2
An
bi
f (t,x), F(t):
g i (x)
L
n
N
T
ti
V
w(x)
amplitude of noise of nth harmonic number

the parameter for A-weighted sound level
parameter to express lug shape
noise in time domain
function of lug shape
length of tire circumference
harmonic number
number of pitches
cycle of rotation
time when lug at x contacts on ground
velocity
weighting function
256 Vol. 125, JULY 2003
W
x
yi
iM
half of tire contact width or lug width

meridian direction on tire
length measured from a reference point
noise amplitude at tire center
noise amplitude at tire shoulder
position of lug in the circumferential direction
lug angle
ith magic angle
References
1 Nakajima, Y., 1980, Theory on Pitch Noise I, Internal Report, Bridgestone
Corporation.
2 Nakajima, Y., 1981, Theory on Pitch Noise II, Internal Report, Bridgestone Corporation.
3 Nakajima, Y., Inoue, Y., and Ogawa, H., 1993, Application of the Boundary
Element Method and Modal Analysis to Tire Acoustics Problems, Tire Sci.
Technol., 21, p. 66.
4 Koike, H., Oshino, Y., and Tachibana, H., 1998, Identification of Source of
Tire/Road Noise in Japanese, Souonseigyo, 22, p. 11.
5 Mukae, T., et al., 1978, Clarification of Mechanism of Tire Noise Generation
in Japanese, Report of JARI, No. 51.
6 JATMA, 1997, On Noise due to Tire and Road, 6th edition in Japanese.
7 Nakajima, Y., and Abe, A., 2000, Application of Genetic Algorithms of Optimization of Tire Pitch Sequences, Japan Journal of Industrial and Applied
Mathematics, 17, pp. 403.
8 Vanderplaats, G. N., 1984, Numerical Optimization Techniques for Engineering Design with Applications, McGraw-Hill.
9 Yoshida, M., and Takayama, M., 1994, Method of Simulating Tire Tread
Noise and Simulator Therefore, USA Patent #4,838,330.
Jeffrey S. Vipperman
Assistant Professor, Associate Mem. ASME,
e-mail: jsv@pitt.edu
Deyu Li
Investigation of the Sound

Transmission into an Advanced
Grid-Stiffened Structure
Graduate Research Assistant
Ilya Avdeev
John Swanson Graduate Fellow,
Student Mem. ASME
Department of Mechanical Engineering,
University of Pittsburgh,
Pittsburgh, PA 15261
Steven A. Lane
Aerospace Engineer,
Air Force Research Laboratory,
Space Vehicles Directorate,
Kirtland AFB, NM 87117
The noise transmission behavior of an advanced grid-stiffened (AGS) composite structure

has been investigated by combining numerical and experimental methods. Structuralacoustic coupling was found to be light, permitting separate analysis of the structure and
acoustic cavity. Finite element analysis permitted the resonant frequencies of acoustic
cavity and structure to be calculated, which play an important role for noise transmission
through the structure. Acoustic mode shapes permitted internal coincidence frequencies to
be estimated and provided insight into modal pressure distributions, when considering
payload location. Experimental structural and acoustic modal analysis permitted the
resonant frequencies and damping ratios for the structure and cavity to be determined,
which in turn were used to corroborate the FEA model. Finally, direct measurement of the
noise transmission was performed based on noise reduction spectrum (NRS), which is
calculated from spatial averages of the RMS acoustic pressures inside and outside of the
shell. It was found that the NRS was dominated by acoustic resonances, which were
marked by sharp dips in the NRS curve. Internal coincidence of the axial wavenumbers
was also found to be a significant mechanism for noise transmission. External coincidence and ring frequencies were found to provide less of an impact on the overall NRS for
the structure. DOI: 10.1115/1.1569511
Introduction
Novel composite structures that have application for use as expendable launch vehicle fairings are being investigated. Composites are already offering higher strength, less weight, and larger
payload envelopes than current systems. Two examples of innovative composite construction are the Advanced Grid-Stiffened
AGS 1 and Chamber Core designs 2, which are additional
contenders to replace the conventional machined-aluminum isogrid structures. Both are fabricated of filament-wound graphite
fiber that has been preimpregnated with an epoxy binder. The
improvements in weight 40 60% decrease and strength 1,000%
increase along the hoop direction are accompanied by a worsening of the sound transmission loss TL behavior of the structure,
which is governed by the physical structural properties mass,
stiffness, and damping and the geometry. The interior acoustic
levels experienced during launch are significant enough to damage
the sensitive cargo.
Initial studies of AGS structures have focused on static or buckling strength, the impact of defects, effects of damage or repairs,
and rudimentary dynamic behavior 1. Modeling studies on improving the transmission loss in fairings have been conducted for
Minotaur-sized launch vehicles where passive 3 and feedback
active structural-acoustic control ASAC 4,5 were numerically
evaluated. Two earlier experimental studies focused on specific
types of transmission control for the AGS mock-scale prototype,
including evaluating the effectiveness of hybrid structural and
acoustic control inputs 6 as well as positive position feedback
PPF controllers 7, which do not require overly complicated
system models and can offer some stability guarantees. An active
control study on a full-scale model has shown that actuator control
authority and power limitations may prove difficult 8. A more
recent study 9 briefly examined an isotropic, truncated, conical
shell and investigated the effects of cavity resonances and ring
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Feb. 2002;
Revised Nov. 2002. Associate Editor: R. F. Keltie.
frequencies on measured noise transmission. These prior studies

illustrate the need for a detailed examination of the physical
mechanisms of noise transmission for composite fairing
structures.
The Sound, Systems, and Structures Laboratory at the University of Pittsburgh is currently engaged in a more thorough characterization of the vibroacoustic behavior of the mock-scale AGS
rocket fairing. Methods used combine experimental structuralacoustic modal analysis, finite element analysis FEA, direct
measurement of the noise transmission, and an investigation of the
effects of the ring frequency and internal and external coincidence. Thorough characterization of the TL behavior of rocket
fairings is necessary in order to understand the dominant physical
mechanisms of noise transmission for these types of structures.
Such knowledge will aid future efforts to improve the noise transmission loss through geometry or property tuning or through the
use of passive or active control techniques.
This paper will present experimental and companion numerical
FEA study results for the mock-scale AGS rocket fairing system.
Measurements include modal analysis of the structure and acoustic cavity and direct measurement of the noise transmission or
reduction for the structure. A novel multi-level finite element
analysis is developed 10 to predict the decoupled resonant frequencies and mode shapes for the in-vacuo structure and rigid
wall acoustic modes. Coupling between the structure and acoustic
cavity was found to be light, as expected, permitting the separate
analyses. The modal analysis results are then related back to the
transmission behavior of the structure, which is governed by resonance and coincidence behaviors. Organization of the manuscript
is as follows. The next sections contain a general description of
transmission loss and structural-acoustic behavior, comments on
the fairing construction, the results of the experimental and numerical modal analyses of the structure and the acoustic cavity,
measurement of the noise-reduction, which is similar to transmission loss, followed by a summary of the findings.
JULY 2003, Vol. 125 257
Fig. 1 Typical transmission loss spectrum for a flat panel
Acoustic Transmission Loss
SA
The transmission loss TL of a panel is defined as
TL10 log10
W inc
1
10 log10
W trans
mn r o pq r dS
(2)
(1)
where W trans is the acoustic power transmitted through the panel

and W inc is the incident power, and the ratio: W trans /W inc is
referred to as the transmission coefficient. Intuitively, the transmission of power will depend upon the reactive impedance mismatch of the panel with air stiffness and mass as well as any
internal dissipation mechanisms damping. These physical properties of the system in conjunction with the geometry determine
the stiffness-controlled, resonance-controlled, mass-controlled,
and coincidence-controlled regions of the TL spectrum. Figure 1
presents a typical TL curve for a flat panel with these regions
labeled. Good TL is often associated with materials having low
stiffness, high mass, and high damping e.g. lead. The panel stiffness dominates the TL behavior at very low frequencies where
there is little inertial effect. Above this region, is the resonancecontrolled region where the lowest resonance frequencies of the
structure will create a peaked nature to the TL spectrum. At approximately twice the frequency of the lowest resonance, the
mass-controlled region begins and it ends at the critical coincidence frequency. Since modes radiate with greatest efficiency
above the coincidence frequency, a dip is observed in the transmission loss before resuming a mass-law type behavior.
The TL behavior of a fairing differs from that of a panel in two
distinct ways. First, the curvature of the fairing stiffens the structure at low frequencies and produces coupling between the axial
and circumferential modes, and second, an interior acoustic cavity
is created. Therefore, the resonance-controlled region of the fairing will not be governed solely by the structurally dominant resonances, but also by the acoustically dominant cavity resonances.
Previous studies have determined that the fairing transmission
characteristics are most strongly influenced by the acoustically
dominant cavity modes, while structurally dominant modes
are seen to play a lesser role 37,11. The terms structurally
dominant and acoustically dominant are used since the coupling between the in-vacuo structural modes and the rigid-wall
acoustic modes is found to be very light. As a result, separate
experimental and numerical modal analyses of the structure and
the cavity are performed in order to predict the mode shapes of the
structure and cavity. Coupling between the structural modes,
mn (r ), and acoustic cavity modes, o pq (r ), is proportional to
the integral of their product over the shared surface, S, between
them 12:
258 Vol. 125, JULY 2003
where m, n, o, p, q are modal indices and r is a position vector.

Such coupling is typically complex and dominated by a few combinations of modes. In particular, internal coincidence between the
structural and acoustic axial wave numbers can occur, as will be
discussed further in the next section.
Note that transmission measurements based on power Eq. 1
would be very difficult to perform for this structure. For example,
if substitution techniques are performed by placing a known
power source inside the acoustic cavity, the acoustic environment
of the cavity itself would be changed. Also, internal sound intensity measurements are impractical. Therefore, an in-situ method of
measuring the noise transmission into the structure based on noise
reduction NR 1317 or noise reduction spectrum NRS was
used. The noise reduction spectrum NRS was computed from
the spatial average of the exterior, p ext( ) 2 , and internal,
p int( ) 2 , pressure measurements as
NRS10 log10
p ext 2
p int 2
(3)
Although this measurement is more closely related to energy

than power, it has been deemed a good indicator of noise transmission behavior and produces results that are similar to TL
measurements.
Mock-Scale AGS Fairing

A mock-scale fairing was constructed to perform vibroacoustic
experimental analyses. The shell of the fairing forms a tapered
cylinder that is 1.8 mm thick and having end diameters of 61 cm
and 38 cm and a height of 56 cm as shown in Fig. 2. Tridirectional
internal ribs with a cross section of 2.513 mm line the inner
surface of the structure. Two sets crisscross at helix angles that
vary between 6 and 10 degrees as the diameter of the fairing
decreases while the third set of 8 are vertically arranged around
the inner circumference of the cone, as will be illustrated later. A
cylindrical coordinate system (r, ,z) is applied as shown in
Fig. 3 and will be used to describe various positions as well as
the structural and acoustic modes of the system. Other parameters
denoted in Fig. 3 include the height, h, the minor and major
end diameters, 2R 1 and 2R 2 , respectively, and the conical half
angle, .
Construction. The following subsection outlines the construction of the fairing structure by the Air Force Research Laboratory. An automated process has been developed 1 for the AdTransactions of the ASME
Fig. 3 Cylindrical coordinate system and dimensions that describe the structure
phone measurement points were achieved with a boom that was

installed within the acoustic cavity and is capable of traversing
both axially (z 10,50 cm), azimuthally 0,360 degrees,
and radially (r 0,28.5 cm). The average mass density of the
composite material was determined by the water displacement
method to be 1,500 kg/m3, which agrees well with calculations.
The resulting mass per unit area is 3.7 kg/m2.
Fig. 2 Picture of a mock-scale AGS fairing test bed
vanced Grid-Stiffened AGS fairings whereby the skin and ribs

are created using a 5-axis filament winding process that requires
only a single cure cycle. The relative simplicity of the construction compared to sandwich style composite structures allows
lower costs and better quality results. Type IM7 graphite fiber
preimpregnated with 977-2 Toughened Epoxy resin was used to
create the shell and ribs of the fairing. A volume fraction of approximately 60 65% resulted. The presence of voids resulted in
reduced stiffness and higher than expected damping in the structure. Theoretical stiffnesses for the graphite-epoxy composite are:
E 11170 GPa, E 2210 GPa, G 125.5 GPa.
Heavy 1.3 cm thick aluminum plates were installed at the top
and bottom of the fairing see Fig. 2 in order to approximate
clamped, rigid end conditions. The larger plate was still found to
couple strongly to the interior acoustics so a heavy I-beam frame
was added to make it more rigid. Ultimately, the end plates
were modeled in the finite element analysis since they possessed
resonances within the pertinent range of the fairing structural
resonances and were found to mildly couple with the fairing. A
Kenwood KFC-W2000 20.3 cm speaker was installed on the
small end of the structure as shown in Fig. 2. The speaker drives
the fairing cavity through a hole of 18 cm diameter and excites all
acoustic modes well. A restrictor plate with a 2 cm hole allows
approximate monopole excitation of the cavity with the speaker,
but was found to provide unsatisfactory results. Internal microJournal of Vibration and Acoustics
Special
Structural-Acoustic
Frequencies. Certain
structural-acoustic frequencies are important in characterizing the
sound transmission behavior of cylindrical structures 12,18,19,
including the ring frequency, f R , the critical coincidence frequency, f C , and the internal cut-off frequencies, f CO . The ring
frequency occurs when the longitudinal wavelength equals the
circumference and it sets the threshold where the stiffness effects
from the curvature no longer impact the vibrational behavior of a
cylinder. For high frequencies ( f f R ), the vibration and associated acoustic radiation behave much like a flat plate. Transmission
is affected since the internal radiation efficiencies tend to be unity
above f R . Below the ring frequency, the circumferential modes of
the cylindrical structure will occur in pairs rotated 90 of each
other, which are identical for axisymmetric structures. A dip in
noise transmission loss is typically observed in the vicinity of f R
due to the high number of internal coincidence phenomena that
occur 12. Computation of f R is accomplished with the following
formula
f R
cl
,
2R
(4)
where both end radii are used to compute two ring frequencies
and c l E/ (1 2 ) is the longitudinal plate wave speed, where
E is the Youngs modulus, is the density of the structural material, and 12 130.245 is the Poissons ratio for the material. Since the structure is anisotropic, average effective values for
the modulus and thickness were determined with FEA and used to
compute the wave speeds. It was determined from the FEA model
that an average isotropic modulus of E eff47 GPa produces structural natural frequencies that are commensurate with the experimentally observed resonance frequencies. Using the value for
E eff , the two ring frequencies corresponding to the two end radii
(R 1 0.19 m, R 2 0.305 m) are estimated to be: f R1 4,840 Hz
and f R2 3,020 Hz. Data presented in a previous study 9 suggests that truncated conical shells do not show dips at or between
f R1 and f R2 , as expected, but may occur above f R2 .
JULY 2003, Vol. 125 259
Table 1 Cut-off frequency bounds for various acoustic modes

Mode Shape
(pq)( ,r)
f CO (Hz)
(R0.19 m)
f CO (Hz)
(R0.305 m)
10
20
01
30
329
546
686
752
529
876
1,100
1,207
Structural Modal Analysis
The critical coincidence frequency, f C , is where external

coincidence between the lowest structural wavenumber k b and the
acoustic wavenumber, k /c, occurs. Stated another way, the
bending wave speed and the acoustic wave speed are equal. Structural modes below f C are termed slow since their radiation efficiencies tend to be below unity 18 and thus contribute less to the
interior acoustic levels. Conversely fast modes ( f f C ) have
unity radiation efficiencies and thus are more efficient at passing
acoustic energy to the interior. As such, the mass-controlled region
of the TL spectrum typically ends around the coincidence frequency, which can be computed as
f C
c 20 3
tcl
(5)
where c 0 343 m/s is the speed of sound in air, and t1.8 mm is

thickness of the shell. Thus, the approximate coincidence frequency for the fairing is f C 6,230 Hz.
Internal coincidence will also occur between the axial wavenumbers of the acoustic and structural modes. Below f R , an excellent estimate of these coincidence frequencies is provided by
the duct cut-off frequencies 12, which can be computed from the
acoustic radial wavenumber k rpq as
f C,int f CO
k rpq c 0
2
end radii and are given in Table 1. In systems with a significant

axial dimension, the lowest cut-off frequency also determines
whether higher-order duct modes can propagate down the cylinder, which is not an issue for this structure.
(6)
where p is the number of diametral pressure nodes and q is the

number of concentric circular pressure nodes. Note that these cutoff frequencies correspond to the natural frequencies for acoustic
modes with no axial variation (k z 0). The bounds for the first
four internal coincidence frequencies are computed using the two
Many of the structural modes of the tapered fairing resemble

the modes of a uniform cylinder and thus the modes will be identified by indices (m,n) corresponding to the axial z and transverse coordinates, respectively. The axial index represents the
number of half wavelengths along the z dimension, and the transverse coordinate represents the number of full waves or nodal
diameters 20 around the perimeter. Although the stiffeners
mildly prevent true axi-symmetry, most of the circumferential
modes were still predicted by the FEA to occur in pairs rotated
by 90 degrees and with proximal resonant frequencies when below f R .
Finite Element Analysis. Multiple levels of FEA models
were developed to aid in calculating the special structural-acoustic
frequencies and to predict the modal behavior, which is related to
sound transmission. The complexity of the structure is very high
due to the taper along the cylinder axis, the anisotropic properties
of the graphite-epoxy material, and the inclusion of the internal
rib structures. Further, there was also a lack of information about
the fiber orientation within the shell and some dynamic interaction
between the composite and rigid end caps was observed. A
modal homogenization of the composite structure permitted an
effective isotropic elastic modulus to be determined for computation of f C and f R in the previous section.
The most complex structural model includes the outer skin Fig.
4a, the rib stiffeners Fig. 4b, and the aluminum boundary
conditions Figs. 4c & d, including four suspension points of
the large plate see Fig. 2. The outer skin of the fairing was
modeled using SHELL93 Shear deformation shell element with
rotational degree of freedom dof, 8 nodes, 6 dof/node and
BEAM189 first order shear deformation, 3-D, 3 nodes, 6 dof/
node type elements in ANSYS finite element analysis FEA software 21. The heavy aluminum boundary plates were found to
possess some modes in the same frequency range as the shell,
Fig. 4 Details of FEA model components
260 Vol. 125, JULY 2003
Fig. 5 First few structural mode shapes from FEA results
which is why they were included in the model. Isotropic homogeneous media was used to model the shell and rib composites, and
a parametric study determined that an effective stiffness of 47 GPa
produces similar modal behavior to the real system. Dynamic behavior from the model was evaluated by comparing the first eight
natural frequencies with those measured in the experiments. The
novel beam-shell mixed FEA model proved to be computationally
effective and accurate.
Figure 5 shows the first three mode shapes predicted by the
FEA model. Only the shell is shown in Fig. 5 since the end caps
would obstruct the cylinder modes. Both the end and side view of
the lowest frequency mode 0,3 repeated is shown at the top
two panes of Fig. 5. The next two modes: 1,0 and 0,2 are
shown in the bottom two panes of Fig. 5. Similarity is noted
between the predicted fairing modes and those for a uniform cylinder. Note that the lowest frequency bending mode for a cylindrical shell does not correspond to the lowest indices: (m,n)
(0,1). It is interesting to observe that the low-frequency modes
of the structure are predicted to respond globally as a shell, rather
than a collection of smaller panels created by the rib stiffeners.
This may or may not be true, since the unidirectional composite
stiffeners are actually stiffer than the isotropic model. A few

modes occured in the end caps before the first structural mode at
612 Hz, but did not cause appreciable normal displacement in the
shell. The mode shapes and frequencies predicted by FEA are
given in the first two columns of Table 2 below.
Experimental Tap Tests. Structural tap tests were conducted
in order to experimentally determine the structural modal properties. A PCB model 086C03 modally tuned hammer was used to
excite the structure while an array of 8 PCB 352B22 0.5-gram
accelerometers measured the response at various locations. An
imaginary measurement grid consisting of 208 points 16 transverse angles and 13 vertical planes z was created. Three
different force input points at z 0.31,0.38,0.46 m were used.
Instead of measuring all 208 response points, representative arbitrarily selected circumferences and vertical slices were measured
along the fairing in order to try to corroborate the FEA results. A
curve-fit was performed on several of the frequency response
functions FRFs between various inputs and response locations in
order to extract global resonant frequency and damping characteristics to validate the FEA models. The complexity of the system
Table 2 Numerical and experimental structural modal properties.

Predicted Mode
Shape (m,n)(z, )
Predicted
Resonant Freq.
r,FEA Hz)
Measured
Resonant Freq.
r,Exp Hz)
%Error Between
r,FEA & r,Exp
Measured
Damping
%
0,3
0,3
1,0
0,2
0,4
0,2
0,4
615
618
702
761
776
818
842
612
627
678
699
775
787
794
0.5
1.2
3.2
8.1
0.2
3.8
4.5
0.8
0.7
0.6
0.8
1.1
1.0
0.5
JULY 2003, Vol. 125 261
Fig. 6 Representative FRF of structure
precluded fitting experimental data for mode shapes. Columns 3

and 5 in Table 2 contain the resonant frequencies and damping
viscous model, respectively, obtained from the curve-fit of the
FRFs. Column 4 of this table shows the error between predicted
and measured resonant frequencies. Errors range from 0.2 8.1%,
which is considered good for a structure of this complexity. The
damping values 0.8 1.1% are perhaps higher than expected for
this type of material due to the existence of voids in the matrix.
Figure 6 depicts a representative FRF and curve-fit taken at measurement point (z, ,r)(39.5 cm, 56.25, 24.4 cm) and drive
point (z, ,r)(18.9 cm, 56.25, 24.4 cm). Some small resonant
responses are observed around 280, 400, and 470 Hz. The FEA
predicted these modes to be corner flapping modes of the large
end cap and two weak accordion-like modes of the structure, respectively. It is the modes with large normal displacement which
couple well to sound, and thus these modes were not of interest
for this study. Note from Table 2 that the lowest structural frequency is 612 Hz, meaning that the resonance-controlled region is
expected to end around 26121,224 Hz. Thus, the mass controlled region is expected to extend from 1,224 Hz to the critical
frequency of f C 6,230 Hz.
Acoustic Modal Analysis

Similar to the fairing shell, the modes of the internal tapered
acoustic cavity conform well to those for a uniform cylindrical
cavity with closed ends. These modes will also be referenced with
respect to the cylindrical coordinate system given by Fig. 3 and
the indices (o,p,q) or (opq). The first modal index, o, represents
the number of half wavelengths along the z-axis, the second, p,
represents the number of nodal diameters occuring along the
transverse coordinate, and q represents the number of radial
nodes along the r direction in Fig. 3. The acoustic mode shapes
are important for three reasons. First, the modal pressure distributions could be examined for problem modes, when considering
the payload location. More importantly, the predicted mode
shapes can be used to estimate the cut off frequencies f CO for the
truncated cone, which will be bounded by the predictions given in
Table 1. Because internal coincidence occurs at these frequencies,
worse transmission loss would be expected, which will be verified
in measurements section. Also, the structural-acoustic coupling
given by Eq. 2 indicates that the relative shape of the in-vacuo
structural modes and rigid-wall acoustic modes will influence the
noise transmission. However, this relationship is of limited value
due to the complexity of the coupling.
262 Vol. 125, JULY 2003
Finite Element Analysis. The rigid-wall acoustic cavity

model of the fairing was created using a uniform tapered cylindrical boundary that neglected the internal ribs, since they would
insignificantly affect the acoustic modal properties of the system.
FLUID30 elements ANSYS 5.7 with properties for air at standard conditions density of 1.21 kg/m3, sound speed of 343 m/s,
and static pressure of 100 kPa were used to model the cavity.
Mesh convergence was performed and the total resulting number
of elements and degrees of freedom were 1,244 and 4,976, respectively. The four degrees of freedom for each node were three
Cartesian particle velocity components (V x , V y , and V z ) and the
pressure, P. A second FEA model included the speaker that was
used for the experimental modal analysis. It was modeled as a
rigid piston moving mass on a spring-damper spider support.
Since both acoustic FEA models yielded similar results, the simpler model without the speaker was adopted for the study. This
model is also consistent with the NRS measurements setup, which
did not contain the attached speaker. Given the high degree of
ideal nature in the model, the predicted acoustic modes are given
a high degree of confidence, particularly at low frequency where
models are known to generally perform well. Figure 7 shows the
first six predicted acoustic mode shapes. The axial-symmetric
mode shapes are repeated rotated by 90 deg, but the repeats are
omitted for brevity. The experimental modal analysis is presented
in the next section.
Experimental Results. The experimental acoustic modal
analysis was performed by exciting the fairing with the attached
speaker at zr0 shown in bottom of Fig. 2. A total of 260
acoustic measurements were taken with the internal microphones
positioned at combinations of the following coordinates: r
3.81,7.62,11.4,15.2 cm, 15,30,45,60,75,90,105,120,135,
150,165,180 degrees, and across vertical planes at z
10,20,30,40,50 cm. The FRF was computed between the input
to the speaker and each of the four array microphones and thus
included the speaker dynamics, which are much lower frequency
than the first acoustic mode of the fairing cavity. Figure 8 shows a
typical measured FRF between the speaker and an internal microphone located at (z, ,r)(19.5 cm, 337.5 deg, 23 cm. Note
also from the figure that the modes and speaker dynamics have
been identified. The dominance by the fundamental 100 mode at
301 Hz is typical of most all of the measured FRFs. However, the
axial zero modes 010, 020, and 001 at 385, 613, and 788 Hz,
respectively, will be shown in the next section to also be very
important for sound transmission. These three modes are used to
estimate the internal coincidence frequencies, the bounds of which
Fig. 7 First few acoustic mode shapes from FEA results
were given in the first three rows of Table 1. Interestingly, although they are significant for sound transmission, many of these
modes appeared insignificant in the acoustic FRFs. The difference
results from the direct excitation by the speaker for the FRF versus excitation via transmission for the NRS measurement. Previous studies focused mainly on the 100 mode as the primary
contributor to noise transmission 6,7.
A curve-fit was performed on an arbitrary selection of 32 of the
260 FRFs so that modal properties could be extracted. Natural
frequencies and damping ratios for the first eight modes are given
in columns 3 and 5, respectively, of Table 3. The predicted mode
shape and natural frequencies from the FEA analysis are given in
columns 1 and 2 of the same table, along with the error between
predicted and measured natural frequencies in column 4, which
are within 0.15.1% in magnitude. Repeated modes that have

nearly identical natural frequencies are indicated 2 in column 1 of Table 3.
Finally, note that when comparing the measured resonant frequencies of the structure and acoustic cavity in Tables 2 and 3, in
general the same modal frequencies do not show up in both tables,
indicating that indeed the structural acoustic coupling is light.
There are however, two frequencies in the vicinity of 612 and 787
Hz that are given in the tables that appear for both the structure
and the acoustic cavity. Although internal coincidence is also predicted to occur for the acoustic modes at these two frequencies, it
is not between the structural modes that have nearly identical
frequency, since internal coincidence can only occur between
modes of the same transverse wavenumber, k , which is related
Fig. 8 Representative FRF between speaker and cavity microphone
JULY 2003, Vol. 125 263
Table 3 Numerical and experimental acoustic modal properties.

Predicted Mode
Shape (opq)(z r)
Predicted
Resonant Freq.
r,FEA Hz)
Measured
Resonant Freq.
r,Exp Hz)
% Error Between
r,FEA & r,Exp
Measured
Damping
%
100
0102
1102
0202
020
2102
001
1202
313
383
550
617
622
765
787
815
301
385
547
583
613
735
788
840
4.0
0.5
0.5
5.6
1.4
3.8
0.1
3.1
0.29
0.23
0.18
0.22
0.27
0.14
0.20
0.33
to modal index n for the structure and p for the acoustic cavity.
When comparing the modal indices n and p from the first column
of Tables 2 and 3, one observes that (n3p2) at 612 Hz and
(n2p1) at 787 Hz.
Measurement of Sound Transmission

NRS measurements are based upon a spatial average of meansquare pressure measurements from the inside and outside of the
structure as given by Eq. 3. To simulate a diffuse field, four
speakers were arranged around the structure and driven with independent white noise sources. Note that a previous work reported
that directional effects in the radiation and reception patterns do
not occur in finite cylinders 13. A wide bandwidth 020 kHz
for the noise sources was desired due to the small scale of the
structure. The speakers and the structure were suspended 0.7 m
above the floor and the room volume was approximately 300 m3.
An external microphone was traversed to 20 different measurement locations which were 1.5 cm from the surface of the shell.
The 20 locations spanned across five different vertical heights: h
10,20,30,40,50 cm, and four different transverse angles:
0,90,180,270 deg. An internal microphone boom was installed 1.5 cm from the edge of the ribs. This microphone recorded 20 internal measurements which were at the same height
and angles as the exterior microphone. Figure 9 depicts a top view
of the measurement setup. Signals were generated and measured
using a Siglab MC20-84 dynamic signal analyzer. Bruel and Kjaer
4190 microphones with B&K 2669 preamplifiers were used for
the acoustic measurements. Marchand PS-24 power amplifiers
drove the four KLH-9912 speakers to generate the external noise
field. Autospectrum measurements from the inside and outside
were post processed and linearly averaged to determine the NRS.
A convergence study on the spatial sampling was performed by

computing NRS using 1, 2, 4, 8, 16, and 20 sets of measurement
points across the surfaces of the shell. The NRS was mostly converged by 8 points, and the difference between 16 and 20 points
was indistinguishable. Thus, 20 was chosen as a suitable number
of points. Effects of angular and vertical position were also examined. Little variation in NRS was noted between angles when
averaging the 5 vertical measurement heights together. Variation
within the vertical measurement planes was more pronounced,
with the sharpest dips in NRS noted near the end caps, which is
consistent with modal pressure maxima and edge radiation effects.
Figure 10 shows the measured NRS for the structure using a
linear frequency axis from 01000 Hz. Note that vertical dashed
and dotted lines have been added at the frequencies of the acoustic
and structural resonances, respectively, which were obtained from
the structural and acoustic modal analyses. In the low-frequency
stiffness-dominated region of the NRS 100250 Hz, the reduction is seen to range from 34 52 dB, with most frequencies above
40 dB. The dip at very low frequency 40 Hz is attributed to the
Helmholtz frequency for the enclosure, which occurs due to leakage. Two dips likely occured because an access door to the structure had to be opened and closed 20 times to arrange the internal
microphone at each measurement location, which may have created nonuniform leakage each time. Commensurate dips in the
NRS are noted at most all of the acoustic resonant frequencies
e.g. 301 Hz (opq)(100), 385 Hz 010, 547 Hz 110, 613 Hz
020, 735 Hz 210, 788 Hz 001, and 840 Hz 120. A Dip in
the NRS is not noted for the acoustic mode at 583 Hz. Most of the
structurally dominant resonances are seen not to influence the
NRS, presumably due to the relatively high level of damping for
this composite structure. The exceptions are at 612, 787, 794, 845,
Fig. 9 Noise transmission measurement setup
264 Vol. 125, JULY 2003
Fig. 10 Measured noise reduction spectrum showing acoustic and structural modes
and 950 Hz. Each of these frequencies also coincides with an

acoustic mode, so it is hard to decipher whether the acoustic,
structural or both contribute to the NRS. There are three frequencies 612, 787794, and 845 Hz where the noise transmission is
actually negative, indicating an amplification of the noise by the
structure. Modes at 612 and 787 Hz have already been associated
with internal coincidence frequencies. The mode at 845 Hz is
assumed to be associated with the 815 Hz 120 mode from the
FEA, but is also near the 030 mode with a predicted natural
frequency of 840 Hz. Note that since the accuracy of FEA models
deteriorate at higher frequency, the measured 840 Hz frequency
may actually be the 030 mode, which would have internal coin-
cidence as well. Again, as noted when examining Fig. 8, these

modes do not have a particularly strong response in this FRF, but
transmit energy well due to coincidence. The 010 mode was also
observed to be a weak mode in most all of the acoustic FRFs, but
is seen to have a significant dip in the NRS at that frequency 385
Hz, again attesting to the importance of internal coincidence for
NRS. The 794 Hz dip appears to be due to the structural mode at
that frequency.
In Fig. 11, the NRS is replotted with a logarithmic frequency
axis which extends from 100 to 20,000 Hz. The approximate
stiffness-, resonance-, and coincidence-controlled regions of the
NRS are indicated in the figure, as well as the probable contribu-
Fig. 11 Measured noise reduction spectrum NRS showing various

regions of transmission loss
JULY 2003, Vol. 125 265
tion by the ring frequency. Referring back to the section on transmission loss, the stiffness-controlled region will begin at 0 Hz and
end at the first resonance, which is an acoustic cavity resonance at
300 Hz. The resonance-controlled region was predicted to continue through approximately twice the first structural resonance
frequency or 26121,224 Hz. Judging from the figure, the
dips in NRS from the resonances are noted to stop around 1,500
Hz and a general upward trend begins until a wide dip begins at
6,000 Hz. The slope in the mass-controlled region is 28 dB per
decade rather than the classical 20 dB/decade mass law. Since f C
was predicted to be 6,230 Hz, this dip is attributed to external
coincidence. The width of the dip is rather large 6,00012,000
Hz, which may be attributed to effects of the stiffeners and/or
ring frequency effects 12. Regardless, the dip is inconsequential
when compared to the levels at the singularities in the NRS that
occur at internal coincidence frequencies and in the region around
where the resonance- and mass- controlled regions meet.
Note that to improve the noise transmission behavior will
largely require the acoustically dominant modes and/or internal
coincidence phenomena to be controlled. Internal absorptive treatments are limited in rocket fairings due to the volume requirements of the payload. The NRS could perhaps also be improved at
612, 787, 794, 845, and 950 Hz by adding damping treatments or
control that targets these structural modes.
Conclusions
The noise transmission behavior for a novel, advanced gridstiffened AGS composite structure has been investigated. This
material is being considered by the US Air Force for constructing
expendable space launch vehicles. A combined numerical/
experimental approach was used to examine the physical mechanisms of noise transmission by the structure. Separate analyses
were performed on the structure and the acoustic cavity since the
coupling between the two was found to be very light. Finite element analysis permitted a smeared or average value for the composite elastic modulus to be determined, which in turn was used to
estimate the ring and critical frequencies for the structure, since
they influence transmission loss. Mode shapes were also predicted
by FEA, which allow the modal pressure distributions to be examined and which also permit investigation of the internal coincidence phenomena. Experimental modal analyses on the structure
and cavity were also performed. Tap tests at various locations
around the structure provided frequency response functions
FRFs that were curve-fit to obtain resonant frequencies and
damping ratios. A speaker was used to drive the acoustic cavity
with a white input while the pressure response was measured inside the cavity. Again, the measured FRFs were fitted to obtain
resonant frequencies and damping ratios for the acoustic modes.
Comparison of measured and predicted natural frequencies permitted corroboration of the FEA models. The resonances for the
structure and especially the acoustic cavity were found to significantly affect the noise transmission into the structure.
Finally, direct measurement of sound transmission was conducted based upon the noise reduction spectrum NRS, which is
computed from the ratio of the spatial averages of acoustic pressures measured along the internal and external surfaces of the
shell. Classic stiffness, resonance, and coincidence controlled regions were identified, which agreed very well with the predictions. The NRS was dominated by the sharp dips that were noted
at most of the acoustic resonant frequencies. Internal coincidence
was found to play a significant role as well, since negative NRS
values amplification was noted at these frequencies. The internal
cut-off frequencies are good estimates of the internal coincidence
frequencies below the ring frequency. These frequencies were determined from the FEA results by identifying the axial-zero modes
(k z 0). External coincidence and possibly ring frequency effects
266 Vol. 125, JULY 2003
also impacted the sound transmission, as noted by the wide dip in

the NRS plot from 6 12 kHz. However, this dip is found to be
inconsequential compared to the singularities found at some of the
acoustic mode frequencies. Future transmission control will provide a significant challenge due to the payload and weight requirements for the launch vehicles. Since the NRS is dominated by the
acoustic modes, control should focus on adding damping to the
acoustic modes or focus on disrupting the internal coincidence
phenomena. Once the effect of the acoustic modes is sufficiently
ameliorated, additional marginal improvements could also be expected by increasing the damping of certain structural modes.
Acknowledgments
The work was sponsored by the Air Force Research Laboratory
Space Vehicles Directorate AFRL/VS. The POC for this effort is
Dr. Steven A. Lane. The authors also acknowledge ANSYS, Inc.
for the John Swanson Fellowship. Thanks also to David J. Belasco, Jr. for help with data collection and reduction.
References
1 Huybrechts, Steven, and Tsai, Stephen W., 1996, Analysis and Behavior of
Grid Structures, Comp. Sci. Technol. 569, pp. 10011015.
2 George, Tommy J., Herman Shen, M. H., Huybrechts, Steven M., Meink, Troy
E., and Wegner, Peter M., 2001, Optimal Design of Composite Chambercore
Structures, Compos. Struct., 523 4, pp. 277286, MayJune.
3 Griffin, Steven, Denoyer, Keith K., and Das, Alok, 1999, Passive Vibroacoustic Isolation for Payload Containers, J. Intell. Mater. Syst. Struct., 10008305, pp. 83 87.
4 Griffin, Steven, Hansen, Collin, and Cazzolato, Ben, 1999, Feasibility of
Feedback Control of Transmitted Sound into a Launch Vehicle Fairing Using
Structural Sensing and Proof Mass Actuators, AIAA 991529, pp. 2592
2602.
5 Griffin, Steven, Hansen, Collin, and Cazzolato, Ben, 1999, Feedback Control
of Structurally Radiated Sound into Enclosed Spaces Using Structural Sensing, JASA, 1065, pp. 26212628.
6 Griffin, Steven, Denoyer, Keith K., and Sciuli, Dino, 1998, Hybrid Structural/
Acoustic Control of a Sub-Scale Payload Fairing, SPIE Conference on Smart
Structures and Integrated Systems, Vol. 3329 of 0277-786X, pp. 237243.
7 Griffin, Steven, and Denoyer, Keith K., 1998, Experimental Sensor and Actuator Location Procedure for Control of Dynamically Complex Smart Structures, SPIE Conference on Smart Structures and Integrated Systems, Vol.
3329 of 0277-786X, pp. 717723.
8 Lane, Steven A., Kemp, J. D., Griffin, Steven, and Clark, R. L., 2001, Active
Acoustic Control of a Rocket Fairing Using Spatially Weighted Transducer
Arrays, AIAA J., 381, 112119.
9 Neple, Pascale, el Hafidi, Ali, and Lesueur, Claude, 2001, On Sound Transmission Through a Truncated Isotropic Conical Shell: Experimental and Preliminary Numerical Results, Internoise 2001, The Hague, The Netherlands,
August 2730.
10 Vipperman, Jeffrey S., Li, Deyu, Avdeev, Ilya, and Lane, Steven A., 2001,
Characterization and Control of Sound Radiation in a Complex Fairing Structure, Internoise 2001, The Hague, The Netherlands, August 2730.
11 Denoyer, Keith K., Griffin, Steven, and Das, Alok, 1998, Passive Vibroacoustic Isolation for Reusable Launch Vehicle Payload Containers, AIAA/ASME/
ASCE/AHS/ASC Structures, Structural Dynamics & Materials Conference,
AIAA-98-1977, pp. 2248 2256, Longbeach, CA, April 2023.
12 Fahy, F., 1985, Sound and Structural Vibration, Academic Press, Inc., New
York.
13 White, Pritchard H., 1966, Sound Transmission Through a Finite, Closed
Cylindrical Shell, JASA, 505, pp. 1124 1130.
14 Manning, J. E., and Maidanik, G., 1964, Radiation Properties of Cylindrical
Shells, JASA, 369, pp. 16911698.
15 Cheng, L., 1994, Fluid-structural Coupling of a Plate-ended Cylindrical
Shell: Vibration and Internal Sound Field, JSV, 174, pp. 641 654.
16 Koval, L. R., 1976, On Sound Transmission into a Thin Cylindrical Shell
Under Flight Conditions, JSV, 48, pp. 265275.
17 Koval, L. R., 1978, Effects of Cavity Resonances on Sound Transmission
into a Thin Cylindrical Shell, JSV, 59, pp. 2333.
18 Blake, William K., 1986, Mechanics of Flow-Induced Sound and Vibration,
Vol. 1, General Concepts and Elementary Sources, Academic Press, Inc., London.
19 Blake, William K., 1986 Mechanics of Flow-Induced Sound and Vibration,
Vol. 2, Complex Flow-Structure Interactions. Academic Press, Inc., London.
20 Blevins, Robert D., 1979, Formulas for Natural Frequency and Mode Shape,
Van Nostrand Reinhold Co., New York.
21 ANSYS, Inc., 2001, Canonsburg, PA 15317.
Jianhui Luo
Hae Chang Gea
Department of Mechanical and Aerospace
Engineering,
Rutgers, The State University of New Jersey,
Piscataway, NJ 08855
Optimal Stiffener Design

for Interior Sound Reduction
Using a Topology Optimization
Based Approach
A topology optimization based approach is proposed to study the optimal configuration of
stiffeners for the interior sound reduction. Since our design target is aimed at reducing the
low frequency noise, a coupled acoustic-structural conservative system without damping
effect is considered. Modal analysis method is used to evaluate the interior sound level for
this coupled system. To formulate the topology optimization problem, a recently introduced Microstructure-based Design Domain Method (MDDM) is employed. Using the
MDDM, the optimal stiffener configurations problem is treated as a material distribution
problem and sensitivity analysis of the coupled system is derived analytically. The norm of
acoustic excitation is used as the indicator of the interior sound level. The optimal stiffener design is obtained by solving this topology optimization problem using a sequential
convex approximation method. Examples of acoustic box under single frequency excitation and a band of low frequency excitations are presented and discussed.
DOI: 10.1115/1.1569512
Introduction
The reduction of low frequency noise is of great interest in

designing transportation vehicles because the 20-200 Hz low frequency noise has an important influence on the product satisfactory quality. In the automobile passenger compartment, the low
frequency noise can be generated by engine vibration or induced
by road roughness; in the aircraft cabin, the low frequency noise
can be originated from engines, propellers or aerodynamic forces.
Although active vibration control is a fast growing research field,
passive vibration controls from structural optimization techniques
are often used in practice due to their low implementation cost.
Studies of noise reduction by structural modification are found in
the literature. Hagiwara et al. 1 investigated the reduction of
vehicle interior noise with shell thickness redistribution using the
sensitivity information semi-analytically. Huff, Jr. and Bernhard
2 used a parametric shape optimization method in the reduction
of the sound pressure level. Wodtke and Koopmann 3, Constants
and Belegundu 4 introduced the structural modifications by placing optimally sized point masses in order to minimize the radiated
sound power of vibrating structures.
In this paper, a topology optimization based approach is proposed to study the optimal configuration of stiffeners for interior
sound reduction. Since our design target is aimed at reducing the
low frequency noise, a coupled acoustic-structural conservative
system without damping effect is considered. Modal analysis
method is used to evaluate the interior sound level for this coupled
system 5. To formulate the topology optimization problem a
recently introduced Microstructure-based Design Domain Method
MDDM 6 is employed. Using the MDDM, the optimal stiffener configurations problem is treated as a material distribution
problem and sensitivity analysis of the coupled system is derived
analytically. The optimal stiffener design is obtained by solving
this topology optimization problem using a sequential convex approximation method called Generalized Convex Approximation
7.
The remainder of the paper is organized as follows: Section 2
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received June 1997;
Revised October 2002. Associate Editor: R. L. Clark.
introduces the modal analysis approach to calculating the frequency response for the coupled acoustic-structural system; Section 3 discusses the sensitivity analysis of frequency response for
this system. In Section 4, topology optimization problem is formulated and the solution procedure of identifying the optimal configurations of stiffeners for the reduction of the interior sound
level is described. In Section 5, four numerical examples of an
acoustic box under external excitations are presented. Conclusion
remarks and future work are discussed in the final section.
Frequency Response of Coupled Systems
In this section, frequency response of coupled acousticstructural systems is derived from the modal analysis method.
Consider that a coupled acoustic-structural system consists of an
enclosure surrounded by an elastic body and the elastic body is
subjected to both force and displacement boundary conditions as
shown in Fig. 1. Using the finite element method to discretize the
coupled system yields the following equations 8
M ss
M as
M aa
K ss
u e
p e
0
K sa
K aa

ue
fs
pe
0
(1)
where the upper equation refers to the structural system and the
lower equation refers to the acoustic system. In these equations,
u e is the vector representing displacement components at the grid
points of the structural finite element model, p e is the vector representing sound pressures at the grid points of the acoustic finite
element model. f s represents the external forces applied to the
structure, such as mechanical excitations. M ss and K ss are structural mass and stiffness matrices, M aa and K aa are acoustic mass
and stiffness matrices. M as and K sa are acoustic-structural couT
. The
pling matrices, and they have the property of K sa M as
structural and acoustic equations of motion are coupled through
the matrix M as which transforms the structural accelerations to
acoustic excitations of the interior cavity, and through the matrix
K sa which transforms the acoustic pressures to loads acting on the
structure.
If the external force has a harmonic form of f s e i t , then the
structural and acoustic responses can be expressed as u e e i t and
JULY 2003, Vol. 125 267
where s a T denotes the right eigenvector matrix of the

coupled system, is the coefficient matrix.
Inserting Eq. 7 into the governing equations of Eq. 2 and
T
premultiplying both sides by left eigenvector matrix
T T
s a , the use of M-orthonormal condition leads to

T
2 I
fs
sT f s
(8)
sT f s p, the coefficient matrix can be computed in the

Denote
following form
3
p e e i t , where is the excitation frequency. Introducing these expressions to Eq. 1 yields the frequency response equation as
K ss 2 M ss
K sa
2 M as
K aa 2 M aa

ue
fs
pe
0
(2)
There are two different approaches to computing frequency response in Eq. 2: the direct method and the modal analysis
method. The direct method is a rather straightforward approach
that solves Eq. 2 directly but it is very computationally expensive. On the other hand, the modal analysis method is extremely
attractive because both structural and acoustic modal density are
relatively low under the low excitation frequencies. Therefore, the
modal analysis method is adopted in this study.
Since the mass and stiffness matrices are unsymmetric in the
coupled acoustic-structural system, the left eigenvectors are not
the same as the right ones. Using the modal analysis method, the
right eigenvalue problem can be formulated as
K ss
K sa
K aa
M ss
s
a
M as
0
M aa

s
a
(3)
where is the eigenvalue matrix; s and a are the right eigenvector matrices corresponding to the structural and acoustic fields
respectively. Similarly, the left eigenvalue problem is represented
as
sT
Ta
K ss
K sa
K aa
sT

Ta
M ss
M as
M aa
(4)
s and
a denote the left eigenvector matrices of the
where
structural and acoustic fields. The M-orthonormal condition of the
coupled system is stated as follows:
sT
Ta
M ss
M as
M aa
s
I
a
a
a
(6)
Suppose that the forced frequency responses are expanded by

the right eigenvectors of the coupled system as
ue

pe
268 Vol. 125, JULY 2003
(7)
Sensitivity Analysis of Coupled Systems
When an optimization problem is solved by mathematical programming algorithms, we must determine the effect resulting from
a small perturbation in the current design on the objective and
constraint functions. This is known as the sensitivity analysis. In
this study, frequency response sensitivities are the derivatives of
structural response and the sound pressure with respect to the
structural modification design variables. With the correct frequency response sensitivity information, one can use various optimization methods to arrive at a modified structure with reduced
interior sound level. The frequency response sensitivity,
u e ,p e T , can be derived by differentiating Eq. 2 as
K ss 2 M ss
K sa
M as
K aa M aa
2 M ss
K ss
u e
p e
0
0
0
uc
pe
(10)
and M ss
are the derivatives of the structural stiffness
where K ss
and mass matrices with respect to the design variable, respectively.
2 M ss
u e , Eq. 10
If we define a pseudo load g s K ss
can be rewritten as
K ss 2 M ss
K sa
M as
K aa 2 M aa

u e
gs
0
p e
(11)
Eq. 11 has the same form as the frequency response governing

equations in Eq. 2 except the excitation force has been changed
from f s to g s . Therefore, the solution of Eq. 11 can be obtained
in the same way as shown in the previous section.
Assume response sensitivities can be decomposed as,
u e ,p e T with being the undetermined coefficient matrix.
Put this expansion into Eq. 11 and premultiply both sides by
T , we have
(5)
For large coupled acoustic-structural systems, it is impractical

to solve Eq. 3 and Eq. 4 directly because of the high computational cost. Luo and Gea 5 proposed a symmetrization approach to effectively extract eigenmodes with less computational
time and storage. They also proved that the right eigenvectors and
the left eigenvectors of a coupled system can be related as
(9)
where i , p i and i are the i th component of , p and . Combining Eq. 9 with Eq. 7, the response of the coupled acousticstructural system is recovered.
Fig. 1 A coupled acoustic-structural system
pi
i 2
T
2 I
gs
sT g s
(12)
sT g s q, then coefficient matrix can be determined in the

Denote
indices form
qi
i 2
(13)
With Eq. 13, the frequency response sensitivity, u c ,p e , can be

calculated easily.
Solution Procedures
In this section, the solution procedures of identifying the optimal configuration of stiffeners placement for the reduction of the
interior sound level are discussed.
The overall interior sound level for the acoustic field is evaluated as
S PL
1
m
10 log
10 p i
i1
/p 0 2
(14)
where p i denotes the sound pressure at the i th node, m is the total

number of nodal points in the acoustic field, p 0 is a reference
pressure, normally, chosen as 2105 Pa.
Although SPL is a direct indicator of sound noise, it inherits
strong localized effects from individual nodal pressure. Therefore,
if this function is used as the objective function in the optimization process, search algorithm will exhibit unstable oscillations
and consequently have difficulty to converge. Consider the fact
that the level of interior sound is fully determined by the acoustic
excitation, that can be measured by the magnitude of 2 M as u e
from Eq. 2 for a given excitation frequency. The minimization of
the norm of 2 M as u e can lead to the reduction of SPL. Therefore, the minimization of the acoustic excitation is chosen as the
objective function. In our numerical examples, iteration histories
of SPL are also listed for the purpose of comparing the sound
pressure level reductions from the optimization process.
When a structure is subjected to a band of frequency excitations
0 , 1 , the objective function can be modified as
Minimize
1
1 0
2 M as u e d
(15)
ume fraction, c (1i ) , is treated as the design variables in the stiffener

topology optimization formulation: as c 1(i ) 0 no stiffener is re and
quired and c (1i ) 1 stiffener exists. Using these relation, K ss
under the pseudo load can be directly calculated.
M ss
Generally, the total allowable weight of the structure is considered as a design constraint that cannot exceed a prescribed
amount. The optimization problem of minimizing acoustic excitation is stated as,
Minimize 2 M as u e
N
1
1 0
1
m
i1
i1
10 log10 p i /p 0 2 d
(16)
Recently, the stiffener layout optimization problem is tackled

by material distribution formulations, in which a given amount of
artificial stiffener material is dispersed optimally to the base
structure for the best structure performance. The material model
for the combination of base material and artificial stiffener material is often borrowed from various theories of composite materials. This method is called the topology optimization because the
topology of stiffener design is to be optimized. Diaz and Kikuchi
9 used this technique on solving stiffener layout design optimization under natural frequency and they applied the homogenization theory 10 to evaluate the material properties of the composite consisting of the base and stiffener material. In this paper, a
Microstructure-based Design Domain Method MDDM introduced by Gea 6 is applied to model the add-on stiffener. In the
MDDM, material is treated as a composite consisting of matrix
and spherical inclusion materials. The effective material properties
of the ith element is expressed in the form
i 0 1
i 0 1
c 1i 1 0
1c 1i 0 1 0 0
c 1i 1 0
1c 1i 0 1 0 0
(17)
(18)
with
1 1 0
3 1 0
(19)
2 45 0
15 1 0
(20)
0
0
where subscripts 0 and 1 represent matrix and inclusion materials;

and denote the bulk modulus and the shear modulus. 0 is
Poissons ratio of the matrix, and c 1(i ) is the volume fraction of the
inclusion material in the ith element. Using this model, the volJournal of Vibration and Acoustics
1 W
(22)
where denotes the material density, v i is the volume of ith

element, c i1 is the design variable as described previously, N is the
is the upper limit of total weight.
number of elements, and W
In order to search for the optimal stiffener topology, the number
of design variables is related to the number of finite elements used
in the structural model. Therefore, the optimizer used here must
be very efficient in handling a large number of design variables. In
this paper, the Generalized Convex Approximation GCA method
introduced by Chickermane and Gea 7 is used to formulate and
solve the topology optimization problem. In the GCA, functions
are approximated as the sum of a series of separable functions of
the design variables as
n
v 1c
Subject To:
And, the overall interior sound level from Eq. 14 can be defined
as
SPL
(21)
f x f xk
b x d b x d
i
k
i
ri
ri
(23)
where b i , d i and r i are a set of approximation parameters to be

determined, and f (x k ) is the value of the original function at the
kth design. Function values and first order sensitivity information
from the current and previous design is utilized to determine the
values of the approximation parameters. The approximation problem is solved iteratively using mathematical programming to generate the next design. This process continues until a satisfactory
design is reached.
Numerical Examples
Optimal stiffener designs of an aluminum alloy 1100-H14 box

using the proposed topology optimization based method are presented in this section. The aluminum box of 30 cm40 cm
50 cm is fixed at four bottom corners and under various external
excitations. The aluminum plate has thickness 0.2 cm, Poissons
ratio 0.33, Youngs modulus 6.91010 Pa, density 2.7
103 kg/m3 and the box without any stiffeners weights 5.076 Kg.
The artificial stiffen material has the same density as the base
material but with 100 times higher strength in order to simulate
the strong rigidity effect produced by stiffeners. The upper weight
limit of the add-on artificial stiffen material is set to be 10% of
the total base weight. That gives the total allowable weight of the
whole structure be 5.5836 Kg. In the initial design, the add-on
artificial stiffen material is uniformly distributed to the box in
order to produce an unbiased starting point. From the analyses of
the initial design, we found 20 structural eigenmodes and 3 acoustic eigenmodes including one Helmholtz mode for the uncoupled
system in the range of 0 to 500 Hz. Then, a coupled acousticstructural analysis was performed. We found that the eigenfrequencies of the uncoupled systems were shifted only slightly after
being coupled in the same frequency range. In Table 1, eigenfrequencies in the range of 0 to 500 Hz for uncoupled and coupled
systems are listed for comparison.
Since we are only interested in the low frequency sound reduction, damping effect for the coupled acoustic-structural system
was neglected. In the structural field, 1536 4-node plate elements
are used and each of them is also treated as design variable for the
stiffener design optimization formulation. In the acoustic field,
4096 8-node solid elements are constructed. To reduce the interior
JULY 2003, Vol. 125 269
Table 1 Coupled and uncoupled eigenfrequencies of a 30 cm

40 cm50 cm aluminum box
Mode
number
Structural
frequency
Hz
Acoustic
frequency
Hz
Coupled
frequency
Hz
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
114.95
130.72
179.50
195.32
195.36
231.06
242.49
257.62
291.94
302.71
323.13
335.41
387.39
389.28
411.79
418.62
443.03
462.87
484.05
494.91
0.00
344.05
430.06
-
0.00
117.02
129.92
178.67
194.99
196.09
239.07
241.66
254.44
291.35
302.71
304.21
330.40
370.61
382.70
386.75
411.72
418.62
443.02
443.33
452.73
482.94
495.40
sound level, the norm of acoustic excitation is chosen as design

objective as we previously discussed. Four optimal design examples under different external excitations are presented below.
5.1 Case 1. In the first case, a unit harmonic excitation with
frequency of 50 Hz is applied at the center of the top plate as
shown in Fig. 2. Using the proposed methodology, the norm of
acoustic excitation is reduced from 8.609643e-7 to 4.758937e-7 in
ten iterations and the overall interior sound decreased from 91.89
dB to 68.97 dB. The iteration history is showed in Fig. 3 and the
trend of SPL reduction is listed in Table 2. The final stiffener
configurations is shown in Fig. 4. It was found that the stiffener is
mainly concentrated in the region where the external force applies, which is reasonable because the excitation frequency is
lower than resonance frequencies of the system. The uncoupled
structural frequencies and the coupled system frequencies in the
range of 0 to 200 Hz are listed in Table 3. Comparing to the initial
design with uniformly distributed stiffener material, we found that
the final design also exhibits great changes of natural frequencies
and mode shapes.
Fig. 2 A box under an excitation loading at the center of the

top plate
270 Vol. 125, JULY 2003
Fig. 3 Iteration history of the objective function in case 1
5.2 Case 2. In the second case, the unit harmonic excitation

with the same excitation frequency as in the case 1 is applied to
the center point of the flank plate as shown in Fig. 5.
Iteration history of the objective function is shown in Fig. 6 and
we found the acoustic excitation is decreased from 6.246733e-7 to
9.847646e-8 with the overall interior sound level reduction from
76.73 dB to 51.92 dB in ten iterations. The result of the optimal
stiffeners placement is shown in Fig. 7 and we can see the stiff-
Table 2 Trend of interior SPL reduction in case 1

Iteration
number
Sound Pressure Level

dB
Weight
Kg
initial
1
2
3
4
5
6
7
8
9
10
91.89
89.62
80.87
77.05
75.64
74.33
72.08
70.88
70.52
70.26
68.97
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
Fig. 4 Optimal configuration of the stiffeners placement in

case 1
Table 3 Natural frequencies of the final design in case 1

Mode
number
Structural frequency
Hz
Coupled frequency
Hz
1
2
3
4
5
6
7
8
9
10
70.18
120.74
120.87
124.97
138.17
147.54
177.84
185.15
190.37
0.00
72.31
120.27
120.53
125.10
146.12
148.25
177.51
184.02
187.38
eners are also mostly located in the external force applied region.
Table 4 shows the trend of interior SPL reduction throughout the
iterations. The corresponding uncoupled structural frequencies
and coupled system frequencies in the range of 0 and 200 Hz are
listed in Table 5. The final design also experiences changes in
natural frequencies and mode shapes from the initial design.
5.3 Case 3. In the third case, a unit harmonic excitation
with higher frequency is applied at the center of the top plate. The
excitation frequency is chosen as 150 Hz. We found the objective function reduced from 2.811064e-5 to 1.548597e-5 and the
overall interior sound decreased from 80.13 dB to 67.03 dB after
ten iterations. The iteration history is showed in Fig. 8 and the
trend of SPL reduction is showed in Table 6. Fig. 9 shows the

case 2
optimal stiffener location. Unlike the first two cases, in this case
the locations of stiffeners are spread out to different region of the
box. This is due to the fact the excitation frequency is in the
middle of several resonance frequencies of the system. To minimize the acoustic excitation, the add-on stiffeners try to reduce the
effects from all these resonance modes as much as possible. The
natural frequencies below 200 Hz of the final design are also
listed in Table 7 for reference.
5.4 Case 4. In the last example, the stiffener design under a
band of frequency excitation is studied. A unit harmonic excitation
is applied at the center of the top plate with a band of low excitation frequencies ranged from 20 to 100 Hz. A modified objective
function, Eq. 15, is used here. Since the excitation frequency
range falls below any resonance frequency, the optimal stiffener
location turns out to be very similar to that of the first case as Fig.
4. We found the objective function is reduced from 3.106249e-7 to
1.842361e-7 and the overall interior sound defined in Eq. 16 is
decreased from 93.53 dB to 75.16 dB after ten iterations. Fig. 10
shows the interior SPL comparison between the initial design and
the final design within the exciting frequency range.
Table 4 Trend of SPL reduction in case 2

Fig. 5 A box under an excitation loading at the center of the
flank plate
Iteration
number

dB
Weight
Kg
initial
1
2
3
4
5
6
7
8
9
10
76.73
72.20
61.03
58.29
57.33
56.62
56.90
55.69
53.92
52.12
51.92
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836

Mode
number
Hz
Coupled frequency
Hz
1
2
3
4
5
6
104.94
117.86
158.14
167.27
188.17
0.00
107.27
117.13
157.36
167.43
195.37
JULY 2003, Vol. 125 271
Table 6 Trend of interior SPL reduction in case 3

Iteration
number

dB
Weight
Kg
initial
1
2
3
4
5
6
7
8
9
10
80.13
79.59
77.39
74.91
75.88
73.40
71.20
70.46
68.99
67.63
67.03
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
5.5836
Fig. 10 Interior SPL comparison between the initial design

and the final design
Conclusion and Future Work
In this paper, optimal stiffener design for interior sound reduction of coupled acoustic-structural system is studied. Forced frequency response and its sensitivity of this coupled system were
solved by the modal analysis method. Using a topology optimization based approach, the optimal stiffener placement problem was
converted into an optimal material distribution problem. Instead
of using the overall interior sound level directly, the norm of
acoustic excitation was used as the objective function and from
our numerical examples, it showed the acoustic excitation is a
good indicator for sound reduction. At low frequency excitations,
the optimal stiffeners might be applied to the region near the
external force source, however when the excitation frequency is
among several resonance frequencies, stiffener might be placed
throughout the structure in order to compensate the resonance
modes involved. In this work, a conservative system with no
damping effect is considered, therefore the proposed approach is
only valid for single frequency excitation or a band of low
frequency excitations without any resonance frequency included.
More general method to handle damping is currently under
investigation.
References
case 3

Mode
number
Hz
Coupled frequency
Hz
1
2
3
4
5
6
88.77
115.09
161.26
166.55
173.40
0.00
90.39
115.62
160.88
166.00
173.79
272 Vol. 125, JULY 2003
1 Hagiwara, W., Kozukue, W., and Ma, Z. D., 1993, The Development of
Eigenmode Sensitivity Analysis Methods for Coupled Acoustic-Structural Systems and Their Application to Reduction of Vehicle Interior Noise, Finite
Elem. Anal. Design, 14, pp. 235248.
2 Huff, J. E., Jr., and Bernhard, R. J., 1995, Acoustic Shape Optimization
Using Parametric Finite Elements, ASME 1995 Design Engineering Technical Conferences, DE-Vol. 84-2, pp. 577584.
3 Wodtke, H. W., and Koopmann, G. H., 1995, Quieting Plate Modes with
Optimally sized Point MassesA Volume Velocity Approach, ASME 1995
Design Engineering Technical Conferences, DE-Vol. 84-2, pp. 647 654.
4 Constans, E., and Belegundu, A., 1996, Minimizing Radiated Sound Power
from Vibrating Shells, The 6th AIAA/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Bellevue, WA, Paper 96-4111, pp. 1106
1116.
5 Luo, J. H., and Gea, H. C., 1997, Modal Sensitivity Analysis of Coupled
Acoustic-Structural Systems, J. Sound Vib., 119, pp. 545550.
6 Gea, H., 1996, Topology Optimization: A New Micro-Structure Based Design Domain Method, Comput. Struct., 615, pp. 781788.
7 Chickermane, H., and Gea, H. C., 1996, A New Local Function Approximation Method for Structural Optimization Problems, Int. J. Numer. Methods
Eng., 39, pp. 829 846.
8 Craggs, A., 1971, The Transient Response of a Coupled Plate-Acoustic System Using Plate and Acoustic Finite Elements, J. Sound Vib., 15, pp. 509
528.
9 Diaz, A. R., and Kikuchi, N., 1992, Solutions to Shape and Topology Eigenvalue Optimization Problems Using a Homogenization Method, Int. J. Numer. Methods Eng., 35, pp. 14871502.
10 Bendse, M. P., and Kikuchi, N., 1988, Generating Optimal Topologies in
Structural Design Using A Homogenization Method, Comput. Methods Appl.
Mech. Eng., 71, pp. 197224.
JULY 2003, Vol. 125 273
Yuji Ohue
Associate Professor,
Department of Intelligent Mechanical
Systems Engineering,
Kagawa University,
Hayashi, Takamatsu, 761-0396, Japan
e-mail: ohue@eng.kagawa-u.ac.jp
Akira Yoshida
Professor
Mem. ASME,
Okayaya University,
Tsushima-naka, Okayama, 700-8530, Japan
e-mail: akira-y@mech.okayama-n.ac.jp
New Evaluation Method on Gear

Dynamics Using Continuous and
Discrete Wavelet Transforms
The aim of this study is to propose a new evaluation method of gear dynamics using the
continuous and discrete wavelet transforms. The wavelet transform (WT) is a method for
the time-frequency analysis of signals. In order to evaluate the difference in the gear
dynamics due to the gear materials, which are sintered and steel ones, the dynamic
characteristics of gears were measured using a power circulating gear testing machine.
The gear dynamics were analyzed in a time-frequency domain by the continuous and
discrete WTs. The new evaluation method using the WTs proposed in this paper was more
useful compared with the conventional one to investigate the damping characteristic and
the dynamic condition of the gear equipment. DOI: 10.1115/1.1547465
Introduction
Test Gear and Testing Machine
Generally, the dynamic signals in the field of engineering problems such as vibration, sound and so on, have been analyzed using
the fast Fourier transform FFT. The FFT has been the most
common method to analyze the frequency properties of the signals. In the FFT analysis, the signal as the function of time is
converted to the power spectra in a frequency domain. However,
except for a special case, the frequency components of the most
signals encountered in the engineering problem change with time.
Based on the FFT alone, it is hard to investigate whether the
frequency components of the signals vary with time or not, even
though the phase of the Fourier transform relates to time shifting.
The FFT analysis is no more adequate for those applications.
Therefore, it is significant and important to adopt the timefrequency analysis for those signals varying with time 13.
It is possible by using the time-frequency analysis to investigate
how the frequency components of the signal vary with time. The
time-frequency analysis can provide more beneficial information
about the frequency compared with the FFT. During the last two
decades, a new mathematical technique for the time-frequency
analysis, which has been called the wavelet transform WT,
has been extensively developed 1,2. The WT has some features different from those of the short-time fast Fourier transform or the Wigner-Ville distribution. Applications of the WT
are actively studied in a variety of the fields of engineering
science 4 7.
Gears have been employed in automobiles, motorcycles, industrial robots and so on. The gears are an important machine element to transmit the power and to change the speed of the equipment. The FFT is generally used for analyzing the gear dynamics.
However, it is important to evaluate the gear dynamic performance more precisely using WT. Therefore, in order to evaluate
the difference in the gear dynamics due to the gear materials, the
dynamic performances of sintered and steel gears were measured
at gear rotational speeds of 1600 rpm to 10,000 rpm, using a
power circulating gear testing machine. The gear dynamics concerned with tooth root strain, vibration acceleration of gear box
and sound near gear box, were analyzed in the time-frequency
domain by the continuous WT, and the signals of the gear dynamics were decomposed and reconstructed by the discrete WT. The
validity of the new evaluation method by the WT is discussed.
Table 1 and Fig. 1 show the specification and the shapes of gear
pair, respectively. The module and the pressure angle of the gear
pair are 5 mm and 20 deg., respectively. The contact ratio of the
gear pair is 1.246. The gear has 16 teeth and a face width of 6
mm. The materials of the test gears are sintered and steel JIS;
SCM440 ones. Those test gears were induction-hardened before
grinding of tooth surface. The material of the mating pinion is
steel JIS: SCM415. The chemical composition of the sintered
material is almost the same as that of the steel SCM440. Many
pores existed in the sintered great. Generally, the sintered materials have higher damping ratio due to the pores, compared with the
steel ones. The density of the sintered gear in this study is 6.8
g/cm3. The Youngs modulus and Poissons ratio of the sintered
gear are 152 GPa and 0.25, and those of the steel gear are 206

Revised Sept. 2002. Associate Editor: R. P. S. Han.
274 Vol. 125, JULY 2003
Table 1 Specification of gear pair

Pinion
Module
Pressure angle
Number of teeth
Addendum modification
coefficient
Tip circle diameter
Center distance
Face width
Contact ratio
Accuracy*
Tooth surface finishing
mm
deg.
Gear
5
20
15
0.571
mm
mm
mm
16
0.560
90.71
94.60
82.55
18
6
1.246
Class 1
Class 1
Grinding
*JIS B 1702
Fig. 1 Shapes and dimensions of gear pair
Wavelet Transform
Continuous Wavelet Transform. The continuous wavelet
transform CWT of a function f (t) is defined as follows.
W f b,a a 1/2
Fig. 2 Tooth profile errors of gear pair
GPa and 0.3, respectively. Figure 2 shows the tooth profile errors
of test gear pairs. The tooth surfaces of both gears and pinion were
finally ground. The tooth profile errors of both sintered and steel
gears and almost the same. The accuracy of the test gear pairs are
in class 1 according to Japanese industrial standard JIS B 1702.
The test apparatus shown in Fig. 3 is a power circulating type
gear testing machine with a center distance of 82.55 mm. The test
gear was driven at an interval of 200 rpm within a range of rotational speeds n 2 of 1600 rpm to 10,000 rpm. The loading torque
was statically applied to the gear shaft by a lever loading with
dead weights and a torsional coupling. The torsional torque was
taken as a scale for loading. In this experiment, the value of the
torque was 30 Nm, which corresponds to a load per unit face
width of 142 N/mm in the normal direction to the tooth surface.
The gear pair was lubricated with a gear oil by pressure feeding at
a flow rate of 750 ml/min and an oil temperature of 313 K.
The dynamic tooth root strain was measured using a dynamic
strain amplifier and a wire strain gage with a gage length of 0.3
mm, which was bonded on a compression side of the root fillet of
the gear. The vibration acceleration of the gear box was detected
by a piezo-electric pickup through an amplifier. The pickup has a
maximum response frequency of 20 kHz. The sound pressure near
the gear box was detected by a condenser type microphone having
a maximum response frequency of 12.5 kHz. The microphone was
located at a distance of 300 mm from the side of the gear box.
These signals were stored into a data recorder with a synchronous
signal obtained by a phototransistor and two disks with slits which
were fixed on both shifts of the gear and the pinion. These stored
signals were analyzed using a computer with an A/D converter
board.
f t
tb
dt
a
(1)
Where, the (t) indicates the conjugate of a mother wavelet function (t). The a and b indicate the parameters on frequency and
time. The Gabor function defined by Eq. 2 was adopted as the
mother wavelet function (t) in this study
t 1/4
1/2
exp
1 p
2
i p t
(2)
Where, p is a center of angular frequency, and is a constant

and was set to be (2/ln 2) 1/25.336. Figure 4 shows an example
of the Gabor function and its Fourier spectrum. The integral for
Eq. 1 was calculated using the trapezoidal rule and the calculation results are expressed using the following equation.
W f Re W f 2 Im W f 2
(3)
Where, Re(W f ) and Im(W f ) are the real and the imaginary
parts in Eq. 1. To express the wavelet map, that is the intensity
map in a time frequency domain, the intensity calculated by the
CWT was normalized by the maximum intensity in the map.
Discrete Wavelet Transform. When the coordinates b, a of
the CWT shown in Eq. 1 are discretized to the coordinates
(2 j k,2 j ) using two integers j and k, the discrete wavelet transform DWT is defined as follows.
d kj 2 j/2
f t 2 j tk dt
(4)
Where, d (kj ) is equal to (W f )(2 j k, 2 j )). j is called level. The

inverse discrete wavelet transform IDWT is defined as
f t
g t
j
(5)
The function g j (t) on the wavelet component is given by
Fig. 3 Power circulating gear testing machine and measurement system of dynamic performance of
gear pair
JULY 2003, Vol. 125 275
Fig. 5 Scaling and wavelet functions based on cardinal

B-spline In order of m 4
f j t f j1 t g j1 t g j1 t g j2 t g j3 t
(12)
Fig. 4 Gabor function
g j t
d
k
j
j
k 2 tk
(6)
Suppose that f j (t) is the function at a level j, f i (t) is satisfied with

the following relation.
f j t
c
k
j
j
k 2 tk
(7)
Where, c (kj ) is a sequence at a level j, and (t) is a scaling

function. The scaling function (t) and the mother wavelet function (t) are satisfied with the two-scale relations as follows.
t
t
p 2tk
k
q 2tk
k
c kj
d kj1
p
l
j
12k c l
b 12k c l j
j1
q k2l d l j1
k2l c l
k0
m
k
m1
tk
(13)
Figure 5 shows the scaling function (t) and the mother wavelet
function (t) based on the function N 4 (t), that is the cardinal
B-spline function in order of 4. The function (t) is equal to the
function N 4 (t). From this figure, it is understood that the function
(t) is very smooth and is very similar to sinusoidal functions.
The sequences a k , b k , p k and q k are given in the reference 1.
Conventional Evaluation of Gear Dynamics

(9)
Figure 6 shows the examples of the dynamic and static tooth

root strains or stresses measured in this experiment. Generally,
in order to evaluate the dynamic load of the gear pair, the dynamic load ratio dmax / smax , that is the ratio of the maximum
dynamic tooth root stress dmax to the maximum static tooth
(10)
(11)
Where, a k and b k are the sequences to decompose the function f j (t) at a level j. Therefore, the decomposition and the reconstruction on the function f (t) are concluded into the calculations
using the sequences a k , b k , p k and q k . The functions
g j (t) and f j (t) at a level j are able to be found by using Eqs. 6
and 7. The function f j (t) decomposed into the function g j (t) on
the wavelet component is satisfied with the following relation.
276 Vol. 125, JULY 2003
1
N m t
m1 !
(8)
Where, p k and q k are two-scale sequences. The algorithms of

the decomposition and the reconstruction using the DWT and the
IDWT on the function f (t) are given as Eqs. 10 and 11.
c kj1
In this study, the scaling and the mother wavelet functions

based on the cardinal B-spline function in order of m4 was
adopted. The cardinal B-spline function N m (t) in order of m is
defined as follows 1.
Fig. 6 Dynamic and static tooth stresses
Fig. 7 Relation between dynamic load ration and n 2
root stress smax , has been used. In this study, the tooth root
stress at a rotational speed n 2 of 6.6 rpm was adopted as the static
stress smax .
Figure 7 shows the dynamic load ratio dmax / smax plotted
against the rotational speed n 2 of gear. The frequency f z indicates
the tooth mesh frequency. The natural frequency of the gear pair
including the shafts, which support the gear pair, was calculated
using a program given in reference 8. The equation of motion
for the calculation of the natural frequency of the gear pair is
given by
2
M x Dx K t,x xW
K t,x e t
i
(14)
where, M is the equivalent inertia mass of gear pair along the line
of action, D is the damping coefficient between gear pair including the shafts, K is the stiffness of gear pair, W is the static load,
e i is the tooth profile error and x is the relative displacement along
the line of action. Subscript i indicates the pinion (i1) and the
gear (i2). Then, the natural frequency f e is calculated using the
following equation.
1
f e
2
1
Tz
Tz K t,x
dt
(15)
Where, T z is the tooth mesh period. In this experiment, the calculated value of f e was approximately 2.56 kHz. The gear box was
hit by a hammer to measure its natural frequency in the horizontal
direction of the gear box using the FFT. As the result, the natural
frequency of the gear box was about 5.0 kHz. From Fig. 7, for
both the sintered and steel gears, the ratio dmax / smax increases
with fluctuation as n 2 increases. The ratio dmax / smax has three
peak values at n 2 4200 rpm ( f z 1.12 kHz, 6000 rpm ( f z
1.60 kHz and 9000 rpm ( f z 2.40 kHz, since the tooth mesh
frequency and its harmonics coincide with both the natural frequencies of the gear pair and the gear box. Namely, the gear pair
for both the sintered and steel gears became the resonance conditions at their speeds. The ratio dmax / smax of the sintered gear is
mostly smaller than that of the steel one against n 2 . Therefore, it
is obvious that the sintered gear is superior to the steel one from
the viewpoint of the damping characteristics.
Analysis of Gear Dynamics Using WT

In the previous section, the conventional evaluation method
was employed for evaluating the gear dynamic load. The conventional evaluation method, however, has not taken into account the
information on the time and the frequency contained in the gear
Fig. 8 Tooth root strain, WT map and FFT at n 2 1800 rpm
dynamic load and used only the maximum value of the gear dynamic load. Thus, the information of the gear dynamic load in the
time-frequency domain is important to evaluate the great performance in detail. In this section, the dynamic tooth root strains are
analyzed in the time-frequency domain using the WT.
Figure 8 shows the waveform of the tooth root strain top of the
figure, its WT map lower left of the figure and the analyzed
result using the FFT lower right of the figure at a rotational
speed n 2 of 1800 rpm. The WT maps of both the gears were
obtained using the CWT. The WT map shows the intensity of the
tooth root strain in the time-frequency domain. In the WT map,
not only the frequency components but also the change of the
intensity of each frequency in the tooth root strain with the time
can be provided more precisely, compared with the FFT analysis.
In the WT maps of the sintered and steel gears, each intensity in
two regions above and below f z 480 Hz indicates each peak. The
contact time of one tooth pair at n 2 1800 rpm is approximately
2.6 ms. Suppose that the wave of the tooth root strain is roughly a
sine wave with a period of 5.2 ms, the frequency of the sine wave
is 192 Hz. It can be considered that the frequency components
below f z depend on the contact time of the one tooth pair. On the
other hand, it can be considered that the frequency components
above f z are caused by the torsional vibration of the gear pair. The
relative intensities at the region above f z of the sintered gear are
JULY 2003, Vol. 125 277
Fig. 9 Two main frequencies contained in tooth root strain
lower than those of the steel one. It can be understood that two
main frequency components above and below f z exist in the dynamic tooth root strain.
Figure 9 shows the relationship between the two main frequencies contained in the tooth root strain of the steel gear and the
rotational speed n 2 . The tooth root strain consists of two main
frequencies above and below f z at each rotational speed of gear.
Especially, the relation between the frequency f low below f z and
n 2 is given by a following relation.
f low
fz
1 z 2n 2
1
1 a 60
2 a 2 a T z
(16)
Where, a is the contact ratio of the gear pair, z 2 is the number

of the gear teeth, T z is the tooth mesh period. Substitution of
the values of a 1.246 and z 2 16 shown in Table 1 into Eq.
16 yields f low 0.4 f z . The relation between f low and f z of
the sintered gear was the same as that of the steel one. It is obvious that the tooth root strain consists of two main frequencies
due to the tooth contact below f z ) and due to the torsional vibration of the gear pair above f z ), independent of the material of
the gear.
In order to evaluate the tooth root strain consisting of both the
frequencies due to the tooth contact below f z ) and due to the
torsional vibration of the gear pair above f z ), the tooth root strain
is decomposed by DWT, and is reconstructed by IDWT in the two
regions above and below f z . Figure 10 shows the results of the
decomposed tooth root strains at n 2 1800 rpm by DWT. The
waveform f j becomes smooth as the absolute value of j becomes
larger. The relation between the waveforms f j and g j is given by
Eq. 12, that is, the waveform f j is decomposed into the waveforms f j1 and g j1 . At levels j3 and 4 corresponding to
the frequencies above f z , the amplitude of the waveforms g j of
the sintered gear is smaller than that of the steel one. On the other
hand, at a level j6 corresponding to the frequency below f z ,
the maximum value of the waveform f 6 of the sintered gear is
larger than that of the steel one, since the Youngs modulus 152
GPa of the sintered gear is smaller than that 206 GPa of the
steel one. Figure 11 shows the reconstructed tooth root strains and
the subtracted strain n2 6.6 at n 2 1800 rpm in the case of the
sintered gear. The reconstructed strain g j above f z is almost the
same as the subtracted strain n2 6.6 in shape. The strain obtained by subtracting the static strain 6.6 from the dynamic strain
n2 can be supposed to be the strain due to the dynamic load.
Therefore, the reconstructed strain g j above f z is dependent not
on the static load but on the dynamic one.
Fig. 10 Decomposed tooth root strain
278 Vol. 125, JULY 2003
Fig. 11 Reconstructed tooth root strains
New Evaluation of Gear Dynamics

Figure 12 shows the conventional dynamic load ratio top of the
figure, the maximum strain value of the decomposed strain f j
below f z second of the figure, the strain amplitude of the reconstructed strain g j above f z third of the figure and the new
dynamic load ratio bottom of the figure plotted against the rotational speed n 2 . The maximum values of the decomposed strains
f j second of the figure at n 2 higher than 1600 rpm are almost
constant and close to that of the static strain at n 2 6.6 rpm. The
fluctuation of the amplitude of the reconstructed strain g j third
of the figure is similar to that of the conventional dynamic load
ratio top of the figure against n 2 . Therefore, it is clear that the
values of the reconstructed strains above f z and the decomposed
strains below f z depend on the damping ratio and the Youngs
Fig. 13 Example of decomposed vibration of gear box
modulus, respectively. The new dynamic load ratio is given by

dividing dmax by the stress, to which the maximum value of the
reconstructed strain f j is converted. The fluctuation of the new
dynamic load ratio bottom of the figure is similar to that of the
conventional one against n 2 . The new dynamic load ratio proposed in this paper can be calculated using only the dynamic tooth
root stress without the static tooth root stress. In some equipment with gear sets, it is occasionally difficult to operate the
equipment at very low speed and to measure the static stress.
Therefore, the proposed new dynamic load ratio, which can be
calculated using only the dynamic tooth root stress, seems to be
beneficial in engineering.
Figure 13 shows the decomposed vibration acceleration of the
gear box at n 2 1800 rpm during one revolution of gear in the
case of the sintered gear. Both the vibration acceleration and the
sound pressure are reconstructed in two regions above and below
f z . Figures 14 and 15 show the root-mean-squares of the vibration acceleration and the sound pressure plotted against n 2 . The
root-mean-square R.M.S. of the vibration acceleration and the
sound pressure was given by the following equation.
R.M.S.
Fig. 12 Conventional and new dynamic load ratios
1
Tr
Tr
v t 2 dt
(17)
Where, T r is one revolution period of the gear and v (t) is a signal.

The values of both the vibration and the sound reconstructed
in the region below f z is proportional to n 2 to the power two,
and is independent of the material of the gear. On the other hand,
the values of both the vibration and the sound reconstructed in
the region above f z for the sintered gear are smaller than those of
the steel one. This tendency is the same as that in the original
vibration and sound. The vibration due to the tooth mesh propagates to the gear box through shafts and bearings supporting the
gear pair. The vibration of the gear box and the sound radiated
from the gear box are caused by the torsional vibration of the gear
pair due to the tooth mesh. The torsional vibration depends on the
JULY 2003, Vol. 125 279
root-mean-square values of the vibration acceleration and the

sound pressure below f z were independent of the gear material
and were proportional to the rotational speed to the power two.
The behavior above f z was influenced by the difference of the
gear material.
3. The new evaluation method using the continuous and discrete wavelet transforms proposed in this paper was very beneficial compared with the conventional one, in order to investigate the damping characteristic. The dynamic condition of the
gear equipment due to the unbalance of the gear sets, the tooth
profile error and so one could be well assessed by analyzing
the vibration and the sound of gear sets in two regions above and
below f z .
Acknowledgments
Fig. 14 Root-mean-square of vibration acceleration above and

below f z
frequency components above f z . In short, the fluctuations of both

the vibration and the sound are caused by the frequency components above f z .
Conclusion
In order to evaluate the difference in the gear dynamics with
different gear materials, the gear dynamics concerning tooth root
strain, vibration and sound were measured using a power circulating gear testing machine in a range of rotational speeds of 1600 to
10,000 rpm. The gear materials employed were steel and sintered
ones. The gear dynamics were analyzed in a time-frequency domain by the continuous and discrete wavelet transforms.
1. The dynamic tooth root strain indicated indirectly the condition of the torsional vibration of the gear pair. The value of
the tooth root strain reconstructed at the region below the tooth
mesh frequency f z depended on the Youngs modulus of the gear
material. On the other hand, the amplitude of the strain reconstructed at the region above f z depended on the damping ratio of
the gear pair.
2. The behavior of the vibration and the sound of the gear box
against the rotational speed of the gear could be also divided
obviously to two different behaviors above and below f z . The
Fig. 15 Root-mean-square of sound pressure above and

below f z
280 Vol. 125, JULY 2003
The authors would like to thank Sumitomo Metal Industries,

Ltd. and Japan Energy Co. Ltd. for providing the sintered material
and the lubricating oil, respectively. This research was supported
financially in part by the scientific research fund of the Japanese
Ministry of Education, Culture, Sports, Science and Technology
and the Mazda Foundation to which the authors express their
gratitude.
Nomenclature
D damping coefficient between gear pair
K(t,x) stiffness of gear pair
M equivalent inertia mass of gear pair along line
of action
N m (t) cardinal B-spline function in order of m
Tr one revolution period of gear
Tz tooth mesh period
W static load of gear pair
(W f )(b,a) wavelet transform
a frequency parameter
a k sequence for decomposition by discrete wavelet transform
b time parameter
b k sequence for decomposition by discrete wavelet transform
c (kj ) sequence at level j
d k ( j ) sequence at level j(W f )(2 j k, 2 j )
e(t,x) tooth profile error
f (t) signal
f j (t) signal at level j
g j (t) wavelet component of signal f (t) at level j
f e natural frequency of gear pair
f low f z /2 a
f z tooth mesh frequency z 2 n 2 /60
j level
k integer for time parameter
m order of cardinal B-spline function
n 2 rotational speed of gear
p k , q k two-scale sequences for reconstruction by discrete wavelet transform
t time
v (t) signal of vibration acceleration or sound pressure
x relative displacement along line of action
z 2 number of gear teeth
a contact ratio
(t) scaling function
(t) mother wavelet function
dmax maximum dynamic tooth root stress
smax maximum static tooth root stress
constant for Gabot function
p center of angular frequency
References
1
2
3
4
Chui, C. K., 1992, Introduction to Wavelet, Academic Press.

Mallat, S., 1998, A Wavelet Tour of Signal Processing, Academic Press.
Cohen, L., 1995, Time-Frequency Analysis, Prentice-Hall PTR.
Newland, D. E., 1993, Random Vibrations, Spectral & Wavelet Analysis,
Longman Scientific & Technical.
5 Staszewski, W. J., and Tomlinson, G. R., 1994, Application of the Wavelet
Transform to Fault Detection in a Spur Gear, Mech. Syst. Signal Process.,
83, pp. 289307.
6 Mori, K., Kasashima, N., Yoshioka, T., and Ueno, Y., 1996, Prediction of
Spalling on a Ball Bearing by Applying the Discrete Wavelet Transform to
Vibration Signals, Wear 195, pp. 162168.
7 Yoshida, A., Ohue, Y., and Ishikawa, H., 2000, Diagnosis of Tooth Surface
Failure by Wavelet Transform of Dynamic Characteristics, Tribol. Int., 33,
273279.
8 JSME Research Group on Accuracy and Design in Gears, 1977, Rep. Fujita,
K., Research Report on Accuracy and Design in Gears, RC-SC38, JSME, in
Japanese.
JULY 2003, Vol. 125 281
J. Antoni
Lecturer,
Roberval UMR CNRS 6066,
University of Technology of Compie`gne,
France
R. B. Randall
Professor,
School of Mechanical and Manufacturing
Engineering,
The University of New South Wales,
Sydney 2052, Australia
A Stochastic Model for Simulation

and Diagnostics of Rolling
Element Bearings With Localized
Faults
This paper addresses the stochastic modeling of the vibration signal produced by localized faults in rolling element bearings and its use for diagnostic purposes. The aim is
essentially to provide a better understanding of the recognized envelope analysis technique as classically used in the diagnostics of rolling element bearings, and incidentally
give theoretical proofs for the specific features of envelope spectra as obtained from
experimental data. The proposed model may also prove useful for simulation purposes.
First, the excitation force generated by a defect is modeled as a random point process and
its spectral signature is derived analytically. Then its transmission through the bearing is
investigated in detail in order to find the spectral characteristics of the resulting vibration
signal. The analysis finally gives sound justification for squared envelope analysis and
the type of spectral indicators that should be used with it. DOI: 10.1115/1.1569940
Introduction
Most frequent faults in rolling element bearings include defects

such as cracks, pits and spalls on the inner race, outer race, or
rolling elements. Such defects are usually very localized during
their early stages which is precisely when they need to be detected. As a result, intensive vibrations are produced by the repetitive impacts of the moving parts of the bearing on incipient defects. Numerous techniques have been proposed over the past
three decades to diagnose rolling element bearing in the case of
localized from the vibration signals they produce. Inter alia, the
so-called envelope analysis or high frequency resonance
technique is probably one of the most valuable and is currently
well established in vibration based condition monitoring
1,2,3,4,5. It is based on the idea that repetitive impacts on a
defect excite some resonanceusually in a high frequency range
where the signal-to-noise ratio is highwhich acts as a modulation carrier. Demodulation of the vibration signal around the carrier then yields the envelope of the signal whose spectral content
has been shown to be very relevant in identifying the location of
the fault in the bearing, and the shaft on which the bearing is
mounted. Since the early and heuristic foundations of the envelope analysis technique, many papers have tried to explain its
actual virtues and unbeaten successes when applied to rolling element bearings. These efforts have first focused on proposing a
proper model for the vibration signal generated by localized
faults. It must be said that the objective of such a model is not to
explain the physics of bearing failures but to describe its consequences as observed by the experimenter, i.e. it is phenomenological.
Most likely, the first valuable model for the vibration signal
produced by a localized defect is due to McFadden & Smith 2,3.
Therein the repetitive impacts generated by a defect were modeled
as a periodic train of Dirac delta functions with period T. Consequently the resonance characteristic in the Fourier domain was
sampled at regular intervals 1/T Fourier series. Moreover, McFaddens model had the advantage of explicitly including different
sources of amplitude modulations the radial load distribution, the
moving location of impact forces thus giving a good understandContributed by the Technical Committee on Vibration and Sound for publication
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received April 2002;
revised January 2003. Associate Editor: M. I. Friswell.
282 Vol. 125, JULY 2003
ing of the spectral content of the envelope of the resulting vibration signal. This model was later refined by Ho & Randall who
pointed out that actual rolling element bearings experience some
random slip in their operation so that the train of impacts is
slightly random instead of periodic 4. However small these effects, Ho showed that the resonance characteristic is no longer
sampled in the Fourier domain but rather resembles a continuous
spectral density where all the harmonics tend to smear over each
other. Hos model resulted in a significantly better description of
bearing vibration spectra as observed in the real world, and was
next used by Randall, Antoni & Chobsaard to show that bearing
signals are quasi-cyclostationaryi.e. their statistics have quasiperiodicity 5. Incidentally, this observation offered an elegant
way for justifying the envelope analysis method from the theory
of cyclostationary processes. In a following paper, Antoni & Randall refined their results after specifying that signals from localized faults are not exactly quasi-cyclostationary since the random
slips are non-stationary in their nature 6. However they concluded that the bearing signals could still be treated as pseudocyclostationary as a first approximation.
The purpose of this paper is to attempt a complete treatment of
the stochastic modeling of bearing vibrations as produced by localized faults, putting together a number of unpublished results
and putting the heuristic considerations of 6 on a firmer mathematical foundation. The aim is twofold. Firstly, it is to provide
the mechanical community with a model that has proven very
satisfactory in describing actual vibration signals and, in particular, their spectra and envelope spectraincluding some typical
features that have never been explained elsewhere. Secondly, it is
to demonstrate how this model permits a proper formalization of
the envelope analysis technique as classically used in the diagnostics of rolling element bearings. The paper is organized as follows.
A first section addresses the accurate modeling of the nonstationary impacting process as generated by a localized defect on the
inner race, the outer race, or on a rolling element. The spectral
characteristics of this process are then derived from the theory of
regular point processes and important results are deduced concerning the nature of spectral harmonics. In a second section these
results are used to investigate the spectral properties of the resulting vibration signal after the impacts have propagated through the
system, i.e. as measured on the housing. In particular, the general
spectral signature due to a localized defect is found and its mani-
Fig. 1 The impacting process viewed as a point process

Fig. 2 Product density of degree one for T 130
festations in a number of spectral indicators the Fourier transform, the power spectral density, the spectral correlation density,
the Fourier transform of the squared signal and the envelope spectrum are investigated in detail. The relative effectiveness of these
spectral indicators in diagnostics is finally discussed in the light of
the new results.
Note the degeneracy occurring at 0, where f 2 (t,0)

f 1 (t) ( ). By using the terminology of stochastic point processes, explicit solutions will now be found for the first two moments of the impacting process.
2.2 Analytical Forms of the Product Densities. The product density of degree one in Eq. 3 can be expanded into
Modeling the Impacting Process
2.1 Regular Point Process. At the outset, consider the process generated by the repetition of impact forces when a defect in
one surface strikes a mating surface. We shall refer to it as the
impacting process F(t). For a localized defect, each impact may
be well described by a Dirac delta function (t) provided the
measured signal is sampled at a rate well below the impact spectral bandwidth 2,3,4. At this stage it is assumed that all impacts
have equal magnitudes; magnitudes and signs of impacts will be
accounted for later in the text by modulating the impacting process with a suitable time-varying function.
Without loss of generality, the reference time t0 is chosen to
coincide with an arbitrary impact which defines the point from
which the process is starting to be observed. Hence,
F t
tT
i0
where T 0 0
(1)
The stochastic process T i governing the arrival of the impacts

can be defined in a variety of ways. However it was argued in 6
that an adequate assumption for rolling element bearings is where
the inter-arrival times T i T i T i1 are independent and identically distributed random variables see Fig. 1. In turn, this can be
shown to define a stationary Markov process T i , that it is to say
in which each arrival is only influenced by its immediate predecessor and irrespectively of its index:
P T i t i /T i j t i j , j1, . . . ,i P T i t i /T i1 t i1
P T 1 t 1 /T 0 0
(2)
It can easily be checked that under these conditions the arrival

time process T i has a stationary mean E T i iT but a nonstationary covariance function Co v T i ,T j 2 min(i,j) where
2 Var T i . It is specifically this non-stationarity that has not
properly been recognized before and actually explains distinctive
features of the vibration signal ensuing from a faulty bearing 6.
Now return to Eq. 1 and define N(t) as the number of
impacts that have occurred in the interval 0,t , so that dN(t)
denotes the number of impacts in the infinitesimal interval t,t
dt . Hence, Eq. 1 simplifies to F(t)dN(t)/dt which defines
an ordinary point process ordinary means one in which the initial
impact occurs at zero time 7. For the physical case of interest, it
is a sound assumption that the probability of occurrence of an
impact in dt is proportional to dt while the probability of more
than one occurrence is negligibly smaller than dt. This property
of regularity ensures the use of the product density technique to
obtain the moments of F(t) 8. Specifically, we define f 1 (t) and
f 2 (t, ) as the product densities of degree one and two which may
be interpreted as the instantaneous mean rate of impacts respectively at time t and at time t plus . Then,
E dN t n f 1 t dt,
n0 in N
E dN t dN t f 2 t, dtd ,
(3)
(4)
f 1 t dt
P tT tdt/T 0 t dt
i
i0
i0
(5)
where 1 (t) is the probability density function of the ith impact

conditioned to the fact that the zeroth occurred at t0 and
0 (t) (t).
Similarly, the product density of degree two is
f 2 t, dtd
P tT tdt,t T t d /T
i
i0 ji
(6)
for 0. Remembering that the arrival time T i is a stationary

Markov process,
f 2 t, dtd
i0
i t dt
d,
k1
(7)
and finally, adding the degeneracy case f 2 (t,0) f 1 (t) ( ) arising

when 0, one gets the simple expression,
f 2 t, dtd f 1 t f 1 dtd ,
(8)
In short Eqs. 5 and 8 give the explicit solutions for first two
moments of the impacting process from which those of the vibration signal will later be derived. For the physical process under
consideration, it is noteworthy that the product density of degree
one f 1 suffices to describe it at least up to the second order because f 2 factorizes into a product of f 1 terms, thus assigning to
the instantaneous mean rate of impacts a major role in this paper.
As a matter of fact, the exact shape of f 1 (t) is worthy of further
investigation. In view of Eq. 5, the first peak in f 1 (t) happens to
be the probability density function 1 (t) of the first time of occurrence T 1 , the second peak the probability density function of
T 2 and so on. Therefore the ith peak is the first one convolved
with itself i times, i.e.
(9)
Then, under mild conditions, the bandwidth of the ith peak as
measured by its standard deviation is i with the standard
deviation of 1 (t). As the peaks slowly enlarge, their amplitudes
decrease accordingly so as to maintain a unit area. This is illustrated in Fig. 2. In the limit, the peaks completely vanish and f 1 (t)
tends to the constant value f 1 ()1/T, that is the mean overall
rate of occurrence.1 However the rate of convergence is extremely
small: considering a percentage of random fluctuation of x/100
/T, then two peaks completely overlap when their band1
An heuristic proof to this result is that, in the limit, the area under each probability density function is still unity whereas the mean paving is one probability
density function per T units of time.
JULY 2003, Vol. 125 283
Fig. 3 Fourier transform modulus of the product density of

degree one for T 130
width is greater than their mutual spacing, that is when i

T or i10000/x 2 . For random fluctuations of typically a few
percent, this means it takes a few thousands peaks for f 1 (t) to
reach its limit.
Similarly, the product density of degree two can be verified to
tend to f 2 (, ) f 1 ( )/T. This supports the assertion of reference 6 where it was pointed out that the nonstationary impacting
process F(t) ultimately tends to stationarity, yet so slowly that the
phenomenon is hardly noticeable in practice over a finite time of
observation.
2.3 Spectral Characteristics. Since most processing involved in the diagnosis of rolling element bearings is performed in
the Fourier domain, it is now necessary to derive the formulas for
the Fourier transforms of the product densities f 1 (t) and f 2 (t, ).
(a) Fourier Transform of the Product Density of Degree One
Combining Eq. 5 and Eq. 9, the Fourier transform of f 1 (t) is
readily found to yield a continuous density,
F 1
1
2
f 1 t e j t dt 1 * 1
(10)
where is the characteristic function of the first time of arrival

T 1 or equivalently of the independent and identically distributed
inter-arrival time process T i . Equation 10 is known as a
renewal type equation in the theory of stochastic processes, the
study of which requires the exact knowledge of the probability
law governing T 1 . The Gamma law would probably be a good
candidate here as it produces strictly positive inter-arrival times
with a peaked probability around the mean value T. However
when its variance 2 is small w.r.t. its mean T, the Gamma distribution is well approximated by the Normal distribution with the
same mean and variance, thus making the calculations more tractable. Under these assumptions,
exp
1 2 2
j t
2
(11)
from which Eq. 10 is readily found to yield a pole at 0 and

a series of finite-energy peaks equi-spaced by 1/T, with maxima
and minima respectively on k/T and k/T1/2, kZ. Figure 3 depicts the behavior of the modulus F 1 ( ) where the
percentage of random fluctuation /T has been set rather large
for sake of demonstration. In contrast to the time domain, note
firstly i that the magnitude of the peaks falls off more rapidly in
the Fourier domain and secondly, ii that the bandwidth of the
successive peaks of F 1 ( ) remain more or less constant since
they are bounded by 1/T.
(i) Fall off of the peaks. For 2 T, the relative magnitude of
the ith peak w.r.t. the first one decreases almost as fast as 1/i 2 , that
is a slope of 40 dB per decade. Since F 1 ( ) ultimately tends
towards a constant amplitude density F 1 ()1, this means that
there exists a cut-off radian frequency c 2/ after which all
the peaks have faded. Considering the percentage of random fluctuation x/100 /T, it is found that F 1 ( ) becomes almost constant after i c 22.5/x. For example, for a random fluctuation of
2%, this means as few as i c 11 peaks.
284 Vol. 125, JULY 2003
Fig. 4 Double Fourier transform modulus of the product density of degree two for T 130
(ii) Quality factor. The apparent increase of the bandwidth

resulting from the fall off of the peaks may be quantified by the
quality factor Q i T 2 /(i ) 2 ratio of the maximum to the
minimum of the ith peak. This is naturally found to strongly
depend on the percentage of random fluctuation /T.
(b) Double Fourier Transform of the Product Density of Degree Two
The formula for F 1 ( ) can now be used to compute the double
Fourier transform of f 2 (t, ):
F 2 ,
1
42
R2
f 2 t, e j t e j dtd
(12)
Distinguishing three cases 0, 0 and 0 and after some

algebra, one finds
F 2 , F 1 F 1 F 1 1
(13)
This defines a continuous spectral density with marked ridges running along the -variable and centered on all k/T, kZ. For
large values of , F 2 ( , ) ultimately tends to F 1 ( ) and thus
resembles a pattern of parallel and horizontal ridges along the
-axis rapidly falling off on each side of 0, as illustrated in
Fig. 4. The presence of these parallel ridges distinctively characterizes the second-order spectral signature of a random train of
impact forces and consequently, that of a faulty rolling element
bearing. Moreover, the distance between the ridges indicates the
mean rate of occurrence of the fault, thus enabling its identification in the mechanical system.
2.4 Discussion. At this stage, it is instructive to review
some former models proposed in the literature in light of the derived results. Clearly, for the deterministic model f 1 (t) is a perfectly periodic train of Dirac deltas XT (t) and f 2 (t, ) a twodimensional version of it, viz f 2 (t, )XT (t)XT ( ). The same
applies to their respective Fourier transforms, viz F 1 ( )1/T
X1/T ( ) and F 2 ( , )1/T 2 X1/T ( )X1/T ( ). The limitation of these formulas arises from the experimental evidence that
actual data do not have line spectra especially in the vicinity of
the high frequency resonance where they are usually demodulated. On the other hand, the simplified stochastic model proposed
by Randall & Antoni in 5 leads to f 1 (t)XT (t) * 1 (t), that is
a periodic train of Dirac delta functions low-pass filtered by the
probability function 1 (t). Therein f 2 (t, ) turns out to be periodic and low-pass filtered in the t-variable while transient in the
-variable. These low-pass filter and transient effects give a
better explanation for the continuity of experimental spectra in the
vicinity of a high frequency resonance 4. The refined stochastic

model proposed herein obviously leads to almost identical properties, yet on the basis of more accurate physical considerations.
Of particular concern are the facts that f 1 (t) and f 2 (t, ) are no
longer periodic functions neither in the t nor in the variables and
that the low-pass filter effect is now replaced by a rapid falloff effect. In other words, the Fourier transforms F 1 ( ) and
F 2 ( , ) are now purely continuous functions in both and
except at 0) where all the former discrete lines harmonics
have been replaced by distributed peaks, gradually broadening
with increase in . Actually, this fact is always observable in
envelope spectra, and was one of the main reasons for modifying
the stochastic model first mentioned in reference 5.
P2: b(t) has an effective duration shorter than the mean interarrival time T, or equivalently its spectral bandwidth is larger
than the mean rate of impacts 1/T.
Therefore, the overall impulse response of the system is obtained from cascading the amplitude modulation function A(t)
with the time-varying impulse response r(t, ) and finally with the
band-pass filter b(t). This is illustrated in Fig. 5b.
In this procedure, A(t) and g(t, )b(t)r(, ) have some
important properties which will make the computation of the
input-output relationship tractable. Specifically, because A(t) encompasses all the periodic modulations with possible stochastic
effects, it has first and second-order statistics given by:
m A t E A t m A t
Spectral Statistics of the Vibration Signal
In the preceding section, expressions were derived which describe the spectral signature of a train of pulses as produced by a
localized defect. This section now discusses how this spectral signature is transformed after the impacts have propagated through
the system, i.e. as it is likely to be measured on the bearing housing by an accelerometer.
3.1 Response of a Rolling Element Bearing to a Random
Train of Impacts. Following classical models, the vibration signal produced by a faulty rolling element bearing may be viewed
as the response of a linear system driven by the impacting process
F(t) 2,7. For this input-output relationship to be fully comprehensive, we now show that the impulse response of the system
should be time-varying and should also accommodate some degree of stochasticity.
At the outset, the impacting process should be modulated by a
periodic and positive function A(t) to account for the variations in
the impact magnitudes as the defect enters and exits the load zone
2. Some random modulation might be incorporated in A(t) due
to the dependence on the position and the number of the rolling
elements in the load zone at time t, but also due to rolling and slip
on possibly rough surfaces especially after a defect has appeared
and spread to some extent.
Next, let us define r(t, ) the structural response at time t of the
system subjected to an impulse at time . In contrast to a
static structure, the impulse response r(t, ) of a rolling element
bearing is time-varying for a variety of physical reasons, the most
obvious of which being the variations in the transmission path as
the coordinates of the point of impact move w.r.t the location of
the sensor, and the variations in the relative angle between the
impact forces and the axis of the sensor. For a system operating at
constant speed, these variations periodically affect the magnitude,
the sign and the phase of the impulse response r(t, ) with a
period depending on whether the defect lies on the inner race, the
outer race or on a rolling element 2. In addition, r(t, ) might
have some small random fluctuations to account for unpredictable
effects such as contact non-linearities. Figure 5a gives a schematic illustration of how the impacting process A(t)F(t) is transformed into a vibration X(t) after passing through the impulse
response r(t, ).
In practice, the vibration signal X(t) produced by a local fault
cannot be observed totally because it is contaminated by other
vibrations from a multitude of neighboring sources in the system.
Therefore, it is customary to filter it in a frequency band where the
signal-to-noise ratio is maximum so that virtually no other sources
than that stemming from the faulty bearing are measured by the
experimenter. This is usually done by designing a band-pass filter
b(t) around a high-frequency resonance of the structure or the
sensor that is excited by the impacts 1 4. In order to retain the
diagnostic information, the band-pass filter b(t) must have the
following properties:
P1: b(t) is a band-pass filter with central frequency 0 much
higher than the mean rate of impacts 1/T,
and
a k e jt
kZ
R A t, E A t A * t R A t
(14)
2
R Ak e jt
,
kZ
(15)
Equations 14 and 15 define a second-order cyclostationary

process, i.e. a stochastic process with periodic mean and autocorrelation function of intrinsic period 2/. Therein is either
equal to the speed of the inner race, that of the outer race or that
of the cage relative to the load vector whether the fault is on the
inner race, the outer race, or on a rolling element.
Similarly, g(t, ) being a periodic causal Greens function describing the periodically varying transmission path, it expands
into:
g t,
g t
2
,
g k t e jt ,
kZ
(16)
From the above expansion, the mechanism relating the impacting

process F(t) to the band-pass vibration signal Y (t) can finally be
obtained from the following Stieltjes stochastic integral
Y t
g t, A dN
jt
kZ
g k t A dN
(17)
in which each Fourier coefficient g k (t) is to be interpreted as a
linear, causal and homogeneous impulse response.
3.2 Spectral Characteristics of the Vibration Response.
From Eq. 17, the spectral characteristics of the band-pass vibration signal can now be derived and applied to a number of potential indicators for use in diagnostics, namely the Fourier transform
of the expected signal, the power spectral density, the spectral
correlation, the Fourier transform of the expected squared signal
and the power spectral density of the squared signal.
(a) Fourier Transform of the Expected Response
From Eqs. 1, 14 and 17, the expected value ensemble
average of the vibration signal is
m Y t E Y t
g t, m A f 1 d
(18)
from which the Fourier transform is found to be:
M Y
1
2
m Y t e j t dt
kZ
Y k k
1
M Y k G k F
F 1 M A * F 1
(19)
a F l
lZ
JULY 2003, Vol. 125 285
Fig. 5 a Generation of the resulting vibration signal. A t : magnitude of the

impacts; r t , : time-varying stochastic structural impulse response at time t .
b Scheme of the overall impulse response. b is a band-pass filter that extracts the bearing signal where its signal-to-noise ratio is the highest
where G k ( ) and M A ( ) are respectively the Fourier transforms

of g k (t) and m A (t). In view of Eq. 19, M Y ( ) is a superposition
of shifted functions M Y k ( ); each of them being in turn constructed from shifted and scaled replicas of F 1 ( ) and then
weighted by the frequency response G k ( ). The construction of
M Y ( ) results in a mixed spectrum with a family of infiniteenergy pseudo-harmonics around 0, all equi-spaced by the rotation speed . The number of these pseudo-harmonics directly
depends on the number of Fourier coefficients in g(t, ) and
m A (t). Note that this specific pattern repeats around all the peaks
of F 1 ( ) at k/T, kZ, but with finite-energy peaks in place
of pseudo-harmonics for any k0. Obviously, this makes the very
distinctive spectral signature of a localized defect as it is expected to appear in a faulty rolling element bearing see Fig. 6.
Most importantly its detection forms the main basis of diagnostics
since it contains the key characteristic frequencies 1/T and that
enable the identification and the localization of faults in complex
systems.
However, the problem in the spectral indicator of Eq. 19 is
that the frequency support of the band-filters G k ( ) is very likely
to be higher than the frequency support of the spectral signature
F 1 ( ) as shown schematically in Fig. 7. In fact, it was already
pointed out that F 1 ( ) falls off by 40 dB per decade down to a
cut-off radian frequency c of about &/ . In order for G k ( )
286 Vol. 125, JULY 2003
to overlap with this support, its central radian frequency 0 resonance frequency chosen for demodulation should be such that
0 T100&/x with x/100 /T the percentage of fluctuation.
Or equivalently, with i c the number of peaks in F 1 ( ) before it
dies to 1, 0 should be such that 0 T2 i c . In most instances
this condition would not be satisfied if a good signal-to-noise ratio
were to be maintained, thus justifying the poor performance anticipated from the Fourier transform of the vibration signal.
(b) Power Spectral Density of the Response
From Eqs. 4, 15 and 17, the autocorrelation function of the
vibration response is:
Fig. 6 Typical spectral signature in the vibration signal for

T 130 and T 10
Fig. 7 Illustration of the low-pass filter effect
R Y t, E Y t Y * t
h t ,
Fig. 8 Scheme of the spectral correlation density
h * t, R A , f 2 , dd
(20)
This is a bivariate function since the vibration signal resulting

from the impacting process F(t) is nonstationary. In order to compute the power spectral density, let us first denote by R y ( ) the
stationarized autocorrelation function
1
W
W
R Y lim
R y t, dt
(21)
whose Fourier transform then yields the explicit expression for the
power spectral density:
SY
1
2
R Y e j d
kZ
y k k

S Y k G k 2 F
2
(22)
F 2 Re F * S
2
1
A
T
where S A ( ) is the Fourier transform of the stationarized version

R A ( ) of R A (t, ) in Eq. 15. The set of Eq. 22 indicate that the
principle of construction of S Y ( ) is similar to that outlined for
the Fourier transform in Eq. 19 because S A ( ) contains the
same discrete harmonics as F A ( ) and G k ( ) 2 obviously acts in
the same frequency band as G k ( ). Therefore, the same conclusion holds in regard to the expected performance of the power
spectral density as a diagnostic indicator.
(c) Spectral Correlation Density of the Response
We now demonstrate that the aforementioned shortcomings due
to the non-intersection of the low-pass and band-pass filters see.
Fig. 7 can be solved by considering the double Fourier transform
of the autocorrelation function R Y (t, ). This yields a quantity
called the spectral correlation density2 5, very similar to the
generalized spectrumwithin a simple change of variable
used by Lin 9. The spectral correlation density,
SY ,
1
42
R Y t, e j e j t d dt
(23)
is found to have explicit expression
SY ,
k,lZ2
S y k y l k, l
2 ,
S Y k Y l , G k G l* F
F 2 ,
(24)
F p, * S
pZ
p
A
where S Ap ( ) is the Fourier transform of R Ap ( ) in Eq. 15. Although involving two frequency variables, the construction of
S Y ( , ) is again similar to that outlined in Eqs. 19 and 22.
Nevertheless, there is now a domain in the frequency plane ,
2
There is a simple relation between the spectral correlation density and the power
spectral density, viz S Y (0, )S Y ( ) ( )
where the diagnostic information is totally preserved. Specifically,

for small values of within twice the bandwidth of g(t, ) and
large values of around the resonance frequency 0 , S Y ( , )
clearly displays the spectral signature of a defect because the
band-pass filters G k ( ) and G l ( ) are band-passing in this
area. This is a direct consequence of property P2, a schematic
illustration of it being shown in Fig. 8. Note that in the domain of
concern S Y ( , ) is markedly ridged in the horizontal -direction
just as F 2 ( , ) was in Fig. 4.
(d) Fourier Transform of the Expected Squared Response
Because it is bivariate, the spectral correlation density may be
difficult to compute and therefore it has been suggested to replace
it by its integrated version over the -variable while preserving
the diagnostic information. In reference 5, this was shown to be
equivalent to the Fourier transform of the expected squared signal,
i.e.
M Y 2
SY , d
1
2
E Y t 2 e j t dt
(25)
This equation is easily found to be identical in structure to Eq.

19 where the coefficients a 1 are replaced by R Al (0) defined in
Eq. 15, and where the band-pass filter G k ( ) is replaced by the
low-pass filter P k ( ) p G k ( )G kp ( ). The fact that P k ( )
is now necessarily a low-pass filter comes from the convolution of
G k ( ) by itself and this is exactly the reason why M Y 2 ( ) can
preserve the diagnostic information whereas M Y ( ) cannot. Indeed, under property P2 the support of P k ( ) necessarily overlaps
with that of the spectral signature of the fault, contrary to the
scheme of Fig. 7.
A last point to consider is whether to take the square of the raw
signal in Eq. 25 or the squared magnitude of its analytic version.
Strictly speaking, the analytic signal should be used so that the
expectation of its squared magnitude truly gives the squared envelope. However, minor differences would be found when using
the real signal provided it is properly band-pass filtered around a
resonance. This point was also addressed in a lot of detail in
reference 4.
(e) Power Spectral Density of the Squared Response
In light of the previous demonstration, one can expect the
power spectral density S Y 2 ( ) of the squared signal to perform
just as well as M Y 2 ( ). Indeed, S Y 2 ( ) is the exact definition of
the spectrum of the squared envelope as was proposed in 4. In
order to prove this result, let us invoke property P1 in conjunction
with the assumption that the point process dN(t) is regular see
section 2.1. Then, the following approximation holds
Y t 2
p t, A 2 dN
(26)
with p(t, ) g(t, ) 2 . After taking the Fourier transform of the

stationarized autocorrelation function of Y (t) 2 , the envelope
spectrum is found to be identical to Eq. 22 where S A ( ) is
replaced by S A 2 ( )the Fourier transform of the stationarized
JULY 2003, Vol. 125 287
Fig. 9 Power spectral density of a vibration signal in case of no fault continuous line and an
inner race fault dotted line
A 2 ( ) of the squared process

autocorrelation function R
A(t) 2 and where G k ( ) 2 is replaced by P k ( ) 2 . The fact
that P k ( ) 2 acts as a low-pass filter demonstrates again that
S Y 2 ( ) is also a usable diagnostic indicator. Here again, the analytic version of the signal may be preferred in Eq. 26 in order to
estimate the power spectral density of the true squared envelope.
3.3 Discussion. It has been proven in some depth why the

Fourier transform and the power spectral density generally are
poor indicators for diagnosing rolling element bearings in the case
of localized faults, a fact that the authors have regularly observed
on experimental data. Indeed, even though classical spectral
analysis may perform very well in detecting a faulte.g. through
monitoring the relative energy levels in some frequency bandsit
rarely helps in recognizing its type nor its locationand this is
exactly what diagnostics asks for. For example, Figure 9 compares
the power spectral densities of a vibration signal measured on a
gearbox before and after one of the rolling element bearings 12
balls, 7.12 mm, pitch circle 38.5 mm was purposely damaged
by machining a small slot on its inner race. The frequency resolution is 12 Hz. Note that the presence of the fault only shows up
at high frequencies. The fact that there is no difference at low
frequencies is due to the extremely poor signal-to-noise ratio in
that band observe that most of the sources there relate to harmonics from the gears. Of interest also is the fact that in spite of its
increase, the spectral density at higher frequencies is continuous
and therefore gives no indication of a fault producing repetitive
impacts.
In clear contrast with the Fourier transform and the power spectral density, the same transformations applied on the squared signal or its analytic version have been shown to solve the problem
in a surprisingly simple manner. For example, Fig. 10 displays the
power spectrum of the squared magnitude of the analytic signal
after band-pass filtering in the frequency band 1.8; 2.2 kHz with
a frequency resolution of 2 Hz. Now the specific spectral signa-
Fig. 10 Power spectral density of the squared envelope

Table 1 Comparison of five spectral indicators in terms of
their ability of detecting and identifying localized faults.
Spectral indicators
Fourier transform of expected signal
Power spectrum of the signal
Spectral correlation density 2-D of the signal
Fourier transform of the expected squared signal
Power spectrum of the squared signal
288 Vol. 125, JULY 2003
Diagnostic skills
ture of the inner race fault shows up in good accordance with Fig.
4, with 1/T71 Hz ball pass frequency on the fault and
10 Hz shaft rotation.
From a theoretical point of view, it is worth mentioning that the
Fourier transform of the squared signal preserves the diagnostic
information by exploiting the non-stationarity of the signal, while
the power spectral density exploits its non-Gaussianity. In fact, the
power spectral density of the squared signal is implicitly a fourthorder stationarized statistic. Strictly speaking, these two indicators have different theoretical justifications and this is supported
by recalling that the former only requires property P2, while the
latter requires the more stringent condition P2P1. However,
both are inclined to provide envelope analysisor squared envelope analysiswith a strong formal justification.
Comparison of the five spectral indicators, which have been
assessed so forth, is summarized in Table 1.
Conclusion
A comprehensive stochastic model has been proposed for describing and simulating the vibration produced by localized faults
in rolling element bearings. Sources of stochasticity were modeled
in both the impacting force processby means of a regular point
processand in the transmission pathby means of a cyclostationary process, thus encompassing a large range of physical situations. These refinements proved very valuable in explaining
some of the actual features observed on experimental data. The
spectral signature of a localized fault was derived analytically and
new results were deduced concerning the nature of spectral harmonics produced by the impacting process. These were shown to
be distributed and equi-spaced by the mean rate of impacts
peaks with a rapid fall-off that could be quantified as a function of
the percentage of stochastic fluctuations. Next, the spectral signature of a defect was shown to duplicate when it propagates
through the structure with shifts equal to the rotation speed of the
defect, thus generating additional families of pseudo-harmonics.
These results finally helped in investigating the effectiveness of a
number of spectral indicators dedicated to the diagnostics of rolling element bearings. From simple considerations on band-pass
and low-pass filtering operations, it was demonstrated that both
the Fourier transform and the power spectral density of the
squared signal are the most relevant indicators, thus bringing new
supports in favor of squared envelope analysis.
References
1 Darlow, M. S., and Badgley, R. H., 1975, Applications for Early Detection of
Rolling Element Bearing Failures Using the High-Frequency Resonance Technique, ASME Paper 75-DET-46.
2 McFadden, P. D., and Smith, J. D., 1984, Model for the Vibration Produced
by a Single Point Defect in a Rolling Element Bearing, J. Sound Vib., 911,
pp. 69 82.
3 McFadden, P. D., and Smith, J. D., 1985, The Vibration Produced by Multiple Point Defects in a Rolling Element Bearing, J. Sound Vib., 982, pp.
69 82.
4 Ho, D., and Randall, R. B., 2000, Optimization of Bearing Diagnostics Techniques Using Simulated and Actual Bearing Fault Signals, Mech. Syst. Signal
Process., 145, pp. 763788.
5 Randall, R. B., Antoni, J., and Chobsaard, S., 2001, The Relationship Between Spectral Correlation and Envelope Analysis in the Diagnostics of Bearing Faults and other Cyclostationary Machine Signals, Mech. Syst. Signal
Process., 155, pp. 945962.
6 Antoni, J., and Randall, R. B., 2002, Differential Diagnosis of Gear and
Bearing Faults, ASME J. Vibr. Acoust., 127, pp. 17.
7 Roberts, J. B., 1966, On the Response of a Simple Oscillator to Random
Impulses, J. Sound Vib., 41, pp. 51 61.
8 Srinivasan, S. K., et al., 1967, Response of Linear Vibratory Systems to
Non-Stationary Stochastic Impulses, J. Sound Vib., 62, pp. 169179.
9 Lin, Y. K., 1965, Nonstationary Excitation and Response in Linear Systems
Treated as Sequences of Random Pulses, J. Acoust. Soc. Am., pp. 453 460.
JULY 2003, Vol. 125 289
Nabeel Shabaneh
Graduate Student
Jean W. Zu
Associate Professor
Department of Mechanical & Industrial
Engineering,
University of Toronto,
5 Kings College Road,
Toronto, Ontario, Canada, M5S 3G8
Nonlinear Dynamic Analysis

of a Rotor Shaft System With
Viscoelastically Supported
Bearings
This research investigates the dynamic analysis of a single-rotor shaft system with nonlinear elastic bearings at the ends mounted on viscoelastic suspension. Timoshenko shaft
model is utilized to incorporate the flexibility of the shaft; the rotor is considered to be
rigid and located at the mid-span of the shaft. A nonlinear bearing pedestal model is
assumed which has a cubic nonlinear spring and linear damping characteristics. The
viscoelastic supports are modeled using Kelvin-Voigt model. Free and forced vibration is
investigated based on the direct multiple scales method of one-to-one frequency-toamplitude relationship using third order perturbation expansion. The results of the nonlinear analysis show that a limiting value of the internal damping coefficient of the shaft
exists where the trend of the frequency-response curve switches. Also, the primary resonance peak shifts to higher frequencies with the increase of the bearing nonlinear elastic
characteristics, but with a flattened curve and hence lower peak values. A jump phenomenon takes place for high values of the bearing nonlinear elastic characteristics.
DOI: 10.1115/1.1547684
Introduction
An effective means for controlling and reducing vibrations in

rotating machinery is the use of external damping and elastic elements often provided via flexible bearings and/or bearing supports. Viscoelastic materials have emerged recently as external
mechanical dampers and bearing supports, offering simplicity in
design, operational enhancement and low cost in comparison with
the traditionally used squeeze film dampers 1.
Modeling of the dynamic behavior of vibrating systems incorporating viscoelastic elements has not been possible until recently
due to the elaborate work of identifying the characteristic behavior of viscoelastic materials. Although there is much research on
modeling systems for structural applications, the work on rotating
shaft bearing systems mounted on viscoelastic supports is limited.
In the few works reported on linear analysis, Dutt and Nakra
2 4 studied the stability and the unbalance vibration response of
a Jeffcott rotor system mounted on viscoelastic supports where a
discrete model of the shaft and linear elastic bearings were assumed. Using the finite element method based on the EulerBernoulli beam model, Kulkarni et al. 5 investigated the unbalance response and the stability of a rotating system with
viscoelastically supported bearings. In further investigations, Shabaneh and Zu 6 8 studied the dynamic behavior of a rotating
disk-shaft system with linear elastic bearings at the ends mounted
on viscoelastic supports. Timsohenko model was assumed for the
shaft, and thus, the gyroscopic effect of the shaft was incorporated
in the analysis. Free and forced vibration analysis were carried out
and the effects of the system characteristics were presented.
Routh-Hurwitz criterion for polynomials with complex coefficients and the variational analysis were used to perform the stability analysis of the system. Furthermore, experimental investigations were performed to verify the obtained theoretical results.
Bearings in rotor shaft systems may possess nonlinear characteristics. For example, nonlinearity is inherent in ball bearings due
to Coulomb friction and the angular clearance between the roller
and the ring 9. Ji and Zu 10 performed free and forced vibration analysis to calculate natural frequencies for nonlinear rotor
bearing system with cubic nonlinearity. However, incorporating
nonlinear bearing characteristics in rotor-shaft systems with viscoelastic supports is very limited. A preliminary investigation was
reported by Bhattacharyya and Dutt 11. They studied the unbalanced response and stability of a rotor shaft system mounted on
nonlinear rolling element bearings with viscoelastic supports. The
shaft was assumed to be massless with linear elasticity and internal damping. In addition, the nonlinear restoring force of the bearings was linearized by the method of effective linearization, enabling an approximate stability analysis using the Routh Hurwitz
criterion. The effects of gravity on the system and nonlinearity
were discussed. In the above work, the system was discretized and
the restoring force was approximated by a linear relation.
The aim of this research is to investigate the nonlinear dynamic
behavior of a continuous rotor shaft system with viscoelastically
supported bearings. The nonlinearity occurs at the boundaries due
to nonlinear characteristics of the bearings. Timoshenko shaft
model is used for the shaft, Kelvin Voigt model is utilized for the
viscoelastic supports, and typical roller bearings with cubic nonlinearity are employed. In the analysis, free and forced vibration
are developed based on the direct multiple scales method of oneto-one frequency-to-amplitude relationship of nonlinear system
using the third order perturbation expansion. This approach does
not require the selection of an orthogonal basis; however, it requires additional eigenfunction solutions at every level of approximations. Moreover, the direct multiple scales method yields better
results than the discretized one for finite mode truncations and for
systems having quadratic and cubic nonlinearities 12. The results of the nonlinear analysis indicate that a limiting value of the
internal damping coefficient of the shaft exists where the trend of
the frequency-response curve switches. Also, a jump phenomenon
takes place for high values of the nonlinear elastic coefficient of
the bearings.
2
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received February
2002; Revised October 2002. Associate Editor: G. T. Flowers.
290 Vol. 125, JULY 2003
Equations of Motion
Consider a continuous shaft-rotor system, as shown in Fig. 1,

where the frame oxyz is the inertial frame. A uniform shaft of
Fig. 1 Rotor shaft system with viscoelastically supported

bearings
length 2l is mounted on viscoelastically supported bearings. The

rotor is considered as a rigid disk. For convenience, the viscoelastic element is represented by its equivalent complex stiffness k *
and is located at both ends of the shaft rotor system underneath
the bearings as support.
Adopting a continuous model of the shaft based on the Timoshenko beam theory, and defining u x , u y , x , and y as the transverse deflections along the ox and oy directions and the corresponding bending angles in the oxz and oyz planes, respectively,
the complex variables u and are assumed as
uu x iu y ,
x i y
(1)
The general equations of motion of a uniform Timoshenko shaft

system with internal damping are
Q 1,t
2 u 1,t
1
M d
M d e 2 exp it
2
t2
(9)
1,t 0
(10)
where M d and e are the mass of the disk and the eccentricity of
the rotor, respectively. F l and F n are the linear and nonlinear
terms exerted by the bearings on the shaft and the viscoelastic
supports. Assuming that the bearings possess cubic nonlinear restoring forces in addition to linear damping forces, the associated
forces are
F l F xl iF yl k b1 u v c b
u v
t t
(11)
Au AG u C i u iu 0
(2)
F n F xn iF yn k b3 u x v x 3 i u y v y 3
I s iJ s EI s AG u 0
(3)
Equation 14 can be rewritten in terms of the complex transverse deflections u and v , and their corresponding complex conjugates u and v , as
where ( )D t / t and ( ) D z / z. A, I, J s and are the

cross-sectional area, transverse moment of inertia, polar mass moment of inertia per unit length and mass density of the shaft. C i is
the internal viscous damping coefficient and is the spin rate of
the shaft. E, G and are Youngs modulous, shear modulous and
shear coefficient, respectively.
The viscoelastic support is assumed to have a mass of M 2 and
is modeled using Kelvin-Voigt model. Thus, the equivalent complex stiffness is defined as k * k v (1i ). Denoting the displacements of the viscoelastic mass M 2 along x and y directions by v x
and v y , respectively, the complex deflection of the viscoelastic
element is given by v v x i v y .
Introducing the nondimensional space variable z/l, the relationship between the stress resultants and displacements in a
complex form are
M ,t M x ,t iM y ,t
EI s ,t
l
1 u ,t
Q ,t Q x ,t iQ y ,t AG
,t
l
Q 0,t F l 0,t F n 0,t

M2
2v
k v 1i v F l 0,t F n 0,t
t2
4 u 3 3u
2v 3u
v 2 v 3
(13)
Introducing a set of nondimensional quantities given in Appendix A, the nondimensional equations of motion with the associated
boundary conditions can be written as
2u * * 2u *
u*
C i
i * u * 0
t *2
2
t*
(14)
*
2 *
u*
2 *
2
3 *
0
2 i 1 *
t*
t*
2
(15)
M * 0,t 0
(16)
* u * v * c *
Q * 0,t * k b1
b
(5)
where M ( ,t) and Q( ,t) are the transverse bending moment and
the shear force at each cross section along the shaft.
Due to symmetry of the system, only the left half of the rotorshaft system, i.e., 0zl, is considered. Thus, the associated
boundary conditions can be defined as
M 0,t 0
F n k b3 43 u 2u 2uu
v v 2u u 2v 2u vv
v 2v
(4)
(12)
u* v*
k *
b3 W *
t* t*
0,t *
(17)
v * 0,t
k v* 1i v * 0,t
t *2
2
M 2*
k *
b1 u * v * c b*
u* v*
* W*
k b3
t* t*
0,t *
(6)
(18)
(7)
u * 1,t 1
1
M*
F * * 2 exp i * t * (19)
2 d t *2
2
(8)
Q * 1,t
* 1,t 0
(20)
JULY 2003, Vol. 125 291
Method of Multiple Scales
The direct multiple scales method is used in this section to

solve for the free and forced vibration of the system. Expansions
of the displacements u * , v * and the shear deformation * are
assumed as
u * u 1 ,T 0 ,T 2 u 3 ,T 0 ,T 2 . . .
(21)
v * v 1 ,T 0 ,T 2 v 3 ,T 0 ,T 2 . . .
(22)
* 1 ,T 0 ,T 2 3 3 ,T 0 ,T 2 . . .
(23)
where u i , v i and i (i1,3, . . . ) are of order O(1), is a small

dimensionless parameter, and T n n t * (n0,2, . . . ) is the nth
order time scale. T 0 is a fast-time scale characterizing motions
occurring at the spin rate * and natural frequency *
n of the
rotor bearing system; n* n l 2 / G. Furthermore, T 2 is a
slow-time scale characterizing the modulation of the amplitude
and phase due to nonlinearity, damping and resonance.
Considering the one-to-one primary resonance of the system,
the damping forces, the forcing term and the excitation frequency
are assumed as
c b* 2 c ,
C i* 2 C i ,
F * 3 F ,
* n* 2
(24)
where O(1) is a detuning parameter. The assumptions in Eq.

24 takes into account the interaction of damping terms with the
nonlinear forces at the same level of approximation, which is a
necessary condition for a nontrivial solution of the governing
equations of motion. Thus, the effect of the nonlinearity of the
system can be balanced with the effect of the system damping
at the same level of approximation, i.e., the third order
approximation 13.
The small dimensionless parameter may take any value between 0 and 1.0. However, by assuming a value of zero, the solution will be trivial. Moreover, in order to obtain a formal solution of the equations of motion 1420, should be equated to
1.0. In such a case, it is stipulated that the nonlinear terms are
themselves small.
Using the chain rule for the partial derivatives with respect to
both time scales T 0 and T 2 , substituting Eqs. 2124 into Eqs.
1420, and equating coefficients of like power of , the following approximations of the system are obtained:
Order :
D 20 u 1
1 u1
C i D 0 u 1 i n* u 1 0
2
2
2 1
u1
D 20 1 i 1 *
0
2 3 1
n D 0 1 2
at 0
1
0,
u1
* u 1 v 1 0,
1 k b1
2
*
M*
2 D 0 v 1 k *
v 1i v 1 k b1 u 1 v 1 0
at 1
1 0,
(25)
(27)
M*
d
(28)
2D 2 D 0 u 1 C i D 2 u 1 i u 1
(30)
* W 1 , and
c D 0 u 1 v 1 k b3
2
*
M*
2 D 0 v 3 k *
v 1i v 3 k b1 u 3 v 3
2M 2* D 0 D 2 v 1 c D 0 u 1 v 1 k *
b3 W 1
(31)
at 1
3 0,
M*
u3
d
1
2
D 20 u 3 M *
3
d D 2D 0u 1 2 F *
n
2
exp i n* 2 t *
where
W 1 41
(32)
u 21u 1 2u 1u 1 v 1 v 21u 1 u 21v 1 2u 1 v 1v 1 v 21v 1
41 u 31 3u
21v 1 3u
1v 21 v 31 .
3.1 First Order Equations. Noting that the equations of
motion and the associated boundary conditions are linear at order
, the first order approximation of the deflections and deformations can be assumed in the form
u 1 ,t * A T 2 e i *n T 0 Y u ,
1 ,t * A T 2 e i *n T 0 Y ,
v 1 t * A T 2 e i *n T 0 Y v
(33)
Substituting Eq. 33 into Eqs. 25 and 26 leads to a set of
ordinary differential equations given by
a 1 Y u a 2 Y u a 3 Y 0
(34)
b 1 Y b 2 Y b 3 Y u 0
(35)
where ( ) d/d . The coefficients in Eqs. 34 and 35 are defined as

2
a 2 *
n ,
a 1 1,
b 1 2 ,
a 3 1
b 2 1 1 *
n 3 ,
(36)
b 3 3
(37)
Decoupling Eqs. 34 and 35 and assuming an exponential

solution form for Y u and Y yield the general solutions
4
Ae
j1
s j
Be
j1
s j
(38)
where s j ( j1, . . . ,4) are the roots of the quartic auxiliary equation 7. Only four of the arbitrary complex constants A j and B j
( j1, . . . ,4) are independent, since they are related by the
relation
a 1 s 2j a 2
a 3s j
j1, . . . ,4
(39)
Substituting Eqs. 33 and 38 into the boundary condition

Eqs. 27 and 28 yields the eigenvalue problem represented in
matrix form as
M*
n Z 0
3 2u 3
C i D 0 u 3 i n* u 3
3
u3
* u 3 v 3
0, 3
3 k b1
B j jA j ,
u1
1
D 20 u 1 0
292 Vol. 125, JULY 2003
at 0
Y u
and
2 3
u3
3 3
2
2D 2 D 0 1 i 1 *
n D 2 1 D 0 1
(26)
Order 3 :
D 20 u 3
D 20 3 i 1 n* D 0 3 2
(40)
in which the eigenconstants are defined in vector notation as

(29)
Z A 1 A 2 A 3 A 4 Y v T , and the 55 coefficient matrix M is

defined as
*
m 1k 3 s j j k b1
*
m 15k b1
m 2k j s j
m 250
*
m 3k k b1
2
*
m 35M *
2 *
n k *
v 1i k b1
m 4k e s j 3 s j j M d* n* 2 /2
m 450
m 5k j e
m 550
sj
To obtain the natural frequency *

n , the determinant of the
coefficient matrix M must be zero. Considering that the coefficient function A(T 2 ) in Eq. 33 is arbitrary and assuming A 1
1, the normal modes Y u ( ), Y ( ) and Y v are uniquely defined
by equation 40 for the eigenvalue *
n .
3.2 Third Order Equations. To achieve a third order solution of the system, assume the complex displacements u 3 , 3 and
v 3 to be
u 3 ,t * u ,T 2 e i *n T 0 ,
3 ,t * ,T 2 e i *n T 0 ,
v 3 t * v T 2 e i n* T 0
(42)
Substituting Eq. 42 into the set of Eqs. 2932 yields

u
2
1 1 *
n 3
2
(43)
(44)
at 0
0,
u
* u v
k b1
c i *
n A T 2 Y u 0 Y v
3
k * A 2 T 2 A T 2 W 0 ,
4 b3
and
* 1i v k b1
* u v
n* M *
2 v k v
2i n* M 2* A T 2 Y v i n* c A T 2 Y u 0 Y v
* A T 2 A T 2 W 0
k b3
2
(45)
at 1
In order to obtain a non-trivial solution for the homogeneous

part of Eqs. 43 and 44, the nonhomogeneous Eqs. 43 and
44 must satisfy a solvability or compatibility condition. As discussed in 13, the solvability condition demands that the right
hand side of the nonhomogeneous equations be orthogonal to every solution of the corresponding adjoint homogeneous problem.
Since the corresponding homogeneous part of Eqs. 43 and
44 is self-adjoint, the solvability condition can be derived as
1
u 3 2i n* C i A T 2 Y u iC i A T 2 Y u
0,
u
2

*
n u
2
1
F * 2 exp i T 2
2 n
(46)
u 0,t * c i n* A T 2 Y u 0 Y v
1
2
i *
n M*
d A T 2 Y u 1 2 F *
n exp i T 2 (47)
where u and are the solutions of the adjoint homogeneous system of Eqs. 43 and 44 and the boundary condition Eqs. 31
and 32 with a zero right hand side.
It can be seen that the homogeneous adjoint system has an
identical form to the first order approximation given by Eqs. 25
28; hence, the solutions for u , and v should have the same
form as u 1 , 1 and v 1 .
Substituting the solutions of u and into Eq. 47, the solvability condition takes the compact form
b 1 2i *
n C i A T 2 b 1 iC i A T 2
b 2 i n* 2 1 A T 2 b 2 1 *
n A T2
1
2
2
i *
n M*
d A T 2 Y u 1 2 F *
n exp i T 2 Y u 1 (48)
where b 1 3 10 Y 2u ( )d , b 2 3 10 Y 2 ( )d , A is the complex

conjugate of the coefficient A, and A is the derivative of A with
respect to the slow time scale T 2 .
The coefficient A can be expressed in a polar form such that
A T 2 21 a T 2 e i T 2
M*
d
i n* M d* A T 2 Y u 1
i *
n 2 1 A T 2 Y 1 *
n A T 2 Y d
3
2
2
* A 2 T 2 A T 2 W 0 Y u 0
c i *
n A T 2 Y u 0 Y v 4 k b3
3
4
Solvability Condition
* A 2 T 2 A T 2 W 0 u
43 k b3
1,t *
i n* 2 1 A T 2 Y 1 n* A T 2 Y
2i n* C i A T 2 Y u iC i A T 2 Y u
(41)
where ( ) / T 2 , A (T 2 ) is the complex conjugate of the coefficient A(T 2 ), and W 0 Y 3u (0)3Y 2u (0)Y v 3Y u (0)Y 2v Y 3v .
Note that the above terms in Eqs. 45 and 46 are collected only
for the primary resonance, i.e., exp(in*T0).
2 u
2 *
n u
k1, . . . ,4
(49)
where a(T 2 ) and (T 2 ) represent the amplitude and phase angle

of the response, respectively. To eliminate the explicit dependence
on T 2 in the solvability condition and thus, transforming Eq. 48
from a nonautonomous system to an autonomous system, the new
dependent variable is introduced as
T 2
(50)
JULY 2003, Vol. 125 293
Substituting Eqs. 49 and 50 into Eq. 48 and separating the

real and imaginary parts, the modulation is given by
(51)
terms and all the damping terms are eliminated. Such frequencies
are called free frequencies of the nonlinear system, and they are
dependent among others on the initial conditions and strictly applicable at steady state. However, these frequencies should not be
confused with the linear system natural frequencies.
(52)
b 3 a b 4 a b 5 a b 6 a 3 b 7 aF n* 2 Y u 1 cos
b 4 a b 3 a b 8 a b 6 a b 9 aF *
n Y u 1 sin
2
where the real coefficients b 3 , . . . ,b 9 are defined in Appendix B.
Free Vibration Response
The free undamped frequencies of the nonlinear system can be

obtained by letting the quantities C i , c and F equal to zero in
Eqs. 51 and 52. Separating the variables and noting that b 7
b 9 0, Eqs. 51 and 52 become
c 1 a c 2 ac 3 a 3 0
(53)
b3 a b6 2
b5
a 1

b4 a b4
b4
(54)
where
c 2 (b 8 b 4 b 5 b 3 ) and c 3 b 6 b 4 b 6 b 3 .
Equations 53 and 54 can be solved in conjunction with the
initial condition aa 0 and 0 at T 2 0 to yield either of the
following three cases:
Case 1: c 2 /c 3 0
c 1 b 24 b 23 ,
a c 2 a 21 e 2c 2 T 2 /c 1 / c 3 a 21 c 3 e 2c 2 T 2 /c 1 c 3
T 2 0
b3 1
ln
b4 2
a 21 e 2c 2 T 2 /c 1
2
a 1 e 2c 2 T 2 /c 1 1
(55)
1 b3
ln
2 b4
a 21 1
a 21
b6
c1
ln a 21 e 2c 2 T 2 /c 1 1
b 4 2c 3 a 21 1
2a 21 c 2
c1
T 2 ln a 21 1
b5
T2
b4
T 2 0
b5
T2
b4
where a 22 (a 20 c 2 /c 3 )/a 20 .
Case 3: c 2 0
T 2 0
b3
2c 3 2
ln 1
a T
2b 4
c1 0 2
c1 b6
2c 3 2
b5
ln 1
a 0 T 2 1
T2
2c 3 b 4
c1
b4
(59)
(60)
The frequency of the nonlinear system *

nn is related to the
frequency of the linear system *
n through the relation
* n* 2
nn
(61)
where is the phase angle derivative with respect to the slow

time scale T 2 .
In a quasi-harmonic nonlinear system, the frequencies of the
nonlinear system can be assumed periodic if the external forcing
294 Vol. 125, JULY 2003
Numerical Results and Discussions
E2.041011 Pa
v 0.3
7750 kg/m3
d s 0.0286 m
c b 400 kg/s
0.68
l0.15 m
0.01
k b3 1107 N/m3
k v 1109 N/m
M 2 0.5 kg
(57)
(62)
Numerical simulations are performed for a typical rotor shaft

system with nonlinear bearings. The physical parameters of this
system are given as
(56)
(58)
a a 20 / 12c 3 a 20 T 2 /c 1
k b1 110 N/m
2c 2
b6
T ln a 22
ln a 22 e 2c 2 T 2 /c 1 1 ln a 22
c1 2
2b 4
b 14
1
b 2 4b 13b 15
2b 13 2b 13 14
where the coefficients of the detuning parameter b 13 , b 14 and b 15

are defined in Appendix C.
Equation 62 defines the frequency-response curve of the nonlinear system, and each point on the curve corresponds to a specific singular point on a state plane.
b3
c1
ln a 22 e 2c 2 T 2 /c 1 1
2c 3 b 4 a 22 1
C i 200 kg/m.s
where a 21 a 20 /(a 20 c 2 /c 3 ).
Case 2: c 2 /c 3 0
a c 2 / c 3 a 22 c 3 e 2c 2 T 2 /c 1 c 3
Forced Vibration Response
The forced vibration of the system can be obtained by including

the damping terms, C i , c and the forcing term F in Eqs. 51
and 52. To obtain the steady state response with periodic motion,
the rate of change of the amplitude and the transformed phase
shift should be zero, i.e., a 0.
Eliminating the transformed phase shift from both of the
modulation Eqs. 51 and 52, the detuning parameter can be
found as a function of the amplitude a
d d 0.13 m
M d 55 kg
e0.01 m
Assume that the shaft has a circular cross section, hence 1

2 and J s 2 I s . In addition, the shaft rotates at a typical speed
300 rad/s. Free and forced nonlinear vibration analysis are
carried out to this system to investigate the influence of nonlinearity in the bearings on the system characteristic frequencies and
frequency response.
Figure 2 shows the free nonlinear oscillations of this system for
different values of the elastic coefficient k b3 of the nonlinear bearing. The initial transverse displacement u(0) is assumed as 10
mm and the initial phase angle 0 is zero. The fundamental natural frequency of the above system is 458.71 rad/s. It can be seen
from the waveforms presented that increasing the nonlinearity of
the system, i.e., increasing k b3 , shifts the frequency of the nonlinear system to a higher value and increases the steady state
amplitude of the oscillation.
It is noted that the detuning parameter in the nonlinear free
vibration analysis is related to the nondimensional spinning frequency * through the relation ( * n* )/ 2 . Also, the nondimensional amplitude a can be related to the transverse deflection of the shaft u at the first order approximation as
u(T 0 ,T 2 , )l/2a(T 2 )e i ( *n T 0 ( T 2 )) Y u ( ). This relation can be
simplified for a zero initial phase angle 0 to become a
2u/(l).
Assuming that mass eccentricity of the disk has no contribution
to the nonlinear forcing terms, the forcing term F * defined in Eq.
24 will be equal to zero, and subsequently F will be equal to
zero. The effect of the bearing nonlinear elastic coefficient k b3 on
the free frequency of the nonlinear system versus the amplitude a
of the transverse deflection is illustrated in Fig. 3. It is shown that
as the amplitude increases, the corresponding shift in the nonlinear natural frequency remains minimal for low values of the elasTransactions of the ASME
Fig. 2 Waveform of the transverse deflection for various values of bearing

nonlinear elastic coefficient k b 3
tic coefficient k b3 of the nonlinear bearing; i.e., for weak nonlinear system. As k b3 increases, the change in the natural frequency
of the nonlinear system increases significantly. The increase of the
frequencies of the nonlinear system and the bend of the curve
towards the higher frequencies are typical trends of hardening
nonlinearities.
Figures 4 and 5 show the effect of the internal viscous damping
coefficient C i of the shaft on the free frequency of the nonlinear
system versus amplitudes. Figure 4 depicts the trend for low C i
values of 100 kg/m.s to 200 kg/m.s, whereas Fig. 5 shows the
trend for higher C i values of 270 kg/m.s to 300 kg/m.s. Figure 4
shows that increasing the internal viscous damping of the shaft
shifts the free frequency of the nonlinear system to higher values
while almost maintaining the curve characteristic. However, this
trend has a limiting value of C i 270 kg/m.s beyond which any
increase in C i will result in a shift of the curve to lower frequencies, as shown in Fig. 5. The limiting value of the internal damp-
ing coefficient, C i 270 kg/m.s, indicates the increased efficiency

of the system to dissipate energy. This may be attributed to the
optimum interaction between the nonlinear restoring forces, external damping provided by the bearing supports and the viscoelastic
elements, and the internal damping of the shaft.
It should be pointed out that since a zero forcing term corresponding to mass eccentricity of the disk is assumed for the results of both Figs. 4 and 5, the free frequencies of the nonlinear
system can be obtained. Also, the internal damping force is assumed to be linear; hence, proportional to the strain velocity,
which is determined by the difference in whirling angular velocity
and the rotational speed of the shaft. In general, the internal damping force hysteretic damping is nonlinear and roughly independent of the frequency. In addition, it is proportional to the square
of the vibration amplitude, to the shaft stiffness, and to the number
of reversals of the sign of the strain velocity 14.
Including the effect of the mass eccentricity and hence the ex-
Fig. 3 Free frequency of the nonlinear system versus amplitudes for various values of bearing nonlinear elastic coefficient k b 3
JULY 2003, Vol. 125 295
Fig. 4 Free frequency of the nonlinear system versus amplitudes for various values of internal viscous damping coefficient C i of the shaft
citation frequency , the steady-state frequency-response curve is

shown near the primary resonance in Fig. 6. The effect of the
elastic coefficient k b3 of the nonlinear bearing on the response
curve is illustrated. At k b3 equals zero, the response curve is identical to the linear response curve. As k b3 increases, the response
curve bends to the right side with lower and flattened peak. This
trend continues until a jump phenomenon takes place due to the
multi-valued phase points, indicating the bifurcation in the system. Also, for low values of the elastic coefficient k b3 (1
107 N/m3 ) of the nonlinear bearing, and thus weak nonlinearity,
the frequency response curve deviates slightly from the frequency
response curve of the corresponding linear system.
It is worth noting that for large values of the nonlinear elastic
coefficient k b3 (5108 N/m3 ), the perturbation analysis using
the multiple scales method starts to deviate from the basic assumption of weak nonlinearity. Thus, the error in predicting the
frequency response curves increases.
Summary and Conclusions
The multiple scales method is adopted to analyze the free and

forced vibration of rotor shaft system with viscoelastically supported nonlinear bearings. Timoshenko shaft model is assumed for
the shaft, two-element viscoelastic model is utilized for the supports, and typical roller bearings with cubic nonlinearity are employed. The free and forced vibration is developed based on the
direct multiple scales method. Third order perturbation expansion
with a one-to-one frequency-to-amplitude relationship of the non-
Fig. 5 Free frequency of the nonlinear system versus amplitudes for higher
values of internal viscous damping coefficient C i of the shaft
296 Vol. 125, JULY 2003
Fig. 6 Frequency response curves for various values of the bearing nonlinear elastic coefficient k b 3
linear system is used. The natural frequencies of the nonlinear

system and the steady state response are obtained.
Numerical simulations of nonlinear free and forced vibration on
a typical system are performed. The results show that a limiting
value of the internal viscous damping coefficient of the shaft exists where the trend of the frequency-amplitude curve switches.
Also, the primary resonance peak shifts to higher frequencies with
the increase of the bearing nonlinear elastic characteristics, but
with a flattened curve and hence lower peak values. A Jump phenomenon takes place for high values of the bearing nonlinear
elastic characteristics.
This research is financially supported by a research grant from

the Natural Sciences and Engineering Research Council of
Canada.
Appendix A
G
t,
l2
u
u * ,
l
C i*
Js
1
,
Is
k v*
l
A G
,
l
M*
d
b 4 2b 1r n* b 1m C i b 2r n* 1 2 b 12r M d* n*
b 5 2b 1r n* 2b 2r n* b 12r M d* n* , b 6 3/16 k *
b3 b 10r ,
b 7 b 11m c n* , b 8 2b 1m n* 2b 2m n* b 12m M d* n* ,
b 9 b 11r c n* ,
b 11Y 2u 0 Y u 0 Y v
b 12Y 2u 1 , and b j b jr ib jm
j1,2,10,11,12 .
The coefficients of the detuning parameter are defined as
*
k b3
l
M ,
Is d
l5
k ,
GI s b3
M*
2
l2
,
G
v
,
l
l
M ,
Is 2
*
k b1
c b*
l3
k ,
GI s b1
l2
I s G
F *
cb ,
eM d
,
Is
W * 43 u * 2u * 2u *u * v * v * 2u * u * 2v * 2u * v *v *
v * 2v * 41 u * 3 3u
* 2v * 3u
*v * 2 v * 3
b 13b 25 b 28 ,
b 142b 6 b 5 b 8 a 2 2 b 5 b 7 b 8 b 9 ,
b 15 b 6 a 2 b 7 2 b 6 a 2 b 9 2
*
l
M *
M 2
,
GI s
u*
l2
Q 3
* ,
GI s
l3
k ,
GI s v
v *
Ci ,
Al 2
3
,
Is
E
,
2
G
Q *
b 3 2b 1m *
n b 1r C i b 2m *
n 1 2 b 12m M *
d *
n ,
Appendix C
First set of nondimensional quantities
The coefficients of the modulation equations
b 10W 0 Y u 0 Y 4u 0 3Y 3u 0 Y v 3Y 2u 0 Y 2v Y u 0 Y 3v ,
Acknowledgment
t *
Appendix B
1 2
F *4Y 2 1
a2 n u
References
1 Darlow, M., and Zorzi, E., 1981, Mechanical Design Handbook of Elastomers,
NASA CR3423.
2 Dutt, J. K., and Nakra, B. C., 1992, Stability of Rotor Systems with Viscoelastic Supports, J. Sound Vib., 1531, pp. 8996.
3 Dutt, J. K., and Nakra, B. C., 1993, Vibration Response Reduction of a Rotor
Shaft System Using Viscoelastic Polymeric Supports, ASME J. Vibr. Acoust.,
115, pp. 221223.
4 Dutt, J. K., and Nakra, B. C., 1995, Dynamics of Rotor Shaft System on
Flexible Supports with Gyroscopic Effects, Mech. Res. Commun., 226, pp.
541545.
5 Kulkarni, P., Pannu, S., and Nakra, B. C., 1993, Unbalance Response and
Stability of a Rotating System with Viscoelastically Supported Bearings,
Mech. Mach. Theory, 283, pp. 427 436.
6 Shabaneh, N. H., and Zu, Jean W., 1999, Vibration Analysis of Viscoelastically Supported Rotor-Bearing Systems, Asia-Pacific Vibration Conference
1999 A-PVC99, Singapore, December.
7 Shabaneh, N. H., and Zu, Jean W., 2000, Dynamic Analysis of Rotor-Shaft
Systems with Viscoelastically Supported Bearings, Mech. Mach. Theory,
359, pp. 13131330.
JULY 2003, Vol. 125 297
8 Shabaneh, N. H., and Zu, Jean W., 2000, Dynamic and Stability Analysis of
Rotor-Shaft Systems with Viscoelastically Supported Bearings, Trans. Can.
Soc. Mech. Eng., 241B, pp. 179189.
9 Yamamoto, T., Yasuda, K., and Nagasaka, I., 1976, Ultra-Subharmonic Oscillations in a Nonlinear Vibratory System, Bull. JSME, 19138, pp. 1442
1447.
10 Ji, Z., and Zu, J. W., 1998, Method of Multiple Scales for Vibration Analysis
of Rotor-Shaft Systems with Non-Linear Bearing Pedestal Model, J. Sound
Vib., 2182, pp. 293305.
11 Bhattacharyya, K., and Dutt, J. K., 1997, Unbalance Response and Stability
Analysis of Horizontal Rotor Systems Mounted on Nonlinear Rolling Element
298 Vol. 125, JULY 2003
Bearings with Viscoelastic Supports, ASME J. Vibr. Acoust., 119, pp. 539
544.
12 Nayfeh, A. H., Nayfeh, J. F., and Mook, D. T., 1992, On Methods for Continuous Systems with Quadratic and Cubic Nonlinearities, Nonlinear Dyn. 3,
pp. 145162.
13 Nayfeh, A. H., 1981, Introduction to Perturbation Techniques, Wiley, New
York.
14 Yamamoto, Toshio, and Ishida, Yukio, 2001, Linear and Nonlinear Rotordynamics: A Modern Treatment with Applications, Wiley Series in Nonlinear
Science.
Animesh Chatterjee
Assistant Professor,
Visvesvaraya National Institute of Technology,
Nagpur, India-440011
e-mail: animeshch@rediffmail.com
Nalinaksh S. Vyas
Professor,
Indian Institute of Technology,
Kanpur, India-208016
e-mail: vyas@iitk.ac.in
Nonlinear Parameter Estimation

in Rotor-Bearing System Using
Volterra Series and Method of
Harmonic Probing
Volterra series provides a structured analytical platform for modeling and identification of
nonlinear systems. The series has been widely used in nonparametric identification
through higher order frequency response functions or FRFs. A parametric identification
procedure based on recursive evaluation of response harmonic amplitude series is presented here. The procedure is experimentally investigated for a rotor-bearing system
supported in rolling element bearings. The estimates of nonlinear bearing stiffness obtained from experimentation have been compared with analytical values and experimental
results of previous works. DOI: 10.1115/1.1547486
Introduction
System identification is an inverse problem of determination of

the mathematical structure of physical system from input-output
measurement. Volterra series provides a structured analytical platform for modeling and identification of nonlinear systems. The
series represents nonlinear response through a set of multidimensional kernels, known as Volterra kernels. Fourier transforms of these kernels provide the definition of higher order kernel transforms, or higher order frequency response functions
FRFs 1. For harmonic excitations, the response can be conveniently expressed in terms of the first and higher order FRFs.
Extensive research work has been done in the area of nonparametric system identification through measurement of first and
higher order kernel transforms or FRFs. Boyd, Tang and Chua 2
developed a second order kernel transform measurement procedure using multi-tone harmonic probing. Chua and Liao 3 extended the procedure for third and higher order kernel transforms.
Gifford and Tomlinson 4 developed a multi-degree-of-freedom
curve fitting procedure for estimating the higher order FRFs.
However, non-parametric system representation, in terms of
higher order kernels, often requires extensive computational work
and needs large number of experiments either with harmonic or
stochastic excitation.
Parametric system identification requires a-priori information
about the topology of the mathematical structure of the system.
Chatterjee and Vyas 5 have suggested an identification procedure for classification of systems between polynomial and nonpolynomial form of nonlinearity. Further classification is made
between symmetric and asymmetric forms of nonlinearity. For a
polynomial form nonlinearity, the series structure is further
identified through a peak ratio comparison test under multi-tone
excitation.
Higher order FRFs, in case of polynomial form nonlinearity,
can be synthesized with first order FRF and nonlinear parameters.
This provides a generalized basis for nonlinear parameter estimation through measurement of first and higher order kernel transforms. Lee 6 extracted the response components of first harmonic through component separation technique. First and higher
order kernel transforms were then computed from the separated
response components and nonlinear parameters were estimated
using the relationship between higher order and first order kernel
revised October 2002. Associate Editor: M. I. Friswell.
transforms. Chatterjee and Vyas 7 have suggested a recursive

iteration technique, which computes the first and higher order kernel transforms recursively from the measured response harmonic
amplitudes. The response series is considered with an optimum
number of terms governed by the convergence criterion.
In rotor-bearing systems, stiffness and damping nonlinearity
play a significant role in determining system behavior. Rolling
element bearing analyses presented by Harris 8 and Ragulski
et al. 9 point out that the stiffness force follows a symmetric
nonlinearity relationship with deflection. Hydrodynamic bearings
operating on fluid film lubrication also exhibit nonlinear relationship in the stiffness and damping forces. Investigations on rotor
vibration, using nonlinear modeling of bearing coefficients, have
been carried out by Bannister 10 and Choy et al. 11. Garibaldi
and Tomlinson 12 have presented a method for identifying the
nonlinear bearing coefficients in a rigid rotor using Hilbert transform. Khan and Vyas 13 have developed parameter estimation
procedures for both rigid rotor and flexible rotor model. The procedure employs Gaussian random excitation and Wiener kernel
transforms are extracted from the input-output data. Volterra kernel transforms are then obtained from the measured Wiener kernel
transforms and used for nonlinear parameter estimation.
The parameter estimation procedure described here employs
harmonic probing technique in conjunction with Volterra series
response representation. Investigation is carried out on a rotor
supported in rolling element bearings. The system is analyzed as a
single-degree-of-freedom system and linear and nonlinear stiffness parameters are estimated from the response measured at the
bearing housing. The estimates are compared with theoretical values obtained from the analytical work of Ragulski and also with
some earlier experimental results.
Volterra Series Response Representation
Volterra series response for a general physical system with f (t)

as input excitation and x(t) as output response is represented by
x t x 1 t x 2 t x 3 t . . . x n t . . .
(1)
with
x n t
...
h n 1 , . . . , n f t 1 . . . f t
nd1 . . . dn
(2)
JULY 2003, Vol. 125 299
h n ( 1 , . . . , n ) is the nth order Volterra kernel and its Fourier

transform provides the nth order frequency response functions
FRFs or Volterra kernel transforms as
H n 1 , . . . , n
...
where
X 0
h n 1 , . . . , n
j ii
i1
d1 . . . dn
(3)
with
For a single-tone harmonic excitation
i n 2
A
A
f t A cos t e j t e j t
2
2
(4)
the expression for the nth order response component, following

Eq. 2, can be obtained as
x n t
A
2
pqn
C q H np,q e j p,q t
n1
x t
n1
pqn
A
2
H 1 p,q
n
Cq
k3
(7)
p ,q 1
n1C q1H n 1
p i q i n i
n 1 n 2 n 3 n
i1, . . . .N
(9)
n1
(11a)
4
X 3
i 3
A3
i2
q ,q 2
p ,q
p ,q
p ,q
C q 1 H n 1 1 * n 2 C q 2 H n 2 2 * n 3 C q 3 H n 3 3
1
2
3
(13)
(11b)
(11c)
where the higher order kernel transforms are related to the lower
order kernel transforms through nonlinear parameters Chatterjee
and Vyas, 14 as
The excitation level A i is varied for constant response amplitude

X( i ) for keeping the series approximation error low and nearly
uniform over the frequency range 14. Standard curve fitting procedure Ewins 15 is employed to obtain the best fit FRF curve
and preliminary estimation of linear parameters is made.
300 Vol. 125, JULY 2003
ni1,i1
C i1 H n2i2
2
X 2
i 2
A2
i2
H 3 , , ,
* n2C q2H n2
Higher order series terms are initially neglected in Eqs. 11ac

and kernel transforms H 1 ( ), H 2 ( , ) and H 3 ( , , ) are extracted from measured harmonic amplitudes. Equation 12 is employed to estimate the second and third order nonlinear parameters k 2 and k 3 . Higher order series term contribution is then
considered and the procedure is iterated till the estimates converge
within a specified limit. For a typical Duffing oscillator, the estimation algorithm can be structured through following steps.
Step-I: System is excited at frequencies, i , with i varying
over a frequency range including the natural frequency, n , of the
system. Response x(t) is measured and harmonic amplitude,
X( i ), is filtered to give preliminary estimate of first order kernel
transform as
H 1 i X i /A i ,
n2i2
1
H 1
X
i
A
i2
(6)
X 3 cos 3 t 3 ..
H np,q
(8)
n2i2
H 2 ,
x t X 0 X cos t 1 X 2 cos 2 t 2
and n X n
mx t cx t k 1 x t k 2 x 2 t k 3 x 3 t A cos t (10)
The response series given in Eq. 6 can be written in terms of its

harmonics as
p i q i n i
n 1 n 2 n
C n H n,n
2n
first three response harmonic series, after re-arranging and truncating Eq. 8 to a finite number of terms, k, can be expressed as
C q H np,q e j p,q t
k2
i1
The total response of the system, then becomes

n
2n
Parameter Estimation
(5)
A
2
2n
For a system with polynomial form of nonlinearity under harmonic excitation given by
(5)
where the following brief notations have been used
X n
for n1
(12)
Step-II: System is excited at frequencies close to one-third of

natural frequency, for distinct measurability of third harmonic
X(3 ) Ref. 14. Employing Eq. 12, preliminary estimate of
nonlinear parameter, k 3 , is obtained through regression between
the estimated third order kernel transform and its synthesized kernel factor, 3 ( ), using the following relationship
H 3 , , k 3 3
(14)
3 H 31 H 1 3
(15)
where
k
i ( ) is computed with the H 1 ( )
Step-III: The series i2
values taken from the best fit curve estimated in Step-I and the
nonlinear parameter, k 3 , estimated in Step-II and substituted in
equation 11a to obtain new estimates of linear parameters.
k
Step-IV: The series i2
i (3 ) is computed and substituted in
Eq. 11c to refine the estimate of the nonlinear parameter k 3 .
Iteration is continued till the estimate of nonlinear parameter, k 3 ,
converges within a specified limit.
Fig. 3 Response acceleration spectrum from rap test
The sign of the nonlinear parameter, k 3 , is identified through

observations of change of sign of real part of measured third harmonic amplitudes during transition of excitation frequency across
one-third of the natural frequency. From Eqs. 14, 15, one can
note that for n /3, both H 1 ( ) and H 1 (3 ) have positive real
parts, which means that the kernel factor 3 ( ) will have a negative real part. Consequently, the third order kernel transform
H 3 ( , , ) will also have a negative real part for positive k 3 .
For n n /3, H 1 ( ) will have a positive real part while the
real part of H 1 (3 ) will bear negative sign. The kernel factor,
3 ( ), and H 3 ( , , ) will have a positive real part for positive
value of the nonlinear parameter k 3 .
Fig. 1 a Experimental set up along with instrumentation b

Close up view of exciter mounting arrangement and impedance
head
Experimental Investigation
Experimental studies have been carried out on a test rig Fig. 1

consisting of a 10 mm diameter shaft supported in ball bearings
SKF 6200. A single-degree-of-freedom model is considered and
cross-coupling stiffness parameters are neglected. Harmonic excitation force is applied on the bearing housing in vertical direction through an electro-dynamic shaker. Measurement for excitation force and resultant vibration are made at the bearing housing
through an impedance head attached between the shaker and bear
Fig. 2 Schematic diagram of rotor bearing test rig and instrumentation
JULY 2003, Vol. 125 301
Fig. 4 Response component spectra for 330 Hz
ing cap. A schematic diagram of the instrumentation is shown in

Fig. 2.
4.1 Nonlinearity Structure Identification. An initial measurement of the natural frequency of the system is obtained
through a rap test, using an impact hammer. The rap test response
spectrum Fig. 3 shows that the natural frequency is in the vicinity of 1010 Hz. An initial test is done to find the form of stiffness
nonlinearity through ordered component separation method presented in 5. The system is harmonically excited at 330 Hz near
one-third of natural frequency at excitation levels 4N, 3N and
2N and the first three response components x 1 (t), x 2 (t) and x 3 (t)
are separated. The response components are filtered to investigate
presence of various ordered harmonics. For an asymmetric nonlinearity structure both even and odd harmonics will be present in
the response spectrum, whereas for symmetric form, only odd
harmonics will be present. Also for polynomial form nonlinearity,
spectra of response components will have ordered presence of
harmonic, i.e., odd response components will have odd harmonics
and even response components will have even harmonics refer
5. Fig. 4ac show the spectra of first three response components separated from the measured response. The figures indicate
that while the stiffness nonlinearity is symmetric, its form is nonpolynomial as the response component spectra do not exhibit ordered harmonic characteristics. Theoretical formulations based on
Hertzian contact theory, Harris 8, Ragulski 9 also indicate
that the nonlinear stiffness associated with rolling elements follows a fractional power relationship and not a polynomial form.
302 Vol. 125, JULY 2003
Fig. 5 Excitation level variation, response amplitude, X ,

and preliminary estimate of H 1
Ragulski 9, however, observed that for a small deformation x

g, where g is the pre-load in the bearing, the stiffness function
is given by the form k(x)abx 2 . This enables one to model the
bearing forces through cubic terms and represent the rotor-bearing
system as
mx t cx t k 1 x t k 3 x 3 t A cos t
(16)
4.2 Estimation of Stiffness Parameters. For estimation of

first order kernel transforms, H 1 ( ), an excitation frequency
range 9251035 Hz is selected. Keeping in view, the singledegree-of-freedom treatment of the rotor-bearing set-up, a wider
excitation frequency range is avoided in order to stay clear of
some other structural modes, which may influence measured vibration data. In order to keep the series approximation error low,
excitation level is varied over the frequency set to obtain response
amplitude at a constant level, which in this case was selected as
1.0108 m. The variation of excitation level is plotted in Fig.
5a and corresponding response harmonic amplitude, X( ), is
shown in Fig. 5b. Preliminary estimate of the first order kernel
transform, H 1 ( ), is shown in Fig. 5c. Curve fitting of the
H 1 ( ) graph gives following estimates of natural frequency and
damping
n 1012.2 Hz,
0.01156.
Estimation of nonlinear stiffness parameter k 3 is done through

measurement of third response harmonic amplitude X(3 ). The
1019 N/m3 . Assuming that both the bearings are identical and
act in parallel, the nonlinear stiffness parameter of each bearing
then becomes 0.8751019 N/m3 . Figure 8b shows the final estimate of the first order kernel transform along with its preliminary estimate. It is seen that the successive iterations have improved the FRF values significantly near and around the natural
frequency. Final estimates of natural frequency and damping become
n 1011.47 Hz,
0.01078.
The linear stiffness parameter, k 1 , of the bearing system is computed from the natural frequency and the equivalent rotor mass
per bearing and is found to be 1.585107 N/m.
Similar exercises are carried out with excitation level of 3 N
Case II and 2 N Case III corresponding to 5% and 3% measurability respectively. The convergence trend in the estimate k 3 is
shown respectively in Fig. 9a, b. The estimates of nonlinear
stiffness parameter of each bearing is found to be 1.4285
1019 N/m3 Case II and 1.7401019 N/m3 Case III respectively. Damping and linear stiffness parameter are found as
Fig. 6 a Typical response spectrum with excitation at

330 Hz b Measurability of third response harmonic at different excitation levels
displacement amplitude of the third response harmonic is low and

therefore measurements were made for acceleration amplitudes.
Figure 6a shows the acceleration amplitudes of the first and third
order harmonics, for an excitation frequency of 330 Hz near onethird natural frequency. The ratio between the amplitudes at third
harmonic and first harmonic is defined as measurability index,
which provides the basis of selecting the excitation level and frequency range for measurement of third response harmonic amplitude X(3 ). For measurement of third harmonic one can select
excitation level corresponding to a certain measurability index.
Figure 6b shows measurability indices for three different excitation amplitudes of 2N, 3N and 4N over a frequency range of 305
Hz355 Hz. Corresponding peak measurability indices can be
seen to be approximately of the order of 3%, 5% and 10%. Excitation frequencies are selected at 330 Hz, 335 Hz, 340 Hz and 345
Hz, as measurability is relatively higher in this range. Third response harmonic amplitude X(3 ) is then measured at these selected frequencies for three different excitation levels as
Case I: Excitation level4 N
Case II: Excitation level3 N
Case III: Excitation level2 N
For case I corresponding to 10% measurability, the measured response acceleration spectra at the four selected frequencies are
shown in Fig. 7ad. A preliminary estimate of nonlinear parameter k 3 is made and the iteration procedure is repeated till estimated value of the nonlinear parameter converges within a limit,
which in this case is specified as 0.1%. Figure 8a shows the
estimated values of nonlinear parameter k 3 over the stages of
successive iteration. Final estimate of k 3 is found to be 1.75
Case II:
0.01015
k 1 1.535107 N/m.
Case III
0.01002
k 1 1.586107 N/m
The estimates for different cases are summarized in Table 1 given

below.
It can be seen from the estimates that, while the values of the
estimated linear parameters are very accurate, the estimate of the
nonlinear parameter k 3 , varies with the excitation amplitude. This
is due to the fact that the system nonlinearity does not adhere to
polynomial form and therefore the cubic coefficient of equivalent
polynomial form would be amplitude dependent.
Figure 10 shows the sign variation of real part of X(3 ), over
the excitation frequency range 330345 Hz for excitation levels
4N, 3N and 2N. The sign of real part of X(3 ) is seen to change
from positive values to negative values during frequency transition through one-third natural frequency. This indicates that the
nonlinearity is negative in sign.
Validation of Estimates
For validation of the experimentally estimated nonlinear stiffness parameter, analytical formulations of Harris 8 and Ragulski
et al. 9 are employed. These formulations are based on Hertzs
theory of elastic contacts and treat the bearings in isolation of the
shaft. Figure 11 shows a typical isolated ball bearing configuration, in which external forces act along x-axis. i is the angle
between load axis i.e., x-axis and the radial direction of a typical
ith ball element. For a displacement x, y of the moving ring
along the respective axes, total elastic force acting in radial direction at the point of contact of ith ball is given by
F i k n gx cos i y sin i 3/2
(17)
and its projections along x and y axes respectively are

F xi k n gx cos i y sin i 3/2 cos i
(18)
F yi k n gx cos i y sin i 3/2 sin i
(19)
where g is the radial pre-load between the ball and the races, k n is
a coefficient of proportionality depending on the geometric and
material properties of the bearing. Total restoring force of the
bearing is equal to sum of elastic forces of all the elements,
n
i.e.,
F
i1
xi
(20)
where n is the number of rolling elements in the bearing in the

loading zone.
JULY 2003, Vol. 125 303
Fig. 7 Acceleration response spectra for Case I: Excitation amplitude4 N
Using the condition of zero elastic force along y-axis since no

external force acts along this axis and following Eq. 19, deformation y is obtained as
gx cos
i
i1
i1
sin i ;
gx cos
i1
3/2
sin2 i
3/2
gx cos i 3/2 sin i
i1
(21)
gx cos
i
i1
1/2
gx cos
gx cos
i1
k x F/ x
(22)
Substituting Eq. 20 in Eq. 22, taking into account Eq. 21 the

bearing stiffness is expressed as a function of deformation as
3/2
sin i cos i
(24)
sin2 i
Equations 18 and 21 are used in Eq. 20 and the bearing

stiffness is determined as a function of deformation x as
3/2
sin2 i cos i
It can be seen that the bearing stiffness is critically dependent on

the pre-load. Table 2 summarizes the theoretical bearing stiffness
parameters for different pre-load along with experimentally obtained stiffness parameters. The stiffness variations are also shown
in Fig. 12 along with experimental results of earlier researchers.
k x K n
gx cos A/Bn sin

i1
AD n1 / Bn 2 sin i cos i
where
304 Vol. 125, JULY 2003
cos i CBn
1/2
(23)
Conclusion
The stiffness parameter estimates from the experiment show

reasonably good agreement with those from available analytical
formulations for isolated ball bearings. The analytical formulation
is dependent on the amount of preload and hence exact compariTransactions of the ASME
Fig. 8 a Iterative estimates of k 3 , Case I: Excitation

amplitude4 N b Final estimate of first order kernel transform, H 1 Case I: Excitation amplitude4 N
Fig. 9 a Iterative estimates of k 3 , Case II: Excitation

amplitude3 N b Iterative estimates of k 3 , Case III: Excitation amplitude2 N
Table 1 Linear and nonlinear estimates at different measurability cases

Nonlinear parameter Natural Linear stiffness
k 3 (1019 N/m3 ) frequency k 1 (107 N/m) Damping
Case I: 4N
Case II: 3N
Case III: 2N
0.8750
1.4285
1.7400
1011.47 Hz
1012.47 Hz
1011.90 Hz
1.585
1.535
1.586
0.01078
0.01015
0.01002
Fig. 10 Variation in sign of real part of X 3 around n 3
Fig. 11 Schematic diagram of a loaded ball bearing
JULY 2003, Vol. 125 305
Table 2 Estimated and theoretical Ragulskis et al. 9; Harris, 8 bearing stiffness

parameters
Theoretical Stiffness N/m
Pre-load m
k(x) from equation 22
0.2
0.3
0.4
0.5
0.6
1.20107 4.011019x 2
1.47107 2.181019x 2
1.69107 1.421019x 2
1.89107 1.021019x 2
2.08107 0.611019x 2
Estimated Stiffness N/m

(k 1 3k 3 x 2 )
Case i: 1.585107 2.6251019x 2
Case ii: 1.535107 4.2851019x 2
Case iii: 1.586107 5.2401019x 2
k1
k2
k3
m
x n (t)
3( )
i (n )
linear stiffness parameter

square non-linear stiffness parameter
cubic non-linear stiffness parameter
mass of the system
nth order response component
Third order synthesized kernel factor
damping factor
ith term in the response harmonic series,
X(n )
References
Fig. 12 Comparison of estimates of stiffness parameters 15:
Theoretical values with pre-load 0.2, 0.3, 0.4, 0.5 and 0.6 m
respectively. Harris 8 and Ragulski et al. 9 6,7,8: Present
experimental estimates for cases I, II and III respectively. 9: Experimental estimate of Tiwari 16 10: Experimental estimate of
Khan 13
son of theoretical values with experimental estimates is difficult.

While, the manufacturer, at times may provide the preload range,
the exact value of preloading of the bearing balls in the shaftcasing assembly, especially during operations which have involved wear and tear, would be difficult to determine. The suggested procedure gives good estimates of nonlinear parameter as
well as damping. The method is based on the assumption of linear
viscous damping model. It can however be extended for identification and estimation of damping nonlinearity.
Acknowledgment
The authors wish to express their thanks to the financial aid
being provided by the Propulsion Panel of Aeronautical Research
and Development Board, Ministry of Defense, Government of India, in carrying out the study.
Nomenclature
H n( 1 , . . . , n)
X(n )
c
f (t)
g
h n( 1 , . . . , n)
nth order Volterra kernel transform

response amplitude of nth harmonic
damping coefficient
excitation force
bearing pre-load
nth order Volterra kernel
306 Vol. 125, JULY 2003
1 Bedrosian, E., and Rice, S. O., 1971, The Output Properties of Volterra Systems Nonlinear System with Memory Driven by Harmonic and Gaussian
Input, Proc. IEEE, 5912, pp. 1688 1707.
2 Boyd, S., Tang, Y. S., and Chua, L. O., 1983, Measuring Volterra Kernels,
IEEE Trans. Circuits Syst., CAS-308, pp. 571577.
3 Chua, L. O., and Liao, Y., 1989, Measuring Volterra Kernels II, Int. J. of
Circuit Theory and Applications, 17, pp. 151190.
4 Gifford, S. J., and Tomlinson, G. R., 1989, Recent Advances in the Application of Functional Series to Nonlinear Structures, J. Sound Vib., 1352, pp.
289317.
5 Chatterjee, A., and Vyas, N. S., 2001, Stiffness Nonlinearity Classification
through Structured Response Component Analysis using Volterra Series,
Mech. Syst. Signal Process., 152, pp. 323336.
6 Lee, G. M., 1997, Estimation of Nonlinear System Parameters using Higher
Order Frequency Response Functions, Mech. Syst. Signal Process., 112, pp.
219228.
7 Chatterjee, A., and Vyas, N. S., 2002, Nonlinear Parameter Estimation
through Volterra Series using Method of Recursive Iteration, accepted for
publication in J. Sound Vib.
8 Harris, T. A., 1984, Rolling Bearing Analysis, Wiley, New York.
9 Ragulskis, K. M., Jurkauskas A. Y., Atstupenas, V. V., Vitkute, A. Y., and
Kulvec, A. P., 1974, Vibration in Bearings, Mintis Publishers, Vilnius.
10 Bannister, R. H., 1976, A Theoretical And Experimental Investigation Illustrating the Influence of Nonlinearity and Misalignment on the Eight Film Coefficients, Proc. Inst. Mech. Eng., 190, pp. 271278.
11 Choi, F. K., Braun, M. J., and Hu, Y., 1992, Nonlinear Transient and Frequency Response Analysis of a Hydrodynamic Bearing, ASME J. Tribol.,
114, pp. 448 454.
12 Garibaldi, L., and Tomlinson, G. R., 1988, A Procedure for Identifying Nonlinearity in Rigid Rotors Supported in Hydrodynamic and Ball/Roller Bearing
System, I. Mech. Proc. on Vibrations in Rotating Machinery, 4, pp. 229234.
13 Khan, A. A., and Vyas, N. S., 2001, Application of Volterra and Wiener
Theories for Nonlinear Parameter Estimation in a Rotor-Bearing System,
Nonlinear Dyn., 243, pp. 285304.
14 Chatterjee, A., and Vyas, N. S., 2000, Convergence Analysis of Volterra
Series Response of Nonlinear Systems Subjected to Harmonic Excitations, J.
Sound Vib., 2362, pp. 339358.
15 Ewins, D. J., 1984, Modal Testing: Theory and Practice, Research Studies
Press, England.
16 Tiwari, R., and Vyas, N. S., 1995, Estimation of Nonlinear Stiffness Parameters of Rolling Element Bearings from Random Response of Rotor Bearing
Systems, Journal of Sound Vib. 187 2, pp. 229239.
Y. S. Ho
H. Liu
The Hong Kong Polytechnic University,
Hong Kong SAR, P.R. China
L. Yu
Theory of Lubrication and Bearing Institute,
Xian Jiaotong University,
Xian, Shaanxi, P.R. China
Effect of Thrust Magnetic Bearing

on Stability and Bifurcation of a
Flexible Rotor Active Magnetic
Bearing System
This paper is concerned with the effect of a thrust active magnetic bearing (TAMB) on the
stability and bifurcation of an active magnetic bearing rotor system (AMBRS). The shaft
is flexible and modeled by using the finite element method that can take the effects of
inertia and shear into consideration. The model is reduced by a component mode synthesis method, which can conveniently account for nonlinear magnetic forces and moments
of the bearing. Then the system equations are obtained by combining the equations of the
reduced mechanical system and the equations of the decentralized PID controllers. This
study focuses on the influence of nonlinearities on the stability and bifurcation of T
periodic motion of the AMBRS subjected to the influences of both journal and thrust
active magnetic bearings and mass eccentricity simultaneously. In the stability analysis,
only periodic motion is investigated. The periodic motions and their stability margins are
obtained by using shooting method and path-following technique. The local stability and
bifurcation behaviors of periodic motions are obtained by using Floquet theory. The
results indicate that the TAMB and mass eccentricity have great influence on nonlinear
stability and bifurcation of the T periodic motion of system, cause the spillover of system
nonlinear dynamics and degradation of stability and bifurcation of T periodic motion.
Therefore, sufficient attention should be paid to these factors in the analysis and design of
a flexible rotor system equipped with both journal and thrust magnetic bearings in order
to ensure system reliability. DOI: 10.1115/1.1570448
Introduction
Active magnetic bearing AMB has found wide applications in

fields such as aerospace, petrochemical and power generation industries due to its advantages over traditional rolling or sliding
bearing. Like conventional mechanical bearings, the AMBs are
nonlinear in nature. These nonlinear characteristics cause nonlinear motion of the rotor. Various nonlinear studies have been carried out on the AMBRS 13. However, the effect of thrust active
magnetic bearing TAMB on the stability and bifurcation of the
AMBRS has received little attention in system modeling and
analysis.
The TAMBs are used to balance axial loads in system and usually their effects are often neglected when in the process of the
controller adjustment and the dynamic analysis of system. Using
linear analysis method, researchers have revealed that hydrodynamic thrust bearings have great effects upon the linear stability
of rotor systems 4 5. Because of constraints of the linear stability theory, only the effects of thrust bearings on the linear stability near the equilibrium point of a balanced rotor system has
been studied. Because a rotor always has mass eccentricity that
affects the stability of system and this effect cannot be taken into
account in linear analysis, a nonlinear analysis method has to be
used. Due to the complexity of nonlinear analysis, nonlinear models of rotor system are often simplified using assumptions such as
a symmetrical rigid rotor or a Jeffcott rotor 6 9. However, such
simplified models do not truly represent the complex system accurately. Therefore several investigators introduced the order reduction methods of system to study the periodic solutions of a
high order unbalanced rotor systems 1013. Bearings in these
studies are traditional hydrodynamic bearings, although few inContributed by the Technical Committee on Vibration and Sound for publication
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received November
2002; Revised October 2002. Associate Editor: J. Cusumano.
vestigators deal with the effect of thrust bearing upon the nonlinear stability of T period motion of unbalanced rotor systems.
This paper is concerned with the effect of a TAMB on the
stability and bifurcation of an AMBRS supported by both journal
and thrust active magnetic bearings. The rotor is flexible, and
system equations are formulated by combining the equations of
motion of the rotor and the equations of the decentralized PID
controllers. This is typically a mechanical-electrical system consisting of a large number of linear components such as flexible
shaft elements without disk and bearing and a few of nonlinear
components nonlinear active magnetic bearings. Finite elements
method 14,15, and order reduction techniques are used to reduce
the order of the model 10,11,13. Floquet theory 16 18, shooting method and path-following technique 12,19 are used to
analyses the stability and bifurcation of T periodic motion of the
AMBRS equipped with both journal AMBs and TAMBs. The
effects of the TAMB and the mass eccentricity are discussed
especially.
Formulations of Model and Reduction
2.1 Finite Element Equations of the Flexible Shaft. An

AMBRS is depicted in Fig. 1 when the rotor is flexible. It can be
divided into linear components the flexible shaft elements without disk and bearing and nonlinear components nonlinear active
magnetic bearings. The finite element method is used to form the
equations of the motion of linear components. Here a 2-node Timoshenko shaft element model with 8 degrees-of-freedom
14,15, as shown in Fig. 2, which can account for the effects of
inertia and shear, is adopted. Equations of the lateral motions of a
flexible shaft can then be written as
MS xS GS xS KS xS QS f S
(1)
where M , G , K R
and Q , f R are the mass matrices,
gyroscope matrices, stiffness matrices, external force vector inS
nn
JULY 2003, Vol. 125 307
Here, the columns of the (n,n k ) matrix i with these elastic

eigenmodes kept are the mass normalized solutions ( Tj MS j
1) of the undamped eigenproblem for k (0, cut )( j
1, . . . ,n k ):
2j MS KS j 0
(8)
The columns of the (n,n b ) matrix b with the residual flexible

modes is defined as follows:
1
T
b KS i 2
kk i
Fig. 1 Structure of a flexible rotor-active magnetic bearing

system
xS x 1 y 1 1 1 x p y p p p T
(2)
where x j , y j , j and j ( j1,2, . . . ,p) are the lateral translations

and tilting angles of the jth nodal point along the horizontal and
vertical directions, respectively. The nonlinear force vector can be
written as
f s 0 f x j f y j 0 0 0 0 M x k M y k O T
MSbb
S
Mbi
S
Mib
MiiS
GSbb
xSb
S
xiS
Gib
QSb
f Sb xSb ,xSb
S
Qi
0
S
Gbi
GiiS
KSbb
xSb
S
xiS
Kib
S
Kbi
KiiS

xSb
xiS
(4)
xSb x 1 y 1 1 1 x m y m m m T
(5)
f Sb f x j f y j 0 0 0 0 M x k M y k T
(6)
where T1 b i
bb
xSb
S
xi
ib
bk pSb
S
ik pk
(10)
1
pSb
bb
pSk
0kb
S
1
bb bk xb
p T2 p
p Sk
Ikk
(11)
This results in the following total transformation

xS Tp
TT1 T2
(12)
Applying the transformation 12, the reduced component equations become:

TT MS TpTT GS TpTT KS TpTT QS TT f S
(13)
After reduction, the local nonlinear forces of bearing and the

unbalance force of disk can be easily added to the reduced linear
equation since the interface degrees-of-freedom is available in the
reduced equation. The equation of motion of the shaft in the axial
direction is given by
(7)
(14)
From Eqs. 10 and 11, the equations of motion of the reduced

mechanical system is given by
MqGqKqQ
(15)
with
For reducing the order of the degree-of-freedom of linear components, xS can be written as a linear combination of n c columns:
x T1 p
mz F z
If the rotor has m points on which journal and thrust active

magnetic bearings act, xSb and fSb R nn (n b 4m) can be written
as
(9)
Starting from Eq. 10, the degrees-of-freedom of pSb are replaced

by the interface degrees-of-freedom of xSb using a coordinate
transformation matrix T 2 to permit simple coupling of the reduced
component equations:
(3)
where f x i and f y j are the horizontal and vertical of magnetic forces

of the journal active magnetic bearing acting on the jth point; M x k
and M y k are the magnetic moments of the thrust active magnetic
bearing acting on the kth point. To simplify notations, the order of
the vector components is rearranged and Eq. 1 can be partitioned
as
Ibb
0ib
where kk is a (n k ,n k ) diagonal matrix with the kept angular

eigenfrequencies lower than or equal to c 20. Then
xS T1 p
cluding weight force and unbalance force and nonlinear force

vector of bearings respectively. For a shaft with p nodal points,
the displacement vector can be written as
q xSb
TT MS TMd
pSk
zT
TT GS TGd
TT KS TKd
Q TT QS 0 T TT f S 0 T FTdex 0 F z T
where Md , Gd and Kd are the mass, damping and stiffness matrices of disks; m is the mass of the rotor; Fdex is unbalance forces
caused by mass eccentricity of disks. The behavior of this nonlinear system depends on the rotating frequency of the shaft and
the eccentricities e explicitly.
The eigenfrequencies cut of Eq. 13 are accurate for the
case QS 0 and GS 0, with cut the highest cut-off eigenfrequency in the reduction, and higher eigenfrequencies will be inaccurate. Because nonlinear systems can generate frequencies
higher than their excitation frequency, cut has to be chosen
higher than the maximum excitation frequency.
Fig. 2 Shaft finite element model
308 Vol. 125, JULY 2003
2.2 Electromagnetic Force of Journal Bearing. An eightpole journal AMB shown in Fig. 3 is used for analysis. To simTransactions of the ASME
Fig. 4 A thrust magnetic bearing in operation

Fig. 3 End view of an eight-pole journal bearing
h i 1 c z z r i sin r i cos ,
h o1 c z z r o sin r o cos ,
plify the notations, subscripts indicating bearing numbers are ignored. The forces in this journal magnetic bearing can be
expressed as
f x f r f l xy x/c r f t f b
(16)
f y f t f b xy y/c r f r f l
with
f r
f t
0 N r2 A r I 0x i x 2
4
c r x 2
0 N r2 A r
I 0y i y
c r y 2
f 1
f b
0 N r2 A r I 0x i x 2
4
c r x 2
0 N r2 A r
4
h o2 c z z r o sin r o cos ,
It is assumed that the magnetic flux from the point (r i , ) on the
inner annulus runs entirely back to the point (r o , ) on the outer
annulus as shown in Fig. 5. Therefore the relationship between
radii r i and r o is
r o R 4 r i R 1
(17)
I 0y i y
c r y 2
F z1
R1
F z2
M x1
M x2
R2
R1
M y2
R2
R1
M y1
R1
ro
0 N I 0z i z 2
20
h o1 r o /r i h i 1
where
ro
0 N I 0z i z 2
2
20
h o r o /r i h i 2
cos dr i d
R4
sin dr i d
sin dr i d
R3
R4
R3
R4
ro
0 N I 0z i z 2
20
h o1 r o /r i h i 1
ro
0 N I 0z i z 2
2
20
h o r o /r i h i 2
R3
R4
R3
R4
1
o N I 0z i z 2
r o dr o d
20
h o1 r o /r i h i 1 2
1
0 N I 0z i z 2
r o dr o d
2
20
h o r o /r i h i 2 2
R3
cos dr i d
R3
R4
ro
0 N I 0z i z 2
20
h o1 r o /r i h i 1
(20)
M y M y1 M y2
(19)
M x M x1 M x2
2
0 N I 0z i z 2
r o /r i
r i dr i d
20
h o2 r o /r i h i 2
R2
R2
R 4 R 3
R 2 R 1
F z F z1 F z2
2
r o /r i
0 N I 0z i z 2
1
1 r i dr i d
20
h o r o /r i h i
ro
0 N I 0z i z 2
2
20
h o r o /r i h i 2
R1
R2
R1
R2
with
The forces and moments provided by the thrust magnetic bearing can be expressed as 22
where xy is the geometric coupling coefficient with an average

value of 0.16 21. In this study, xy is taken to be 0.16.
2.3 Magnetic Forces and Moments of Thrust Bearing. A
two-pole TAMB in operation is shown in Fig. 4. When tilting of
runner occurs, the right and left air gaps at those points on the
inner and outer annuli are different and can be expressed as
(18)
h i 2 c z z r i sin r i cos ,
ro
0 N I 0z i z 2
20
h o1 r o /r i h i 1
ro
0 N I 0z i z 2
2
20
h o r o /r i h i 2
cos dr o d
cos dr o d
sin dr o d
sin dr o d
(21)
JULY 2003, Vol. 125 309
When state variables X(q,q,Ibb ) T are introduced, the corresponding system equations in state space are
q
M1 QGqKq
Aa As Kp xbb Ki xbb Kd xbb
(28)
The stability and bifurcation problem of the reduced system can

be solved by using shooting method and path-following technique.
3
Fig. 5 Magnetic flux path in a thrust magnetic bearing
Method of Solution
It is assumed that the external load acting on the system is

periodic with period T:
Qex t Qex tT
2.4 Equations of Sensors, Amplifiers and PID Controllers.

The signal flow in a rotor-magnetic bearing system is shown in
Fig. 6. The power amplifier can be modeled as a first-order system, whose transfer function is
Ga s
Aa
1Ta s
(22)
where Aa is gain, and Ta is time constant.

The displacement sensor can also be represented by a first-order
system.
Gs s
As
1Ts s
(23)
where As is gain, and Ts is time constant.

In this investigation, the controller is taken to be a PID controller as it is widely used. The transfer function of a PID controller is
Gc s Kp
Kd s
Ki
s
1Td s
(24)
where Kp is proportional gain, Ki is integral gain, Kd is derivative

gain and Td is time constant, For an ideal system, time constant
Ta Ts Td 0. The transfer function of the whole system is thus
given by
Ki
G s Ga s "Gs s "Gc s Aa As Kp Kd s
s
(25)
and the corresponding differential equation can be written as

Ibb Aa As Kp xbb Ki xbb Kd xbb
(26)
(29)
The steady-state behavior of the system, i.e., the attractor that is

reached after the transient has damped out, might be periodic,
quasi-periodic or chaotic. The periodic solutions of the system can
become unstable in certain intervals of system parameters: such as
the angular speed of the rotor, the mass eccentricity e, and the
axial force F z , etc. Such instability is due to the nonlinear characteristics of the AMBs. Generally, in these intervals, the maximum response of the nonlinear system will increase and also additional vibration frequencies occur compared to the linear
system. This may lead to shaft rubbing which is undesirable in
practice. So for practical application, it is important to determine
the unstable intervals and the kinds of dynamic behaviors of the
rotor that result in these intervals: periodic, quasi-periodic or
chaotic.
3.1 Periodic Solutions. Periodic solutions are calculated by
solving a two-point boundary value problem, which is defined by
Eq. 28 supplemented with the boundary condition X(t)X(t
T). It can be written as:
with
f X,t,
f X,t,
X
X t X tT
q
M1 QGqKq
Aa As Kp xbb Ki xbb Kd xbb
(30)
where is a system parameter. It can be the angular speed , or

the mass eccentricity e, and or the axial force F z , etc.
In the shooting method, Eq. 30 is integrated by means of a
numerical time integration method over the period T and a T periodic solution can be found if the following criterion is satisfied:
H Xs , 0
(31)
with H Xs , Xs t 0 T Xs t 0
and N J is the number of journal AMBs.
2.5 System Equations. Combining the equations for the rotor, the sensors, the amplifiers and the PID controllers, the system
equations are obtained
MqGqKqQ
Ibb Aa As Kp xbb Ki xbb Kd xbb
(27)
Fig. 6 Signal flow in a rotor-active magnetic bearing system
310 Vol. 125, JULY 2003
For a given s , the corresponding solution Xs can be obtained from Eq. 31 using the iterative Newton process. The Jacobian matrix can be obtained as
Table 1 Parameters of journal and thrust AMBs
Bearing type
Parameter
Value
Journal bearings
radial clearance
bias current
winding number
width
diameter
0.0004 m
4A
57
0.08 m
0.16 m
Thrust bearing
total axial clearance

bias current
winding number
diameter
R1
R2
R3
R4
0.0006 m
4A
143
0.1 m
0.04 m
0.065 m
0.08 m
0.1 m
Table 2 Parameters of controllers, sensors and power

amplifiers
Parameter
Value
A a 1/
A 2 V/m
kp
k i 1/s
k d s
1
7800
3.8
200
0.01
H
JI
Xs
Xn1 Xn
H X,
Xs
H X,

(34)
n1 n
Subsequently this periodic solution is corrected by shooting

method at n1 . Here H(X, )/ can be calculated by
numerical time integration of Eq. 34 about the trajectory Xs (t 0
t), and
f t, ,X
f t, ,X
d
S
S
dt
X
(32)
where J is calculated by numerical time integration of equation

30 linearized at those points along the trajectory Xs (t 0 t), and
f
d
S
S
dt
X
consists of a predictor-corrector mechanism. Starting from a

known solution Xn at n , the prediction of n-step is
(33)
with S(t 0 )I and S(t 0 T)J. This is called shooting method

12,19.
3.2 Continuation of Periodic SolutionsPath-Following
Technique. The problem of continuation of solutions in general
is to find the whole solution branches, i.e., to obtain a solution at
j1 when the exact solution at j is known. Several
continuation schemes are available. In the present study, the path
following technique 12 is used to investigate how a periodic
solution is influenced by a change of . In essence, the technique
(35)
with S(t 0 )0 and S(t 0 T) H(X, )/ .

3.3 Local Stability and Bifurcation: Floquet Theory. The
local stability of periodic motion of a nonlinear system can be
determined from Floquet theory 16 18. The procedure involves
calculating the eigenvalues Floquet multipliers of the matrix J,
which is also called monodromy matrix. For a stable periodic
solution, all the Floquet multipliers should be within the unit
circle in the complex plane. When the unit circle in the complex
plane is crossed by
1. a real eigenvalue on the positive real axis, it is a cyclefold or transcritical bifurcation;
2. a pair of complex conjugate eigenvalue, it is a secondary
Hopf bifurcation;
3. a real eigenvalue on the negative real axis, it is a perioddoubling bifurcation.
Here, the eigenvalue above mentioned, which has the largest
magnitude among all eigenvalues of the monodromy matrix, is
called the leading Floquet multiplier.
Fig. 7 A e 0 diagram of Stable and unstable regions of T

period motion
Numerical Results and Discussions
The rotor system depicted in Fig. 1 is analyzed. Parameters of

the journal and thrust AMBs are given in Table 1. The parameters
of sensors, amplifiers and controllers for the five channels are the
same and given in Table 2. The axial load is taken to be 0N and
the rotational speed is 30,000rev/min. The rotor and the journal
bearings are assumed to be coaxial in static state.
The shaft is discretized into 16 elements and 17 nodes. The
cut-off rotating speed cut is chosen to be 120000rev/min which
is greater than two times the expected bandwidth of the response.
The number of retained eigenmodes k of the linear component is
8, 2 of which are rigid body modes, and other 6 modes are elastic
eigenmodes with eigenfrequencies of 4908rev/min, 12666rev/min,
27445rev/min, 38030rev/min, 46473rev/min, 62026rev/min, re-
Fig. 8 Hopf T periodic solution without TAMB a Hopf T periodic solution at

120000 revmin b A amplitude-frequency diagram of x a
JULY 2003, Vol. 125 311
Fig. 9 Hopf T periodic solution with TAMB a Hopf T periodic solution at 70000 revmin b
A amplitude-frequency diagram of x a
Fig. 10 Quasi-periodic solution without TAMB e 0 4 m a Quasi-periodic solution at

113405 revmin b The orbit of the center of the rotor at journal bearing a c Poincare maps d
Time series of x a e Amplitude-frequency diagram of x a
312 Vol. 125, JULY 2003
Fig. 11 Quasi-periodic solution with TAMB e 0 4 m a Quasi-periodic solution at

67345 revmin b The orbit of the center of the rotor at journal bearing a c Poincare maps
d Time series of x a e Amplitude-frequency diagram of x a
spectively. The number of interface degrees-of-freedom is 16.

This include eight displacements and eight tilting degree-offreedom for the four couplings between the disks and the shaft
(x a ,y a , a , a ,x b ,y b , b , b ,x c ,y c , c , c ,x d ,y d , d , d ), in
which 4 degrees-of-freedom are the displacements for two journal
AMBs (x a ,y a ,x c ,y c ) and 2 degrees-of-freedom are the tilting
angles for a TAMB ( d , d ). In present study, only unbalance
forces caused by mass eccentricities of four disks (e a e b e c
e d e 0 ) are studied.
The combined effect of e 0 and is studied with and without
the effects of the TAMB considered. The stable T period motion
harmonic motion and its unstable regions are plotted in an e 0 -
diagram in Fig. 7, where the line is the bifurcation borderline,
with the modulus of leading Floquet multiplier equals to one. For
values of e 0 and below the line, the motion of the system is
stable T period motion. For that above the line, the motion is
Table 3 Leading Floquet multiplier

e0
m
5
10
15
20
22.7
25
Not consider TAMB

leading Floquet multiplier
modulus
Consider TAMB
leading Floquet multiplier
modulus
0.992395i0.0726290
0.995049
0.992400i0.0726408
0.995055
0.992411i0.0726599
0.995067
0.992429i0.0726863
0.995088
0.992444i0.0727040
0.995104
0.992460i0.0727215
0.995121
0.992214i0.0751451
0.995055
0.992219i0.0751573
0.995061
0.992230i0.0751772
0.995074
0.992249i0.0752049
0.995095
0.960353i0.279283
1.00014
0.977981i0.273811
1.01559
JULY 2003, Vol. 125 313
Fig. 12 Stable T periodic motion at 30000 revmin, e 0 10 m a With TAMB b Without

TAMB
Fig. 13 Stable Quasi-periodic motion at 30000 revmin, e 0 22.5 m with TAMB a Stable
Quasi-periodic motion b Poincare maps
quasi-period or unstable. When values of e 0 and cross the line

from below to above, a pair of Floquet multiplier crosses a unit
circle in the complex plane and secondary Hopf bifurcation
occurs.
It can be seen from Fig. 7 that the stability criteria rotating
speed of the system is decreased when TAMB is considered in
modeling. As an example, the critical stability speed of system
with e 0 0 m is decreased by 41% from 119000 rev/min to
69850 rev/min, and the critical stability speed is decreased by
42% from 113100 rev/min to 65870 rev/min when e 0 is
increased from 0 m to 5 m. This indicates that the effects of
TAMB and mass eccentricity e 0 on the stability of system are both
significant.
4.1
odic solutions are different; the amplitude of the Hopf T periodic

solution with TAMB considered is smaller than that with TAMB
not considered.
2 Mass Eccentricity e 0 4 m. The T periodic motion of
the system produces a quasi-periodic solution after the second
Hopf bifurcation at 113400 rev/min when TAMB is not considered. The corresponding quasi-periodic solution at
113405 rev/min is depicted in Fig. 10a, and the locus of the
center of the runner is shown in Fig. 10b. Its Poincare maps
Effects of TAMB and Mass Eccentricity
1 Mass Eccentricity e 0 0 m. When the effect of TAMB is

not considered, the equilibrium point of the system lost stability
and a Hopf T periodic solution appears at 119000 rev/min.
The corresponding Hopf T periodic solution at
120000 rev/min is depicted in Figure 8a, and a corresponding
amplitude-frequency diagram is shown in Fig. 8b. When TAMB
is considered, the Hopf bifurcation of the equilibrium point occurs
at 69850 rev/min, and a stable Hopf T periodic solution appears at 70000 rev/min which is depicted in Fig. 9a. The
corresponding amplitude-frequency diagram is shown in Fig. 9b.
It is noted that both two Hopf T periodic solutions exhibit typically the first mode whirl of system, and the frequencies are both
about 24.4 Hz. However, the motions of these two Hopf T peri314 Vol. 125, JULY 2003
Fig. 14 Stable periodic motion at 30000 revmin, e 0

22.5 m without TAMB
shown in Fig. 10c reveal that the motion is quasi-periodic as the

Poincare points lie on closed curves. The corresponding time series is shown in Fig. 10d, and its corresponding amplitudefrequency diagram is shown in Fig. 10e. When the TAMB is
considered, the second Hopf bifurcation of the harmonic motion
occurs at 67340 rev/min. A stable quasi-periodic solution appears at 67345 rev/min, which is depicted in Fig. 11a, and
the locus of the center of the runner is shown in Fig. 11b. Its
Poincare maps shown in Fig. 11c, reveal that the motion is
quasi-periodic as the Poincare points lie on closed curves. The
corresponding time series is shown in Fig. 11d, and its corresponding amplitude-frequency diagram is shown in Fig. 11e.
These two quasi-periodic solutions have two frequencies. One
is about 24.4 Hz corresponding to the Hopf T periodic solution
frequency of balanced rotor system, and another corresponds to
harmonic frequency. However, the motions in the lateral directions corresponding to above two solutions are very different. The
motion of the quasi-periodic solution considering TAMB is typically the first mode whirl, and that of the quasi-periodic solution
not considering TAMB is typically the second mode whirl.
4.2 Effects of TAMB and Mass Eccentricity on Leading
Floquet Multiplier. The leading Floquet multiplier of the T periodic motion of the system and its modulus for different mass
eccentricity e 0 is shown in Table 3. It can be seen that all the
moduli of leading Floquet multiplier are increased when TAMB is
considered. Figures 12a and 12b show the stable T periodic
motion for e 0 10 m with and without TAMB respectively.
These two T periodic motions are almost the same.
However, for e 0 22.5 m, the moduli of the leading Floquet
multiplier is increased from 0.995104 to 1.00014 when TAMB is
considered, which means the T periodic motion lost its stability.
Figures 13 and 14 show the stable quasi-periodic motion and periodic motion of the system with and without TAMB respectively.
So the effect of TAMB on system stability is significant.
Conclusions
In this paper, the effect of a TAMB on an active magnetic

bearing-flexible rotor system has been investigated. The finite element method was applied to discretize the flexible shaft. Using
the component mode synthesis method, the degrees of freedom of
equation is reduced. The system equations are obtained by combining the equations of the reduced mechanical system and the
equations of the decentralized PID controllers. Shooting method
and path-following technique are used to get the periodic motions
and their stability margins. The local stability and bifurcation behaviors of the periodic motion are obtained by Floquet theory. The
results indicate that TAMB and mass eccentricity have great influence on nonlinear stability and bifurcation of the T periodic
motion of system, and causes degradation of stability and bifurcation of T periodic motion. Therefore sufficient attentions should
be paid to the spillover effect due to the neglect of TAMB in the
dynamic analysis and design of the AMBRS in order to ensure
system reliability.
Acknowledgment
The work described in this paper was fully supported by a grant
from the Research Grants Council of the Hong Kong Special Administrative Region, China Project No. PolyU5115/98E. It is
also supported by the National Natural Science Foundation, China
Project No. 19990511.
Nomenclature
MS
GS
KS
xS
mass matrices of shaft

damping matrices of shaft
stiffness matrices of shaft
displacement vector of shaft
xb interface or local nonlinearities DOFs

of shaft
xi unloaded internal DOFs of shaft
QS external force vector of shaft
f S nonlinear force vector of bearings act on
shaft
R real number
n numbers of displacement DOFs of shaft
n b numbers of interface or local nonlinearities DOFs of shaft
n i numbers of unloaded internal DOFs of
shaft
angular velocity of rotor
angular speed of rotor
cut cut off frequency
x, y, z Cartesian coordinates
, deflection angles of shaft
b residual flexibility modes
k kept elastic eigenmodes
kk diagonal matrix with the kept angular
eigenfrequencies
T,T1 ,T2 coordinate transformation matrix
m mass of whole rotor
Md mass matrices of disks
Gd damping matrices of disks
Kd stiffness matrices of disks
Q force vector of the system
Fdex excited forces of rotor
e 0 mass eccentricity of couple rotor
M mass matrices of the system
G damping matrices of the system
K stiffness matrices of the system
q displacement vector of the system
xy geometric coupling coefficient of journal
active magnetic bearing
A r pole area of journal active magnetic
bearing
N r number of winding turns of journal active magnetic bearing
0 permeability of air, 4 107 N/A2
N number of winding turns of thrust active
magnetic bearing
c r ,c z clearance of journal and thrust active
magnetic bearing
i x ,i y ,i z control current of rotor journal and
thrust active magnetic bearing
f x , f y nonlinear forces of x and y direction for
journal active magnetic bearing
I 0x ,I 0y ,I 0z bias current of rotor journal and thrust
f r , f l , f t , f b nonlinear forces of right, left, top and
bottom direction for journal active magnetic bearing
R 1 ,R 2 ,R 3 ,R 4 radii of inner and outer annulus for
thrust active magnetic bearing
hi( 1 ) ,hi( 2 ) ,h(o1 ) ,h(o2 ) right and left air gaps at a point on the
inner and outer annulus of thrust active
magnetic bearing
F z ,M x ,M y nonlinear forces and moment of thrust
Ga ,Gs ,Gc transfer function vector of amplifier,
sensor and controller for system
Aa ,As gain vector of amplifier and sensor for
system
Ta ,Ts ,Td time constant vector of amplifier and
sensor for system
Kp ,Ki ,Kd proportional gain, integral gain and derivative gain
JULY 2003, Vol. 125 315
xbb nonlinearities DOFs of active magnetic

bearings
Ibb current vector act on nonlinearities
DOFs of active magnetic bearings
N J number of JAMBs
G transfer function vector of the whole
system
X state variables of system
J Jacobian matrix
system parameter
Subscripts
S shaft
1, 2 right and left annuli of thrust active
magnetic bearing
References
1 Virgin, L. N., Walsh, T. F., and Knight, J. D., 1995, Nonlinear Behavior of a
Magnetic Bearing System, ASME J. Eng. Gas Turbines Power, 1177, pp.
582588.
2 Chinta, M., Palazzolo, A. B., and Kascak, A., 1996, Quasiperiodic Vibration
of a Rotor in a Magnetic Bearing With Geometric Coupling, Proc. 5th International Symposium on Magnetic Bearing, Kanazawa, Japan, pp. 147152.
3 Chinta, M., and Palazzolo, A. B., 1998, Stability and Bifurcation of Rotor
Motion in a Magnetic Bearing, J. Sound Vib., 2145, pp. 793 803.
4 Mittwollen, N., Hegel, T., and Glienicke, J., 1991, Effect of Hydrodynamic
Thrust Bearings on Lateral Shaft Vibration, ASME J. Tribol., 1134, pp.
811 818.
5 Yu, L., and Bhat, R. B., 1995, Coupled Dynamics of a Rotor-Bearing System
Equipped with a Hydrodynamic Thrust Bearing, Shock and Vibration 21,
pp. 114.
6 Lund, J. W., and Nelson, H. B., 1980, Instability Threshold of an Unbalanced
Rigid Rotor in Short Journal Bearings, Proc. Second International Conference on Vibration in Rotating Machinery, Cambridge, UK.
7 Brancati, R., Rocca, E., Rosso, M., and Rosso, R., 1995, Journal Orbits and
Their Stability for Rigid Unbalanced Rotors, ASME J. Tribol., 117, pp. 709
716.
316 Vol. 125, JULY 2003
8 Kim, Y. B., and Noah, S. T., 1990, Bifurcation Analysis for a Modified
Jeffcott Rotor with Bearing Clearance, Nonlinear Dynamics, 1, pp. 221241.
9 Choi, S. K., and Noah, S. T., 1994, Mode-Locking and Chaos in a Jeffcott
Rotor with Bearing Clearance, ASME J. Appl. Mech., 61, pp. 131138.
10 Nelson, H. D., Mechan, W. I., Fleming, D. P., and Kascak, A. F., 1983, Nonlinear Analysis of Rotor Bearing System Using Component Mode Synthesis,
ASME paper No. 83-GT-303.
11 Nataraj, C., and Nelson, H. D., 1989, Periodic Solutions in Rotor Dynamic
System With Nonlinear Supports: A General Approach, ASME J. Vibr.
Acoust., 111, pp. 187193.
12 Fey, R. H. B., Van Campen, D. H., and de Kraker, A., 1996, Long Term
Structural Dynamics of Mechanical Systems With Local Nonlinearities,
ASME J. Vibr. Acoust., 118, pp. 147153.
13 Zheng, T., and Hascbe, N., 2000, Nonlinear Dynamic Behaviors of a Complex Rotor-Bearing System, ASME J. Appl. Mech., 67, pp. 485 495.
14 Nelson, H. D., 1980, A Finite Rotating Shaft Element Using Timoshenko
Beam Theory, ASME J. Mech. Des., 102, pp. 793 803.
15 Lalanne, M., and Ferraris, G., 1997, Rotordynamics Prediction in Engineering,
John Wiley, New York.
16 Iooss, G., and Joseph, D. D., 1980, Elementary Stability and Bifurcation
Theory, Springer-Verlag, New York.
17 Parker, T. S., and Chua, L. O., 1989, Practical Numerical Algorithms for
Chaotic System, Springer-Verlag, New York.
18 Seydel, R., 1988, From Equilibrium to Chaos, Practical Bifurcation and Stability Analysis, Elsevier, New York.
19 Sundararajan, P., and Noah, S. T., 1997, Dynamics of Forced Nonlinear Systems Using Shooting/Arclength Continuation Method-Application to Rotor
System, ASME J. Vibr. Acoust., 1191, pp. 1020.
20 Craig, R. R., Jr., 1985, A Review of Time-Domain and Frequency-Domain
Component Modes Synthesis Methods, Combined Experimental/Analytical
Modeling of Dynamic Structural Systems Using Substructure Synthesis, D. R.
Martinez and A. K. Miller, eds., ASCE/ASME, New York, NY, pp. 131.
21 Knight, J. D., Xia, Z., and McCaul, E. B., 1992, Forces in Magnetic Journal
Bearings: Nonlinear Computation and Experimental Measurement, Proc. 3rd
International Symposium on Magnetic Bearing, Alexandria, VA, pp. 441 450.
22 Ho, Y. S., Yu, L., and Liu, H., 1999, Rotor Dynamic Coefficients of a Thrust
Active Magnetic Bearing Considering Runner tilt, Proc. Inst. Mech. Eng.,
Part J: J. Eng. Tribol., 213J6, pp. 451 462.
Chun-Ping Zou
School of Mechanical and Power Engineering,
Shanghai Jiao Tong University,
Shanghai, Peoples Republic of China;
Department of Mechanical and Power
Engineering,
East China Shipbuilding Institute,
Zhenjiang, Peoples Republic of China
Duan-Shi Chen
Hong-Xing Hua
School of Mechanical and Power Engineering,
Shanghai, Peoples Republic of China
Torsional Vibration Analysis of

Complicated Multi-Branched
Shafting Systems by Modal
Synthesis Method
The torsional vibration calculations of the complicated multi-branched system with rigid
connection and flexible connections made up of elastic-coupling parts are very difficult to
perform using conventional methods. In this paper, a modal synthesis method of torsional
vibration analysis for the system is proposed. This approach is an improved method of
Hurtys fixed-interface and Hous free-interface modal synthesis methods. Because of the
introduction of flexible substructure, the improved modal synthesis method can effectively
treat the complicated system in which there exists a rigid connection and a flexible
connection that is formed by an elastic-coupling part. When the calculation is performed,
the complicated multi-branched system is divided into several substructures that are analyzed by FEM (finite element method) except the special elastic-coupling part that is
defined as flexible substructure and treated individually. The efficiency of modal synthesis
is improved by choosing suitable number of lower-frequency modes in modal synthesis. As
an example of an application of this method, the analysis of torsional vibration of a
cam-type engine shafting system is carried out both numerically and experimentally. The
results show that the above kind of multi-branched shafting system can be analyzed
effectively by the proposed method. DOI: 10.1115/1.1569949
Introduction
There are several kinds of methods for torsional vibration calculation of rotating shafting system, namely Holzers method,
transfer matrix method and FEM. The well-known Holzers
method, a simple and systematic approach to calculating the frequencies and mode shapes of systems, was used originally for
analysis of an undamped system. Den Hartog and Li 1 improved
the Holzers method by using complex numbers. Pestel and
Leckie 2 introduced the transfer matrix method with the point
and field transfer matrices. Later, Sankar 3 and Dawson and
Davies 4 improved this approach. Huang and Horng 5 used
complex numbers to extend the transfer matrix method with the
Newton-Raphson technique to analyze the torsional vibration for
damped systems. The method extended by Huang and Horng
eliminates the operation of the inverse matrix because the derivatives of angular displacement and the torque are used directly with
the Newton-Raphson technique to determine the eigenvalues of
torsional vibration system. The major drawbacks of the transfer
matrix approach as noted by Firoozian and Stanway 6 are that
the information obtained is limited to stability assessment and the
prediction of critical speeds, and that inherent numerical instabilities are liable to occur in the analysis of more complex systems.
Li 7 used FEM to analyze the crankshaft torsional vibration of
diesel engine. One of the characteristics of FEM is that a very
large number of degrees of freedom are required. All these methods, namely Holzers method, transfer matrix method and FEM,
are very effective for the calculation of a single-spool shafting
system. However, when these methods are applied to calculate
shafting systems with open-circuit and closed-circuit branches,
they are not very effective.
Modal synthesis method is a modeling method permitting the
representation of a relatively complex structure by a reduced number of degrees of freedom. The concept of substructure synthesis
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received November
2001; revised November 2002. Associate Editor: B. Yang.
can be traced to Hurty 8,9. The general idea is to regard a given

structure as an assemblage of substructures acting together in
some fashion. To guarantee that the various substructures do not
act independently but as parts of the whole structure, Hurty introduced the concept of constraint modes. Generally, the constraint
modes enforce geometric compatibility at a preselected set of
points on the boundary between any two adjacent substructures
10. Hou advocates that substructure modes are not necessarily
constrained and can be free-free 11. The use of unconstrained
modes is also advocated by Goldman and Dowell 12,13. Since
publication of the original work by Hurty and Hou, there have
been many suggestions for improvement 14 19. One of the
most important improvements is that the effect of modes not retained explicitly is considered by MacNeal 19 and by Rubin
16. The procedure described by MacNeal employs hybrid component modes, i.e., modes obtained with some interface coordinates free and others fixed, and uses statically derived modes to
improve accuracy. The procedure suggested by Rubin employs
residual flexibility to a second order of approximation.
The modal synthesis methods discussed above have one aspect
in common, namely they all regard the connection between any
two substructures as a rigid connection. However, flexible connection made up of elastic-coupling parts is widely applied to engineering structure with multi-branched shafting system, e.g., the
elastic coupling of marine propulsion system. The boundary displacement that exists in-between substructures of such structure is
incompatible. Hale and Meirovitch 20 proposed the concept of
incompatible intermediate structure to treat the problem, and
the original structure is substituted by this structure approximately. In this paper, a modal synthesis method improved from
Hurty and Hous methods is presented to accommodate the vibration analysis of multi-branched shafting system. It can effectively
treat the complicated system in which there exists a rigid connection and a flexible connection that are formed by an elasticcoupling part. When the proposed approach is developed, the
elastic-coupling part mentioned above is defined as flexible substructure and treated individually. Uncoordinated boundary dis-
JULY 2003, Vol. 125 317
placement of each substructure is converted into innerdeformation of flexible substructure so that the coordinative
boundary condition between the flexible substructure and other
substructures could be satisfied. The other parts are naturally divided into some substructures and they can be analyzed by FEM.
The lower-frequency normal modes of substructures are retained
and the higher-frequency normal modes are neglected by a frequency truncation criterion. The lower-frequency normal modes
are considered to be the assumed modes of Rayleigh-Ritz analysis
of whole structure. As an example of application of this method,
the analysis of torsional vibration of a cam-type engine shafting
system is carried out both numerically and experimentally.
Modal Synthesis Method
Many structures used in engineering are very complicated in

which there are not only rigid connections but also some kinds of
flexible connections which exist in-between adjacent substructures. The torsional vibration model of a multi-branched shafting
system is shown in Fig. 1. The system is divided into subsystems
in the following principles:
1. Main shaft i.e., a and its direct branches are considered as
first-class substructure i.e., b, c, and d.
2. The branches belonging to first-class substructure are considered as second-class substructure a consists of secondclass substructures 1,2, ,m).
3. The elastic-coupling part should be taken out and defined as
flexible substructure i.e., e1 , e2 ).
There are two kinds of ways of connection between substructures, shown in Fig. 1. One is a rigid connection, e.g., the connection between second-class substructures 1,2, ,m. Another is a
flexible connection, e.g., substructures a and b are connected by e1
and substructures c and d are connected by e2 .
2.1 Modal Synthesis Between Second-Class Substructures.
As an example, first-class substructure a consists of m secondclass substructures, shown in Fig. 1. Each of the second-class
substructures is calculated by FEM. The free vibration equations
of each second-class substructure with zero damping are 15
s K s Q s 0
M s Q
s1,2,m
s s1 , s2 , si sn
where si is the ith order column vector of dominant mode

shape of sth second-class substructure; and n is the number of
degrees of freedom of sth substructure.
The dominant mass matrix is obtained according to the orthogonal
condition
s sT M s s diag M s1 ,M s2 M sn
M
(3)
M si
where
is the ith dominant mass in dominant mass matrix.
The normalized modal matrix from 2 and 3 is
s
s1 s2
,
1
,
2
M s M s
si
,
i
M s
sn
M sn
s1,2,m
(4)
According to vibration theory, the relationship between system

coordinates Q s and normal coordinates P s is
s P s
Q s
(5)
By substituting Eq. 5 into 1 and multiplying the left hand side

sT , the following is obtained.
by
sT M s
s P s
sT K s
s P s 0
(6)
According to the orthogonal condition, we have

sT M s
s I ;
sT K s
s
where I is unit matrix; is a diagonal matrix whose diagonal

element is the square of each natural circular frequency, that is to
say, diags2.
According to the above equations, Eq. 6 becomes
I P s diag s2 P s 0
(7)
The normalized modal matrix 4 is denoted as x s , and it is

divided into two parts.
s x s x s1 ,x s2 ,x si ,x sn
(1)
s , Q s respectively represent the total

where K s , M s , Q
stiffness matrix, total mass matrix, angular acceleration vector,
and angular displacement vector of sth second-class substructure;
and m is the number of second-class substructures.
The dominant mode shape of each second-class substructure from
1 is
s1,2,m
(2)
x sk ,x sa s1,2,m
(8)
x s( i )
where
is ith order of normalized modal column vector of sth
second-class substructure; x sk represents the prior l order retained
modes; x sa represents the truncation modes of the order from (l
1) to n; and n is the number of degrees of freedom of sth
substructure.
Fig. 1 Torsional vibration model of complicated multi-branched shafting system
318 Vol. 125, JULY 2003
The diagonal matrix of squared frequency and normal coordinates

can also be divided into two parts, i.e., retained part and truncated
part respectively;
2
2
s,a
diag s2 diag s,k
(9)

P sk
P s P a
s
(10)
According to frequency truncation criterion, the transformation

relationship between system coordinates Q s and normal modal
coordinates P s of sth substructure can be easily obtained by
substituting Eqs. 8 and 10 into 5.

P sk
Q s x s P s x sk x sa P a x sk P sk
s
(11)
Substituting Eq. 11 into 1 and multiplying the left hand side by

x sk T , the following equation is obtained according to the orthogonal condition.
2
I P sk diag s,k
P sk 0
s1,2,m
Fig. 2 The model of flexible substructure
the other substructures. Assume that the shafting system consist of

first-order substructure a, b and flexible substructure e, and each
elastic unit can be simulated by a set of mass-spring elements. The
flexible substructure e, as shown in Fig. 2, is the combination of
these mass-spring elements. For the ith element of flexible substructure e, the following equation is obtained
(12)
According to equation 12, we have

I P L K L P L 0
(13)
k T
where P L p k1 ,p k2 ,p m
; K L diagdiag21,k ,diag22,k ,
2
diagm,k]. P L is the matrix of normal modal coordinates of
m second-class substructures where some coordinates are independent of each other. Using boundary connection condition, the
UN-independent coordinates can be eliminated
Q sb x sb P s
s1,2,m
(14)
where Q sb is boundary displacement vector; x sb is the matrix

composed of those rows corresponding to boundary connection
points in matrix x s .
The coordinated condition of boundary displacement is
b
(15)
x mm P m x 1 1 P 1
x b P L x b1
that is
x b1 P b x b2 P l 0
(16)
(17)

Pb
x b1 1 x b2
P l T P l
I
where
M I T T ,
T
(19)
K l T K L T
T
Equation 19 is free vibration equation of first-class substructure

a, from which the modal data of a can be solved.
2.2 Model of Flexible Substructure. For the modal synthesis of substructures connected by flexible substructure, noncoordinative boundary displacement has to be converted into innerdeformation of flexible substructure to satisfy the coordinative
boundary condition at the connection of flexible substructure and
q a j q a1 j ,q a2 j ,q ai j ,q an j T
is the boundary displacement of substructure a;
q b j q b1 j ,q b2 j ,q bi j , ,q bn j T
q m q m1 ,q m2 , ,q mi ,q mn T
is the force column vector of elastic coupling unit;
F a j F a1 j ,F a2 j ,F a1 j , ,F an j T
is the boundary force of substructure a;
F b j F b1 j ,F b2 j , ,F bi j ,F bn j T
is the boundary force of substructure b;
is the inner force of elastic coupling unit.

(18)
By substituting Eq. 18 into 13 and multiplying the left hand

side by T T , the following is obtained.
M l P l K I P l 0
is displacement vector of elastic coupling unit;
F m F m1 ,F m2 , ,F mi , ,F mn T
P b x b1 1 x b2 P l ,
P L P
I
Q e q a j T , q m T , q b j T T
F e F a j T , F m T , F b j T T
where P b and P l respectively represent the matrix of unindependent normal modal coordinates and the matrix of independent normal coordinates among P 1 , P 2 .... P m ; and x b1 is a
square matrix.
The following equation is given from Eq. 17
so
(21)
where
is the inner displacement of elastic coupling unit;
By synthesizing Eq. 15, and writing in blocked form

Pb
x b2
0
P1
e K e Q e F e
and in matrix form: M e Q
(20)
is the boundary displacement of substructure b;
x 11 P 1 x 22 P 2 ; x 22 P 2 x 33 P 3
i
K ai q ai j q m
F ai j
i
i
i
i
m l q m K ai q a j q m K bi q bi j q m
0
i
i
i
K bi q b j q m F b j
Under free vibration condition, F m 000 T , i.e., null

force is exerted inside flexible substructure. M e and K e are
the mass matrix and stiffness matrix of flexible substructure
respectively.
2.3 Modal Synthesis of Two Substructures Connected by
Flexible Substructure. The shafting system, as shown in Fig. 1,
can be considered as two parts, i.e., a, e1 , b and c, e2 , d which are
the combination of two first-class substructures connected by a
flexible substructure. The modal synthesis among a, e1 , b is used
to explain the modal synthesis method of two substructures connected by a flexible substructure. To perform vibration analysis of
the whole structure and calculate the response of the system under
external excitation, the vibration equation of substructures a and b
is expressed as follows by Zhang 14
JULY 2003, Vol. 125 319
i K i Q i F i ia,b
M i Q
(22)
i , Q i , and F i represent the mass mawhere M i , K i , Q

trix, stiffness matrix, acceleration, displacement and force vector
of ith substructure respectively.
The dominant mode matrixes of first-class substructure a, b are
given as a and b respectively by using the methods in
Section 2.1. Choosing lower-frequency modes la and lb
and denoting the corresponding mode coordinates as P la and
P lb , one has
Q a la P la Q b lb P lb
(23)
There is transformation relationship
Q T1 P
where
Q Q Ta
T
P P la
, Q Te
, Q Te
(24)
, Q Tb T ,
T T
, P lb
,
Tl
la
lb
According to Eq. 23, the boundary displacement Q j a and

Q j b between substructure a, b and flexible substructure e1 are
given as follows
Q j a j la P la
Q j b j lb P lb
(25)
where, b j c la and b j c lb are composed of the rows corresponding

to the boundary in la and lb . The coordinative boundary
conditions of substructures a, e1 , b are given by
Q e Q j Ta , Q j Tb T
(26)
and the coupling transformation relationship is
P T2q
T
T T
where q P la
, P lb
; T2
jla
jlb
(27)
F T T F * .
For free torsional vibration, the outer vector F 0 obviously.
To solve Eq. 30 and perform transformation of vibration modes
according to 28, the natural frequency and modal shapes of the
total system can be obtained.
3 Torsional Vibration Analysis of Shafting System of

Cam-Type Engine
3.1 Structure of the Shafting System and Its Calculation
Model. The shafting system of cam-type engine is shown in Fig.
3. The combustion chamber 21 is in the left side. There are 5
cylinders located uniformly along the periphery around the axis of
power output shaft. The gas with high temperature and high pressure produced in the combustion chamber enters the cylinder 4
when engine is running. The gas forces piston 5 to move reciprocally, whose movement impels cam-disc 6 to rotate. That is to say,
the axial force is converted into circumferential force, which
drives the inner shaft 9 and rear propeller 13 to rotate around
power output shaft. Simultaneously, the counter-moment of camdisc acting on piston assembles drives cylinder, cylinder-body,
outer shaft, and the front propeller to rotate in the opposite direction. The output moment M i of inner shaft is equal to the output
moment M 0 of outer shaft in magnitude. However, they are in
opposite direction. Gear 3 fixed on cylinder-body 15 drives fuel
pump, cooling water pump, and generator. Besides the reciprocal
movement, the piston assembles rotate around axis of engines
output shaft together with the cylinder. This makes torsional vibration analysis of cam-type engine too difficult to be performed
by conventional methods.
As the number of peak on cam-disc profile m2 and the number of cylinder Z(2m1), the cam-type engine will be selfbalanced because the unbalanced inertial force and moment are
equal to zero. For the example under consideration, the output
torsional moment curve of the engine with m2, Z5 is shown
in Fig. 4. It is seen that there are 5 similar waveforms on the
curve, each of which could be attributed to one piston. The excitation of torsional moment is related to rotation speed of inner
By substituting Eq. 27 into 24, the transformation relationship

of modal synthesis is derived
Q T 1 T 2 q T q
(28)
To assemble substructures a, e1 , b together in form, the following

is given.
K * Q F *
M *Q
(29)
where
M * diag M a , M e , M b ,
K * diag K a , K e , K b ,
and
Fig. 3 Structure of cam-type engine shafting system
F * F Ta , F Te , F Tb T
M i , K i , F i ia,b can be obtained from equation 22 .
Substituting Eq. 28 into 29 and multiplying the left side by

T T , the following is obtained.
q K
q F
M
(30)
where
T T M * T , K
T T K * T
M
and
320 Vol. 125, JULY 2003
Fig. 4 The picture of output moment of torsion
quencies and mode shapes of first-class substructure b and

second-class substructures c, d, e, and f are calculated by using the
FEM. Second, the torsional vibration of first-class substructure a
can be obtained by modal synthesis based on the lower-frequency
modes of the second-class substructure. Last, the first-class substructure a, b and flexible substructure s are synthesized to obtain
the torsional vibration of the whole system. Table 1 shows the
calculated results of natural frequencies according to different
stages of modal synthesis. It is seen that the calculated natural
frequencies of system decrease when the number of the reasonable substructures involved increases.
To find the effect of truncation of higher-frequency modes, the
natural frequencies of the whole system are calculated with different number of modes retained. The results are shown in Table
2. It is seen that when the total number of modes is 67, the relative
error of natural frequencies ranges only from 0.508% to 2.10%
while the number of modes involved decreases from 67 to 15.
That is to say, high accuracy of calculation can be reached with
less lower-frequency modes involved in mode synthesis.
Fig. 5 Calculation model of torsional vibration for cam-type

engine shafting system
shaft (n i 2185 rpm) and outer shaft (n 0 2085 rpm) so that the
fundamental excitation frequency of torsional vibration can be
calculated as Zhao 21:
f 0 mZ n i n 0 /6025 21852085 /60711.67 Hz
In the past, the inner shaft and outer shaft of cam-type engine
were treated separately, and only one of the two natural frequencies of inner shaft and outer shaft was considered as the natural
frequency of whole system so that the results of calculation are
not realistic 22. Owing to the existence of interaction, the inner
shaft and outer shaft should be considered as a whole system in
vibration analysis. The whole system is modeled as two first-class
substructures connected by a flexible substructure. The combination of cylinder-body, outer shaft, front propeller and enginedriven auxiliary machinery is defined as first-class substructure a.
The combination of cam-disc, inner shaft and rear propeller is
defined as first-class substructure b. The piston assembles to link
inner shaft and outer shaft are treated as flexible substructure s in
order to treat the uncoordinated angular displacements. First-class
substructure a is further divided into second-class substructures,
i.e., the combination of cylinder-body, outer shaft and front propeller is defined as second-class substructure c, cooling water
pump as second-class substructure d, fuel pump as second-class
substructure e and generator as second-class substructure f, as
shown in Fig. 5.
Experiment
In order to validate the proposed modal synthesis method and

determine the dominant frequencies of excitation, the experiment
has been conducted in two cases. The experimental set-up of fullsize cam-type engine is shown in Fig. 6. The signal gear for torsional vibration measurement is fixed on outer shaft.
Case 1: Natural Frequencies of Shafting System of Cam-type
Engine in Static Condition.
The simplest test to check the fundamental natural frequencies
of any vibratory system is the rap test. To check the natural frequencies of torsional vibration, the forces exerted on the system
are different from the rap forces in commonly rap test. A circumferential rap force is successfully applied at the shafting system by
a special instrument. The natural frequencies of the rap test and
numerical calculation are listed in Table 3. It can be found that the
calculation results of natural frequencies are in good agreement
with the experimental results. So the modal synthesis method proposed in this paper is very effective for torsional vibration analysis of the complicated multi-branched shafting system of the camtype engine.
Case 2: The Dominant Excitation Frequencies of Shafting System of Cam-type Engine.
To determine the dominant excitation frequencies, the tests of
operation of cam-type engine are performed in two ways, namely
no load test and low-load test. The dominant excitation frequencies are measured under rotating conditions. The inner and outer
shafts are respectively run at 2185 rpm and 2085 rpm, but the two
3.2 Analysis of Torsional Vibration. Based on the model

shown in Fig. 5, the characteristics of torsional vibration for camtype engine shafting system can be determined by using the proposed method which is incorporated into computer software designed by the authors. Natural frequencies of torsional vibration
are the matter of our main concern. When the calculation is carried out, the following steps are followed. First, the natural fre-
Table 1 The natural frequencies of modal synthesis according to substructures involved Hz

Substructure
Substructure c
Substructure a
Substructure b
whole system
404.825
370.310
400.180
198.010
5431.014
1220.826
3577.616
371.843
6296.115
6124.352
5139.323
453.629
10298.600
6675.603
5981.029
1220.560
12164.090
10898.827
6835.063
3576.992
14839.410
12194.780
9408.980
5137.702
Table 2 Natural frequency and error with selected different number of modes retained Hz
Retained No. of modes
Total number 67
Retained number 55
Error %
Retained number 45
Error %
Retained number 35
Error %
Retained number 25
Error %
Retained number 15
Error %
198.010
198.014
0.002
198.101
0.046
198.175
0.083
198.463
0.229
199.016
0.508
371.843
371.848
0.001
372.222
0.102
372.410
0.152
373.589
0.470
374.216
0.638
453.629
453.645
0.003
454.114
0.107
454.983
0.298
455.978
0.518
456.814
0.702
1220.560
1220.748
0.015
1224.303
0.306
1228.936
0.686
1234.871
1.172
1246.213
2.10
3576.992
3577.998
0.028
3583.242
0.175
3588.367
0.318
3597.832
0.583
3604.314
0.764
5137.702
5138.896
0.023
5142.056
0.085
5150.724
0.253
5160.411
0.442
5166.865
0.568
JULY 2003, Vol. 125 321
Fig. 6 Experimental set-up of cam-type engine
shafts run in opposite directions. Results for the tests are shown in
Fig. 7. It is seen that the component of 34.75 Hz is predominant in
Fig. 7, which is corresponding to the rotating speed of engine
outer shaft where signal gear is fixed on. Many frequencies corresponding to the higher-frequency peaks are the integral times of
34.75 Hz. This is caused by the excitation of unbalanced moment
on outer shaft. Another predominant peak is at about 710 Hz,
which is in accordance with the fundamental excitation frequency
of torsional vibration mentioned above. It is caused by the pulsating output moment of torsion pulsation frequency is f
711.67 Hz). The reason for the phenomenon is that each cylinders work of cam-type engine is uneven and that the moment of
torsion amplitude acted on cam-disc is different.
Table 3 Natural Frequencies of shafting system Hz
Calculation
Experiment
Error %
198.010
195.3
1.39
371.843
369.6
0.61
Conclusion
A modal synthesis method is proposed to analyze the torsional

vibration of multi-branched shafting system where rigid coupling
and flexible connection exist. When the proposed approach is developed, a new concept of connecting substructure, namely flexible substructure, is presented. It is used to tackle the elasticcoupling part of multi-branched shafting system. The other parts
of the system are partitioned into some substructures that are analyzed by FEM. The lower-frequency normal modes of substructures are retained while the higher-frequency normal modes are
neglected. The lower-frequency normal modes are regarded as the
assumed modes of Rayleigh-Ritz analysis of whole structure. The
proposed approach is used to analyze the torsional vibration of
cam-type engine shafting system. We can draw the following
conclusions:
1. The concept of flexible substructure and the improved modal
synthesis method presented in this paper are of value for torsional
vibration analysis of complicated system where uncoordinated
boundary conditions exist.
2. For the shafting system of cam-type engine, its fundamental
excitation frequency of torsional vibration 710 Hz is just inbetween the 3rd natural frequency 453.629 Hz and 4th natural
frequency 1220.56 Hz. Furthermore, the dominant excitation frequencies of the cam-type engine are not coincided with natural
frequencies of the shafting system so that there is no resonance
induced. In consequence, torsion vibration of the cam-type shafting system is not harmful during its normal operation.
3. The cam-type engine shafting system that contains inner and
outer shaft assembles should be considered as a whole system as
the natural frequency decreases with the increasing of the number
of substructures involved in modal synthesis. Using the proposed
method, the inner and outer shaft assembles are analyzed simultaneously and the torsional vibration of the whole system is
obtained.
4. The numerical results of natural frequencies are in good
agreement with the experimental results. The reasonable accuracy
of natural frequency can be reached by considering less number of
lower-frequency modes in modal synthesis calculation.
References
Fig. 7 The frequency spectrum of torsional vibration
322 Vol. 125, JULY 2003
1 Den Hartog, J. P., and Li, J. P., 1964, Forced Torsional Vibration with Damping: An Extension of Holzers Method, ASME J. Appl. Mech., 31, pp. 276
280.
2 Pestel, E. C., and Leckie, F. A., 1963, Matrix Method in Elasto Mechanics,
McGraw-Hill, New York.
3 Sankar, S., 1979, On The Torsional Vibration of Branches System Using
Extended Transfer Matrix Method, ASME J. Eng. Ind., Series B, 101, pp.
546 553.
4 Dawson, B., and Davies, M., 1974, An Improved Transfer Matrix Procedure, Int. J. Numer. Methods Eng., 8, pp. 111117.
5 Huang, Y. M., and Horng, C. D., 1999, Analysis of Torsional Vibration Sys-
6
7
8
9
10
11
12
13
14
tems by the Extended Transfer Matrix Method, ASME J. Vibr. Acoust., 121,
pp. 250255.
Firoozian, R., and Stanway, R., 1989, Design and Application of A Finite
Element Package for Modelling Turbomachinery Vibrations, J. Sound Vib.,
134, pp. 115137.
Li, H. Z., 1991, Crankshaft Torsional Vibration Calculation by Finite Element Method, Journal of Internal Combustion Engines, 9, pp. 157162 in
Chinese.
Hurty, W. C., 1960, Vibration of Structural System by Component Mode
Synthesis, J. Eng. Mech. Div., ASCE, 86, pp. 51 69.
Hurty, W. C., 1965, Dynamic Analysis of Structural System Using Component Modes, AIAA J., 3, pp. 678 685.
Craig, Jr., R. R., and Bampton, M. C. C., 1968, Coupling of Substructures for
Dynamic Analysis, AIAA J., 6, pp. 13131319.
Hou, S. N., 1969, Review of Modal Synthesis Techniques and A New Approach, The Shock and Vibration Bulletin, 40, pp. 2539.
Goldman, R. L., 1969, Vibration Analysis by Dynamic Partitioning, AIAA
J., 7, pp. 11521154.
Dowell, E. H., 1972, Free Vibration of an Arbitrary Structure in Terms of
Component Modes, ASME J. Appl. Mech., 39, pp. 727732.
Zhang, H. T., 1990, Free-Interface Mode Synthesis Method for Vibration
15
16
17
18
19
20
21
22
System of Contains Stiffness Coupling Component, Journal of Engineering

Mechanics, 7, pp. 93101 in Chinese.
Xu, K. Q., 1989, Non-coordination Dynamic Substructure Synthesis
Method, Journal of Vibration and Shock, 31, pp. 64 67 in Chinese.
Rubin, S., 1975, Improved Component-Mode Representation For Structural
Dynamic Analysis, AIAA J., 13, pp. 9951006.
Inamura, T., Suzuki, H., and Sata, T., 1994, An Improved Method of Dynamic Coupling in Structural Analysis and Its Application, ASME J. Dyn.
Syst., Meas., Control, 106, pp. 82 89.
Gaganis, B. J., 1999, Modal Analysis of Rotor on Piecewise Linear Journal
Bearings Under Seismic Excitation, ASME J. Vibr. Acoust., 121, pp. 190
196.
MacNeal, R. H., 1971, A Hybrid Method of Component Mode Synthesis,
Comput. Struct., 1, pp. 581 601.
Hale, A. L., and Meirovitch, L., 1982, A Procedure for Improving Discrete
Substructure Representation in Dynamic Synthesis, AIAA J., 20, pp. 1128
1136.
Zhao, L. F., 1991, The Principle of Torpedo Piston Engine, Northwestern
Polytechnic University Press, Xian, China, pp. 258 269 in Chinese.
Zou, C. P., 1994, FEM Analysis of Later Vibration of Torpedo Propulsion
System, Journal of Torpedo Technology, 2, pp. 2232 in Chinese.
JULY 2003, Vol. 125 323
Study on the Dynamics of a Rotor

in a Maneuvering Aircraft
Fusheng Lin
Guang Meng
This paper shows how the dynamics of a rotor in a maneuvering aircraft changes according to the operation of the aircraft. The mathematical model of an unbalanced rotor
system located in the maneuvering aircraft is derived. The dynamic characteristics of the
rotor running at a constant angular speed or a constant acceleration are studied under
the assumptions that the aircraft maneuvers only in a vertical plane and that the pitching
angle and the flight path inclination of the aircraft are equal. The effects of gravity and
unbalance parameter are considered. The results show that the unbalanced response of a
rotor in an aircraft is obviously influenced by the aircrafts flying status.
DOI: 10.1115/1.1576422
e-mail: gmeng@mail.sjtu.edu.cn
State Key Lab of Vibration,
Shock & Noise,
Shanghai 200030, P. R. C.
Introduction
In the previous literatures on rotor dynamics, the supports of

the rotor are assumed to be static and in the same horizontal or
vertical direction 12. This is not practical for rotors located in
aircraft because the aircraft operates under various conditions during maneuvering. Many researchers are focused on studies of only
the maneuvering flight of aircraft 37, but there is no literature
analyzed the dynamics of a rotor running in an operating aircraft.
This paper attempts to determine the dynamic characteristics of a
rotor and to simulate, more realistically, the response of such a
rotor in a maneuvering aircraft.
Equations of Motion
In Fig. 1, OXYZ is a space-fixed and stationary coordinate system, and O 1 X 2 Y 2 Z 2 is an aircraft-carried coordinate system parallel to OXYZ. It is supposed that the pitching angle and the flight
path inclination of the aircraft are equal and defined as , i.e., the
axis of the rotor shaft O 1 Z 1 is in the same direction as the longitudinal axis of the fuselage. It is also assumed that O 1 X 1 Z 1 is in
the same vertical plane with OXZ and O 1 X 2 Z 2 , and the aircraft
moves only in the vertical plane. A Jeffcott rotor model is used.
The displacements of the disk center are all measured in a bodyfixed and rotating O 1 reference frame, where and are in
the principal directions of the cross-section of the shaft and
coincides with the Y 1 axis, as shown in Fig. 2. The flexibility of
the rotor is r. The temporary position of the disk center is O 3 . The
position of the mass center of the disk is C. The angle between the
axes O 1 and O 1 Y 1 is . The coordinate O 1 rotates with
respect to O 1 (O 1 Z 1 ) with an angular speed . It is noted that e 0
is a reference eccentricity, m is the equivalent mass of the heavy
disk mounted at the mid-span of a massless elastic shaft, e is the
eccentricity of the disk, is the angle between the orientation of
the eccentricity and the 1 axis, k is the stiffness coefficient of the
shaft, and c is the external viscous damping coefficient of the shaft
and disk.
The dimensionless differential equations of motion of the rotor
system can be written as follows:
1
2
2
1
1
d 2 2 2 1cos 2 d 2 d
2 sin 2 d U 2 sin
2
Ge
2
2
2 cos U sin cos
x cos z d sin x d cos z d sin cos 0
d
2
1
1
1
2
d d U sin d 2 2 2 1cos 2 d 2 d
2 sin 2 d U 2 cos
2
Ge
2
cos U 2 sin sin

x cos z d sin x d cos z d sin sin 0
2
d
d d U cos
where,
d /e o , d /e o , c k/m, c/2m c , / c ,
G e g/ 2c e o , Ue/e o , x d x/e o , z d z/e o . is the dimensionless viscous damping coefficient, G e is the gravity parameter, U is
the unbalance parameter, and denotes differentiation with respect
to dimensionless time .
If the movement of the aircraft is given, i.e., x( ) and z( ) are
given, then the angle can be calculated by

in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received August 2002;
Revised manuscript received March 2003. Associate editor: G. Flowers.
324 Vol. 125, JULY 2003
tg
dx dx dz
dz dt dt
arctg
(1)
x x x d
,
z z z d
x d
z d
Accordingly, can be obtained.

The numerical response of the system in O 1 coordinate can
be obtained by the Runge-Kutta integration method. Response in
O 1 X 1 Y 1 Z 1 coordinate can also be obtained through a coordinate
transformation.
Equation 1 describes an unbalanced rotor system located in an
aircraft moving in the vertical plane. Assuming that the rotating
motion of the rotor, i.e., and can be obtained, Eq. 1 can be
Fig. 1 Sketch of the rotor system located in an aircraft and the

space-fixed stationary coordinates
simplified and then solved. When the rotor rotates at a constant

speed , then 0 and 0 . If the rotor accelerates at
a constant rate of acceleration C a , then C a 0 and
0 C a , where 0 is the initial angular speed of the rotor.
Numerical Analysis
When the aircraft flies at a constant rate of acceleration, the

dimensionless x d0 and z d0 , x d and z d , and x d and z d are assumed
to be the initial velocity components, the velocity components,
and the acceleration components in the directions of the X and the
z d and x d x d0
x d . If the
Z axes, respectively. Then z d z d0
flight path of an aircraft is a sine curve in the vertical plane OXZ
and the horizontal component of the aircraft velocity is constant,
the motion can be written as x d x d0 sin(zd). Thereby the
acceleration components of the aircraft can be obtained. When the
rotor rotates at a constant speed, the responses at three different
speed ratios (0.8,0.3,1.6) are calculated. When the rotor rotates at a constant acceleration, two cases of the initial speed ratio
( 0 0.8,0.3) are investigated.
The numerical responses of some cases are shown from Fig. 3
to Fig. 7. In these figures the amplitude refers to the maximum
amplitude of the disk center in every revolution and is nondimen-
Fig. 2 The reference frame O 1 and parallel frame O 3 1 1 1
Fig. 3 Influence of accelerating abruptly in vertical direction

constant horizontal velocity component U 1, G e 1,
0.8, z d 0, z d z d 0 100, x d 0 100
sionalized as the ratio of the actual amplitude of the disk center to

the reference eccentricity e 0 . In each figure, the amplitude is
shown in the form of the natural logarithm with base 2.71828 of
the dimensionless amplitude r e 2d 2d /e 0 . The velocity and
the acceleration are assumed to change suddenly at the beginning
of the aircrafts action. In all cases, 0.08. Data was taken for
50 revolutions after the response became stable, following this,
the status of the aircraft changes.
3.1
Constant Rotor Rotating Speed
3.1.1 Influence of the Aircraft Velocity on Rotor Response

The numerical results show that the steady-state amplitudes for
different horizontal velocities are the same, which means that
changing the horizontal velocity z d has no influence on the response amplitude. When the aircraft climbs at a constant velocity
and the vertical velocity component is set to different values, the
amplitude varies with x d . Larger x d corresponds to smaller amplitude. When changes, the difference between the amplitudes
in different vertical velocities changes also. The influence of the
vertical velocity component of the aircraft when 0.3 is weaker
than when 0.8.
3.1.2 Influence of the Aircraft Acceleration on Rotor Response
When 0.8, the amplitude of the disk increases suddenly when
the aircraft accelerates abruptly in the vertical direction but the
horizontal velocity component remains unchanged. The amplitude
decreases with time until reaching a steady value Fig. 3. The
steady-state amplitude is smaller than the value when x d 0. The
increasing amount of response amplitude depends on the value of
x d when the aircraft accelerates in the vertical direction. Larger
values of x d correspond to larger amplitudes. The difference between the steady-state amplitudes for different values of x d is
small. Changing G e hardly affects the steady-state amplitude
when the aircraft is accelerating. However, a larger value of the
parameter U will cause larger steady-state amplitude.
When is small, the increase in amplitude is less and the time
needed to reach steady state is shorter.
3.1.3 Rotor Response When Aircraft Flies in a Sine Curve in
the Vertical Plane OXZ. It is supposed that x d x d0
sin(zd) and the horizontal velocity of the aircraft is unaltered, i.e., z d 0 and z d z d0 constant. In this case, z d denotes
the period of the motion. The maneuvering range in the vertical
direction is reflected by . The following two cases are calculated.
1 changes while and z d are invariant; 2 z d changes while
JULY 2003, Vol. 125 325
Fig. 4 Rotor responses when the aircraft flies in a sine curve 0.8, U 0.5, G e 0.5, z d 0, 0.00002 a Influence
of z d 100 b Influence of z d 50000
and are unaltered. The response when the aircraft moves only in
the horizontal direction and 0 is taken as the basis for comparison. The aircraft first flies in a horizontal direction at a constant velocity. When the free vibration vanishes away and the
steady state amplitude is reached, data of 50 revolutions are taken.
Afterwards, the aircraft maneuvers for several periods. The corresponding period in dimensionless time is 12002 .
If the aircraft maneuvers periodically, its velocity and acceleration will also vary periodically. The amplitude response of the
disk center changes periodically except at the beginning of aircraft
action. In Fig. 4a the amplitude decreases suddenly at the very
beginning of the aircraft maneuver. The vibration response is periodic when the second period of aircraft movement begins, and
the period is the same as that of sine curve (5002 ). For most
part of each period almost 80% the amplitude is less than that
for 0, but for some parts the amplitude is larger than that for
0. The amplitude fluctuates acutely when the aircraft starts to
maneuver. Then the fluctuation decreases. The peak amplitude
value for 100000 is almost 50% greater than that for 0,
and is almost 26% greater than that for 50000 (G e 0.5, U
0.5). In Fig. 4b, larger values of z d correspond to smaller
aircraft maneuvering periods, relatively smaller periods of vibration response and larger peak values. This indicates that quick
maneuvering of the aircraft will cause the vibration response of
the rotor system to change dramatically, although the maneuver
range in the vertical plane is unchanged.
When becomes smaller, the increased amplitude due to the

aircrafts maneuver is less and the fluctuation is reduced.
3.2
Constant Angular Acceleration of the Rotor
3.2.1 Influence of the Aircraft Velocity on Rotor Response

( 0.00025). If the aircraft is static, the steady state unbalance response of the rotor at constant speed acceleration appears
to fluctuate rapidly after the critical rotating speed. When x d 0
and is invariable, different z d lead to obviously different response
at the beginning of the rotor acceleration. With the rotor speed
increasing, the difference between responses becomes smaller and
then vanishes. A smaller value of amplitude corresponds to a large
x d . When z d remains constant, the response curves for different
x d are clearly different at the beginning of acceleration. 0 appears to have no effect on the response.
3.2.2 Influence of the Aircraft Acceleration on Rotor Response
( 0.00025). For different aircraft accelerations, the responses of the rotor at low speed range are also different, but the
responses at high speed range may be almost the same Fig. 5a
or very different Fig. 5b depending on the values of z d0 and
z d . G e has little influence on the response curve, but larger values
of U correspond to larger amplitude. The responses for 0 0.3
are like those for 0 0.8.
Fig. 5 Influence of the aircraft acceleration in horizontal and vertical directions on rotor accelerating response 0
0.8, x d 0, 10, 50, x d z d 100, G e 1, U 1.0 a z d 0 z d 0 b z d 0 z d 100
326 Vol. 125, JULY 2003
Fig. 6 Rotor accelerating response when the aircraft flies in a sine curve 0 0.8, U 0.5, G e 1.5, z d 0 a Influence of z d 100, 0.00002 b Influence of z d 100, z d 0.002
3.2.3 Rotor Response When Aircraft Flies in a Sine Curve in

Vertical Plane OXZ ( 0.00025). When the aircraft flies in a
sine curve and 0 0.8, the rotor response amplitude at the beginning of the rotor acceleration decreases and the response curve
fluctuates. The response is not influenced by the parameters of the
sine curve when the rotor speed is near or at the critical speed.
When the rotor speed is much greater than the critical speed, the
response curve is significantly influenced by aircraft maneuvering
and fluctuates slowly Fig. 6. The fluctuation at the beginning of
rotor acceleration and after the critical speed is reduced if U increases or G e decreases.
When the initial rotating speed is small ( 0 0.3), the responses are different for different values of Fig. 7. Before
reaching the critical speed, the response fluctuates slowly. If
0, the response has many peak values equal to the value on the
0 curve, this means that the amplitude of the rotor with constant acceleration is less than that of the rotor with lower rotating
speed and no aircraft maneuver. Aircraft maneuvering has no influence on the response peak amplitude.
Conclusions
1 The operation of an aircraft affects the dynamics of the rotor

located in it. When the aircraft maneuvers at a constant acceleration in the vertical and/or horizontal directions, the amplitude of
the rotor running at a constant speed increases suddenly at the

very beginning of aircraft acceleration and eventually becomes
stable, except when the aircraft moves and accelerates only in the
horizontal direction. When the rotor rotates at a constant acceleration, the amplitude fluctuates before the critical speed. Different
aircraft acceleration lead to much different rotor responses when
the rotor runs at a constant rotating speed or at a constant acceleration.
2 The response of the rotor rotating at a constant speed is
clearly influenced by the amplitude and the period of the flight
path when the aircraft moves in a sine curve in the vertical plane.
The response curve appears to be periodic except at the beginning
of the aircraft action, the period is the same as that of the sine
curve. The gravity parameter affects the response significantly,
especially the trend of the response curve. The unbalance parameter affects only the amplitude value.
3 When the aircraft flies in a sine curve in the vertical plane,
the response of the rotor running at a constant acceleration is
different from that when the aircraft has no maneuvering, mainly
at the beginning of the acceleration and after the rotor speed
passes through the critical speed. The response curve is obviously
influenced by the gravity parameter when the initial speed is low
in the case of acceleration.
Acknowledgment
The support from China 863 Project No. 2002AA412410
and The Doctor Subjects Research Founding of China University
No. 20020248053 is gratefully acknowledged.
References
Fig. 7 Rotor accelerating response when the aircraft flies in a

sine curve 0 0.3 U 0.5, G e 1.5, z d 100, z d 0,
0.00002, 0, 50000, 100000
1 Lee, A. C., Kang, Y., Tsai, K. L., and Hsiao, K. M., 1992, Transient Analysis
of an Asymmetric Rotor-Bearing System during Acceleration, ASME J. Ind.,
1144, pp. 465 475.
2 Ganesan, R., and Sankar, T. S., 1993, Resonant Oscillations and Stability of
Asymmetric Rotors, Proc. of the 14th Biennial ASME Conference on Mechanical Vibration and Noise, ASME DE, 56, pp. 1922.
3 Spence, A. M., and Cele, R., 1995, Coupled Rotor Fuselage Dynamics and
Aero-Elasticity in Turning Flight, J. Am. Helicopter Soc., 401, pp. 4758.
4 Cao, Y., 1999, Modelling the Unsteady Aerodynamic Forces of a Maneuvering Rotor, Aircraft Engineering and Aerospace Technol., 715, pp. 444 450.
5 Bagai, A., Leishman, J. G., and Park, J., 1999, Aerodynamic Analysis of a
Helicopter in Steady Maneuvering Flight Using a Free-Vortex Rotor Wake
Model, J. Am. Helicopter Soc., 442, pp. 109120.
6 Park, J. S., and Leishman, J. G., 1999, Investigation of Unsteady Aerodynamics on Rotor Wake Effects in Maneuvering Flight, Annual Forum
Proceedings-American Helicopter Society, 1, pp. 467 480.
7 Krothapalli, K. R., Prasad, J. V. R., and Peters, D. A., 2001, Helicopter Rotor
Dynamic Inflow Modelling for Maneuvering Flight, J. Am. Helicopter Soc.,
462, pp. 129139.
JULY 2003, Vol. 125 327
Jinhao Qiu
e-mail: qiu@ifs.tohoku.ac.jp
Junji Tani
Taekyu Kwon
Institute of Fluid Science, Tohoku University,
Katahira 2-1-1, Aoba-ku, Sendai 980-8577,
Japan
Control of Self-Excited Vibration

of a Rotor System With Active Gas
Bearings
This paper presents an experimental study on the active control of self-excited vibrations
in a rotor-bearing system supported on a pair of externally pressurized thrust bearings in
the axial direction and on actively controlled journal gas bearings in the radial direction.
The active journal gas bearings used are of the tilting-pad type with one traditional
passive pad and two active pads with embedded piezoelectric actuators. Feedback control
systems are constructed with gap sensors for measuring the vibration of the rotor, embedded piezoelectric actuators, and PID controllers. The experimental results show that the
self-excited vibration can be effectively suppressed with the designed feedback control
system, if the gains of the PID controllers are properly tuned. DOI: 10.1115/1.1576423
Introduction
Since gas bearings use gas as a lubricant, they can work in a

wide range of temperatures and have been widely used in turbo
compressors and expanders of helium liquefiers 1,2,3. One of
the main problems with gas bearings is their intrinsic fluid-related
hydrodynamic instability as the rotor exceeds twice the natural
frequency of the rotor rigid body mode. The instability is due to
the interaction between the journal and the gas used as the lubricant. The stability of the gas bearings has been greatly improved
with better designs such as spiral grooves, tilting pads and elastic
foils. In these cases, rotors supported by gas bearings can reach
very high rotational speeds 1,4. However, self-excited vibrations
in rotor systems supported by gas bearings have always been an
important issue, and a further enhancement in stability is required
in practical applications.
In some cases, a modification of the configuration or some parameters of the design can prevent gas bearings from undergoing
self-excited vibration 5,6. For example, Ihara 6 designed a
movable pivot with an integrated spring, which can adjust the
pressing force between the pad and the journal, to increase the
damping effect of the bearing system. The new pivot had the
ability to suppress the self-excitation of the rotor system, but the
friction between the pads and shaft also prevented the shaft from
starting rotation when the torque was too small.
Recently, the active control of the self-excited vibration in gas
bearings has begun to attract attention, due to the increasing demands for high rotation speed and stability 711. Recent advances in sensor and actuator technology and electronics have
also made the active control of gas bearings possible. Many studies on the active control of externally pressurized bearings have
also been reported, but no research has been reported on the active
control of tilting-pad journal gas bearings, probably due to their
complicated configuration and vibration mechanism.
In this study, an active pad, supported by a pivot with an embedded piezoelectric actuator, is proposed for active tilting-pad
journal gas bearings and used in the active control of the selfexcited vibration of a rotor-bearing system. The rotor is supported
on a pair of externally-pressurized thrust bearings EPTB in the
axial direction, and on active journal gas bearings in the radial
direction. The radial position of the active pads in the journal
bearing can be actively adjusted by applying a voltage to the
embedded piezoelectric actuators. Experiments were performed
on the active control of the self-excited vibration in the rotorContributed by the Technical Committee on Vibration and Sound for publication
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received March 2002;
Revised March 2003. Associate Editor: G. Flowers.
328 Vol. 125, JULY 2003
bearing system using the control force generated by the two active
journal gas bearings. The effectiveness of the active pads was
confirmed by experimental results.
2 Configuration and Vibration Characteristics of the

Rotor-Bearing System
Figure 1 shows the configuration of the rotor-bearing system,
originally designed for use in a high-expansion-ratio turbo expander of a helium liquefier, shown in Fig. 2. The rotor consists of
a shaft, a turbine impeller at one end of the shaft for the generation of the driving moment, and a blower impeller for braking at
the other end. The rotor of the expander is supported by two
tilting-pad journal gas bearings in the radial direction and a pair of
externally pressurized thrust gas bearings in the axial direction.
The rotor can reach a maximum speed of 35,000 rpm at a driving
air pressure of 0.5 MPa on the turbine impeller.
The rotor is installed vertically and its weight is supported by
the thrust bearings. Since the load in the radial direction is very
small, tilting-pad journal gas bearings are used. The total length of
the rotor is 181.35 mm and the two tilting bearings are located at
23.54 mm above and 26.46 mm below the center of mass of the
rotor, as shown in Fig. 2a. The cross-sectional view of the
tilting-pad journal gas bearing used in the rotor-bearing system is
shown in Fig. 2b and Fig. 3. There are three tilting pads in each
bearing. The width of a pad is 15.4 mm and its spanning angle
is 110 deg. The nominal clearance C r is 20 m and the preload of
the pads is 0.5. The parameters of the rotor-bearing system are
summarized in Table 1. The dynamic characteristics of the tiltingpad gas bearings are discussed in 4.
The externally-pressurized thrust bearing was used to support
the high thrust load resulting from the impeller. The thrust bearings are a pair of the inherently compensated annular bearings
shown in Fig. 4. Their outer and inner diameters are 32 mm and
16.4 mm, respectively. There are 18 nozzles uniformly distributed
in a circle of r a 22.4 mm in each thrust bearing. The diameters
of the upper and lower bearings are 0.3 mm and 0.6 mm, respectively. The parameters of the thrust gas bearings are summarized
in Table 2. The design and dynamic characteristics of the
externally-pressurized thrust bearings are discussed in 1.
The vibration of the rotor is measured using eddy-current gap
sensors with a sensitivity of 0.5 m. The four gap sensors are
divided into two groups and fixed at two planes at 10.84 mm
above and 14.16 mm below the center of mass of the rotor. The
two gap sensors in the same plane are mounted 90 deg apart and
measure the lateral displacement of the shaft in that plain, as
shown in Fig. 3. The positions of the sensors are labeled as LR,
Fig. 1 Configuration of the rotor-bearing system
LL, UR and UL. The LR and LL sensors are installed in the lower
plane and the UR and UL sensors are installed in upper plane.
Moreover, the LL and UL sensors measure the displacement of the
shaft in the y direction, and the LR and UR sensors measure the
displacement in the x direction of the right-handed coordinate systems shown in Figs. 1 and 3.
The characteristics of the rotor vibration are measured using the
following procedure. Firstly, a 40 V DC voltage is applied to all
four piezoelectric actuators as a bias voltage in the vibration control of the rotor system. Secondly, the thrust bearing is activated
by supplying pressurized air. Thirdly, the valves of the gas supplies for the driving impeller and the braking impeller are turned
on to start the rotation of the shaft. The rotation speed of the shaft,
determined by the relative pressure of the air supplied to the driving impeller and the braking impeller, is raised to 30,000 rpm by
increasing the flow rate of the air to the driving impeller and then
reduced to 0 rpm again by decreasing the flow rate. The vibration
of the shaft is measured when the speed of the rotor reduces from
30,000 rpm to 0 rpm.
Figure 5 shows the waterfall plot of the vibration characteristics
in the y direction measured by the LL sensor. It can be seen that
both the synchronous vibration with the frequency equal to the
rotation speed, and the asychronous vibration self-excited vibration, are induced in the rotor system. The sychronous vibration
reaches the maximum amplitude at 11,600 rpm, which equals the
natural frequency of the first rigid mode of the rotor. The frequency of the self-excited vibration is about 57 Hz, regarless of
the rotational speed, but its amplitude also reaches the maximum
value at around 11,600 rpm. The fluid-related instability in both
the journal bearings and the thrust bearings can induce the selfexcited vibration. A dynamic coupling between the journal bearings and the thrust bearings may also occur in the rotor-bearing
system. The objective of this study is to control the self-excited
vibration of the rotor-bearing system.
Active Gas Bearings
In order to control the vibration of the shaft, piezoelectric actuators were embedded in two of the three pivots in each bearing,
as shown in Fig. 3. The piezoelectric actuators are 55
10 mm in size and can generate 6.1 m of displacement at 200
V input voltage. The piezoelectric actuators, which have a wide
frequency bandwidth and can generate a large force output, are
suitable for this kind of application. The positions of the four
actuators are also labeled as LR, LL, UR and UL. The LL and UL
actuators are embedded in the pivots on the y-z plane and the LR
and UR actuators are embedded in the pivots on the plane at 30
degrees clockwise from the x-y plane.
When the shaft rotates, a pressurized gas film is formed between the shaft and each pad due to the viscosity of the gas. The
rotor is supported by the pressure of the gas films in the radial

direction. The pressure distribution of the gas film, and therefore
the resultant force due to pressure, are dependent upon the clearance distribution between the shaft and pads. The resultant force
can be actively controlled by changing the radial position of the
pivot. Therefore, the vibration of the shaft can be actively suppressed when an appropriate voltage is applied to the piezoelectric
actuators.
In order to investigate the static characteristics of the PZT actuators in the rotor-bearing system, the bias displacements of the
rotor are measured at the four sensors when a DC voltage was
applied to the UR actuator during operation. The bias displacements are the DC components of the rotor displacement, that is,
the displacements of the whirling center of the rotor. Figure 6
shows the results measured at 11,600 rpm. As the voltage increased from 0 V to 150 V, the displacements of the shaft measured by the LR and UR gap sensors decreased from 0 m to
4.3 m, while the displacements of the shaft at the LL and UL
sensors increased from 0 m to 2.2 m at a rotation speed of
11,600 rpm. This means that, due to the static displacement generated by the PZT actuator in the UR pad, the rotor moves away
from the LR and UR sensors but closer to the LL and UL sensors.
Control System and Experimental Setup
Figure 7 shows the configuration of the experimental setup of

the rotor system and the feedback control system used. Since it is
difficult to derive a mathematical model of the dynamical system,
a PID control method that does not require a mathematical model
of the control object is used in this study. It is also assumed that
the interaction between the actuators is negligible and that each
actuator can be controlled independently. The input-output relationship of the controller for each actuator is defined by the following equation:
U s
Ki
1
K p K d s K p 1T d s
Ys
s
T is
(1)
where Y (s) is the feedback signal in volts, U(s) is the control

input the voltage applied to the actuators in volts, and K p , K d
and K i are the proportional, derivative and integral gains, respectively. Moreover, T i K p /K i and T d K d /K p are the integral and
derivative constants.
Different gains were selected for different actuators in the experiment. For the actuator installed in the same direction as the
sensor, the output of the sensor was used directly as the feedback
signal Y (s). For the actuators at the LR and UR positions, the
feedback signal Y (s) was estimated from the output of the two
sensors at the nearest measuring plane.
Figure 8 shows a block diagram of the active control system.
The analog signals from the sensors are sampled at a frequency of
JULY 2003, Vol. 125 329
Fig. 2 Configuration of the high-expansion-ratio expander a Longitudinal

cross-section b Transverse cross-section
330 Vol. 125, JULY 2003
Fig. 3 Configuration of the tilting-pad journal gas bearings
10 kHz through an A/D converter and then transmitted to a DSP.

The designed controller was implemented digitally in the DSP,
and the control input calculated on the basis of the sensor signals.
The calculated control input was added to a bias voltage, amplified and then applied to the actuators. The bias voltage was superposed on the input voltage calculated from Eq. 1 in order to
increase the maximum output displacement of the actuators. The
gain of the power amplifier is 20 and the bias voltage after amplification is 40 V. The gain of the sensor and the power amplifier
are not included in the transfer function in Eq. 1.
Results and Discussions
Fig. 4 Configuration of the externally-pressurized thrust gas

bearings a Distribution of nozzles b Upper and lower
clearances
duced. However, the synchronous vibration remained almost unchanged. The small peak at 50 Hz is due to the electromagnetic
noise of the power supplies.
Figure 10 show the rotor frequency response measured by the
four sensors at 11,600 rpm. K d was set to 0.01, and the proportional gains K p of the UL and UR actuators were set to the opti-
Table 2 Parameters of the thrust bearing
Due to the influence of the pressurized air supplied to the driving impeller, the actuators at LL and LR were less effective in the
control of the self-excited vibration than the other two actuators.
Therefore, the gains of these two actuators were fixed at K d
0.01, K i 0.001 and K p 50, while the gains of the other two
actuators at UL and UR were varied to investigate their influence.
Moreover, the integral gain of the two actuators at UL and UR
was also set to 0.001, since it has little influence on the control
effect. Control experiments were performed for different combinations of K p and K d for the actuators at UL and UR. Figure 9
shows the shaft frequency response measured by the sensors at
UR and LR for a rotation speed of 10,000 rpm with gains K p and
K d for the actuators at UL and UR set to 10 and 0.02, respectively.
The asynchronous vibration about 57 Hz was significantly re-
Outer diameter, r 1
Inner diameter, r 0
Diameter of nozzle position circle, r a
Number of nozzles, n
Diameter of nozzles in the lower bearing, r s1
Diameter of nozzles in the upper bearing, r s2
Unloaded clearance of the lower bearing h n1
Unloaded clearance of the upper bearing h n2
Nondimensional axial displacement h n /(h n1 h n2 )
32 mm
16.4 mm
22.6 mm
18
0.6 mm
0.3 mm
23.4 m
16.6 m
0.22
Table 1 Parameters of the rotor-bearing system

Shaft length, l
Shaft mass, m
Shaft radius, R
Bearing width, L
Length of pad,
Clearance, C r
Pivot position p /
Dimensionless preload
181.35 mm
0.28 kg
8 mm
15.4 mm
110
20 m
0.65
0.5
l1
l2
l3
l4
l5
l6
l7
l8
23.54 mm
26.46 mm
10.84 mm
14.16 mm
90.25 mm
91.1 mm
38.74 mm
6 mm
Fig. 5 Waterfall chart of the frequency response of the rotor at

the LL gap sensor experimental result
JULY 2003, Vol. 125 331
Fig. 6 Bias displacement of the shaft versus DC voltage applied to the UR actuator
Fig. 8 Block diagram of the control system
mal values of 150 and 100, respectively. Since the rotor speed
equals the resonant frequency of the rotor-bearing system, the
amplitude of the asynchronous vibration increases significantly at
the uncontrolled state. However, the asynchronous vibration is
almost completely suppressed after control. Despite this, little effect could be observed on the synchronous vibration. The other
components of the asynchronous vibration with frequencies of 25
Hz, 79 Hz and 132 Hz were also suppressed.
The influence of the feedback gain on the control results was
further investigated. Fig. 11 shows the control results when the
proportional gains of the actuators at UL and UR were set to 200,
while the rotor speed and the other gains were kept the same as
those in the former case. The asynchronous vibration at a frequency of 57 Hz was almost completely suppressed. However, a
second asynchronous vibration of 93 Hz was excited.
Figure 12 shows the rotor frequency response measured by the

UR sensor when the K p of the actuators at UL and UR was set to
10, and K d was set to 0.2 for Case a and 0.01 for Case b. In
Case b, the asynchronous vibration of 57 Hz was satisfactorily
suppressed, though a low peak remained due to the small proportional gain. However in Case a, another asynchronous vibration
of 67 Hz was significantly excited. Moreover, the gains used in
Case b are the same as those in Fig. 9, except that K d decreased
from 0.02 to 0.01. This means that the same gains can yield good
control performance for a range of rotor speeds. The equivalent
stiffness and damping coefficients of the closed-loop system corresponding to the proportional and differential gains cannot be
estimated quantitatively since the mathematical model of the control plant is unknown. However, their variation can be observed
from the position and shape of the peak corresponding to the
asynchronous vibration in the power spectra shown in Fig. 912.
The decrease in the stiffness coefficients moves the peak of the
asynchronous vibration to the left side Fig. 9, while the peak
becomes sharper as the damping coefficient decreases Fig. 10.
Fig. 9 Power spectra of the shaft at the LR sensor position

Rotation speed: 10,000 rpm, PID gain: K p of UL act.10, K p of
UR act.10
Fig. 7 Schematic diagram of the active control system
332 Vol. 125, JULY 2003
Fig. 10 Power spectra of the shaft at different sensor positions Rotation speed: 11,600 rpm, PID gain: K p of UL act.150, K p
of UR act.100 a UL sensor position b LL sensor position c UR sensor position d LR sensor position
Fig. 11 Power spectra of the shaft at the UL sensor position

Rotation speed: 11,600 rpm, PID gain: K p of UL act.200, K p of
UR act.200
Fig. 12 Power spectra of the shaft at the UR sensor position

Rotation speed: 11,600 rpm, a PID control: K d of UL Act.
0.2, K p of UL act.10, b K d of UR act.0.01, K p of UR act.
10
JULY 2003, Vol. 125 333
Conclusions
A novel tilting-pad journal gas bearing system with active pads

was studied for the control of self-excited vibration in a rotor
bearing system. Each active pad contains an embedded piezoelectric actuator, used to adjust the clearance between the shaft and
pads. The novel gas bearings were used in the rotor system of a
high-expansion-ratio turbo expander for helium liquefier, and a
feedback control system was constructed with four gap sensors,
embedded actuators and PID controllers. It was confirmed by experiment that the self-excited vibration in the rotor system could
be effectively suppressed with the feedback control if the gains of
the controllers were appropriately chosen.
It was also found that little control was achieved for the synchronous vibration of the rotor system using the present configuration. The active control of synchronous vibration will be the
subject of future work.
Acknowledgment
The authors thank Dr. Yanagi, and Mr. Tsugawa of Mayekawa
Mfg. Co. Ltd., MYCOM Advanced Technology Laboratory,
Ibaraki, Japan, for their effort and cooperation in the design and
manufacturing of the rotor-bearing system.
334 Vol. 125, JULY 2003
References
1 Ino, N., Machida, A., Tsugawa, K., and Hashimoto, H., 1991, Development
of Externally Pressurized Thrust Bearing for High-Expansion-Ratio Helium
Expander, Trans. Jpn. Soc. Mech. Eng., Ser. C, in Japanese 58550, pp.
18851891.
2 Togo, T., 2000, Investigation Report on the Research of Gas Bearings, Japanese Society of Tribologists.
3 Andres, L. S., 1996, Turbulent Flow, Flexure-Pivot Hybrid Bearings for
Cryogenic Applications, ASME J. Tribol., 118, pp. 190200.
4 Qiu, J., Takagi, T., Tani, J., Machida, A., Tsugawa, K., Yanagi, H., and Ino, N.,
1993, Dynamic Characteristics of a Tilting-pad Bearing System for High
Expansion-Ratio Expander, Adv. Cryog. Eng., 39, pp. 909914.
5 Bently, D. E., and Muszynska, A., 1989, Anti-Swirl Arrangements Prevent
Rotor/Seal Instability, Stress, Reliab. Des. 1112, pp. 156 162.
6 Ihara, K., 1987, Gas Expander for Lower Temperature, Journal of Turbine
Machinery in Japanese, 1511, pp. 57 62.
7 Palazzolo, A. B., Lin, R. R., Alexander, R. M., Kascak, A. F., and Montague,
J., 1989, Piezoelectric Pushers for Active Vibration Control of Rotating Machinery, ASME J. of Vibration, Acoustics, Stress, Reliab. Des. 111, pp. 298
305.
8 Palazzolo, A. B., Jagannathan, S., Kascak, A. F., Montague, G. T., and Kiraly,
L. J., 1993, Hybrid Active Vibration Control of Rotorbearing Systems Using
Piezoelectric Actuators, J. Vibr. Acoust., 135, pp. 111119.
9 Horikawa, O., and Shimokohbe, A., 1990, An Active Air Bearing, JSME,
Int. Journal, 331, pp. 55 60.
10 Muszynska, A., Franklin, W. D., and Bently, D. E., 1988, Rotor Active AntiSwirl Control, Stress, Reliab. Des. 1102, pp. 143150.
11 Sun, L., Krodkiewski, J. M., and Cen, Y., 1998, Self-Tuning Adaptive Control of Forced Vibration in Rotor Systems using an Active Journal Bearing, J.
Sound Vib., 2131, pp. 114.
Vibration of Flex Circuits in Hard

Disk Drives
J. A. Wickert
Fellow ASME,
Carnegie Mellon University,
Pittsburgh, PA 15213
A flex circuit connects the stationary electronic components in a hard disk drive to the
rotating arm that carries the read/write heads and positions them above data tracks on
the disk. Flex circuits are conventionally formed as a laminate of polyimide substrate,
adhesive, and copper conductors. Deformation of a flex circuit is discussed in the context
of the following stages: the initial unstressed shape, configurations in which stresses set
and relax in response to elevated temperature, equilibrium, and small amplitude vibration. The model involves displacements of the flex circuit in the directions tangent and
normal to the local equilibrium shape, and those motions couple with the arms dynamics.
Nonlinearity associated with finite curvature, partial elastic springback, and the arms
geometry and inertia properties are incorporated within the vibration model to predict
system-level natural frequencies, mode shapes, and coupling factors between the circuit
and the arm. Laboratory measurements using noncontact laser interferometry validate the
model with respect to the circuits shape, stiffness, restoring moment, and natural frequencies. The primary degrees of freedom for optimizing flex circuit design are the thicknesses
of the individual layers within the circuit, free length, and the locations and slopes of the
circuits attachment points to the arm and electronics block. The models predictions and
trends developed from a case study in free length are discussed with a view toward
reducing coupling between the circuit and arm in certain vibration modes.
DOI: 10.1115/1.1547661
Introduction
The storage density of hard disk drives, as measured by the

number of data bits captured within a unit area, has grown at a
historical rate of about 60% per year, and that rate has accelerated
recently to more than 100% per year. In a similar vein, the cost
associated with storing a megabyte of data has fallen by over four
orders of magnitude during the past two decades. With the physical limit for the density of magnetic disk recording now appearing
on the horizon, higher-precision requirements are being placed on
the vibration of each structural and actuation component.
By way of background and motivation, the construction of a
hard disk drive is shown in the photograph of Fig. 1. Data is
stored as magnetic transitions on the thin media layer that coats
the glass disks. The arm pivots about a fixed bearing, and it is
driven by a voice coil that swings between two permanent magnets. The read/write heads are located at the tip of the arm above
each disk surface, but they are not discernible in Fig. 1 because of
their small size. The particular drive shown in the figure has a
total of ten read/write heads.
The voice coil and its companion servo system slew the heads
to a desired cylinder of data, and follow it in the presence of disk
runout, vibration, windage, and other disturbances. Electrical
leads are routed to each read/write head to carry the recording and
readback signals, and other larger leads power the voice coil. Each
head and coil wire is integrated within a flat and flexible circuit
that conveys all of the electrical signals between the rotating arm
and the stationary electronics on the drives body. This overall
construction is shown schematically in Fig. 2. In a typical embodiment, the flex circuit is a polyimide film laminate having rolled
annealed copper wires in its conductor layer. Flex circuits replace
conventional multi-lead wiring and combine electrical functionality with mechanical flexibility, which in turn can potentially introduce unwanted vibration. Quite aside from hard disk drives, other
types of flex circuits are used in avionics packages, gyroscopes,
hearing aids, and cardiac pacemakers.
revised October 2002. Associate Editor: Chin An Tan.
The mechanical behavior and vibration of a hard drives disks,

motor, spindle, bearings, voice coil, and other elements have been
the subject of substantial engineering development effort since the
first introduction of a hard disk drive in 1956. As drives have
become more sophisticated, sources of vibration that had previously been within tolerance limits have become the focus of further improvement efforts. In particular, the vibration of flex circuits and their coupling with motion of the read/write heads are
now factors in high density recording applications. No longer
viewed as a lightweight appendix to the arm and voice coil, the
flex circuit has dynamics that couple through the arm and produce
track positioning errors.
Reducing flex circuit vibration in certain modes is a potential
strategy for improving transient settling response after a seek operation to a new data track. From the modeling perspective, an
objective of this investigation is to better understand the flex circuits equilibrium shape involving finite deformation and partial
elastic springback, and the character of its small amplitude vibration about equilibrium. Of particular interest are the flex circuits
static shape at a specified arm rotation angle, natural frequencies,
and coupling to the arm and read/write heads in each mode. In
what follows, the roles played by such parameters as the circuits
laminated structure, finite curvature, free length, and boundary
conditions at the arm and electronics block are explored.
Material Characteristics and Loading Sequence
In Fig. 3, the two micrographs of a flex circuits cross-section

describe the internal morphology and dimensions of the several
material layers. In optical diagnostics, a segment of the flex circuit
was embedded in epoxy, diced, and polished smooth. A thin layer
of gold was then sputtered onto the surface of the cross-section to
enhance imaging. The particular circuit of Fig. 3 is formed as a
symmetric sandwich laminate of two h p 31 m thick polyimide
layers on either side of flat copper electrical leads. Kapton is
often used as the substrate material, and it is chosen on the basis
of electrical, mechanical, and thermal properties. In Fig. 3a, a
matrix of epoxy bonds the polyimide layers to the conductors, and
the thickness of the intervening adhesive is h a 13 m. The circuit comprises twenty conductors having common h c 26 m
JULY 2003, Vol. 125 335
Fig. 1 Photograph depicting certain mechanical components within a hard

disk drive used for computer data storage
thickness, and one of four different widths depending on the type

of electrical signal being carried. Some 67% of the flex circuits b11.15 mm width is associated with the conductors, and
the remaining fraction of space between the conductors remains
insulated by adhesive. Figure 3b depicts a second portion of the
same flex circuit, and its image is spaced width-wise relative to
Fig. 3a. Three adjacent conductors, slight variation in the adhesive thickness, and indentation of the conductor into the lower
substrate are visible features. The cumulative thickness is hh c
2(h p h a )114 m, and the flex circuit has linear density
2.38 g/m in the direction along its free length L.
Vibration of the flex circuit depends in part on its equilibrium
curvature and internal loads, which are distinguished from their
Fig. 2 Schematic of the equilibrium and vibration model for

the arm and flex circuit mechanism
336 Vol. 125, JULY 2003
counterparts in the circuits natural state where no reactions are

applied through attachment to the arm and electronics block. Further, the flex circuits stresses, particularly in the adhesive layer,
are known to partially relax with time and temperature. Deformation from the initial free shape, through finite motion to equilibrium, and ultimately to small amplitude vibration are broken down
in the following sequence of loading states:
Initial S( 0 ) The flex circuit is specified to be initially unstressed
and straight. Imperfections that are present as a result of the circuits production or the drives final assembly could be incorporated at this stage by specifying a functional form of curvature
k ( 0 ) (s), where s(0,L) denotes arc length, but such effects are
not considered at this first level of approximation.
Set S( 1 ) Referring to Fig. 2, the flex circuit is bent from S ( 0 ) and
attached to the arm and electronics block. Elements within it are
subjected to tension T(s), shear force N(s), and bending moment
M (s). During the process to S ( 1 ) , the circuit undergoes finite
Fig. 3 a Brightfield and b differential interference contrast

micrographs of flex circuit cross-sections. The images depict
the construction morphology and the thicknesses of the polyimide, conductor, and adhesive layers.
deformation to the new shape of curvature k ( 1 ) . The coordinates

(x L ,y L ) of the contact point and the angle L of approach at the
electronics block are specified. Rotation of the arm is an independent variable in S ( 1 ) , and together with the geometric parameters , , and r defined in Fig. 2, its value sets the circuits
conditions
x 0 r cos
y 0 r sin
(1)
(2)
at its endpoint sL. In S ( 1 ) , the arm is held by an external agent

at an angle that could correspond to the disks inner ID or outer
OD diameters, or the load-unload LU support where the arm is
parked when the drive is not in use.
Springback S( 2 ) During both its production and day-to-day usage,
the flex circuit is exposed to elevated temperature. Certain fabrication stages take place at 80C; an assembly-level wash/dry
cycle occurs at 85C; and the internal environment of a performance drive can reach 50 60C. By comparison, the glass transition temperature of an adhesive commonly used within flex circuits lies in the range 5095C. In the springback stage, while the
flex circuit remains attached to the arm and electronics block in
S ( 1 ) , stresses are allowed to partially relax and the circuit assumes
the new natural state S ( 2 ) . To characterize this process, thermal
cycling experiments were conducted with the arm locked at the
disks inner diameter ( ID64 deg) and the assembly held at
45C for one hour. After having cooled to ambient temperature,
the arm and flex circuit assumed nearly that point as the new
equilibrium configuration. The arm was then pivoted to the disks
outer diameter ( OD32 deg), thermally cycled, and observed to
assume a nearby position in equilibrium. During those processes,
however, the stresses within the flex circuit were only partially
reduced. Indeed, when the circuit was cut free and released, it
assumed a new natural state having shape and curvature somewhat between k ( 0 ) and k ( 1 ) . The empirical factor p is introduced to
quantify the amount of elastic springback that occurs when the
flex circuit is unloaded, and the curvature in S ( 2 ) is defined k ( 2 )
(1p)k ( 1 ) . When p1, the circuit behaves elastically, is unaffected by exposure to elevated temperature, and returns to its initial straight shape when unloaded from S ( 1 ) . On the other hand,
when p0, the flex circuit assumes S ( 1 ) as the new natural state.
In practice, the actual loading process and k ( 2 ) lie between these
two extremes, and they change in a time-, temperature-, and
history-dependent manner as the arm continuously slews between
the disks ID and OD. In the baseline parameters of Table 1, the
value p0.55 was determined by experience in thermally cycling
several disk drives and examining the circuits when they were cut
and released from the arm and electronics block. The flex circuits
were photographed, and their natural shapes were compared to
those predicted by the equilibrium model described below at various levels of springback. Parameter p was then adjusted by cutand-try until the predicted and measured natural shapes were
agreeably close.
Equilibrium S( 3 ) From the natural configuration S ( 2 ) , the flex
circuit is imagined to be re-attached to the arm and electronics
block. With no external torque M o about the pivot point being
applied by the voice coil, the arm rotates to equilibrium in response to the combined influence of the tension, shear force, and
bending moment at s0. At this stage, the arms equilibrium
angle and the circuits curvature are denoted by and k
*
*
k ( 3 ) , respectively.
Deformed S( 4 ) In this loading stage, the flex circuit is elastically
deformed about equilibrium, corresponding to static slew at a certain angle, to small amplitude vibration of the flex circuit about
equilibrium, or to finite amplitude vibration. The tension, shear
force, and bending moment are each incremented relative to their
equilibrium values. Motion of the flex circuit is resolved into the
Table 1 Baseline parameters for the arm and flex circuits

equilibrium and vibration models
Flex Circuit
Width, b
Thickness
Polyimide layer, h p
Adhesive layer, h a
Conductor layer, h c
Cumulative, h
Conductor fraction,
Free length, L
Composite linear density, A
Modulus
Polyimide, E p
Adhesive, E a
Conductor, E c
Springback ratio, p
Composite bending stiffness, EI
Composite axial stiffness, EA
11.1 mm
31 m
13 m
26 m
114 m
67%
31 mm
2.38 g/m
2.75 GPa
1.03 GPa
115 GPa
0.55
4.78(106 ) Nm2
2.45(104 ) N
Arm
Attachment radius, r
Read/write head radius, R
Attachment angle,
Offset angle,
Angles
Load-unload, LU
Disk OD, OD
Disk ID, ID
Inertia, m
Radius of gyration,
10.6 mm
43.8 mm
95
8
23
32
64
17.2 g
11.9 mm
Electronics Block
24, 11 mm
90
Coordinates, (x L ,y L )
Tangency angle, L
directions tangent and normal to the local equilibrium, and those

displacements are denoted by u(s,t) and v (s,t), respectively.
Equilibrium Shape, Loads, and Stresses
The flex circuits shape in S ( 3 ) is defined parametrically by the

coordinates (x(s),y(s)). With the nomenclature () d/ds,
equilibrium in the local tangential and normal directions is governed by the force and moment balances 1
T Nk,
N Tk,
k N/EIk 2
(3)
in which the constitutive relation M EI(kk ) has been embedded. The circuits bending stiffness is given by
(2)
EI
1
1
1
1
E h 3 b E a h 3c 1 b2E a
bh 3 bh h
12 c c
12
12 a 4 a c
h a 2 2E p
1
1
bh 3p bh p h c h p 2h a 2
12
4
(4)
where values for the elastic constants of the conductor E c , polyimide E p , and adhesive E a are listed in Table 1. While Eq. 4 is
specific to the cross-sectional construction of Fig. 3, the treatment
can be adapted for other geometries.
On the basis of measured layer thicknesses and published elastic constants 2 in Table 1, the circuits composite bending stiffness is EI4.78106 Nm2 . The conductor layer contributes
26% to the stiffness, the polyimide layer 71%, and the adhesive
layer only 3%. This value of EI was validated by both static bending and natural frequency measurements conducted with a circuit
segment that was embedded as a cantilever in an epoxy casting. In
the static test, the segment was mounted on a micrometer translation stage, and under specified displacement, the force applied to
its tip was measured. The signal from a planar beam sensor Futek FR-1020 was conditioned and amplified to provide a calibrated and linear force response for loads up to 295 mN. The
JULY 2003, Vol. 125 337
Fig. 4 Predicted
and measured shapes of the flex circuit
in equilibrium with attachment to the arm and electronics
block, and in its natural state. The shapes extracted from photographs of the circuit in the two states are denoted by the
and data points, respectively.
Fig. 5 Equilibrium tension and shear force along the circuits

arc length
along the circuits length from the arm. While the average strain
over the cross-section is only T/EA1.3106 with
stiffness recorded on this basis was 4.60106 Nm2 . In the second validation test, the value 4.91106 Nm2 was determined
by measuring the segments two lowest natural frequencies and
matching them to the values expected for a cantilever. In the calculations described below, the stiffness value determined from Eq.
4 was used.
The circuits equilibrium curvature is determined by integrating
Eqs. 3, and its shape is found subsequently from the kinematic
relations
x cos ,
y sin ,
k s
(5)
subject to position and slope conditions at s0 and L. While

some rotation between the circuit and arm does occur locally at
s0, that effect is neglected in this first approximation, recognizing that the resulting model may overestimate the circuits actual
boundary stiffness.
Solutions to Eqs. 3 and 5 are found through a shooting
method in which the boundary value problem in x, y, and is
converted to an initial value problem, and then solved iteratively.
Estimates for initially unknown N(0), T(0), and k(0) are assigned, and the system of differential equations is integrated numerically. The values x(L), y(L), and (L) which result from
that calculation do not, in general, satisfy the endpoint constraints.
In iteration, N(0), T(0), and k(0) are then adjusted through a
nonlinear minimization scheme until each kinematic condition is
satisfied within desired tolerance.
On the basis of the parameters in Table 1, Fig. 4 depicts simulated and measured shapes of the flex circuit in its natural (S ( 2 ) )
and equilibrium (S ( 3 ) ) states. The data points shown in the figure
represent coordinate locations as extracted from photographs of
the circuit in the two configurations. Curvature in the natural state
was set with the arm held at the disks OD and the springback
constant being p0.55. At the equilibrium angle 26.4 deg,
*
the tension, shear force, and bending moment applied by the flex
circuit to the arm produce no resultant torque about the pivot
bearing. Their variations in response to first-order changes in ,
however, are captured by stiffness 10.2 mNmm/deg.
The equilibrium tension and shear force distributions along the
circuit are shown in Fig. 5. The circuit is compressed over its
entire length with the minimum, mean, and maximum loading
values for T being 55, 33, and 14 mN, respectively. The
zero crossing in shear occurs at a distance approximately 64%
338 Vol. 125, JULY 2003
EAE c h c b 2E a h a bE a h c b 1 2E p h p b
(6)
the peak bending strain h c k k /2 within the conductor layer

*
is 0.12%. For other free lengths, arm positions, or endpoint locations, local yielding could occur within the ductile conductor.
As the arm slews about equilibrium, the flex circuit bends further to S ( 4 ) as shown in Fig. 6. The voice coil applies the static
bias torque
(2)
M o T 0 r sin N 0 r cos M 0
(7)
which increases in Fig. 7 from zero at equilibrium to the maximum value 0.35 Nmm at the disks ID. Multiple measurements
of the torque were made for one disk drive at four different slew
angles, and those results are also shown in Fig. 7. The indicated
variation of M o is representative of such measurements and captures hysteresis in the circuit, and friction in the pivot bearing and
Fig. 6 Variation of the flex circuits static shape for arm positions which range between the disks outer and inner diameters. For each arm position, the locations of the circuits endpoints are denoted by .
Fig. 7 Predicted
and measured dependencies of the static restoring moment for arm positions between the disks outer and inner diameters
supports. The trend in Fig. 7 is agreeably linear over the operating

range, although the circuits stiffness dM o /d does decrease in
the disks mid-diameter range, and grow again for slew angles
near the ID.
Figure 8 depicts the manner in which the flex circuits equilibrium shape changes as a function of its free length. Results are
shown for simulations in which L varies between 70% and 130%
of its nominal value. With the other parameters of Table 1 remaining fixed, the corresponding changes to the arms equilibrium
angle and the circuits static stiffness are listed in Table 2.
4 Natural Frequencies, Vibration Modes, and Coupling

For small amplitude vibration in S ( 4 ) , the shear, tension, and
curvature are expanded
NN N 1
*
(8)
(9)
TT T 1
*
kk k 1
(10)
*
about their equilibrium values, denoted by () . Here 1 is a
*
dimensionless scaling parameter used in the linearization, and the
first-order corrections are written 3,4
N 1 EI v ,ss k u ,s ,s
(11)
*
T 1 EA u ,s k v
(12)
*
k 1 v ,s k u ,s
(13)
*
in terms of the circuits tangential and normal displacements. Here
the comma-subscript notation signifies partial differentiation. The
equations of motion become
Au ,tt T 1,s N k 1 k N 1 0
*
*
A v ,tt N 1,s T k 1 k T 1 0
*
*
where A is the circuits mass-per-unit-length.
Vibration of the flex circuit and arm couple through
(14)
(15)
m 2 1 T 1 0,t r sin N 1 0,t r cos M 1 0,t

(16)
where m is the arms mass, is its radius of gyration about the
pivot bearing, and 1 is the first-order rotation about . The
*
time-dependent incremental loads in Eq. 16 are evaluated at the
circuits connection point to the arm. Although not considered
here, the effects of the voice coils driving torque and damping in
Table 2 Dependence of the equilibrium angle and the circuits

stiffness with respect to free length over a range 70% to 130%
of nominal
Fig. 8 Flex circuit equilibrium shapes predicted for free

lengths which range between 70% and 130% of the nominal
value in Table 1
Length ratio
Length mm
Equilibrium
angle deg
Stiffness
mNmm/deg
70%
80%
90%
100%
110%
120%
130%
21.7
24.8
27.9
31.0
34.1
37.2
40.3
41.3
33.0
28.4
26.4
29.3
36.9
44.2
63.6
40.0
20.2
10.2
5.6
5.0
5.8
JULY 2003, Vol. 125 339
Fig. 9 First six vibration modes

shown superposed on
the equilibrium shape
. Each element of the figure is
annotated with the predicted natural frequency and displacement ratio .
the pivot bearing could be incorporated in Eq. 16 at this stage

for direct numerical simulation or control system studies. Motion
of the circuit and arm are also constrained through
u 0 r sin 1
(17)
v 0 r cos 1 ,
(18)
and the conditions u0, v 0, and v ,s k u0 at sL.

*
The flex circuit is substantially stiffer for displacements in u
than v , and the ratio of longitudinal to flexural stiffness is measured by the dimensionless parameter (EAL 2 )/(EI), which is
O(106 ) for the problem at hand. On the time scales of the lower
transverse modes, tension variations propagate almost instantaneously, and the explicit appearance of u in the equations of motion can be suppressed by approximating T 1 as a function of time
alone. After integrating Eq. 12 in this manner, the longitudinal
displacement field is approximately 5
Fig. 11 Measured spectrum and natural frequencies of one

disk drives flex circuit in modes two, three, and four
k v dse
sr sin 1
*
(19)
where e is the average longitudinal strain. By embedding Eq. 19

into Eqs. 1113 and Eq. 15, the working form of the transverse equation of motion involves only the dependent variables v
and 1 , and it is applied to characterize the lower modes.
The vibration model is discretized in terms of 1 and nodal
values for v that are evenly-spaced along the circuits length, and
the natural frequencies and mode shapes are determined numerically. Figure 9 depicts the lowest six modes on the basis of the
parameters listed in Table 1. Each element of the figure is annotated to indicate the modes natural frequency and the displacement ratio
R 1
max u 2 v 2
(20)
which measures the relative motions of the read/write head and

flex circuit. At only some three Hertz, the fundamental mode is
dominated by motion in 1 and set primarily by the arms inertia
and the circuits static stiffness. Because of its low frequency, this
modes dynamics generally do not contribute tracking errors to the
extent that the modes at 374 Hz, 837 Hz, 1.39 kHz, and higher
frequencies do.
Fig. 10 Test stand used for measuring flex circuit transient

responses and natural frequencies. A 90 deg prism directs the
target beam from the laser head to the flex circuit. A small
patch of retroreflective tape not visible in the photograph was
placed on the circuit to reduce measurement sensitivity to misalignment and rotation of the circuit during vibration.
340 Vol. 125, JULY 2003
Fig. 12 Ring-down of a flex circuit in its second mode, which

for this disk drive was placed at 332 Hz. The damping ratio is
1.1%.
Fig. 13 Transient response in the arm and flex circuits 3.08 Hz

fundamental mode following an impact
Although the coupling ratios are only a fraction of a percent

in modes two and higher, they do have design implications because of the high precision required of this mechanism. A data
track on a typical performance drive may be only 0.5 m wide,
with positioning tolerances for the heads on the track of 5%
25 nm for write and 10% 50 nm for read operations. In
a situation where the flex circuit vibrates at an amplitude corresponding to one h114 m thickness and with the coupling ratio
being 0.23% in the second mode, the head would in turn respond
with amplitude 260 nm, some ten times greater than the write
inhibit limit.
For vibration modes that involve significant flexure of the circuit, the natural frequencies are relatively insensitive to T and
*
N , at least for the ranges of parameters considered here. When
*
those forces are artificially set to zero, for instance, the differences
in the calculated natural frequencies for modes two through five
are only 3.2%, 2.1%, 1.4%, and 0.9%, respectively. Similarly, the
differences in the calculated for those modes are only 3% or
less.
Figure 10 depicts a photograph of the test stand used for measuring the flex circuits natural frequencies. Transverse vibration
was measured by using a Michelson-style laser interferometer
Polytec OFV-3000, and fiber optic leads were used to establish
the paths for the reference and target light beams. The interferom-
eter measured changes in the two light path lengths through the
interference fringes generated by superposition of coherent beams
that reflected from i a stationary reference surface and ii the
moving flex circuit. The target beam was directed onto the flex
circuit by a right-angle prism located on the concave side of the
flex circuit. To ensure that sufficient light was returned by the flex
circuit into the optical head, a small patch of retroreflective tape
was attached to the circuit at the measurement point. Particles
within the retroreflective medium ensured that a portion of the
incident light was returned into the source optical fiber regardless
of the flex circuits potentially large displacement or slope. As
the arm or circuit was impacted, the displacement or velocity
signal was captured on an digital oscilloscope HP 54600A, and
its frequency content was characterized by using an dynamic signal analyzer HP 35665A. Peaks in the autocorrelation record
provided the natural frequencies. With this technique, vibration
measurements were readily made with a strong signal-to-noise
ratio, and with displacement resolution and bandwidth exceeding
the tests requirement.
In the spectrum for one disk drive shown in Fig. 11, the natural
frequencies of modes two, three, and four were measured at 356
Hz, 844 Hz, and 1.23 kHz. Several of the flex circuits torsion
modes were also present in the illustrated 1.6 kHz frequency
range, but their content in Fig. 11 was suppressed by judicious
placement of the impact and measurement points relative to the
torsion modes nodes. Despite its layered construction and attachments to the arm and electronics block, the flex circuit presents a
damping ratio of only 1.1% in the second mode as indicated by
the time record for ring-down in Fig. 12. By contrast, the mechanism is highly damped in the fundamental sway mode, with a
measured frequency at 3.08 Hz. Figure 13 depicts the arm and flex
circuits transient response in that mode, and just over one cycle
of motion occurred following impact.
Figures 14 and 15 show trends for the natural frequencies and
coupling ratios which are predicted in a parameter study of free
length L. For each length, the tension, shear force, and curvature
were determined on the basis of the equilibrium configurations
shown in Fig. 8. The natural frequencies for modes two through
six decrease monotonically in Fig. 14 as L is examined over a
range 30% below, and 30% above, the nominal value. Also for
these modes, Fig. 15 depicts the behavior of the displacement
ratio . The fundamental mode is dominated by the arms sway
Fig. 14 Dependence of the natural frequencies in modes two through six on the flex
circuits free length
JULY 2003, Vol. 125 341
Fig. 15 Dependence of the displacement ratios in modes two through six on the
flex circuits free length
with the flex circuit responding nearly statically, and for that
mode increases gradually within the range 357 412%. In the
even modes two, four, and six, is relatively insensitive to
design changes in L. On the other hand, has a zero crossing for
the odd modes three and five at nearly the same free length (L
26 mm) which corresponds to some 83% of the baseline value.
At that design point, the natural frequencies have increased relative to their values at L31 mm in Fig. 14, and the static stiffness
has likewise grown in Table 2. However, the modal displacement
ratios can be reduced significantly or precisely driven to zero in
modes three and five. That trade-off may be desirable in certain
applications. In short, the equilibrium and vibration model can be
used to advantage for optimizing flex circuit designs with respect
to their vibration and load transmission performance.
study in free length is discussed with a view toward exploring the

mechanisms design space and reducing arm-circuit coupling in
certain vibration modes. Parameter and optimization studies with
respect to the models other degrees of freedom, and the implications of flex circuit vibration for control system design, are subjects of current investigation.
Acknowledgment
This work was supported by a grant from IBM Corporation.
The author appreciates the assistance of Matthew Brake in conducting the static stiffness and natural frequency measurements.
Kapton is a registered trademark of E.I. du Pont de Nemours
and Company.
Summary
The primary degrees of freedom for optimizing the design of

flex circuits are the thickness of the layers within the circuit, its
free length, and the locations and angles of the attachment points
at the arm and electronics block. Those parameters can be selected
with respect to three criteria: i the static bias torque applied by
the voice coil to position the arm at a certain location on the disk;
ii the circuits natural frequencies relative to the control systems
bandwidth; and iii the extent of vibration coupling between the
flex circuit and the arm in certain vibration modes. The parameter
342 Vol. 125, JULY 2003
References
1 Love, A. E. H., 1944, A Treatise on the Mathematical Theory of Elasticity,
Fourth Edition, Dover Publications, New York.
2 Anonymous, 1996, General Specifications, Bulletin GS-96-7, DuPont Films,
E.I. du Pont de Nemours and Company.
3 Perkins, N. C., 1990, Planar Vibration of an Elastica Arch: Theory and Experiment, ASME J. Vibr. Acoust., 112, pp. 374 379.
4 Brush, D. O., and Almroth, B. O., 1975, Buckling of Bars, Plates, and Shells,
McGraw-Hill, New York.
5 Wickert, J. A., 1992, Non-linear Vibration of a Traveling Tensioned Beam,
Int. J. Non-Linear Mech., 27, pp. 503517.
M. Saigo
Research Manager,
Mechanical Engineering Laboratory,
Agency of Industrial Science and Technology,
1-2 Namiki, Tsukuba, Ibaraki 305-8564, Japan
e-mail: saigo@mel.go.jp
K. Tani
Professor,
Gifu University,
1-1 Yanagito, Gifu 501-1193, Japan
e-mail: tani@info.gifu-u.ac.jp
H. Usui
Nippon Steel Corporation,
2-6-3 Otemati, Chiyodaku, Tokyo 100-8071,
Japan
e-mail: usui@tekkai.nsc.co.jp
Vibration Control of a Traveling

Suspended System Using Wave
Absorbing Control
This paper describes vibration control of a suspended system using a wave absorbing
method. Here, we treat a system that accepts a traveling command. This system is called
a traveling system. In the previous paper, we treated a system that performs only the
vibration control, where the support of the suspended system moves only for vibration
control and eventually settles at the original position. This system is called a nontraveling
system. In a traveling system, the support moves both for traveling and for vibration
control. We present a new control strategy for these two different aims by applying the
vibration control method developed in the previous paper. A traveling multiple-pendulum
system and a traveling wire-and-load system are treated. The wire-and-load system has a
small rigid pendulum between the support and the wire. The vibration control is performed by monitoring this small rigid pendulum. The wire-and-load system is extended to
a model crane system that has a motor system to roll up and down the suspended mass
like a real crane. The same program with different parameter values controls these three
systems. Both numerical simulation and experiment have been conducted, and the developed control method has been shown to be quite effective. DOI: 10.1115/1.1569515
Introduction
Recently, traveling-wave control has been studied as an alternative to mode-based vibration control by several researchers.
Wave control has several advantages over mode-based vibration
control; it has no control and observation spillovers that may occur in the vibration control when there are some imperfections in
system modeling or inaccuracy in sensor locations; it has a better
control performance than vibration control at low frequencies; and
it is basically a local control method to which we have paid attention in our studies. The last feature means that we can suppress the
vibration of a system using no more than the information about
the dynamic states of the element nearest the actuator. This is
quite advantageous to the system whose parameters are changeable during control operations.
The studies of traveling-wave control include Vaughan 1, Von
Flotow 2,3, Millar 4, Mace 5, Fujii 6, Tanaka 7 and Utsumi 8. These studies of traveling-wave control of elastic beams
or strings have used the theoretical solution expressed in the form
of traveling wave and derived the nonreflecting condition of
waves at the control point. On the other hand, OConnor 9
treated a mass-and-spring system and presented a method of the
wave-absorption in a discrete vibration system. In these studies,
few attempts have been made to apply the wave control strategy
to practical vibrating systems to show the above-mentioned advantages of traveling-wave control.
In the previous paper 10, we presented a new practical wave
control strategy that is easy to build in a control computer with
on-line calculation of the imaginary wave-propagating system. We
applied the presented method to the vibration suppression of a
multiple-pendulum system and showed the effectiveness of the
method experimentally. There, vibration control by controlling the
support movement was addressed and no support movement for
traveling toward a given target position was considered.
In the present paper, the wave control method presented in the
previous paper is expanded to consider the combination of traveling control and vibration control. In this case, the control system
has to achieve a system displacement to a desired target position
Revised Jan. 2002. Associate Editor: R. L. Clark.
while suppressing the vibration. Both a multiple-pendulum system

and a wire-and-load suspended system are treated. Furthermore,
the method is applied to a model crane system and shown to be
useful for the vibration suppressing of a practical crane.
Equation of Motion
2.1 Multiple-Pendulum System. Figure 1 shows a traveling multiple rigid-pendulum system and a traveling wire-and-load
system. The equations of motion of a traveling multiple-pendulum
system of n degrees of freedom DOF are obtained using the
Lagranges equation of motion. The kinetic energy T k and the
potential energy U k of the k-th pendulum are expressed as
T k m k y 0
jk1
l j j
/2I k 2k /2m k y 0
jk1
U k m k g h k 1 2k /2
jk1
l j 1 2j /2
l j j k h k
where k is the angle of the k-th pendulum numbered from the

free end and assumed to be so small that the terms having powers
higher than second of k are negligible, h k is the distance between
the center of gravity of the k-th pendulum and the axis of the k-th
connecting pin, I k is the moment of inertia of the k-th pendulum
about the axis of the k-th connecting pin, m k is the mass of k-th
pendulum, l k is the distance between the axes of the kth and (k
1)-th connecting pins, and y 0 is the position of the support of
n
(T k U k ) gives
the pendulum system. The Lagrangian L k1
the equations of motion of k-th pendulum as
k1
I k l 2k
i1
h k m k
k1
m i k l k
ik1
j1
j1
lj
i1
k1
m i h j m j j l k
i1
mi
k1
l i i g l k
i1
m i h k m k k 0
(1)
From the equations obtained by replacing k by k1 and by k

1 in Eq. 1, as well as Eq. 1 itself, the following equation is
obtained.
JULY 2003, Vol. 125 343
Fig. 1 Traveling multiple-pendulum system and wire-and-load

system
q 1 p 1 J 1 / r 1 m 1 1 q 2 p 1 2 g m 1 r 2 m 2 p 1 1 2
0
q k1 p k1 k1 q k p k J k l k 1r k1 m k1 p k1 k
k2
q k1 p k k1 g
k1
j1
j1
m j r k1 m k1 p k1 k1
r k1 m k1 p k k1 0
m j r k m k p k p k1 k
j1
mj
k1,n
(2)
q n1 p n1 n1 J n l n 1r n1 m n1 p n1 n
n2
gp n1
j1
Control Strategy
The concept of our vibration control is to connect the system

whose vibration should be suppressed to a virtually infinite system
that can absorb vibration energy endlessly. In the previous paper,
we have presented a control strategy in which the real pendulum
system is suspended by the imaginary energy-absorbing multiplependulum system whose dynamics is simulated by on-line computation. Since we have to use a finite-DOF energy absorbing
system in practice, we introduced initialization methods for the
energy absorbing system. At the time of initialization, the deflections and velocities of all the imaginary pendulums other than the
lowest are set to zero, and the deflection and velocity of the lowest
are set to fit the present position and velocity of the support of the
real system.
In this paper, we treat vibration control of a suspended system
that accepts a traveling command. The traveling command is
given in term of the acceleration of the suspended system y c as a
function of time. Two types of imaginary multiple-pendulum system are possible as shown in Fig. 2. One is the nontraveling
imaginary system NTIS, Fig. 2a and the other the traveling
imaginary system TIS, Fig. 2b. The NTIS does not accept the
traveling command y c and thus its algorithm is the same as that of
the vibration control of the non-traveling system treated in the
previous paper. The vibration control calculated using NTIS is
added with the traveling command to produce position control of
the support of the real system. The influence of the traveling is
actually regarded as a disturbance appearing on the uppermost
pendulum of the real system. With TIS, on the other hand, the
support of the imaginary system is moved according to the traveling command y c and the movement is propagated through the
imaginary system down to the real system. For both cases in Fig.
n
2, the value of x 0 k1
l k k is the distance between the horizontal positions of the support and of the lowest end of the imaginary
system. The initialization is performed based on this value.
Through numerical simulations we have found the control performance using NTIS is better than that using TIS. It is considered
that the initialization using TIS brings about a larger initial deflection and velocity to the lowest imaginary pendulum because the
acceleration of the imaginary system due to y c produces a larger
value of x 0 . This causes the vibration energy flow back into the
controlled real system. Therefore, we use NTIS in the following
work.
n1
m j r n1 m n1 n1
j1
mj
r n m n n y 0
r k h k /l k ,
J k I k /l k ,
p k 1/ 1r k m k r k1 m k1 ,
q k J k h k m k
From the above equation, we see the term of the support movement appears explicitly only in the equation of the uppermost
pendulum.
2.2 Wire-and-Load System. The wire-and-load system
shown in Fig. 1 has a small rigid pendulum between the support
and the wire, and there is a load at the bottom of the wire. The
wire length is fixed. Assuming the equation of the wire is expressed by that of a dangling string and applying the finite difference method to the equation of motion, we obtain a system of
equations similar to that of a multiple simple-pendulum system
see Appendix. In the following numerical simulation, we will
treat the wire-and-load system as a non-homogenous multiplependulum system consisting of the uppermost rigid pendulum and
a large-DOF series of simple pendulums, among them the lowest
having a mass equal to that of the load.
344 Vol. 125, JULY 2003
Fig. 2 N-DOF imaginary system for traveling pendulum system y c : traveling command, x 0 : vibration control a Nontraveling Imaginary System NTIS, b Traveling Imaginary
System TIS
geous first to move the system near the destination with less vibration and then to control the final position accurately. It is an
easy job for the control computer, given the present and target
positions, to cancel the error of the final position.
When the real pendulum system is connected to the imaginary
system, the acceleration of the lowest end of the imaginary system
is represented as
x 0
i 2 /lh g 0 h/l g 1 0
1h/l
(3)
where is the angle of the uppermost rigid pendulum, is the

angle of the lowest imaginary system pendulum, 0 is the ratio of
the mass of the total real system to that of the uppermost rigid
pendulum, is the ratio of the mass of an imaginary system
pendulum to that of the uppermost rigid pendulum, and l, h and i 2
are the length, the distance between the supporting point and center of gravity, and the square of radius of gyration on the supporting point, of the uppermost rigid pendulum, respectively. As stated
for the non-traveling case investigated in the previous paper, x 0 is
used here also as the control of the movement of the support of
the real pendulum system for vibration suppression. A homogenous multiple simple-pendulum system is used as the imaginary
system in Eq. 3 for simplicity same as in the previous paper. A
measured value of , the numerically approximated value of and
the computed value of from the imaginary system give the
vibration control x 0 . Note that Eq. 3 includes no dimensional
parameters of the suspended pendulums except those of the uppermost one. So, it can give vibration control for the multiple
rigid-pendulum system as well as for the wire-and-load system
with an uppermost rigid pendulum. In other words, the control
does not depend on the length of the wire.
The parameter is introduced in Eq. 3 for the adjustment of
the performance of the control system. For a large value of , the
wave propagation in the imaginary system becomes slower and
the control gain for the vibration suppression smaller. A smaller
gain makes the control of the system more stable but less effective. In the experiment, we obtained the practical values for
considering the system stability and the limitations of the actual
DC servo motor system. Thus, the wave propagating characteristics in the imaginary system and the control performance of the
vibration suppression can be designed by changing the values of
. The parameter , the ratio of the length of the imaginary system pendulum to that of the uppermost rigid pendulum, can also
change the wave propagating characteristics in the imaginary system.
Experiment
4.1 Experiments of the Multiple-Pendulum System and

the Wire-and-Load System
Fig. 3 Three types of initializing method for NTIS. a PI
method: initialization when x 0 0, b VI method: initialization
when x 0 0, c VI method: initialization when x 0 0 with support shift
Three types of initializing timing are investigated for NTIS as

shown in Fig. 3. The cases a and b in Fig. 3 are the same as
those used in the previous paper, while c in Fig. 3 is a new
method presented in this paper. In the case a, the PI method,
n
initialization is made when x 0 0( k1
l k k ), and in the case
b, the VI method, initialization is made when x 0 0. In the case
c, the VI method, initialization is made when x 0 0 as in the
case b but also has a position shift of the support of imaginary
system. The vibration control performance in this case is better
than in the cases a and b, except that it may cause an error in
the final support position of the real system. It may be advantaJournal of Vibration and Acoustics
4.1.1 Experimental Apparatus. Figure 4 shows a schematic

diagram of the experimental apparatus. The uppermost pendulum
is connected to and supported by the nut of the ball-screw. The
ball-screw is driven by a 350W DC servomotor to generate the
horizontal movement of the support of pendulum. The DC servomotor is a velocity feedback type with an integrated tachogenerator. A rotary encoder is attached to the uppermost pendulum.
The multiple-pendulum system is made of three same-size aluminum plates connected serially by pins allowing free rotation.
The width and thickness of each plate are 40 mm and 10 mm,
respectively. The distance between the center axes of the connecting pins of each plate is 300 mm. The adjacent plates can be fixed
rigidly with braces to form a pendulum system having less than
three degrees of freedom.
The wire-and-load system is made up of a rigid pendulum, a
wire and a load. The length, width and thickness of the rigid
JULY 2003, Vol. 125 345
Fig. 4 Experimental apparatus wire-and-load system
pendulum are 60 mm, 40 mm and 15 mm, respectively. The diameter of the wire is 1 mm. The length of the wire and the weight
of the load can be changed.
The computation of the control is conducted by a DSP
TMS320C30 for the rigid-pendulum system and a personal computer with 200 MHz CPU for the wire-and-load system. The sampling period of A/D conversion is 0.1 ms for the rigid-pendulum
system and 2 ms for the wire-and-load system. A 10-DOF system
of simple pendulums has been used as the imaginary system.
4.1.2 Experimental Results. The following system movement pattern is used as the traveling command in the experiments;
the acceleration y c is 4.26 m/s2 for the time period between 0 s
and 0.0352 s and 0.01883 m/s2 between 0.0352 s and 8 s. Using
this acceleration pattern the pendulum system should travel the
distance of 0.60 m in 8 seconds.
a The case of the multiple-pendulum system
Figures 5 show the effects of the different initializing methods,
the PI and VI methods, for the 3-DOF traveling rigid-pendulum
system with 3. In these figures, the curve rising to the righthand side is the position of the support and the vibration waveform is the angle of the uppermost pendulum. From these figures,
we can confirm that our method is effective for a traveling pendulum system as well as for a nontraveling system. Figures 6
show the control performance of the PI method and the VI
method on the 1-DOF pendulum system three pendulums are
connected rigidly. Both initializing methods have excellent vibration suppressing effects. Similarly, the control performance on the
non-homogeneous 2-DOF pendulum system the lowest and the
middle pendulums are connected rigidly is well confirmed figures are not shown. The control program for the 2-DOF system is
the same as that for the 3-DOF system, for the uppermost pendulum and the total mass of the system are the same.
b The case of the wire-and-load system
Figures 7 show the effects of the different initializing methods,
the PI and VI methods, for the traveling wire-and-load system
with the wire length l0.5 m, the load weight w12.3 N and
3. In these figures, the curve rising to the right-hand side is the
position of the support and the vibration waveform is the angle of
the uppermost pendulum. From these figures we can see that the
VI method has quite an excellent damping performance Fig.
7c. The PI method is accurate in positioning the pendulum system at the traveling destination, but the vibration control performance is not so good. The performance of the VI method is not so
good figure is not shown. The vibration waveform shown in Fig.
7c resembles well that of the 1-DOF rigid pendulum shown in
Fig. 6b. This means the dynamic characteristics of the experimental wire-and-load system is similar to that of the 1-DOF rigid346 Vol. 125, JULY 2003
Fig. 5 Experimental results of 3 DOF rigid-pendulum system

3, 1. a No control, b Control with PI method, c Control with VI method
pendulum system and the vibration of the wire is practically negligible. Figures 8 show the effect of the wire length l and the load
weight w on the control performance of the VI method when
1. We can see the control performance is independent of the
change of the wire length and the load weight. Besides, the control performances in Fig. 8 for 1 are better than that in Fig.
7c for 3, which confirms that the quick wave propagation in
the imaginary system brings about a better performance in vibration suppression while the system is stable.
Fig. 6 Experimental results of 1 DOF rigid-pendulum system

1, 1. a Control with PI method, b Control with VI
method
Fig. 9 Experimental results of wire-and-load system for final

positioning l 0.3 m, w 12.3 N, 10, 10. a Interruption
of traveling command overrun case, b Position correction
after traveling command
Fig. 7 Experimental results of wire-and-load system l

0.5 m, w 12.3 N, 3, 10. a No control, b Control with
PI method, c Control with VI method
Figures 9 show the results of the operation in which final positioning accuracy is concerned when the VI method is used for
vibration control. As stated earlier, this method can produce a
positioning error, and some measures should be taken when accurate positioning at the target is needed. Figure 9a is an overrun
case where the vibration control is switched to the VI method and
the traveling command is suppressed the instant the support
reaches the exact target position. Figure 9b is a case where the
positioning correction to the exact target position is carried out
using the VI method after the traveling command is over. Either
Fig. 8 Experimental results of wire-and-load system for different wire length and load weight with VI method 1,10. a
l 0.3 m, w 12.3 N b l 0.9 m, w 31.9 N
process can position the support accurately at the target position

with small vibration. Figure 10 shows a non-traveling case where
a combination of initializing methods is used for a large initial
amplitude. This process combines the PI method that absorbs
quickly large amplitudes with the VI method that absorbs quickly
small amplitudes. The PI method was used between 0 s and 6 s
and the VI method was used afterwards. This, combinations of
different initialization methods can be used effectively to adapt to
various situations.
In order to understand the characteristics of the initializing
methods obtained in the experiments, several numerical simulations have been conducted. Figures 11 show the simulation results
for the wire-and-load system corresponding to Fig. 7b and c. In
the simulation, dry friction is assumed at each connecting pin of
the pendulum to represent the wire. The ratio of the mass of the
wire to that of the uppermost pendulum is assumed to be 0.0001.
Small wire vibrations occur at the starting period in the simulation
results, which are not observed in the experiments. In Fig. 11 we
show the waveform of the swing angle of the load in stead of that
of the uppermost pendulum because the latter is affected by the
small wire vibration. There is little difference between the results
in Fig. 11 and Fig. 7. From these figures, we can confirm that the
experiments have been performed successfully and that the angle
of the uppermost rigid pendulum is virtually equal to the swing
angle of the load.
As is easily understood, the vibration of the load would propagate up to the uppermost rigid pendulum. So, our wave-absorbing
system, which attends the uppermost rigid pendulum, is eventually effective in suppressing the vibration of the load.
Fig. 10 Experimental result of nontraveling wire-and-load system for large amplitude with PI and VI methods l 0.3 m, w
12.3 N, 10, 10.
JULY 2003, Vol. 125 347
Fig. 11 Numerical results of wire-and-load system l 0.5 m,

w 12.3 N, 3, 10. a Control with PI method, b Control
with VI method
The VI method shows best vibration suppression among the

three initializing methods. The PI and VI methods have inferior
vibration control performances, especially for a small value of .
Therefore, it is suitable to use a relatively large value of
depressing the vibration control performance for the PI or VI
methods.
4.2
Experiment of a Model Crane
4.2.1 Experimental Apparatus. Figure 12 shows the experimental model of a crane, which has a load suspended by a wire
and a pulley. One end of the wire is fixed to the motor shaft for
winding and the other end is fixed to a small rigid pendulum that
is attached to the nut of the ball-screw with free rotation. The
distance between the axis of the rigid pendulum and the wire
return on the returning pulley is equal to the diameter of the pulley. Then, the angle of the rigid pendulum is practically equal to
the swing angle of the load independent of its height when the
small vibrations of the wire can be ignored. This means the sta-
Fig. 12 Model crane system
348 Vol. 125, JULY 2003
Fig. 13 Experimental results of crane system for raising load

total weight17.6N. a No control, b Control with PI method
10, 10, c Control with VI method 1, 10
tionary direction of the pendulum is always vertical and the vibration suppression strategy for the wire-and-load system is applied
directly by monitoring one half of the load suspension system.
Our experiment has shown that it is possible to neglect the small
wire vibration in the wire-and-load system as well as in most
practical crane systems. The velocity pattern used of winding-up
and rewinding-down is 0.1 m/s and 0.1 m/s, respectively. The
load moves between the vertical positions of 0.9 m and 0.3 m
during the time period between 0 s and 6 s.
4.2.1 Experimental Results. Figures 13 show the experimental results for the case of winding up the load; a in Fig. 13 is the
case where no vibration control is used, b is the case with wave
control in the PI method, and c is the case with wave control in
the VI method. Figures 14 show the cases of rewinding down the
load with controls similar to those in Fig. 13. The system movement pattern is the same as in the case for the wire-and-load
system. In Fig. 13 we can see the amplitude of the vibration in
winding-up without control becomes larger as the wire length becomes shorter, due to the instability in winding-up of a suspended
load. The contrast in these figures demonstrates the effectiveness
of the stabilization using the wave-absorbing control. In addition,
the wave control method presented has shown an excellent control
performance regardless of the wire length. Figures 13b and
14b with the PI method for a large value of show relatively
good results in the final state of the load, that is, accurate final
position and small vibration. Thus, we can use the PI method to
position the system accurately at the target position if its vibration
suppression performance is acceptable. Even if the VI method is
ratio of the mass of an imaginary system pendulum to that of the uppermost rigid pendulum
ratio of the length of an imaginary system pendulum to that of the uppermost rigid pendulum
angle of the uppermost rigid pendulum
angle of the lowest imaginary system pendulum
k angle of the k-th rigid pendulum numbered from
the free end
h k distance between the center of gravity of k-th
rigid pendulum and the axis of k-th connecting
pin
I k moment of inertia of k-th rigid pendulum about
the axis of k-th connecting pin
m k mass of k-th rigid pendulum
l k distance between the axes of the k-th and
(k1)-th connecting pins of the rigid pendulum
k angle of the k-th imaginary system pendulum
numbered from the lowest end
n
x 0 k1
l k k n: degrees of freedom of imaginary
system
y 0 horizontal displacement of the support imaginary
or real
y c traveling command
deflection of the wire
i deflection of the i-th finite element of the wire
z coordinates of the wire measured from the lower
free end
Appendix
We assume the equation of motion of the wire is expressed by
that of a dangling string. By balancing the horizontal component
of forces on an infinitely small element z(zdz), the following
equation is obtained as
Fig. 14 Experimental results of crane system for lowering load
total weight17.6 N. a No control, b Control with PI method
10, 10, c Control with VI method 1, 10
used, the final position errors are not very significant. Figures 13
and 14 have shown our wave absorbing method is useful for the
actual crane system.
Concluding Remarks
In this paper we propose a wave control method using a nontraveling imaginary multiple-pendulum system applied to vibration control of the traveling suspended system. We have shown
that the method presented, using little information about the suspended system states, is quite effective for a traveling suspended
system whose dynamics are changeable during operation. The initializing methods with and without shifting the support of the
imaginary system have their respective merits for the application.
The former realizes accurate final positioning while the latter realizes excellent vibration suppression. Combination of methods
can produce excellent overall performances. This control method
can readily be applied to real crane systems that have a moving
pulley and parallel wiring.
Nomenclature
g Gravitational acceleration
( ), ( ) first and second order differentiation with respect
* *
to time
i 2 square of radius of gyration on the supporting
point of the uppermost rigid pendulum
l length of the uppermost rigid pendulum
h distance between the supporting point and the
center of gravity of the uppermost rigid pendulum
0 ratio of the mass of the total real system to that of
the uppermost rigid pendulum
M
2 1 2
z

0
z g t2
z2
(A1)
where is the lateral deflection of wire, z is the coordinate measured from the lowest end, is the mass of string per unit length
and M is the mass of load.
Appling the finite difference method to the above equation using the following approximations,
2
z
i1 2 i i1
z
i i1
,
z
z
ziz
we obtain
M
2 i i1 i1 i1 2i1 i i i1
z i1
g i
(A2)
where i is the lateral deflection of i-th mesh point numbered

from the lowest end of the wire and z is the finite difference
mesh.
From Eq. A2 and the equation obtained by replacing i by i
1 in Eq. A2, we obtain the following equation of motion by
substituting i ( i i1 )/z,
z
M
i
2 i i1 i1 i 2i i i1 i1
g
z i1
(A3)
The above equation is the same as the equation of motion of a
multiple simple-pendulum system obtained from Eq. 2 if we
regard z as the length of the simple pendulum and M as the
additional mass to the lowest pendulum. This means that we can
simulate the dynamics of wire in the form of a system of simple
pendulums of length z.
JULY 2003, Vol. 125 349
The influence of a wire on the total dynamics of the pendulum

system is quite small as can be seen in the experimental results.
The exact dynamical formulation for a wire is not so important in
our study M is much greater than z). So, we will not try to
develop a more accurate mathematical model of the wire in this
paper.
References
1 Vaughan, D. R., 1968, Application of Distributed Parameter Concepts to
Dynamic Analysis and Control of Bending Vibrations, ASME J. Basic Eng.,
90, pp. 157166.
2 Von Flotow, A. H., 1986, Traveling Wave Control for Large Spacecraft Structures, AIAA J., 9, pp. 462 468.
3 Von Flotow, A. H., 1986, Disturbance Propagation in Structural Networks,
J. Sound Vib., 106, pp. 433 450.
350 Vol. 125, JULY 2003
4 Miller, D. W., and von Flotow, A. H., 1989, A Travelling Wave Approach to
Power Flow in Structural Networks, J. Sound Vib., 128, pp. 145162.
5 Mace, B. R., 1984, Wave Reflection and Transmission in Beams, J. Sound
Vib., 97, pp. 237246.
6 Fujii, H., and Ohtsuka, T., 1992, Experiment of a Noncollocated Controller
for Wave Cancellation, AIAA J., 153, pp. 9397.
7 Tanaka, N., and Kikushima, Y., 1992, Active Wave Control of a Flexible
Beam, JSME Int. J., Ser. III, 351, pp. 236 244.
8 Utsumi, M., 1999, Analytical Implementation of Wave-Absorbing Control
for Flexible Beams Using Synchronization Condition, ASME J. Vibr. Acoust.,
121, pp. 468 475.
9 OConnor, W., and Lang, D., 1998, Position Control of Flexible Robot Arms
Using Mechanical Waves, ASME J. Dyn. Syst., Meas., Control, 120, pp.
334 339.
10 Saigo, M., Tanaka, N., and Tani, K., 1998, An Approach to Vibration Control
of Multiple-Pendulum System by Wave Absorption, ASME J. Vibr. Acoust.,
121, pp. 524 533.
J. R. Banerjee
Professor of Structural Dynamics,
School of Engineering
and Mathematical Sciences,
City University, Northampton Square, London
EC1V OHB
e-mail: j.r.banerjee@city.ac.uk
Dynamic Stiffness Formulation

and Its Application for a
Combined Beam and a Two
Degree-of-Freedom System
This paper is concerned with the dynamic stiffness formulation and its application for a
Bernoulli-Euler beam carrying a two degree-of-freedom spring-mass system. The effect of
a two degree-of-freedom system kinematically connected to the beam is represented exactly by replacing it with equivalent stiffness coefficients, which are added to the appropriate stiffness coefficients of the bare beam. Numerical examples whose results are obtained by applying the Wittrick-Williams algorithm to the total dynamic stiffness matrix
are given and compared with published results. Applications of the theory include the free
vibration analysis of frameworks carrying two degree-of-freedom spring-mass
systems. DOI: 10.1115/1.1569943
Introduction
There are many important engineering structures that can be

modelled as beams carrying one, two or multi degree-of-freedom
spring-mass systems. Examples of such practical applications include components of buildings and machine tools, vehicle suspensions, rotating machinery, accessories of machine structures and
robotics amongst others. Because of these wide-ranging applications the vibration behavior of beams carrying discrete structural
elements such as a one, two or multi degree-of-freedom springmass systems has received considerable attention for many years.
The research has developed from the simpler problem of the free
vibration analysis of beams with lumped masses 19 or spring
supports 1016 to that of beams carrying elastically mounted
masses 1725 or a one or two degree of freedom spring-mass
systems 26 31. For further reading on combined dynamical
systems in general, Refs. 32, 33 are recommended. All these
investigations have shown that the natural frequencies and mode
shapes of a beam carrying spring-mass systems may deviate considerably from those of the beam alone.
The main drawback of the existing solution procedures for such
problems is that by and large they concentrate solely on uniform
and unidirectional beams, and cannot be generally applied to
frameworks. This paper presents an entirely different approach
based on the dynamic stiffness method which is free from such
limitations and yields exact results, some of which can be directly
compared to the solution of the complete governing differential
equations. It does not seem to be widely recognized that the dynamic stiffness method allows an assembly technique similar to
that used in finite element methods that enables exact free vibration analysis of either a single structural element or a combination
of structural elements with different orientations. Furthermore, in
contrast to finite element methods, the results using the dynamic
stiffness method are not only exact, but also independent of the
number of elements used in the analysis and thus offers much
better computational efficiency.
The free vibration analysis of simple beams or frameworks has
been solved using the dynamic stiffness method without difficulty
34,35, because the dynamic stiffness properties of a beam 36
and the solution technique known as Wittrick-Williams algorithm
37 are both well established. The method also allows a concenContributed by the Technical Committee on Vibration and Sound for publication
Revised Dec. 2002. Associate Editor: B. Yang.
trated or lumped mass/inertia, or a spring, to be added at a node or

joint of a structure 38, without compromising the accuracy, and
hence retaining the exactness of results. However, a difficulty
arises when the beam is attached to a discrete structural element
such as a two degree-of-freedom system because the member dynamic stiffness matrix, the basic building block in the method of
such a combined system, is not currently available. This paper
develops the dynamic stiffness matrix of a Bernoulli-Euler beam
to which has been attached a two degree-of-freedom spring-mass
system, which is free to translate and rotate. The resulting eigenvalues of the combined system, which yield the natural frequencies of the combined structure, are determined using the WittrickWilliams algorithm 37. In order to illustrate the method several
cases have been evaluated and some representative results have
been compared with published ones. This paper is thought to be
the first in which the dynamic stiffness matrix of a combination of
continuous and discrete structural elements has been developed.
Although the simple case of the Bernoulli-Euler beam has been
considered here as the parent structural member for illustrative
purposes, the procedure is completely general and can easily be
extended to Timoshenko beams with or without an axial load 39
42 carrying a two degree-of-freedom spring-mass system.
Theory
2.1 Dynamic Stiffness Formulation for a Bernoulli-Euler

Beam. The dynamic stiffness matrix of a Bernoulli-Euler beam
has been available in the literature 34,36 for many years. Banerjee 43 has given a general formulation that leads to the development of the dynamic stiffness matrix of a structural element.
For a Bernoulli-Euler beam the procedure is simple and is described as follows. Readers who are not familiar with the subject
are referred to the Appendix, which gives details of the dynamic
stiffness development. In both axial and transverse motion the
governing partial differential equations in free vibration are derived, and the corresponding total differential equations for harmonic motion established. These are then integrated in terms of
two arbitrary constants for axial motion, and in terms of four more
for the flexural motion. The associated end loads are then obtained
in terms of the arbitrary constants, so that for the longitudinal
motion, the axial force at each end, as well as the axial displacements are both expressible in terms of the two arbitrary constants.
Eliminating the two constants then gives the relationship between
the two end forces and two end displacements. This embodies the
JULY 2003, Vol. 125 351
Fig. 1 Amplitudes of displacements and forces at the ends of

a Bernoulli-Euler Beam in free vibration
frequency dependent dynamic stiffness matrix for axial motion

relating the amplitudes of forces and displacements at the ends of
the freely vibrating Bernoulli-Euler beam.
In a similar way, the shear force and bending moment at each
end of the beam, together with the bending displacement and rotation at each end, are expressible in terms of four constants of
integration when dealing with the free vibration in flexural motion. On eliminating these constants, a relationship between the
four end loads and the four kinematic conditions at the ends, is
established. The resulting 44 matrix linking the two is the frequency dependent dynamic stiffness matrix in flexural motion.
The axial and flexural dynamic stiffness properties can now be
compounded into a 66 dynamic stiffness matrix of the
Bernoulli-Euler beam, relating the amplitudes of the forces to the
corresponding amplitudes of the displacements at the ends of the
vibrating beam see Fig. 1. In the usual notation when the beam
parameters EA, EI, m and L represent the extensional rigidity,
bending rigidity, mass per unit length and length of the beam
respectively, the dynamic stiffness matrix takes the following
form see Appendix for details
a1
F x1
0
F y1
0
M1
F x2
a2
F y2
0
M2
0
a2
d1
d2
d4
d5
d2
d3
d 5
d6
a1
d4
d 5
d1
d 2
d5
d6
d 2
d3

U1
V1
1
U2
V2
2
(1)
(2)
where F and are respectively the force and displacement vectors

and K is the frequency dependent dynamic stiffness matrix whose
elements k(i, j) (i1,2 . . . 6; j1,2, . . . 6). are given by a 1 , a 2
and d 1 d 6 .
In Eq. 1, F x1 , F y1 , M 1 and U 1 , V 1 and 1 are the amplitudes
of the forces and displacements at end 1 and F x2 , F y2 , M 2 and
U 2 , V 2 and 2 are the corresponding amplitudes of the forces and
displacements at end 2, respectively see Fig. 1. The elements of
the dynamic stiffness matrix a 1 , a 2 and d 1 d 6 are as follows
see Appendix for details.
a 1
EA
cot ,
L
a 2
EA
cos ec
L
(3)
with
and
352 Vol. 125, JULY 2003
m 2L 2
EA
d 1 W 3 3 S C h C S h /,
d 2 W 2 2 S S h /,
d 3 W 1 S C h C S h /
d 4 W 3 S S h /,
(5)
d 5 W 2 2 C h C /,
d 6 W 1 S h S /
with
W 1
EI
,
L
W 2
C cos ,
S sin ,
EI
L
W 3
EI
L3
m 2L 4
C h cosh ,
(4)
(6)
(7)
EI
S h sinh (8)
and
1C C h
or
FK
Fig. 2 A two degree-of-freedom system kinematically connected to a beam element
(9)
2.2 Dynamic Stiffness Formulation for a Two Degree-ofFreedom Spring-Mass System Attached to a Beam. Figure 2
shows a rigid mass M having a mass moment of inertia I about
its centroidal axis and mounted at the ends nodes of a BernoulliEuler beam by means of two linearly elastic springs of spring
constant K S1 and K S2 , respectively. The distances of the center of
gravity g of the mass from the location of the springs are D 1 and
D 2 respectively, as shown.
The governing differential equations of motion for the two
degree-of-freedom spring mass system shown in Fig. 2, undergoing free vibration in translation and rotation, are
M
2v M
2 M
t2
f k1 f k2 0
(10)
f k1 D 1 f k2 D 2 0
(11)
t2
where v M and M are the transverse displacement and angular

rotation of the rigid mass, t is time, and f k1 and f k2 are the interactive spring forces between the spring-mass system and the beam
at the two connecting points at nodes 1 and 2 respectively see
Fig. 2. These forces are given by
f k1 K S1 v 1 v M D 1 M
(12)
f k2 K S2 v 2 v M D 2 M
(13)
where v 1 and v 2 are the transverse bending displacements of the

beam at nodes 1 and 2 respectively.
If harmonic oscillation with circular or angular frequency is
assumed, then
v M x,t V M e i t ,
M x,t A M e i t ,
v 1 x,t V 1 e i t ,
v 2 x,t V 2 e i t
(14)
a1
F x1
0
F y1 F k1
0
M1
F x2
a2
F y2 F k2
0
M2
0
and
f k1 x,t F k1 e i t ,
f k2 x,t F k2 e i t
(15)
where V M , A M , V 1 , V 2 , F k1 and F k2 are the amplitudes of free

vibration.
Substituting Eqs. 14 and 15 into Eqs. 1013 gives
M 2 K S1 K S2 V M K S2 D 2 K S1 D 1 A M
K S1 V 1 K S2 V 2 0
K S1 D 1 K S2 D 2 V M I 2 K S1 D 21 K S2 D 22 A M
(17)
Equations 16 and 17 are solved for V M and A M to give

V M K S1 K S2 D 2 D 1 D 2 I 2 V 1
K S2 K S1 D 1 D 1 D 2 I 2 V 2 /Z
From Eqs. 12 and 13 the spring force amplitudes are given by

F k1 K S1 V 1 V M D 1 A M
(21)
F k2 K S2 V 2 V M D 2 A M
(22)
With the expressions for V M and A M given by Eqs. 18 and 19,

Eqs. 21 and 22 can now be written in terms of V 1 and V 2 only,
in the form
(23)
where, after some reduction, e 1 , e 2 and e 3 are given by

e 1 K S1 2 M I 2 K S2 I D 22 M /Z
(24)
e 2 K S1 K S2 I D 1 D 2 M /Z
(25)
e 3 K S2 M I
2
K S1 I D 21 M /Z
(26)
Equation 23 provides the required dynamic stiffness matrix of

the two degree-of-freedom system see Fig. 2 where e 1 , e 2 and
e 3 are the equivalent stiffnesses linking forces which will be transmitted to the beam at nodes 1 and 2 to the associated displacements V 1 and V 2 . The stiffnesses account for the attached two
degree-of-freedom system properly, and no approximation is involved in their derivation.
The above analysis enables a two degree-of-freedom spring
mass system to be represented exactly by essentially a pair of
springs, loading the beam at nodes 1 and 2 with stiffnesses varying dynamically according to Eq. 23.
2.3 Dynamic Stiffness Matrix of the Combined BernoulliEuler Beam and the Two Degree-of-Freedom Spring-Mass
System. The dynamic stiffness matrix of the combined
Bernoulli-Euler beam and the two degree-of-freedom system can
now be obtained by incorporating the dynamic stiffness forcedisplacement relationship of the two degree-of-freedom system of
Eq. 23 into the dynamic stiffness force-displacement relationship of the Bernoulli-Euler beam given by Eq. 1. This is realized
by adding respectively, e 1 to k(2,2) term, e 2 to k(2,5) term, e 2 to
k(5,2) term and e 3 to k(5,5) term of the stiffness matrix K of Eqs.
12. Thus the modified force displacement relationship of the
combined dynamical system can be written as
d2
d3
d 5
d6
a1
d 4 e 2
d 5
d 1 e 3
d 2
d5
d6
d 2
d3
(27)
(28)
(19)
(20)
e2 V1
e3 V2
d5
where K, F, and are the modified dynamic stiffness matrix, and

force and displacement vectors respectively.
ZM I 4 2 K S1 I D 21 M K S2 I D 22 M
e1
F k1
F k2
e2
d 4 e 2
FK
or
where
K S1 K S2 D 1 D 2 2
a2
(18)
A M K S1 D 1 M 2 K S2 D 1 D 2 V 1
K S2 K S1 D 1 D 2 D 2 M 2 V 2 /Z
0
d2
U1
V1
1
U2
V2
1
(16)
K S1 D 1 V 1 K S2 D 2 V 2 0
0
d 1 e 1
Application of the Dynamic Stiffness Matrix
The dynamic stiffness matrix of Eqs. 2728 can now be

used to compute the natural frequencies and mode shapes of either
a single beam with various end conditions or an assembly of
beams, for example a plane or space frame, carrying one or more
two degree-of-freedom spring-mass systems. When several elements are to be used the overall dynamic stiffness matrix of the
complete structure must be assembled. The associated natural frequencies and mode shapes are then extracted using the WittrickWilliams algorithm, see Refs. 37 and 34,35. The algorithm
guarantees that no natural frequency and its associated mode are
missed. This is, of course, not possible in the conventional finite
element method. Interested readers will fill the procedure detailed
in Refs. 34,35,37. Note that the algorithm requires information
about the clamped-clamped natural frequencies of all elements
34,35,37 within the structure to provide exact results. For the
present problem the determination of the clamped-clamped natural frequencies of the structural element shown in Fig. 2 will
involve the calculation of natural frequencies of the BernoulliEuler beam as well as those of the two degree-of-freedom system
when both ends nodes 1 and 2 are built-in. The frequency equation for a Bernoulli-Euler beam with clamped-clamped end conditions can be derived using standard procedure 44 whereas for
the two degree-of-freedom system connected to the BernoulliEuler beam shown in Fig. 2, the corresponding clamped-clamped
natural frequencies can be obtained by substituting Z0 of Eq.
20. In order to ensure that no natural frequencies of the structure
are missed the actual requirement for applying the WittrickWilliams algorithm 37 is to identify how many such clampedclamped natural frequencies exist below an arbitrarily chosen trial
frequency rather than calculating them precisely 34,35,43.
Numerical Results and Discussion
To illustrate the above theory two examples are given. The first
is taken from Ref. 28 in which a beam clamped at both ends
carries a two degree-of-freedom spring-mass system as shown in
Fig. 3. The data used in the analysis are as follows:
EA3.01010 N,
m1179 kg/m,
L 3 1.4 m,
EI6.25108 Nm2 ,
L 1 1.4 m,
M 200 kg,
D 1 D 2 0.6 m
and
L 2 1.2 m,
I 70.833 kgm2 ,
K S1 K S2 1010 N/m
The first three natural frequencies of the beam in flexural vibration

are shown in Table 1 together with those reported in Ref. 28 and
JULY 2003, Vol. 125 353
Fig. 3 A clamped-clamped beam carrying a two degree-offreedom spring-mass system
also with those obtained with the two degree-of-freedom springmass system missing. Except for the second natural frequency the
results agreed completely with those of Ref. 28 in which a
Laplace transformation technique with respect to the spatial variable is used, and is expected to give exact results. The author was
unable to pin-point the exact reason for the small discrepancy
less than 0.04% in the second natural frequency, but this may be
attributed to the accuracy used in the input data of Ref. 28. In
Ref. 28 the built-in end conditions were simulated by using very
large linear and rotational springs of the order of 1020) and are
thought unlikely to alter the results. The mode shapes corresponding to these first three natural frequencies are shown in Fig. 4,
which agree very well with the ones illustrated in Fig. 3 of Ref.
28.
Although the present theory is capable of determining the natural frequencies in axial vibration, these were suppressed by disallowing the axial deformation of the beam, so that the results are
directly comparable with those of Ref. 28. The axial natural
frequencies are uncoupled from the bending ones and they are not
of interest as they will remain unaffected by the presence of the
two degree-of-freedom spring-mass system. If the spring connections are oblique or the two degree-of-freedom system is connected to a framework a coupling would occur in which case the
axial motion of the beam would have to be included in the analysis. The results of Table 1 show that the presence of the two
degree-of-freedom spring-mass system has changed the first three
natural frequencies by 3.43%, 3.97% and 0.05%, respectively. Of
course, any change in the natural frequencies will be dependent on
the parameters of the two degree-of-freedom spring-mass system
and its location on the beam.
The next set of results for this example was obtained to demonstrate the effect of the spring constants K S1 or K S2 ) on the
fundamental natural frequency of the beam. Note that K S1 and
K S2 are equal for the present problem. Figure 5 shows the variation of the fundamental natural frequency ( 1 ) of the beam
against the variation of the spring constant K S1 or K S2 ). For
smaller values of stiffness coefficients the fundamental natural
frequency increases with the increase in spring constant, but beyond a large stiffness constant it approaches a steady constant
value and remains virtually unaltered. This is to be expected and
is in accord with an earlier investigation 28. Other parameters
Fig. 4 The first three natural frequencies and mode shapes of

the clamped-clamped beam carrying a two degree-of-freedom
shown in Fig. 3
Table 1 Natural frequencies of a combined beam and a two

degree-of-freedom system
Natural frequencies rad/s
Combined beam and two DOF system
Frequency Number
Present theory
Ref. 28
Beam alone
1
2
3
983.19
2695.1
5499.0
983
2694
5499
1018.1
2806.4
5501.8
354 Vol. 125, JULY 2003
Fig. 5 The effect of spring stiffness on the fundamental natural frequency of the clamped-clamped beam carrying a two
degree-of-freedom shown in Fig. 3
Table 2 Natural frequencies of a plane frame carrying a two

degree-of-freedom system
Natural frequency rad/s
Frequency
Number
Frame with
two DOF system
Frame without
two DOF system
%
difference
1
2
3
176.62
229.40
246.09
224.76
245.70
267.38
21.4
6.63
7.96
degree-of-freedom system is ignored. The results shown in Table

2 indicate that the natural frequencies vary significantly as a result
of the attached two degree-of-freedom system.
5
Fig. 6 A plane frame carrying a two degree-of-freedom spring
mass system
such as the lengths of the beam segments (L 1 , L 2 and L 3 ) and/or

the mass and mass moment of inertia M and I ) of the two
degree-of-freedom system, see Fig. 3, can also be varied and their
subsequent effects on the natural frequencies can be studied. A
detailed investigation of this type is beyond the scope of this
paper because it will no-doubt take it further than it is intended.
However, it should be stressed that the proposed method enables a
parametric study to be made to position appropriately the natural
frequencies within or without a certain band. This may sometimes
be necessary to avoid resonance, for example in the mounting of a
machine.
The second illustrative example is a plane frame consisting of
beam members as shown in Fig. 6. One of the structural members
carries a two degree-of-freedom spring-mass system as shown.
The natural frequencies of the frame without the two degree-offreedom spring-mass system have been obtained by Howson and
Williams using a published computer program 34 based on the
dynamic stiffness matrix method. However, when the two degreeof-freedom spring-mass system is included the problem becomes
very different and cannot generally be solved using earlier approaches. The application of the present theory is particularly useful for solving such problems. As in Ref. 34 it is assumed that
each member beam of the frame shown in Fig. 6 has the same
properties which are as follows:
EA8.0108 N,
EI4.0106 Nm2
and
m30 kg/m
The length of each member of the frame can be deduced from the
geometry shown. The parameters of the two degree-of-freedom
system connected to one of the members in the frame as shown in
Fig. 6 are
M 1500 kg,
I 75 kgm2 ,
K S2 1.25108 N/m,
K S1 108 N/m,
Conclusions
A dynamic stiffness matrix for a combined beam and a two

degree-of-freedom spring-mass system is developed starting from
the classical differential equation theory of its free natural vibration. By applying the Wittrick-Williams algorithm the resulting
dynamic stiffness matrix is used to determine the free vibration
characteristics of a beam and a plane frame which are kinematically connected to a two degree-of-freedom spring-mass system.
The results for the beam carrying a two degree-of-freedom springmass system agreed with published results. The theory allows any
number of two degree-of-freedom spring-mass systems to be
added to a structure consisting of beams. This useful extension of
the dynamic stiffness method to combined system to cover frameworks is the context in which the theory is expected to be most
advantageous, particularly in the avoidance or attenuation of unwelcome vibration. The method used is exact and can be applied
to validate finite element or other approximate methods.
Appendix
Derivation of the Dynamic Stiffness Matrix of a BernoulliEuler Beam. The axial and bending deformations for a
Bernoulli-Euler beam are uncoupled. Therefore, the derivation of
the dynamic stiffness matrix for each of these two cases can be
carried out separately, and later can be combined together.
Axial Stiffnesses. A uniform Bernoulli-Euler beam of length L
is shown in Fig. 7 in a rectangular right handed Cartesian coordinate system with the X-axis coinciding with the axis of the
beam. The governing differential equation of motion of the beam
in free axial or longitudinal vibration is given by 36,44
EA
2u
2u
t2
m
2
(A1)
where EA and m are the axial or extensional rigidity and mass

per unit length of the beam respectively, and u(x,t) is the axial
displacement of the cross-section at a distance x, and t is time.
D 1 D 2 1.5 m.
The nodes of the frame are numbered as shown. Following the

usual procedure generally used in the finite element method, the
dynamic stiffness matrix of the complete frame is assembled from
the dynamic stiffness matrices of all the individual elements in the
frame. Of course, for the element connecting the nodes 2 and 4,
the dynamic stiffness theory developed in this paper is used. The
first three natural frequencies of the frame including and excluding the two degree-of-freedom system were obtained using the
Wittrick-Williams algorithm and are shown in Table 2 together
with the percentage difference between the two sets of results.
The published program of Howson and Williams 34 gives the
same natural frequencies as the present theory when the two
Fig. 7 Coordinate system and notation for a Bernoulli-Euler

beam
JULY 2003, Vol. 125 355
a1
F x1
F x2
a2
a2 U1
a1 U2
(A11)
where the frequency dependent elements of the 22 dynamic

stiffness matrix are given by
a 1
EA
cot ,
L
a 2
EA
cos ec
L
(A12)
Bending Stiffnesses. The governing differential equation of

motion of the Bernoulli-Euler beam shown in Fig. 7 in bending
or flexural natural vibration is given by 36,44
EI
Fig. 8 End conditions for a Bernoulli-Euler beam in free vibration
Assuming harmonic oscillation with angular or circular frequency so that

u x,t u ,t U e i t
(A2)
2 U0
(A3)
d 4V
d4
m 2L 2
EA
(A4)
U C 1 cos C 2 sin
(A5)
where C 1 and C 2 are constants.

The expression for axial force F(x) in the positive direction of
X see Fig. 7 can be expressed with the help of Eq. A5 as
dU
EA dU
dx
L d
At x0 i.e. 0 ,
At xL i.e. 1 ,
UU 1
UU 2
and
and
FF x1
and
C1
cos C 2
m 2L 4
EI
(A16)
C 6 cosh
M x M
EI d 2 V
L d
2
EI
L2
(A18)
2 C 3 cos C 4 sin
C 5 cosh C 6 sinh
(A19)
S x S
(A9)
(A15)
1 dV
C 3 sin C 4 cos C 5 sinh
L d L
(A7)
EI d 3 V
L d
3
EI
L3
3 C 3 sin C 4 cos
C 5 sinh C 6 cosh (A20)
(A10)
The constants C 1 and C 2 can now be eliminated from Eqs. A9

and A10 to give the dynamic stiffness formulation of the axially
vibrating beam relating amplitudes of the forces and displacements at its ends as follows
356 Vol. 125, JULY 2003
FF x2 (A8)
C1
sin C 2
0
EA
F1
F2
L
sin
4 V0
where C 3 C 6 are constants.

The expressions for rotation , bending moment M and
shear force S consistent with the positive direction of the coordinate system at the left hand end of the beam can be written as
Substituting Eqs. A7 and A8 into Eqs. A5 and A6, the following matrix relationships can be obtained
1
U1
U2
cos
(A14)
V C 3 cos C 4 sin C 5 cosh C 6 sinh

(A17)
(A6)
Now referring to Fig. 8a, the boundary or end conditions for

displacements and forces can be applied as follows.
(A13)
The solution of the differential Eq. A15 is given by 36,44
EA
C 1 sin C 2 cos
L
where
The solution of the differential Eq. A3 is given by
F x F EA
t2
where V( ) is the amplitude of bending or flexural vibration and

x/L. Substituting Eq. A14 into Eq. A13 gives
where
v ,t V e i t
d2
2v
m
4
where EI and m are the bending or flexural rigidity and mass per
unit length of the beam respectively, and v (x,t) is the bending or
flexural displacement of the cross-section at a distance x and t is
time.
Assuming harmonic oscillation with angular or circular frequency so that
where U( ) is the amplitude of axial vibration and x/L.

Substituting Eq. A2 into Eq. A1 gives
d U
4v
Now referring to Fig. 8b, the boundary or end conditions for

the displacements and forces can be applied as follows
At x0 i.e. 0 ,
and
At xL i.e. 1 ,
1 ,
VV 1 ,
SF y1
M M 1
VV 2 ,
2 ,
(A21)
SF y2
and
M M 2
(A22)
Substituting Eqs. A21 and A22 into Eqs. A17A20, the following two matrix equations can be obtained for displacements
and forces, respectively, in terms of the constants C 3 C 6 .
1
V1
0
1
V2
C
2
S /L
/L
/L
C h
S h
C /L
S h /L
C h /L

C3
C4
C5
C6
0
W 3 3
0
W 3 3
F y1
2
2
W 2
0
W 2
0
M1
3
3
3
F y2
W 3 S W 3 C W 3 S h W 3 3 C h
M2
W 2 2 C W 2 2 S W 2 2 C h
W 2 2 S h
C3
C4
C5
C6
(A24)
where
C cos ,
S sin ,
W 1
EI
,
L
C h cosh ,
W 2
EI
L
W 3
S h sinh ,
EI
(A25)
L3
The constants C 3 C 6 can now be eliminated to give the 44

dynamic stiffness matrix of the Bernoulli-Euler beam in flexural
vibration as follows
d1
F y1
d2
M1
F y2
d4
M2
d5
d2
d4
d5
d3
d 5
d6
d 5
d1
d 2
d6
d 2
d3

V1
1
V2
2
(A26)
where
d 1 W 3 3 S C h C S h /,
d 2 W 2 2 S S h /,
d 3 W 1 S C h C S h /
d 4 W 3 3 S S h /,
(A27)
d 5 W 2 2 C h C /,
d 6 W 1 S h S /
(A28)
1C C h
(A29)
with
Combination of Axial and Bending Stiffnesses. A simple superposition is now possible to put the axial and bending dynamic
stiffnesses together in order to express the force-displacement relationship of the Bernoulli-Euler beam. Superposing Figs. 8a
and 8b, see Fig. 1, and using Eqs. A11 and A26, one obtains
the dynamic stiffness relationship of a Bernoulli-Euler beam in
free vibration, as follows
a1
F x1
0
F y1
0
M1
F x2
a2
F y2
0
M2
0
a2
d1
d2
d4
d5
d2
d3
d 5
d6
a1
d4
d 5
d1
d 2
d5
d6
d 2
d3
FK
(A31)
where F and are respectively the force and displacement vectors

and K is the frequency dependent 66 dynamic stiffness matrix
whose elements k(i, j) (i1,2 . . . 6; j1,2, . . . 6) are given by
a 1 , a 2 and d 1 d 6 defined in Eqs. A12, A27 and A28. Note
that K is symmetric as expected.
References
(A23)
and
or
U1
V1
1
U2
V2
2
(A30)
1 Chen, Y., 1963, On the Vibration of Beams or Rods Carrying a Concentrated

Mass, ASME J. Appl. Mech., 30, pp. 310311.
2 Pan, H. H., 1965, Transverse Vibration of an Euler Beam Carrying a System
of Heavy Bodies, ASME J. Appl. Mech., 32, pp. 434 437.
3 Laura, P. A. A., Pombo, J. L., and Susemihl, E. A., 1974, A Note on the
Vibration of a Clamped-Free Beam with a Mass at the Free End, J. Sound
Vib., 37, pp. 161168.
4 Parnell, L. A., and Cobble, M. H., 1976, Lateral Displacements of a Vibrating
Cantilever Beam with a Concentrated Mass, J. Sound Vib., 44, pp. 499511.
5 Gurgoze, M., 1984, A Note on the Vibrations of Restrained Beams and Rods
with Point Masses, J. Sound Vib., 96, pp. 461 468.
6 Burch, J. C., and Mitchell, T. P., 1987, Vibration of a Mass-Loaded ClampedFree Timoshenko beam, J. Sound Vib., 114, pp. 341345.
7 Laura, P. A. A., Filipich, C. P., and Cortinez, V. H., 1987, Vibrations of
Beams and Plates Carrying Concentrated Masses, J. Sound Vib., 117, pp.
459 465.
8 Wu, J. S., and Lin, T. L., 1990, Free Vibration Analysis of a Uniform Cantilever Beam with Point Masses by an Analytical-and-Numerical-Combined
Method, J. Sound Vib., 136, pp. 201213.
9 Abramovich, H., and Hamburger, O., 1991, Vibration of a Cantilever Timoshenko Beam with a Tip Mass, J. Sound Vib., 148, pp. 162170.
10 Massalas, C., and Soldatos, K., 1978, Free Vibration of a Beam Subjected to
Elastic Constraints, J. Sound Vib., 57, pp. 607 608.
11 Davies, H. G., and Rogers, R. J., 1979, The Vibration of Structures Elastically Constrained at Discrete Points, J. Sound Vib., 63, pp. 437 447.
12 Lau, J. H., 1981, Fundamental Frequency of a Constrained Beam, J. Sound
Vib., 78, pp. 154 157.
13 Verniere, P., Ficcadenti, G., and Laura, P. A. A., 1984, Dynamic Analysis of
a Beam with an Intermediate Elastic Support, J. Sound Vib., 96, pp. 381389.
14 Lau, J. H., 1984, Vibration Frequencies and Mode Shapes for a Constrained
Cantilever, ASME J. Appl. Mech., 51, pp. 182187.
15 Maurizi, M. J., and Bambill de Rossit, D. V., 1987, Free Vibration of a
Clamped-Clamped Beam with an Intermediate Elastic Support, J. Sound Vib.,
119, pp. 173176.
16 Rao, C. K., 1989, Frequency Analysis of Clamped-Clamped Uniform Beams
with Intermediate Elastic Support, J. Sound Vib., 133, pp. 502509.
17 Jacquot, R. G., and Gibson, J. D., 1972, The Effects of Discrete Masses and
Elastic Supports on Continuous Beam Natural Frequencies, J. Sound Vib., 23,
pp. 237244.
18 Laura, P. A. A., Maurizi, M. J., and Pombo, J. L., 1975, A Note on the
Dynamic Analysis of an Elastically Restrained-Free Beam with a Mass at the
Free End, J. Sound Vib., 41, pp. 397 405.
19 Laura, P. A. A., Susemihl, E. A., Pombo, J. L., Luisoni, L. E., and Gelos, R.,
1977, On the Dynamic Behavior of Structural Elements Carrying Elastically
Mounted, Concentrated Masses, Appl. Acoust., 10, pp. 121145.
20 Bapat, C. N., and Bapat, C., 1987, Natural Frequencies of a Beam with
Nonclassical Boundary Conditions and Concentrated Masses, J. Sound Vib.,
112, pp. 177182.
21 Ercoli, L., and Laura, P. A. A., 1987, Analytical and Experimental Investigation on Continuous Beams Carrying Elastically Mounted Masses, J. Sound
Vib., 114, pp. 519533.
22 Larrondo, H., Avalos, D., and Laura, P. A. A., 1992, Natural Frequencies of a
Bernoulli Beam Carrying an Elastically Mounted Concentrated Mass, Ocean
Eng., 19, pp. 461 468.
23 Abramovich, H., and Hamburger, O., 1992, Vibration of a Cantilever Timoshenko Beam with Translational and Rotational Springs and with Tip
Mass, J. Sound Vib., 154, pp. 67 80.
24 Rossi, R. E., Laura, P. A. A., Avalos, D. R., and Larrondo, H. O., 1993, Free
Vibrations of Timoshenko Beams Carrying Elastically Mounted, Concentrated
Masses, J. Sound Vib., 165, pp. 209223.
25 Gurgoze, M., 1996, On the Eigenfrequencies of a Cantilever Beam with
Attached Tip Mass and a Spring-Mass System, J. Sound Vib., 190, pp. 149
162.
26 Jen, M. U., and Magrab, E. B., 1993, Natural Frequencies and Mode Shapes
of Beams Carrying a Two-Degree-of-Freedom Spring-Mass System, ASME
J. Vibr. Acoust., 115, pp. 202209.
27 Wu, J. S., and Huang, C. G., 1995, Free and Forced Vibrations of a Timoshenko Beam with any Number of Translational and Rotational Springs and
Lumped Masses, Int. J. Commun. Num. Meth. Eng., 11, pp. 743756.
28 Chang, T. P., and Chang, C. Y., 1998, Vibration Analysis of Beams with a
Two Degree-of-Freedom Spring-Mass System, Int. J. Solids Struct., 35, pp.
383 401.
29 Wu, J. S., and Chou, H. M., 1998, Free Vibration Analysis of a Cantilever
Beam Carrying any Number of Elastically Mounted Pointed Masses with the
JULY 2003, Vol. 125 357
30
31
32
33
34
35
36
Analytical-and-Numerical-Combined Method, J. Sound Vib., 213, pp. 317

332.
Wu, J. S., and Chou, H. M., 1999, A New Approach for Determining the
Natural Frequencies and Mode Shapes of a Uniform Beam Carrying any Number of Sprung Masses, J. Sound Vib., 220, pp. 451 468.
Wu, J. J., and Whittaker, A. R., 1999, The Natural Frequencies and Mode
Shapes of a Uniform Beam with Multiple Two-DOF Spring-Mass Systems, J.
Sound Vib., 227, pp. 361381.
Dowell, E. H., 1979, On Some General Properties of Combined Dynamical
Systems, ASME J. Appl. Mech., 46, pp. 206 209.
Nicholson, J. W., and Bergman, L. A., 1986, Free Vibration of Combined
Dynamical Systems, J. Eng. Mech., 112, pp. 113.
Howson, W. P., and Williams, F. W., 1977, Compact Computation of Natural
Frequencies and Buckling Loads for Plane Frames, Int. J. Numer. Methods
Eng., 11, pp. 10671081.
Howson, W. P., Banerjee, J. R., and Williams, F. W., 1983, Concise Equations
and Program for Exact Eigensolutions of Plane Frames including Member
Shear, Adv. Eng. Software, 5, pp. 137141.
Kolousek, V., 1973, Dynamics in Engineering Structures, Butterworths, London.
358 Vol. 125, JULY 2003
37 Wittrick, W. H., and Williams, F. W., 1971, A General Algorithm for Computing Natural Frequencies of Elastic Structures, Q. J. Mech. Appl. Math., 24,
pp. 263284.
38 Anderson, M. S., Williams, F. W., Banerjee, J. R., Durling, B. J., Herstrom, C.
L., Kennedy D., and Warnaar, D. B., 1986, User Manual BUNVIS-RG: An
Exact Buckling and Vibration Program for Lattice Structures, with Repetitive
Geometry and Substructuring option, NASA Tech. Memo. 87669.
39 Cheng, F. Y., 1970, Vibration of Timoshenko Beams and Frameworks, J.
Struct. Div. ASCE, 96, pp. 551571.
40 Wang, T. M., and Kinsman, T. A., 1971, Vibration of Frame Structures According to the Timoshenko Theory, J. Sound Vib., 14, pp. 215227.
41 Howson, W. P., and Williams, F. W., 1973, Natural Frequencies of Frames
with Axially Loaded Timoshenko Members, J. Sound Vib., 26, pp. 503515.
42 Cheng, F. Y., and Tseng, W. H., 1973, Dynamic Stiffness Matrix of Timoshenko Beam Columns, J. Struct. Div. ASCE, 99, 527549.
43 Banerjee, J. R., 1996, Dynamic Stiffness Formulation for Structural Elements: A General Approach, Comput. Struct., 63, pp. 101103.
44 Clough, R. W., and Penzien, J., 1975, Dynamics of Structures, McGraw-Hill,
Singapore.
Effect of Damping on Asymmetric

Systems
Paolo Gallina
Associate Professor,
Department of Energetics,
University of Trieste,
Via A. Valerio 10-34127 Trieste, Italy
e-mail: pgallina@units.it
This paper addresses the phenomenon of the destabilizing effect of slight damping on
asymmetric linear systems. Previous works had showed that the destabilizing effect, regarded for a long time as a paradox, depends upon the ratio of the damping coefficients. This work extends those results to n-dof systems. In fact, conditions for a general
asymmetric n-dof slightly damped system to be stable are obtained. Also, a useful sufficient condition is carried out. This practical design tool gives optimum damping ratios
and takes into consideration the parameter uncertainties as well. The solution is based on
the eigenvalue sensitivity analysis. Moreover, a formal physical explanation of the destabilizing effect of damping is given. Eventually, the theory is validated by means of a
simple example. DOI: 10.1115/1.1569945
Introduction
The dynamic behavior of a general linear discrete undamped

system can be described by a vector differential equation of the
form
Mx t Kx t 0
(1)
where M and K are the mass and stiffness matrices respectively.

If mass and stiffness matrices are symmetric positive definite,
the system is stable in the BIBO sense, where the acronym BIBO
stands for bounded input bounded output. In most cases, the
damping matrix is neglected as far as stability analysis is concerned. In fact, the structural damping of the real system would
have the effect of moving the eigenvalues of the undamped system toward the negative complex plane.
If the mass and/or the stiffness matrices are asymmetric, the
undamped system can be either stable or unstable. This type of
dynamic systems can be found in aeronautics, active control of
large structures, vehicle dynamics and milling machining 1,2. In
these circumstances, neglecting the damping matrix is no longer
possible.
Because of the asymmetry of the matrices, a destabilizing effect
due to the damping matrix could arise. This atypical phenomenon,
regarded for a long time as a paradox 3 has been physically
explained by Semler et al. 4. Using a 2 d.o.f. example, Semler
showed that damping introduces two effects: it dissipates energy
and it modifies the modes of vibration. The latter could cause the
work done by the external forces to increase. If the work done by
the external forces becomes greater than the energy dissipated by
damping, the system will be unstable.
Bolotin 5 showed how the destabilizing effect of slight damping is related to the ratio of the damping coefficients. This result
was carried out for a 2 d.o.f. asymmetric system. Studying a linearized double pendulum with a follower force, Herrmann et al.
6,7 achieved similar results. Using an approximated method,
Nemat-Nasser 8, showed the destabilizing effect of damping on
a cantilevered continuous pipe conveying fluid. The approach he
adopted can be extended to 3 d.o.f systems. Unfortunately, all
those works focus on particular examples with few degrees of
freedom. This paper extends the theory to a general n-dof asymmetric linear slightly damped system. If the undamped system is
stable, all the eigenvalues are located on the imaginary axis. The
introduction of slight damping modifies the location of the eigenvalues of the system on the complex plane. By means of an eigenvalue sensitivity analysis 9, one can foresee which direction
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Sept. 2001;
Revised Jan. 2003. Associate Editor: J. Cusumano.
the eigenvalues will move. We proved that destabilizing effect due

to slight damping occurs when a system of homogeneous linear
inequalities is not satisfied. This general condition is an extension
of the results obtained by Bolotin 5, Herrmann et al. 6
The paper is divided into three parts: in the first one a detailed
analysis of slight damping effects on a general n-dof stable undamped system is carried out. A stability condition is obtained.
Moreover, using the cone algorithm 10, a simplified sufficient
condition for the slight damped system to be stable is obtained.
This result is of practical interest because it gives the optimum
damping ratios for the system to be stable. Damping coefficient
uncertainties are taken into account as well.
In the second part, a physical explanation of the destabilizing
effect of damping is given for a general n-dof system, which extends the result obtained by Semler 4 Eventually, the proposed
method is explained and validated by means of a simple example.
Theory
2.1
Sensitivity Eigenvalue Analysis. Consider the system

Mx t Cx t Kx t 0
(2)
where M and K are real nn asymmetric matrices. The damping

matrix C can be decomposed into a product of a matrix U and a
scalar c, namely CcU, where c C and UC/C. C is assumed symmetric and U inherits this symmetry.
Note that the system 2 can be always arranged so that the
damping matrix is symmetric. In fact, if C was not symmetric, it
is always possible to decompose the damping matrix using the
V, where S
singular value decomposition method, namely CSC
is a symmetric matrix. Therefore,
and V are square matrices and C
x(t)C
x(t)K
x(t)
the system 2 can be written in the form M
1
1
1
1
0 where MS MV , KS KV and C is symmetric.

By means of the decomposition cU, one can calculate the eigenvalues of the system as a function of c. If c0, Eq. 2 represents the equation of the undamped system. As c increases, the
damping matrix will affect the dynamic behavior of the system. In
this way one can study the effect of a slight damping by introducing a small c. Since U i j C i j /c, each component of U represents a damping ratio. Note that, for definition, all the damping
ratios are constant. For the sake of clarity, we will call U damping
ratio matrix and c damping.
In the following, we answer this simple but fundamental question: If the undamped system is stable, will the introduction of the
damping destabilize the system? In order to answer the question,
one can find where the eigenvectors move on the complex plane
as c changes.
JULY 2003, Vol. 125 359
The right eigenvalue problem associated with Eq. 2 can be

represented by
M i2 C i K ui 0
(3)
where i , ui are respectively the ith eigenvalue and the ith right
eigenvector of the system. In the same way, the left eigenvector
problem is given by the equation
viT M i2 C i K 0
(4)
where vi is the ith left eigenvector of the system.

Differentiating Eq. 3 with respect to the damping c, one obtains
2M i
d i
dui
d i
U i C
0 (5)
ui M i2 C i K
dc
dc
dc
Fig. 1 Mechanical example of a n d.o.f. system. Nonconservative forces that cause the mass matrix andor the damping matrix to be asymmetric are not depicted in figure. A damping
element can be inserted between two masses andor between
one mass and the frame.
Premultiplying by the left eigenvalue viT , it yields
viT 2M i
vT1 Uu1
d 1 d 2
T
0.5263
dc
dc
2v1 Mu1
d i
dui
d i
U i C
u vT M i2 C i K
dc
dc i i
dc
viT 2M i
d i
d i
u 0
U i C
dc
dc i
Therefore the derivative of the ith eigenvalue calculated for c

0 is
viT U i ui
d i
T
dc
vi 2M i C ui
c0
viT Uui
Re d i /dc 0,
(8)
a small c will move the ith eigenvalue toward left, causing the
system to remain stable. Therefore, for small c, the damping ratio
matrix will have a stabilizing effect only if Re(di /dc)0 i
1, . . . ,n. Note that the eigenvectors of an undamped system are
always real. Therefore, for this reason and for Eq. 7
d i
d i
Re
dc
dc
(9)
We now apply the condition 8 to the following simple example.

Consider the system MxcUxKx0, with the matrix
5
4
1.5
10
(10)
The Undamped system (c0) is stable in the sense of BIBO

stable. Its eigenvalues are 1 1.4684i, 2 1.4684i, 3
1.7074i, 3 1.7074i. The right and left eigenvectors of the
undamped system are
0.7327
u1
,
0.3825
u4
0.6466
,
v1
0.5148
0.7327
u2
,
0.3825
0.6526
0.5645
0.6466
v2
,
0.5148
v4
0.2400
0.8289
Therefore
360 Vol. 125, JULY 2003
In conclusion, for a small value of c, the stable undamped system

will remain stable
0.2400
v3
,
0.8289
2.2 An Alternative Simpler Way to Express Stability Conditions. Instead of Eq. 7, by expanding matrix U, we can obtain a simplified formulation. For the sake of clarity, we introduce
Fig. 1, where part of a discrete mass-spring-damped system is
represented. In this way one can visualize the physical meaning of
the elements of the matrix U. In other words, Fig. 1 shows the
damping coefficients one can introduce in the mechanical system.
The forces that cause the stiffness matrix and/or the mass matrix
to be asymmetric are not depicted in figure.
We can now study the structure of the damping ratio matrix U
and give its physical interpretation. This provides the designer
with an important tool to select the best combination of damping
elements. Basically, one has two different ways to introduce a
damping element on a mechanical system:
a damping element can be connected between one mass/
inertia and the frame;
a damping element can be connected between two masses/
inertias.
The former is represented in the figure by a physical damping
element attached between the mass m l and the frame. Its numerical damping value is c ll . We call damping element coefficient
the adimensional coefficient ll .
In this case, if c ll was the only damping element on the mechanical system, the matrix U would be
UUll
0.6526
u3
,
0.5645
(12)
(7)
2viT Mui
Eq. 7 is a simple tool for foreseeing the effect of damping. In

fact, if Re(di /dc)0, a small c will move the ith eigenvalue i of
the undamped system ( i belongs to the complex axis since the
undamped system is supposed to be stable in a BIBO sense toward the right complex plane, causing the system to become unstable. Conversely, if
vT3 Uu3
d 3 d 4
T
0.1162
dc
dc
2v3 Mu3
(6)
(11)
Ull o, p 0
o,pl
Ull o, p ll
opl
0
]
]
0
ll

0
]
]
0
(13)
where the notation o, p indicates the element of the matrix on

the oth row and pth column.
Since the system has n degrees of freedom, it is possible to
connect up to n damping elements between each mass/inertia and
the frame.
If c ls was the only damping elements introduced between the
mass m l and the mass m s the matrix U would be
Uls o,p 0 ol,os,pl,ps

Uls l,l Uls s,s ls
UUls
Uls l,s Uls s,l ls
0
]
ls
0
]
ls
0
0
]
0
ls
ls
In order the derivatives of the eigenvalues to be negative, the

following system of linear inequalities has to be satisfied
A0
where AR
(14)
In a n d.of. system, the number of possible damping elements

attached between one mass and another one is (n(n1))/2.
Therefore the sum of the number of the possible damping elements attached to the frame plus the number of the possible damping elements attached between two masses is n(n(n1))/2
(n(n1))/2. In many cases, the number of damping elements
in a real mechanical system is less than (n(n1))/2.
Using a compact notation and considering all the damping element coefficients at the same time, one can compose the damping
ratio matrix
n
rr
r,t1
rt
Urt
r,t1
rt
Urt
(15)
Multiplying by the ith right eigenvalue and premultiplying by the

ith left eigenvalue, it yields
n
viT Uui
r1
viT Urr ui
r,t1
rt
i r
viT Urt ui
n
v u
r1
i r rr
r,t1
rt
vi r ui r vi r ui t
vi t ui r vi t ui t rt
(16)
where the symbol r means the rth component of a vector. It is

clear from the Eq. 16 that viT Uui is a linear combination of the
damping element coefficients r,r , r,t (r,t1, . . . ,n,rt).
Therefore, one can write
viT Uui aiT
(17)
where
ai
vi 1 ui 1
]
vi n ui n
vi 1 ui 2 vi 1 ui 2 vi 2 ui 1 vi 2 ui 2
]
vi n1 ui n vi n1 ui n vi n ui n1 vi n ui n
and the vector R r (r((n(n1))/2)) gathers the damping coefficients
T 11 , 22 , . . . , nn , 12 , 13 , . . . 1n , 23 , . . . , n1 n
(18)
By replacing Eq. 17 into 7, one obtains the derivative of the ith
eigenvalue as a function of the damping element coefficient vector
d i
dc
is
aiT
2viT Mui
(19)
Therefore
d i
0sign viT Mui aiT 0
dc
(20)
(21)
sign vT1 Mu1 aT1

]
A
sign vTn Mun aTn
(22)
In conclusion, instead of Eq. 8, the stability condition is now

replaced by the system of inequalities 21. They present the advantage that damping element coefficients appear explicitly.
Therefore damping elements can be easily chosen in order maintain the system stable.
Example:
Considering again the numerical example 10, where 112,
221, one obtains
r1
nr
and Eq. 20 becomes
0.4737
0.1969
0.4737
0.1969
0.1566
0.4679
0.1566
0.4679
A 11 22 T 0.7505 0.7505 0.1547 0.1547 T 0
(23)
Eq. 23, according to Eq. 12 confirms again that the slightly

damped system is stable
2.3 A Simpler Sufficient Condition for Stability Solution
Cone. Although condition 21 is useful to directly foresee the
stabilizing or destabilizing effect of the damping element coefficients, it is not of practical use for the mechanical designer.
The main reason is that system 21 has an infinite number of
solutions. A method to choose the best solution of 21 among
the possible alternatives is desirable.
We now introduce the notation ai sign(viT Mui )ai . From a
geometrical point of view, all the infinite solutions of 21 can be
obtained in the following way. Each vector ai can be thought of as
the normal vector to an homogeneous hyperplane in R r . The hyperplane bounds the halfplane S i aiT 0,R r . The intersection of the r halfplanes generates the solution set S of system
r
21 S i1
S i . It is well known that the solution region is a
convex polyhedral cone in R r 11. Introducing the definition of
solution cone one can find a subset of solutions of system 21.
A hypercone 10 with axis w and angle in R r is defined as
C w, w, , R r
(24)
where w,arccosw"/w is the angle between w and .

A solution cone is a hypercone enclosed by the solution set S. One
can find the largest solution cone enclosed by the solution set. The
set of all the angles of the cones enclosed by S is
sc C w, S,wR r , R
(25)
Therefore, the angle of the largest solution cone is sc

maxsc and its axis is
wsc w C w, sc S,wR r
(26)
The largest solution cone can be found using the cone algorithm
proposed by Wan 10. A modified version of the cone algorithm
is proposed in appendix. Since the largest solution cone is the
largest cone included in S, it gives a set of solutions of system
21. Of course, it does not represent all the possible solutions; it
defines a subset of the solutions of the system of inequalities. In
spite of that, one can note that, while system 21 is made up of r
inequalities, the largest solution cone is defined by means of only
one inequality:
JULY 2003, Vol. 125 361
arccos
T
wsc
sc
cos sc
wsc
wsc
T
wsc
(27)
where /. Moreover, only two parameters are necessary to

express the condition: wsc , sc . Equation 27 is a powerful and
simple tool for designing the damping matrix. The damping matrix will stabilize the undamped system under the hypothesis of
slight damping if Eq. 27 is satisfied. If one could choose the
damping element coefficients, the best solution would be
wsc
o pt
wsc
(28)
T
/ wsc
In fact, when o pt the value of the scalar product wsc
1 of Eq. 27 is maximum.
Not only can the designer choose the best solution, but he also
can analyze the robustness of the solution.
In a practical scenario model errors occur. Therefore, the real
value of defers from the theoretical optimum solution o pt . One
can take into account the uncertainties in the model by means of
the relationship
o pt
(29)
By replacing Eq. 29 into Eq. 27, we can express the stability

condition as a function of the uncertainty vector
oT pt 1cos sc
(30)
Using Eq. 30, given the optimum choice for the damping coefficients, one can analyze the effect of the uncertainties as far as
robustness of the system is concerned. In other words, as long as
uncertainties satisfy Eq. 30, the damped system is stable.
Resuming the example 10, suppose now that the damping
element coefficients 11 , 22 have to be set by the designer. The
optimum solution provided by the cone algorithm is
o pt 0.7262 0.6875 T ,
sc 24.3
(31)
and the damping element uncertainties have to satisfy Eq. 30

0.7262 110.6875 220.0886
(32)
3 A General Physical Explanation of the Destabilizing

Effect of Damping
In this section we explain in a elegant way the destabilizing
effect of damping from a physical point of view. We show that
there exist modes of vibration that extract energy from the reservoir of the system. In order to be excited these modes need an
energy sink where the system dissipates part of the energy by
means of damping. We will prove that, when the slightly damped
system is unstable, the energy acquired by these modes is greater
than the dissipated energy.
The mass and stiffness matrices of system 2 can be always
decomposed in symmetric positive definite matrices and non symmetric matrices, namely MMs M and KKs K. Therefore the system 2 can be rewritten in the form
Ms x t cUx t Ks x t Mx t Kx t
(33)
Since Ms and Ks are positive definite, system 33 can be thought

of as a conservative system subjected the external forces
(Mx(t)Kx(t)).
Given the ith eigenvalue and right eigenvector of the system
i , ui , one mode of vibration of the system is given by
x
1
i e i t
u e i t u
2 i
(34)
where ui , i are the complex conjugated of ui , i i i i . To

gain a better understanding of the energy exchanges, consider
only the imaginary part of the eigenvalue in order to neglect the
exponential term of the mode of vibration. In this way we con362 Vol. 125, JULY 2003
sider only the steady periodic part of the solution. In the following
we calculate the energy the system exchanges during each cycle.
x
1
i e i i t
u e i i t u
2 i
(35)
The energy dissipated by damping during one cycle is

W d
2/i
xT Cxdt
(36)
while the energy introduced by the external forces is

W e
2/i
xT Mxdt
2/i
xT Kxdt
(37)
Consider two cases:

1 Undamped system.
When c0 undamped system the system does not dissipate
energy, namely W dU 0. Since the undamped system is stable
hypothesis assumed in the sensitivity eigenvalue analysis also
the work done by external forces is null W eU 0. The subscript
indicates that the energies are related to the undamped system.
2 Unstable slightly damped system (d i /dc0).
When slight damping is introduced, the system starts dissipating energy W dD 0. This could lead one to believe that, if the
undamped system is stable, all the more so, the slightly damped
system should be stable. This paradox can be explained by means
of the following formula presented and proved in the appendix
dW e
dW d
d i
i
dc
dc
dc
(38)
In fact, the work done by the external force W eD 0 for small

damping c is
W eD W eU
dW e
dW e
c
c
dc
dc
(39)
Replacing Eq. 38 into 39, it yields

W eD
dW d
d i
d i
i
c W dD i
c
dc
dc
dc
(40)
Eq. 40 states that, when the damped system results unstable

(d i /dc0) W eD W dD 0. Therefore, the external forces
introduce an amount of energy at each cycle that is not completely
dissipated. The net energy at each cycle increases the internal
energy causing the system to become unstable. Note that dissipation energy is always required in order to have a positive work
done by external forces. The physical explanation of the paradox
of damping provided in this section has general validity.
Conclusions
This paper deals with asymmetric linear systems. These systems can be found in many problems of interest in aeronautics and
control. In many cases, stability is studied by neglecting damping.
Unfortunately, even slight damping can destabilize the system. We
introduced a new criterion for foreseeing the destabilizing effect
of damping. The criterion asserts that damping does not destabilize the system if and only if a system of homogeneous linear
inequalities is satisfied. Moreover, we proved that the system of
inequalities can be replaced with a simpler inequality that gives a
sufficient condition. A simple example shows the theory step by
step. Eventually, a general physical explanation of the destabilizing effect of damping is given, using energy considerations.
Appendix
Description of the Modified Cone Algorithm. The cone algorithm finds a solution cone C(ws , s ), given an angle s . It
starts with an arbitrary axis w0 . The vector is then modified according to the following rule
i 90 s
if wi ,a
i
wi1 wi a
wi1 wi
(41)
i 90 s
if wi ,a
2M
Wan 10 proved that, if there exists a solution cone with angle

s , the algorithm will converge to the solution in a finite number
of steps. The algorithm can be repeated with different s . In fact,
one can start from s 0. Then, by increasing the angle s in fixed
increments, one can find a series of solution cones. Each time the
algorithm converges, the angle s is increased. The algorithm will
not converge when s overcomes the angle sc of the largest
solution cone C(wsc , sc ).
Energy Theorem. The energy introduced by the external
forces in one period of oscillation is
W e
2/i
xT Mxdt
2/i
xT Kxdt
2/i
ui e
iit
ie
u
i i t
dt
2/i
c0
2/i
dxT
Cxdt
dc
(42)
xT C
2/i
2/i
xT Uxdt
2/i
iT e i i t U ui e i i t u
i e i i t dt
uiT e i i t u
(43)
2/i
i2 MK
ii
i e i i t T i2 MK
ui e i i t u
4
i
dui i t du
e i
e i i t dt
dc
dc
2/i
2/i
i i t
i i i t i t dui
i
e i e
4
dc
i i i t i t T
dui i t
i u 2 MK
e i e
e i
i
i
4
dc
dt

dui
dc
i2 MK ui uiT
dui
dc
It is reminded that the derivative 44 is evaluated at c0. Conjugating Eq. 44 and considering the fact that d i /dcR, and
that the conjugate of the product of two complex numbers equals
the product of their conjugates, it yields
dui
dc
(46)
Premultiplying Eq. 3 by uiT and differentiating, one gets
iT
du
d i
d i
U i C
ui
M i2 C i K ui
dc
dc
dc
iT M i2 C i K
u
(44)
i uiT i2 MMT KKT
uiT 2M i
d i
dui
U i i ui i2 MK
0
dc
dc
i
i i dui i t du
e i
e i i t
4 dc
dc
i2 MK
dui
d i
U i ui i2 MK
0
dc i
dc
i2 i t i t T
i u Uu e i i t e i i t dt
e i e
i
i
4
d
M i2 cU i K ui 0
dc
2M
2/i
d
M i2 C i K ui 0
dc
xT Uxdt
In the last step we used the fact that ui ui R when c0 6.

Moreover, differentiating Eq. 3 with respect c and replacing i
i i
2M
i
d i
du
U i i ui i2 MK
0
dc
dc
i2 MK ui e i i t e i i t dt
dxT
dt
dc
i uiT Uui
2M
i2
i
du
d i
U i iui i2 MK
dc
dc
2/i
2M
i e i i t dt
ui e i i t u
The increment of the energy dissipated by the damping elements

for each cycle due to an increment of c is
dW d
dc
dui
d i
U i i ui i2 MK
dc
dc
i /dc
Comparing formula 45 and 44, one obtains du
dui /dc
The increment of the energy given by the external forces for
each cycle due to an increment of c is
i i T i t
iT e i i t i2 MK
ui e i u
4
2/i
2M
(45)
dW e
dc
xT MxKx dt
dui
d i
U i i ui i2 MK
dc
dc
dui
0
dc
(47)
Using the normalization 2uiT Mui 1 and the fact that C c0 0

one can obtain for the eigenvalue i i i
i iuiT Uui i i
iT
d i
dui du
iT i2 MK
u
i2 MK ui
dc
dc
dc
i i uiT Uui i i
d i
dui duiT
uiT i2 MK
dc
dc
dc
i2 MK ui 0
(48)
Since Ms MsT and Ks KsT

JULY 2003, Vol. 125 363
uiT i2 MK
dui duiT
i2 MK ui
dc
dc
uiT i2 MMT KKT
dW e
dW d
d i
i
dc
dc
dc
dui
dc
uiT i2 MMT KKT
dui
dc
(49)
Replacing Eq. 48 into Eq. 47 one obtains

i i uiT Uui i i
d i
dui
uiT i2 MMT KKT
dc
dc
0
Multiplying by i one obtains
i uiT Uui i
d i
iuiT i2 MMT
dc
KKT
dui
0
dc
(50)
Therefore, comparing Eqs. 46 and 43 with 49 the substitution

gives
dW d dW e
d i
i
dc
dc
dc
(51)
Since the derivative of the dissipative forces is always negative,

we conclude
364 Vol. 125, JULY 2003
(52)
References
1 Gasparetto, A., 1998, A System Theory Approach to Mode Coupling Chatter
in Machining, ASME J. Dyn. Syst., Meas., Control, 120, pp. 545547.
2 Inman, D. J., 1983, Dynamics of Asymmetric Nonconservative Systems,
ASME J. Appl. Mech., 50, pp. 199203.
3 Kounadis, A. N., 1992, On the Paradox of the Destabilizing Effect of Damping in Non-Conservative Systems, Int. J. Non-Linear Mech., 27, pp. 597
609.
4 Semler, C., Alighanbari, H., and Padoussis, M. P., 1998, A Physical Explanation of the Destabilizing Effect of Damping, ASME J. Appl. Mech., 65, pp.
642 648.
5 Bolotin, V. V., and Zhinzher, N. I., 1969, Effects of Damping on Stability of
Elastic Systems Subjected to Non-Conservative Forces, Int. J. Solids Struct.,
16, pp. 965989.
6 Herrmann, G., and Jong, I. C., 1966, On Nonconservative Stability Problems
of Elastic Systems with Slight Damping, ASME J. Appl. Mech., 32, pp.
125133.
7 Herrmann, G., and Jong, I. C., 1965, On the Destabilizing Effect of Damping
in Nonconservative Elastic Systems, ASME J. Appl. Mech., 32, pp. 592597.
8 Nemat-Nasser, S., Prasad, S. N., and Herrmann, G., 1966, Destabilizing Effect of Velocity-Dependent Forces in Nonconservative Continuous Systems,
AIAA J., 4, pp. 1276 1280.
9 Adhikari, S., and Friswell, M. I., 2001, Eigenderivative Analysis of Asymmetric Non-conservative Systems, Int. J. Numer. Methods Eng., 51, pp. 709
733.
10 Wan, J. S., 1994, Cone Algorithm: An Extension of the Perceptron Algorithm, IEEE Trans. Syst. Man Cybern., 24, pp. 15711576.
11 Murty K., 1976, Linear and Combinatorial Programming, Wiley, New York.
Marty E. Johnson
Assistant Professor,
Vibration and Acoustic Labs,
Mechanical Engineering,
Virginia Tech,
Blacksburg, VA 24061-0238
Luiz P. Nascimento
Sao Paulo State University,
Sao Paulo, Brazil
Mary Kasarda
Rotor Dynamics Group,
Virginia Tech,
Blacksburg, VA 24061-0238
Chris R. Fuller
Vibration and Acoustic Labs,
Virginia Tech,
Blacksburg, VA 24061
The Effect of Actuator and Sensor

Placement on the Active Control
of Rotor Unbalance
This paper investigates both theoretically and experimentally the effect of the location and
number of sensors and magnetic bearing actuators on both global and local vibration
reduction along a rotor using a feedforward control scheme. Theoretical approaches
developed for the active control of beams have been shown to be useful as simplified
models for the rotor scenario. This paper also introduces the time-domain LMS feedforward control strategy, used widely in the active control of sound and vibration, as an
alternative control methodology to the frequency-domain feedforward approaches commonly presented in the literature. Results are presented showing that for any case where
the same number of actuators and error sensors are used there can be frequencies at
which large increases in vibration away from the error sensors can occur. It is also shown
that using a larger number of error sensors than actuators results in better global reduction of vibration but decreased local reduction. Overall, the study demonstrated that an
analysis of actuator and sensor locations when feedforward control schemes are used is
necessary to ensure that harmful increased vibrations do not occur at frequencies away
from rotor-bearing natural frequencies or at points along the rotor not monitored by error
sensors. DOI: 10.1115/1.1569946
Introduction
The Active Magnetic Bearing AMB is a feedback mechanism

that supports a spinning shaft by levitating it in a magnetic field.
Patents associated with passive, active, and hybrid magnetic bearings go back over 150 years and there are many thousands of
commercial applications of AMBs in the field today 1. Compared to conventional rolling element and hydrodynamic bearings,
magnetic bearings have the capability for high surface speeds with
low power losses. The AMB also has the added capability for
active vibration control allowing for the reduction of rotor
vibrations.
Reduction in rotor vibration is sometimes obtained by the application of an open-loop or feedforward control strategy superimposed on the closed-loop control strategy necessary for
rotor support. Researchers have successfully implemented these
open loop control strategies with active magnetic bearings to
adaptively reduce synchronous rotor unbalance response and there
are considerable references in the literature regarding control
strategies for achieving this. Many of these control strategies are
frequency domain methods and sometime require the use of
look-up tables in order to track changes in rotor speed. Also,
there has been little or no discussion on the impact of actuator and
sensor location on the performance of these active control systems. In the work presented here the filtered X-LMS algorithm,
which is a time domain based adaptive feedforward approach used
widely in the active control of sound and vibration, will be tested
as an alternative control approach for reducing rotor vibration.
The X-LMS is a least mean squares approach where the reference
signal, typically denoted by x, is filtered before the LMS operation is performed
The goal of this paper is twofold: i to discuss the impact and
optimization of sensor and actuator location on the reduction of
synchronous vibrations in a high-speed laboratory rotor when active control strategies are used and ii to present the time domain
filter X-LMS algorithm as an alternative control approach for the
control of rotor vibration. Results of both an analytical and an
in the Journal of Vibration and Acoustics. Manuscript received August 2001; Revised
November 2002. Associate Editor. J. Wickert.
experimental investigation into the effect of sensor and actuator

location are presented. The effect of the number of sensors and
number of actuators, respectively, used for the Multiple Input
Multiple Output MIMO feedforward scheme is also addressed.
The motivation for the work is to examine the optimal set up for
actuators and error sensors on a rotor for achieving local or global
vibration reductions.
Review of Open Loop Control Strategies for Rotor Control.
A considerable amount of literature exists on control algorithms
used for reduction of rotor vibration. Knospe 2 discussed adaptive on-line balancing using digital control where algorithms in
the frequency domain were used for the feedforward control
scheme. Knospe 3 used look up tables and recommended the
convergence coefficient approach if a good off-line model of
the system is available because the computational burden is quite
small. In this study, the authors used three inputs where they
weighted the rotor midspan input sensor the sensor closest to
source of unbalance at 40% and weighted the other two sensors
at 30% each. Mid-span vibration levels were reduced but outboard
and inboard levels actually increased. Rutland 4 discussed the
transient response of a feedback H-infinity controller in terms of
controlled rotor response. The authors compared this controller
system with a synchronous feedback controller and concluded that
the performance of the second strategy was slightly inferior to the
H-infinity controller. Matsumura et al. 5 addressed the reduction
of vibration due to rotor unbalance using a gain scheduled
H-infinity robust controller. Hisatani et al. 6 presented a new
method for suppressing the unbalance vibration using two kinds
of adaptive filtering techniques. The first filter separated the rotor
displacement signal into rotation-synchronous and rotationasynchronous components. The synchronized component of the
rotor system, which is due to the rotor unbalance, can be cancelled
by the output of the second filter. The second filter was a disturbance estimator which used the output of the first filter together
with the impulse response of the plant measured on-line at highspeed rotation. Industrial applications also exist. Hope 7 discussed the use of two types of Active Vibration Control or
AVC on a 6-stage hydrogen process compressor operating at
20.6 Mpa 3000 psi. One of the AVC modes is an open-loop
JULY 2003, Vol. 125 365
rotating magnetic flux that is superimposed on top of the control

fluxes. The rotating AVC flux effectively creates a force that is
counter to the rotating unbalance force. The AVC flux is adaptive and can be applied based on the desired reduction of vibrations at a particular location, either along the machine shaft or on
a base plate to reduce transmitted forces. The second AVC
mode is a synchronous current reduction approach which makes
the rotor revolve about its inertial center, reducing the control
effort associated with unbalance to nearly zero. The filtered
X-LMS algorithm has been used successfully with magnetic bearings by Piper and Calvert 8 but in their application the control
system was used to actively control fluid borne noise from a centrifugal pump. They used a tachometer as a reference signal and a
downstream hydrophone to generate an error signal.
Review of Active Vibration Control in Beams. While there
has been considerable work involving control strategies for reducing unbalance responses in rotors, very little work has been presented addressing the effects of actuator and sensor location relative to the force source. There has also been little discussion
concerning the number of sensors required to effectively reduce
local and global vibration in a rotating machine. There is however
a substantial amount of published work on the active control of
bending vibration in finite and infinite beams that does address
these points 9,10. The vibration of a beam can be used as a
simplified model of a vibrating rotor and is useful for analyzing
the potential performance of control systems on rotors.
Fuller et al. 9 discussed both feedback and feedforward control strategies for controlling the vibration on beams. While Fuller
et al. did not specifically discuss actuator and sensor placement,
Nelson and Elliott 11 gave a general discussion about actuator
location in their book on the active control of sound. They showed
that control is always most successful when the control actuator is
placed close to the source of the disturbance. They also showed
that a set of discrete sensors can be used to approximate the total
vibrational energy sound in their case in a system with a larger
number of sensors leading to a better approximation. Fuller et al.
12 looked at simultaneous control of flexural and extensional
waves in a beam and considered the control of noise radiation
from beams 13. The most complete discussion of the active control of vibration in beams is given in a publication by Brennan
et al. 10 where the authors compared different wave control and
vibrational power minimization strategies for the control of vibration on both finite and infinite beams. They showed that in finite
beams, global control i.e. reduction achieved everywhere on the
beam is relatively easy to achieve at beam resonances or critical
speeds but difficult when away from resonance conditions. This
point has considerable relevance to the control of rotor vibrations
since rotors are normally run at speeds away from resonance conditions. The paper also showed that active control using a single
secondary actuator resulted in vibration reduction downstream of
the secondary actuator but is not very effective upstream of the
actuator often making the vibration larger. Post and Silcox 14
showed that the minimization of vibration over a section of a
beam can actually lead to significant increases in vibration away
from the error sensor. This has implications for error sensor
location.
Theory
In this section the filtered X-LMS algorithm is briefly described

and a simple, one dimensional model of a rotor is used to investigate the performance of an active control system. The model
developed in this section is a beam model i.e. does not include
gyroscopic effects and is intended to act as a tool through which
the actuator and sensor locations can be investigated and is not
intended to accurately predict the behavior of the test rotor. In the
following section a rotor finite element package that includes gyroscopic effects is used to more accurately model the behavior of
the test rotor and to validate the results of the simplified model
presented here.
366 Vol. 125, JULY 2003
Fig. 1 Control flow chart showing a filtered X-LMS adaptive

control system
Filtered X-LMS. This section will give a brief review of the

filtered X-LMS algorithm but the reader should refer to text books
such as that by Widrow and Sterns 15 or by Nelson and Elliott
16 for a more in depth description.
The filtered X-LMS is a time domain algorithm see the control
flow chart Fig. 1 that uses a reference signal x, such as a tachometer signal, to drive a set of secondary actuators in order to affect
the system under control. The reference signal is first digitally
sampled and then passed through a finite impulse response FIR
control filter W before being converted back into an analog signal
y and used to drive the control actuators. Another set of sensors,
called error sensors, are used to monitor the behavior of the system error e and are used to automatically adapt the control FIR
filter using the LMS algorithm. The plant represents the transfer function between the input to the actuators magnetic bearing
currents in this case and the vibration detected at the error sensors eddy current proximity probes in this case. The disturbance
d is the vibration at the error sensors due to the unbalance in the
system. This control architecture differs from the LMS in that the
reference signal needs to be first filtered by a model of the plant G
i.e. filtered X before being used by the LMS algorithm. The
model of the plant is usually stored as an FIR filter see below
and is measured in a system identification stage before the control
system is turned on.
If a signal, such as x, has been sampled at discrete time intervals then it can be considered as a sequence x(n) where n can
only take integer values. The current output of an FIR filter, such
a W, is the weighted sum of the previous inputs. For example the
current output from the control filter y(n) can be expressed as a
series of previous inputs of the reference signal y(n)
I1
i0
w i x(ni)where the control filter has I filter weights (w i ).
The LMS algorithm updates the filter coefficients in W using the
most recent error signal e(n) and the past I filtered reference
signals r(ni).
w i n1 w i n e n r ni
(1)
All I filter coefficients can be updated this way. The coefficient

is the convergence coefficient and determines how rapidly the
control system converges. must be large enough such that the
convergence time is small but cannot be too large since this can
cause instability. Ideally, this algorithm converges to a solution
where the time averaged sum of the squared error signals are
minimized. In principle, only 2 coefficients are necessary to
achieve good control if the disturbance is at a single frequency. If
multiple frequencies need to be controlled, for example harmonics
of the rotor speed, then more than two coefficients are necessary.
1-D Model of Beam for Active Control Analysis. In this
section a simplified model of the rotor vibration or velocity in
the vertical direction will be described. The rotor will be considered as a free beam or rod with gyroscopics neglected with
masses attached and supported by bearings modeled as a pair of
springs with damping included as shown in Fig. 2. Both unbalance forces and control forces have been introduced onto the
At a single frequency the velocities at a number of locations can

be described in matrix form as,
uiTijfj
(5)
where the column vector ui describes the velocities at positions xi

also a column vector due to a number of forces fj acting at
locations xj . Each element in the matrix Tij is calculated using
Eq. 4.
Fig. 2 Schematic of rotor system used for experiments and

beam used in the theoretical model. The bearing supports and
masses are attached at three locations along the beam. Sensors and actuator axes for the experimental system are also
shown.
beam to determine the optimal active control performance. The

velocity of the rotor in the horizontal direction will not be considered in this model and will be assumed to be independent of the
vertical velocity. Any velocity in the horizontal direction can be
controlled in the same way as the vertical velocity using another
uncoupled active control system producing similar results.
The velocity, u, of a beam or rod can be described as a series or
summation of modes 17.
u x,
a x
n0
(2)
where the nth mode shape is a function of position x and is given

by n (x) and the complex mode amplitude of the nth mode is
given by a n ( ). If the frequency range of interest is limited then
it is possible to accurately describe the beams behavior using a
finite set of modes. The mode shapes are given in Appendix A.
The mode amplitude a n ( ) is a function of angular frequency
and can be calculated by considering all of the forces acting on the
rotor. The mode amplitude due to a single force F( ) acting at a
single point x j on the rotor is given by,
j
2
a n
x j F
2
2
m r n 2 j n n
(3)
where the natural frequency and damping ratio for the nth mode
are given by n and n respectively. The damping ratio is typically low for a steel rotor, on the order of 0.005. Most of the
damping in this system will be provided by the bearings. The total
mass of the rotor is m r lS where S is the cross sectional area of
the rotor, is the density of the rotor material and l is the rotor
length. The natural frequencies of the modes can be calculated
using the equations presented in Appendix B.
By combining Eq. 2 and Eq. 3, point and transfer mobility
terms can be calculated. The transfer mobility T i j is defined as the
velocity of the beam at position x i due to an input force at position
x j . The point mobility is simply the velocity due to a force operating at the same position on the beam i.e. x i x j .
2
j
T i j
xi x j
2
2
m r n0 n 2 j n n
(4)
Matrix Impedance Method for Modeling Bearings. The

above theory and Eq. 5 can only be used to model the motion of
the rotor itself but does not include any bearing stiffness, damping
or mass that may be supporting or supported by the rotor. These
elements will be included in the model using a matrix impedance
method 18. In this paper this method will be used to describe the
addition of stiffness, damping and mass at three discrete locations
along the rotor but can be extended to include any number of
mass, stiffness and damping locations. As shown in Fig. 2 the
rotor interacts with external loads at three distinct points along the
beam, namely, at the two bearing positions where two small
masses m 1 and m 2 are attached and the rotor is supported by two
springs of stiffness k 1 and k 2 and in the middle of the beam where
a mass m 3 , containing a slight unbalance, is applied. The bearings
are also considered to have viscous dampers c 1 and c 2 included in
them. The mass and stiffness attachments create reaction forces
when the rotor is moved and these forces can be described using
an impedance matrix Z by,
f r1
frZum f r2
f r3
j m 1 c 1
k1
j
k2
j m 2 c 2
j
jm3

u m1
u m2
u m3
(6)
Where um is the vector of velocities at the three mass locations
and the reaction force is given by the vector fr . The velocity
vector um can be considered as the combination of the velocity
umr due to the reaction forces fr and the velocity umj due to external input forces to the rotor fj . The external input forces in this
model are created by the mass unbalance and by the active control
forces. Using Eq. 5 and Eq. 6 leads to an expression for the
reaction force in terms of the external forces,
umumjumrTmjfjTmmfrTmjfjTmmZum
um ITmmZ 1 Tmjfj
frZumZ ITmmZ
(7)
1
Tmjfj
(8)
The two matrices Tmm and Tmj contain rotor mobilities describing
the velocity at the three mass locations due to forces acting at the
mass locations and the locations of the external forces respectively. The elements in these matrices can be calculated using Eq.
4. Equations 7 and 8 can be used to calculate a new set of
ij that include the reaction force fr of
augmented rotor mobilities T
the masses, dampers and stiffnesses. From this augmented set of
equations the velocity at any point on the rotor due to a force
applied at any other point on the rotor can be calculated. We can
now re-write the mobility equation Eq. 5 to include reaction
forces.
ijfj
uiTijfjTimfrT
ij TijTimZ ITmmZ 1 Tmj
T
(9)
JULY 2003, Vol. 125 367
Table 1 Physical properties of rotor and bearings used for the simulation
Physical Properties
Rotor
Bearing 1
Legth
Diameter d
Modulus of
elasticity E
Density
Poissons
ratio
18 in or
457 mm
0.375 in or
9.52 mm
207 GNm2
3
7800 kgm
0.3
Bearing 2
0.25 kg
Mass
0.25 kg
Mass
0.8 kg
Stiffness
k1
Damper
c1
Position
x1
16000 Nm1
Stiffness
k2
Damper
c2
Position
x2
16000 Nm1
Position
xm
6.5 in or
165 mm
50 Nm
4.5 in or
114 mm
The two matrices Tim and Tij contain rotor mobilities describing
the velocity at the observation locations xi due to forces acting at
the mass locations and the locations of the external forces. Using
this augmented set of equations the performance of an active control system can be evaluated.
Active Control Performance. In this section a method for
calculating the control performance of an active control system
will be presented. An active control system using the filtered
X-LMS system will attempt to minimize the sum of the squared
outputs from the error sensors. If a control system has J control
actuators and L error sensors then at a single frequency the L
length vector of errors e can be written in terms of the vector of
primary disturbance forces fp i.e. rotor unbalance and the J
length vector of secondary control forces fc .
epfpTecfc
eT
(10)
The two matrices Tep and Tec contain augmented rotor mobilities
describing the velocity at the error locations xe due to forces acting at the locations xp of the unbalance forces primary disturbance and the locations xc of the control forces. The sum of the
squared error signals can be calculated using the Hermitian transpose or conjugate transpose denoted by H) as eHe and is minimized when the secondary control forces fc are given by 9,11,
fc
opt
H
1 T
H
T
ecTec
ecTepfp
(11)
Using these optimal forces the velocity at any set of observation

locations xi along the rotor can then be calculated before ui b and
after ui a optimal control as,
ui bTipfp
(12)
ui aTipfpTicfc
Unbalance mass
Mass
opt
50 Nm
14.5 in or 368
mm
can in principle perfectly control the vibration at a single position

on the rotor or beam but it may cause the vibration levels at
other locations to be greatly increased.
In this section the results for a single representative configuration
of mounts and unbalance mass will be presented. The physical
parameters used for these simulations are given in Table 1 and are
similar to the known values of the experimental system see section 4. The unbalance in the system will be generated at the mass
mounted on the rotor shaft between the active magnetic bearings
but closer to the inboard left bearing x6.5 in or 0.165 m. The
magnetic bearings act as both mounts and actuators and are located at x4.5 in or 0.114 m and at x14.5 in or 0.368 m. x
0 refers to the beginning of the rotor at the motor end see Fig.
2 and Fig. 3.
The results in this section Fig. 4 and Fig. 5 will be presented
for excitation frequencies between 0 and 200 Hz such that the
vibration level plotted at each frequency represents the vibration
level as if the rotor was run at that frequency and the excitation
was caused by a unit unbalance force. It is not intended to show
the control at all of these frequencies simultaneously since this is
not how the rotor would be excited in practice.
Mode Shapes and Natural Frequencies. The augmented
mobility of the rotor given in Eq. 9 can be considered to contain
augmented natural frequencies and augmented mode shapes. The
mode shapes and natural frequencies of the first three modes are
shown in Fig. 3. It can be seen that all three of these modes have
components that are due to bending in the rotor and components
due to bouncing or rocking on the two mounts.
Limitations Due to Actuator Location. In this section the
limitation due to actuator location will be investigated. To achieve
this it will be assumed that the control system is designed to
minimize the global vibration level. For the purposes of investi-
(13)
ic contain augmented rotor mobilities

The two matrices Tip and T
describing the velocity at the observation locations xi due to
forces acting at the locations xp of the unbalance forces primary
disturbance and the locations xc of the control forces. This process can be repeated for a range of frequencies.
Theoretical Analysis
The performance of a control system implemented on a rotor

will be limited by: i the control architecture, ii the actuator
placement and iii the sensor placement. In this section the limitations caused by actuator placement and sensor placement will be
investigated using the theory presented in section 2. The term
global vibration level will be used in this section and refers to
the mean velocity squared level averaged over the entire length of
the rotor at 20 equally spaced points. Although it is often more
important to reduce the vibration levels at particular positions
along the rotor, the global vibration level will provide some indication as to the overall performance of the system and highlight
potential problems. Since the excitation is tonal, a single actuator
368 Vol. 125, JULY 2003
Fig. 3 The mode shapes and natural frequencies of the first

three modes of the rotor predicted using a beam model.
Fig. 4 The global vibration level, when the global vibration

level is the cost function, using actuator 1, actuator 2 and then
both actuators.
gating the effect of actuator location alone it is assumed that a

perfect error sensor, that can directly measure the global vibration level, is used not realizable in practice.
In the first example the magnetic bearing to the left of the
unbalance mass is used as a control actuator and is driven such
that the vibration averaged over the entire shaft at twenty points
for this simulation is minimized. In this example the unbalance
mass is located closer to the left mount at x6.5 in or 0.165 m.
Figure 4 shows the global vibration level averaged over twenty
points over the entire shaft with and without thick solid line
control. Three cases are considered, i when actuator 1 close to
the unbalance mass is used dashed line, ii when actuator 2 is
used dotted line and iii when both actuators are used simultaneously thin solid line. The results are plotted for all frequencies
between 0 and 200 Hz and this represents the best control possible
at each of these frequencies i.e. limitations in performance are
due purely to the actuator location and unbalance location.
The first thing to be noted from these results is that the actuator
near to the unbalance mass is much more effective than the actuator farther from the unbalance mass. This demonstrates that it is
important to locate the actuator as close to the source of disturbance as is possible. It can also be seen that at frequencies away
from resonance or critical frequencies, for example at 80 Hz using
actuator 1, performance can be very poor. Since rotors are often
Fig. 5 The global vibration level when actuator 1 is used. The

two cases shown are with error sensor 1 and then with both
error sensors used as part of the cost function.
Fig. 6 Velocity along the rotor at 100 Hz with and without control using actuator 1 and a single error sensor e1 , actuator 1
and both error sensors and then both actuators and both error
sensors.
deliberately run at speeds away from resonance there are likely to

be control problems using single actuator systems. These difficulties can be overcome if two actuators are used. Near resonance
conditions reductions are possible using a single actuator but the
location is very important. For example near the first resonant
mode, actuator 2 has very poor performance.
The performance of an active control system is often dependent
on the number of actuators vs. the number of degrees of freedom
that are important in the system under control or complexity of
the system. This complexity can be roughly approximated by the
modal overlap factor 11 that describes the number of modes
acting within a given bandwidth. For beams this term does not
increase very rapidly with frequency as compared to two and
three dimensional structures and even while coupled to the
mounts the rotor seems to maintain this modal simplicity i.e. the
modal overlap factor remains low. It is therefore likely that two
well positioned actuators will be sufficient to achieve good vibration control over a fairly large frequency band.
Using Discrete Error Sensors. In any realistic control system only a limited number of sensors can be used and it is therefore important to investigate the use of discrete error sensors on
the performance of the control system. For this study, sensor locations were determined based on accessibility to the shaft as
dictated by physical constraints. Figure 5 shows the global vibration level when the vibration is cancelled at a single point (x
11 in or 0.279 m downstream of the unbalance mass using a
single upstream actuator dashed line. Unless the system is excited close to a resonance frequency the global vibration level can
be substantially increased after control. For example the vibration
at 120 Hz increases globally by 6 dB. If a second error sensor is
introduced at the end of the rotor (x18 in or 0.457 m then the
control system minimizes the average squared vibration level at
the two error sensor locations dotted line and this helps to condition the control system and reduces the likelihood of increased
vibration. The introduction of the second error sensor greatly improves the performance of the system, and global performance is
similar to the case with the ideal global error sensor see Fig. 4.
As will be shown later this type of system can also be used to
reduce the vibration over an area on the rotor.
Figure 6 shows the vibration along the rotor with and without
control at 100 Hz using both a single error sensor and then both
sensors. Since this vibration has a complex value the absolute
value of the velocity is plotted. Before control heavy solid line
the vibration is dominated by the second mode that is a rocking
JULY 2003, Vol. 125 369
Rotor Dynamic Analysis Using a Transfer-Matrix Analysis.

In order to more accurately predict the experimental rotors dynamic behavior, the rotor-bearing system was modeled by using a
transfer-matrix method based rotor analysis code 19. Figure 2
shows the schematic of the experimental rotor. In the model the
shaft was considered to be a flexible beam divided into elements
having distributed mass and elasticity. The unbalance mass or
disk was modeled as a lumped mass, and its properties were
added to the element where it is attached. The equivalent stiffness
and damping of the magnetic bearings is dependent on the
closed loop or feedback control system PID controller used to
position the rotor in the magnetic field. These stiffness and damping values can be calculated as a function of the overall system
transfer function G(i ), and the position stiffness K x and the
current stiffness K i of the bearings. The overall controller transfer
function expresses the relationship between the output control current and input shaft position, and can be written as,
G i a G ib G
(14)
This transfer function is composed of components from the position sensor, low pass filter, PID filter and power amplifier. The
equivalent stiffness and damping are then given by,
Fig. 7 The mode shapes and natural frequencies for the first
three modes of the rotor for the three different experimental
mass configurations.
type mode note that the absolute value is plotted. If actuator 1 is

used to cancel the vibration at a single error sensor 1 the overall
vibration level is not significantly controlled and vibration problems can still occur at positions away from the error sensor
dashed line. If two downstream error sensors are used then global vibration reduction is still not large dotted line but the downstream vibration section, where the error sensors are located, has
the lowest vibration levels. This implies that the use of multiple
sensors can allow vibration reduction over a region even if global
control is not possible. When both actuators are used to cancel the
vibration at the pair of downstream error sensors light solid line
very effective isolation of this section of the rotor is achieved. It
has been shown 10 that for infinite beams using wave theory
all downstream wave propagation can be removed if the velocity
and gradient of velocity in the x-direction are simultaneously reduced to zero at one point along the beam. These results imply
that even in circumstances where global control is difficult to
achieve, a pair of actuators can be effective in isolating sections of
a rotor.
370 Vol. 125, JULY 2003
K eq K x K i a G ;
(15)
C eq K i b G /
(16)
The position stiffness and current stiffness are given as a function

of the physical characteristics of the magnetic bearing 20. Since
the transfer function of the control circuit has a frequency dependence, the equivalent stiffness and damping is also given as a
function of the frequency, so that the average values of these
parameters can be calculated in the frequency band of interest,
using the procedure presented by Clements 21.
Finding the modal characteristics of the system provides useful
information that can help optimize the active control configuration. There is great interest in being able to control the vibration of
the rotor near critical speeds or resonances and the vibration
mode shapes may also indicate the best location for placement of
the errors sensors. Figure 7 shows the first three vibration mode
shapes of the rotor with the unbalance mass placed at the three
experimental locations along the length of the rotor: at midspan,
close to the inboard bearing, and close to the outboard bearing.
Also these simulations note: results presented in Hz show that
considerable changes in the natural frequencies and vibration
mode shapes of the rotor occur as a function of the unbalance
mass location. These simulation results validate the theory presented in section 2 since there is a very close match in natural
frequencies and mode shapes with the Transfer-Matrix model
note: results presented in rpm.
Fig. 8 Picture of experimental rig used in this study. The two

magnetic bearings and unbalance mass are shown.
Table 2 Different Error Sensors and Actuators Locations for Unbalance Control
Cases
Error Sensors
Actuators
Unbalance Mass Position
S 1 and S 2
V 1 and W 1
S 1 and S 2
V 2 and W 2
S 1 , S 2 , S 3 and S 4
V 1 and W 1
S 1 , S 2 , S 3 and S 4
V 1 , W 1 , V 2 and W 2
S 1 and S 2
V 1 and W 1
S 1 and S 2
V 2 and W 2
S 1 and S 2
V 1 and W 1
S 1 , S 2 and SV 1
V 1 and W 1
S 1 , S 2 and SV 1
V 2 and W 2
Experimental Results
Experimental Set-Up. Figure 8 shows a picture of the small

high-speed test rotor used in this study. The rotor consists of a 457
mm 18 in. long shaft that is 9.52 mm 0.375 in. in diameter and
is driven by an electric motor through a flexible coupling. The
rotor is capable of operating at rotating speeds as high as 10,000
rpm and its rotor dynamic characteristics mimic that of large-scale
turbomachinery. A steel disk with holes for the placement of unbalance weights has been attached to the shaft at various locations
to vary the amount and location of rotor unbalance forces. The
disk is 76.2 mm 3 in in diameter, 25.4 mm 1 in thick and
weighs 811 gr. 1.78 lbs. In addition to the large unbalance disk
there are two additional masses 250 gr. which are the bearing
ferromagnetic rotors that are also attached to the shaft. The rotor
is supported in two 8-pole heterepolar Active Magnetic Bearings
AMBs with a 12 lbs force capacity. Each of the AMBs consists
of a radial rotor, a stator, a position sensor ring and a backup
bearing. The radial rotor and the stator are used to levitate the
rotor while the sensor ring measures radial position of the shaft
for feedback. The backup bearing provides support to the rotor
when the bearing is turned off or in case of a power loss. Each
Radial AMB has two axes of control, each utilizing a PID control
algorithm to maintain support of the shaft. The control axes are
oriented at 45 degrees to the horizontal denoted V and W as
shown in Fig. 2.
The sensors available for use as error sensors are the four eddy
current sensors located at two points of the shaft listed as S 1 , S 2 ,
S 3 and S 4 as well as the displacement sensors already used for
feedback control of each of the support bearings shown as V 1 ,
W 1 , V 2 , and W 2 in Fig. 2. All sensors, even if they are not used
as error sensors, are used for rotor displacement monitoring purposes. A single eddy current sensor mounted at the drive end of
the shaft is used as a key phase reference signal to drive the
control system denoted K p ). The positions of the sensors were to
some extent limited by physical constraints.
The multi-channel filtered X-LMS active control algorithm was
implemented on a quad TMS C40 board made by Spectrum Signal, running five C40 DSP chips in parallel. This system was
developed at Virginia Tech and can accommodate a wide variety
of control configurations i.e. various numbers of reference signals,
actuators and errors signals.
Experimental Results of Unbalance Control. In this work
nine different control configurations were used where different
numbers and locations of sensors and actuators were used with
different locations of the unbalanced disk. Both inboard and outboard magnetic bearings were used as actuators. Each bearing had
two uncoupled active perpendicular axes and were considered as
two independent actuators. A summary of the different control

configurations used for the nine cases is shown in Table 2. It
should be noted that using two actuators i.e. one magnetic bearing acting in one plane and two error sensors also in one plane
in the experimental case would be similar to using a single sensor
actuator pair in the theoretical analysis presented in section 3 because of the added degree of freedom in the real rotor i.e., both
horizontal and vertical motion is permitted.
Prior to control, the model of the plant G was measured by
injecting white noise into the actuators and measuring the output
of the error sensors. This allowed the system to operate, if necessary, over a range of frequencies. The experimental tests for unbalance control were carried out at the rotating speed of 3420 rpm
or 57 Hz, which is close to the first critical speed of the rotor, as
calculated in the rotor dynamic analysis. Figure 9 shows the autospectrum of the displacement measured at the four sensor locations with and without control for Case 1. As shown in Fig. 9,
large reductions in unbalance vibration at the error sensor positions is achieved. The 48 dB in attenuation at these positions
shows that the filtered X-LMS system is extremely effective a
250 fold decrease in vibration level. To analyze the effect of
control, the resulting vibration reduction for all 9 cases at all
measurement positions is summarized in Table 3. For Cases 1 6
all eight sensor signals were recorded and for cases 79 three
sensor signals were recorded. Also presented in Table 3 is an
estimate of the global reductions in vibration that have been
calculated by considering the sum of the squared outputs from all
eight proximity probes.
Included in Table 3 is a list of the control performance predicted using the computational model presented in section 2 and 3
for Cases 1 to 6. A number of control trends match those of the
theoretical model.
i. If a pair of error sensors is positioned at a single location
along the rotor then very good local control can be achieved. This
is clearly illustrated by cases 1, 2, 5 and 7 where typically 40 dB
reduction in vibration level was measured.
ii. Good local control does not correspond to global reductions. This is well illustrated by case 6 and especially case 7 where
a large reduction at the error sensor location is achieved while
other sections of the rotor undergo an increase in vibration i.e.
shown as negative reduction in Table 3.
iii. Actuators positioned close to the disturbance achieve better
control. This is clearly illustrated by comparing the results of case
5 to case 6. In case 5, where the actuator closest to the unbalance
mass is used, large reductions are achieved across the rotor 15.1
JULY 2003, Vol. 125 371
Conclusions
A theoretical and experimental investigation into the effect of
the location and number of sensors and actuators on both global
and local vibration reduction along a rotor using a feedforward
control scheme with magnetic bearing actuators has been completed. Theoretical approaches developed for the active control of
beams have been shown to be useful as simplified models for the
rotor scenario for optimization studies involving sensor and actuator location and number. Also presented is the time-domain filtered X LMS feedforward control strategy, used widely in the
active control of sound and vibration, as an alternative control
methodology to the frequency-domain open loop approaches commonly presented in the literature. This control scheme performed
extremely well with reductions in vibration amplitudes as high as
51.7 dB a 380 fold decrease.
The rotor under analysis consists of a small 9.52 mm diameter
high-speed laboratory rotor supported in active magnetic bearings
with one large disk that can be placed in different locations. Both
experimental results and simulation results are presented in Table
3. In general,
Fig. 9 The measured vibration level at the four measurement

sensors S1, S2, S3 and S4 for case 1. Huge reductions in the
vibration level at the error sensors S1 and S2 were achieved
48 dB or 250 fold reduction.
dB global whereas in case 6 the reductions are much lower only

3.5 dB global and increases in vibration were measured at one
point along the rotor.
iv. Using a larger number of error sensors than actuators results
in better global reduction of vibration but worse local reduction.
This is illustrated by comparing the results from Case 1 and Case
3. Local reduction at the error sensor drops from 48 dB to 28 dB
but the global attenuation increases from 16.8 dB to 17.8 dB implying a more even reduction in vibration along the rotor.
v. Using a larger number of actuators improves performance:
This is illustrated comparing Case 4 to Case 3 where improved
vibration reductions were achieved 19.1 dB as compared to 17.8
dB. However, it should be cautioned that in any case where the
same number of actuators as error sensors are used square system there may be frequencies at which large increases in vibration away from the error sensors can occur. In addition, note that
there is not a dramatic improvement between Case 3 and Case 4
indicating that a configuration with more sensors and a reduced
number of actuators may be advantageous from a reduced hardware or reduced control burden standpoint.
1. A reduced model simulation method was demonstrated to be

successful at predicting overall behavior trends on a rotor when
examining the impact of sensor and actuator configuration
schemes.
2. Very good reduction of local vibrations can be achieved at
the error sensor when a single error sensor is used over 50 dB
but not without risk of increased vibration at other locations and
hence poor global attenuation.
3. Using a larger number of error sensors than actuators results
in better global reduction of vibration but decreased local reduction. For example, by increasing the number of error sensors from
2 to 4 cases 1 and 3 the global attenuation increased from 16.8
dB to 17.8 dB. Although some local reductions were not as large
the lower global attenuation reflects the fact that the control is
more evenly spread along the rotor.
4. For any case where the same number of actuators as error
sensors are used there can be frequencies at which large increases
in vibration away from the error sensors can occur.
5. There may be advantages for hardware reduction and/or reduced control burden to go with a scheme of more error sensors
and reduced actuators with little reduction in overall performance
at frequencies near resonance.
Overall, the study demonstrated that an analysis of actuator and
sensor location when active control schemes are used to reduce
rotor vibration is necessary at the design phase to insure that
harmful increased vibrations do not occur at points along the rotor
not monitored by error sensors or unexpectedly at frequencies that
do not correspond to rotor-bearing natural frequencies.
Acknowledgments
The authors would like to acknowledge the National Science
Foundation NSF for supporting the work presented in this paper
Table 3 Unbalance Frequency Reduction dB-3420 RPM 57 Hz

Simulation results: Attenuation
dB
Experimental results: Attenuation dB

Cases
SV 1
SW 1
S1
S2
S3
S4
SV 2
SW 2
Global
SV 1 , SW 1
S1 , S2
S3 , S4
SV 1 , W 1
1
2
3
4
5
6
7
8
9
13.2
3.5
10.4
15.2
8.7
3.5
3.8
2.2
25.6
15.7
11.7
14.5
19.8
14.1
6.5
48.8
49.4
23.2
34.8
41.6
8.1
41.2
14.1
30.6
48.1
51.7
28.5
30.5
50.8
16.4
41.7
11.4
29.6
17.8
11.3
20
29.2
18.9
6.2
22
15.8
26.6
29.3
28.1
4.4
6.8
3.4
8.9
7.8
11.2
9.3
14.3
3.4
15.1
9.4
20.3
8.2
16.8
9.8
17.8
19.1
15.1
3.4
7.4
16.1
13.0
16.7
18.0
4.5
22.1
16.8
16.1
23.3
19.9
3.0
11.1
4.8
6.8
17.6
7.2
20.8
372 Vol. 125, JULY 2003
under Grant No. DMI-9733124. The authors would also like to

thank Dr. Gordon Kirk and the Virginia Tech Rotordynamics
Laboratory for the use of equipment and rotor analysis codes.
Besides, the authors would like to acknowledge the FAPESP
o Paulo, Brazil,
Fundacao de Amparo a Pesquisa do Estado de Sa
which granted funds to Luiz P. Nascimento for a Visiting Scholar
Program at Virginia Tech.
where k n is the wavenumber of the n th mode, E is the modulus of

elasticity, I is the second moment of area and d is the rotor diameter. Small correction factors can be added to equation 23, for the
lower order modes, to improve the accuracy of the calculation
17.
Appendix A: Mode Shapes for a Free-Free Beam
References
The first two modes mode of a free-free beam are rigid body
modes with the higher modes given as combinations of sine, cosine and hyperbolic functions 17.
0 x 1
1 x 1
n x cos
(17)
2x
(18)
x
x
0.5 k cosh 0.5
for n2,4,6,
(19)
where
k
and tan tanh 0
2
2
sinh
2
sin
n x sin
x
x
0.5 k sinh 0.5
for n3,5,7,
(20)
where
k
and tan tanh 0
2
2
sinh
2
sin
Appendix B: Natural Frequencies of the Modes of a

Free-Free Beam
For a free-free rod the natural frequencies of the modes are
given by 17,
n k 2n
0 0
(21)
1 0
(22)
EI n0.5 2 2 d
S
4 2
for
n2
(23)
1 Kasarda, M. E. F., 2000, An Overview of Active Magnetic Bearing Technology and Applications, invited review paper, Public Admin. Rev., 322, pp.
9199.
2 Knospe, et al., 1993, Adaptive On-Line Rotor Balancing Using Digital Control, Proceedings of 93 ROMAG, pp. 153164.
3 Knospe et al., 1998, Synthesis Techniques for Robust Adaptive Vibration
Control, 6th International Symposium on Magnetic Bearings, Cambridge,
Mass, p. 183.
4 Rutland et al., 1994, Comparison of Controller Designs for Attenuation of
vibration in a Rotor-Bearing System Under synchronous and Transient conditions, 4th International Symposium on Magnetic Bearings, Zurich, Switzerland, pp. 107112.
5 Matsumura et al., 1994, Elimination of Unbalance Vibration in AMB Systems Using Gain Scheduled H Robust Controllers, 4th International Symposium on Magnetic Bearings, Zurich, Switzerland, pp. 113118.
6 Hisatani et al., 1997, Adaptive Filtering for Unbalance Vibration Suppresion, Proceedings of MAG97, Alexandria, VA, pp. 125130.
7 Hope, R. W., Tessier, L. P., Knospe, C., and Miyaji, T., 1998, Adaptive
Vibration Control of Industrial Turbomachinery, 98-GT-405, International
Gas Turbine & Aeroengine Congress & Exhibition.
8 Piper, G. E., and Calvert, T. E., 1995, Active Fluidborne Noise Control of a
Magnetic Bearing Pump, NCA-Vol. 21, IMECE, Proceedings of the ASME
Noise Control and Acoustics Division, pp. 5576.
9 Fuller, C. R., Elliott, S. J., and Nelson, P. A., 1996, Active Control of Vibration,
Academic Press, London.
10 Brennan, M. J., Elliott, S. J., and Pinnington, R. J., 1995, Strategies for the
Active Control of Flexural Vibration on a Beam, J. Sound Vib., 1864, pp.
657 688.
11 Nelson, P. A., and Elliott, S. J., 1993, Active Control of Sound, Academic
Press.
12 Fuller, C. R., Gibbs, G. P., and Silcox, R. J., 1990, Simultaneous Active
Control of Flexural and Extensional Waves in Beams, J. Intell. Mater. Syst.
Struct., 1, pp. 235247.
13 Guigou, C., and Fuller, C. R., 1993, Active Control of Sound Radiation from
a Simply Supported Beam: Influence of Bending Near-field Waves, J. Acoust.
Soc. Am., 935, pp. 2716 2725.
14 Post, J. T., and Silcox, R. J., 1990, Active Control of the Forced Response of
a Finite Beam, Proc. Of Noise-Con90, Austin, Texas, pp. 197202.
15 Widrow, B., and Sterns, S. D., 1985, Adaptive Signal Processing, Prentice
Hall, Englewood Cliffs, New Jersey.
16 Elliott, S. J., Stothers, I. M., and Nelson, P. A., 1987, A Multiple Error LMS
Algorithm and Its Application to the Active Control of Sound and Vibration,
IEEE Trans. Acoust., Speech, Signal Process., 35, pp. 14231434.
17 Warburton, G. B., 1954, The Vibration of Rectangular Plates, Proc. Inst.
Mech. Eng., 168, p. 371383.
18 Bishop, R. E. D., and Johnson, D. C., 1960, The Mechanisms of Vibration,
Cambridge University Press.
19 Childs, D., 1993, Turbomachinery Rotordynamics Phenomena, Modeling, and
Analysis, John Wiley & Sons, Inc., New York.
20 Allaire, P. E., et al., 1993, Magnetic Bearings, STLE Handbook of Tribology
and Lubrication, Vol. III.
21 Clements, J. R., 2000, The Experimental Testing of an Active Magnetic
Bearing/Rotor System Undergoing Base Excitation, Master Thesis, Virginia
Tech.
JULY 2003, Vol. 125 373
Hemanshu R. Pota
e-mail: h-pota@adfa.edu.au
Ian R. Petersen
School of Electrical Engineering,
UNSW at the Australian Defense Force Academy,
Canberra ACT, 2600, Australia
Atul G. Kelkar
Iowa State University,
Ames, IA 50011
e-mail: akelkar@isu.edu
Robust Control of a 2D Acoustic

Enclosure*
This paper reports experimental results in the application of feedback control of acoustic
noise in a 2D enclosure. It is shown that the feedback control of 2D enclosures poses
peculiar problems compared to simple 1D ducts. The chief among them are the need for
explicit budgeting of uncertainties and dealing with high model orders. This paper demonstrates a practical way to design controllers for such systems using minimax LQG
methods. It is shown that an important step in the controller design is the proper choice
of a weighting function. The results presented in this paper are impressive and they can be
further improved by proper choice of actuator and sensor placements.
DOI: 10.1115/1.1570447
Introduction
Active reduction or control of acoustic noise has significant

practical applications 1. Successful active noise control schemes
mostly use adaptive feedforward control 13. Feedforward control is ideal in the situations where a signal strongly correlated
with the noise can be directly measured. There are many practical
applications, e.g., structure induced vibrations, where its difficult
to obtain a signal correlated with acoustic noise which can be used
to effect a feedforward control scheme. In these situations it is
fruitful to apply feedback control for active noise reduction.
The essentials of modeling and control of a 1D duct have been
discussed in 4,5. Experimental results using feedback have been
obtained for 1D ducts; see 6,7 for results without robust control
approach while the work in 5,8 10 uses robust control theory to
design controllers. It is well-known 11 that for feedforward control, perfect model matching is essential to provide satisfactory
noise cancellation. In feedback control, a mismatch in the model
can lead to an unstable system. This is the main reason for the
reluctance to use feedback control for acoustic noise reduction.
This means that the proper use of robust control theory is essential
to obtain results which have practical applications. For example,
the significant improvement of feedback controller performance in
9 over the other reported feedback controllers is due to its explicit accounting of modelling uncertainties.
All the results in the literature using feedback control 57,9
are reported for 1D ducts only. In principle, control of 1D ducts is
no different from 2D enclosures. But in practice there is much
difference. Firstly its a lot easier to get an excellent match between identified models and experimental data for 1D ducts. Secondly the model order of 1D ducts is significantly lower. In this
paper the practical aspects of noise reduction in 2D enclosures are
highlighted based on experimental results. The minimax LQG
feedback control 12,13 is used to design practical controllers for
a 2D acoustic enclosure at the Iowa State University ISU shown
in Fig. 1.
Experimental Setup and Modeling
2.1 Experimental Setup. A picture of the experimental

acoustic enclosure to be considered in this paper is shown in Fig.
1. A schematic of the enclosure is shown in Fig. 2.
The enclosure is constructed from aluminum sheets. One side
of the enclosure with the largest cross-section has two doors on
*This work was supported by the Australian Research Council H. R. Pota & I. R.
Petersen, the NSF through Grant No. CMS:9713846, and NASA through Grant No.
NAG-1-01039 A. G. Kelkar
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received December
2001; Revised February 2002. Associate Editor: R. F. Keltie.
374 Vol. 125, JULY 2003
hinges which provide access to the inside of the enclosure to

position the microphones. The experiment is so set that there are
two speakers and one microphone. One speaker is used as a disturbance and the other speaker is the control speaker. There is one
microphone to sense the acoustic disturbance and provide feedback to the controller. The feedback control system is a SISO
system with the disturbance coming from an independent speaker
located on one end of the enclosure.
The 2D enclosure is 1850 mm320 mm1240 mm. Let the
origin 0, 0, 0 be at the bottom left corner when one is facing the
door side of the enclosure; z-axis is up, y-axis is into the enclosure, and the x-axis is along the length of the enclosure as seen in
Fig. 2. The disturbance speaker is located at one end of the enclosure with the center at 0, 127, 406 mm; the control speaker is
located at 600, 320, 430 mm; and the microphone is located at
588, 200, 420 mm.
2.2 System Identification and Nominal Modeling. Experimentally identified model is used in this paper to design the controller. The difficulty in obtaining an analytical model for enclosures with high modal densities is well-known 14. The use of
finite element and other numerical methods is an active area of
research 15,16. From the results in 15,16 it can be clearly seen
that the accuracy of these models have a long way to go before
they can be used for feedback controller design. The error is so
large that robust controllers designed to maintain closed-loop stability will be ineffective. Most analytical and numerical methods
predict the modes quite accurately. The main difficulty in modelling is the uncertainty in the damping terms which makes accurate
modelling for high modal density enclosures very difficult. In 1D
enclosures like acoustic ducts or enclosures with low modal density, analytical models can be used to design very effective controllers 4. This is the reason why many experimental setups to
demonstrate active control for 3D enclosures choose small enclosures 17, this keeps the modal density low hence the modes well
apart. The problem of obtaining accurate acoustical models with
unknown dampings and high modal densities is a very difficult
one which is currently an active research area.
Significant advance has been made in active cancellation of
acoustic noise using feedforward techniques 18,19. These methods need an accurate estimate of the error path transfer function.
This is the path from the secondary control sources to the error
microphones or the locations where the acoustic noise has to be
reduced. There are two major differences between the use of the
error path transfer function and the acoustical model for feedback
control design. Firstly, most practical implementations of feedforward noise cancellations work best for tonal noise and this means
that the model needs to be accurately known only at a few particular frequencies. Most of the digital filters implementing these
models are adaptively tuned too. The second difference is that the
Fig. 1 The experimental 2D enclosure
order of these error path transfer functions can be chosen to be

very high. Its common to have an order of 256 and up. It must be
realised that the design of feedback controllers with such highorder models is nearly impossible. For these practical reasons experimentally identified models are used in this paper.
To identify the system model, two separate frequency responses
are recorded using Stanford Researchs SR785 spectrum analyzer.
One response is taken from the disturbance speaker to the microphone and the other from the control speaker to the same microphone. The subspace identification technique 20 is then used
to get a state-space representation of this two-input-one-output
system.
After the identification process, we end up with a nominal
transfer function matrix of the form
P s P 1 s P 2 s
where P 1 (s) represents the transfer function from the disturbance
speaker input to the microphone output and P 2 (s) represents the
transfer function from the control speaker input to the microphone
output.
2.3 Uncertainty Modeling. The system representation
which forms the starting point in the minimax LQG controller
design is shown in Fig. 3. In the figure y is the microphone output,
u is the control speaker input, z is the uncertainty output, w
rep-
Fig. 3 Uncertain system representation
resents the noise input and it is assumed that the effects of the
uncertainty show up at the sensor output through the disturbance
channel. The minimax LQG method requires that the inputs due to
uncertainty i.e., the output of the (s) block and the noise enter
through the same channel. Figure 3 shows one particular uncertainty model but the minimax LQG method is not restricted to this
model alone and alternative representations are possible.
The block (s) in Fig. 3 is chosen such that
j 1 ,
P 2 s P 2 s / P 1 s s W s .
To restrict ( j ) 1 , the weighting function needs to be

chosen such that
Fig. 2 Schematic of the experimental 2D enclosure
(1)
and W(s) is a stable frequency weighting transfer function. The

controller design section will discuss the importance of choosing
this weighting function properly.
Let the true transfer function from the control speaker to the
microphone be given by P 2 (s). From the block diagram in Fig. 3
it can be seen that
P 2 j P 2 j
W j .
P 1 j
(2)
The bound 2 is an inequality bound on the magnitude of W(s)

and there are several functions which will satisfy this bound. In
this paper the function on the left-hand-side of the bound 2 is
computed in the frequency range of interest from the experimental
measurements and the identified system models. From these functions a magnitude envelope is constructed and finally that magnitude envelope is matched by a transfer function obtained using the
Yule-Walker method 21,22.
The weighting function W(s) in the experiments reported in
this paper is mainly used to account for the modelling uncertainties. In general it can include parameter uncertainties too. In the
case of acoustical system this would mean the change in temperature and the enclosure geometry. From the experimental and identified models in Figs. 7 and 8 it can be seen that the modelling
uncertainty is much larger than the uncertainty due to parameter
variations. As a matter of fact these experiments were run over
several days. It is common for the microphones to be slightly
JULY 2003, Vol. 125 375
shifted during the experiments. The speakers and the enclosure

geometry was fixed though. The experiments were conducted under realistic conditions giving confidence in the robustness of the
controller to uncertainties.
Minimax LQG Control
This section presents a brief description of the minimax LQG

robust controller synthesis method. A more complete and rigorous
description of this method can be found in the references 12,13.
The minimax LQG method is applied to uncertain systems of the
form shown in Fig. 4. In this figure, the nominal system is described by the following stochastic state equations:
x AxB 1 uB 2 B 2 w
,
Fig. 5 The scaled H control problem
zC 1 xD 1 u,
,
yC 2 xD 2 D 2 w
y 0 0,
(3)
m
In the above equations, x(t)R is the state, u(t)R is the

control input, w
(t) is a unity covariance white noise input, z(t)
Rq is the uncertainty output, (t)Rp is the uncertainty input
and y(t)Rl is the measured output.
The uncertainty block can be any dynamical system satisfying a
general uncertainty constraint; see 12,13. In particular, this uncertainty constraint is satisfied by the uncertainty block (s) in
equation 1.
It is assumed that the cost function under consideration is of the
form E. is the expected value
1
E
T 2T
J lim
x t Rx t u t Gu t dt,
It is well known that the solution to this H control problem can

be obtained in terms of the following pair algebraic Riccati equations e.g., see 23:
AB 2 D 2 D 2 D 2 1 C 2 Y Y AB 2 D 2 D 2 D 2 1 C 2
Y C 2 D 2 D 2 1 C 2
B 2 ID 2 D 2 D 2 1 D 2 B 2 0
1
1
X AB 1 G 1
AB 1 G X R G
R 1/2x
,
G 1/2u
(5)
then the minimax LQG control problem can be solved by solving

the scaled H control problem represented in Fig. 5; see 12,13.
In this H control problem, the nominal system is described by
Eq. 3 and 5 and the controller is to be constructed such that the
closed loop system is stable and the transfer function from w
(t) to
(t) satisfies the H norm bound
T w j 1 .
(6)
and
(4)
where R0 and G0. The minimax LQG control problem involves finding a controller which minimizes the maximum of this
cost function where the maximum is taken over all uncertainties
satisfying the uncertainty constraint 1. If we define a variable
1
R Y

X B 1 G 1
B 1
1
B B X 0,
2 2
(7)
where the solutions are required to satisfy the conditions Y 0,

X 0, I1/ Y X 0 and R G 1
0. Here R ,R
C 1 C 1 , G ,G D 1 D 1 and , C 1 D 1 . In order to solve
the minimax LQG control problem, the parameter 0 is chosen
to minimize the cost bound W an upper-bound on J in 4
defined by
Y C T2 B 2 D T2 D 2 D T2 1
T
W ,tr C 2 Y D 2 B 2 X IY X 1 .
Y R
Then, the minimax LQG controller is defined by the state

equations
x AB 1 G 1
B 1 G 1
x
B 1
1
Y X

y C 2
376 Vol. 125, JULY 2003
1
B B X x
2 2
Y C 2 B 2 D 2 D 2 D 2 1
1
D B X x
2 2
u G 1
.
B 1 X x
Fig. 4 Stochastic uncertain system
(8)
(9)
Controller Design
The appropriate state-space representation in Eq. 3 are arrived

at from P 1 (s), P 2 (s), and W(s) as discussed previously. Note
that the theory of 12,13 requires that D 2 D 2 0 in 3. This is
achieved by adding a small measurement noise to the system in
addition to the process noise w
(t). We choose the matrix R in the
cost function 4 as RC 2 C 2 . That is, the term x(t) Rx(t) in the
cost function 4 corresponds to the norm squared value of the
nominal system output. The term u Gu in the cost function 4 is
treated as a design parameter affecting controller gain. However,

in all cases it was found that setting G to the small value of G
108 did not lead to excessive controller gains.
Note that with the above choice of plant model 3 and cost
function 4, the nominal LQG problem essentially amounts to the
problem of minimizing the noise energy at the microphone position when the system is subject to a white noise disturbance entering the system through the control input channel.
The minimax LQG controller is synthesised by first choosing
the constant 0 to minimize the quantity W defined in Eq. 8.
With this value of the parameter , the controller is constructed
according to the formula 9. The order of this controller will be
the sum of the order of the nominal plant model P(s) together
Fig. 6 Two dimensional duct feedback controller setup
Fig. 7 Experimental and identified disturbance speaker response
Fig. 8 Experimental and identified control speaker response
JULY 2003, Vol. 125 377
Fig. 9 Uncertainty
P 2 j P 1 j
bound
2 j
dashed-20 log 10 W j ,solid-20 log 10 P
with the order of the weighting filter W(s). Typical order of controllers designed in this paper is around eighty. Such a high order
controller may lead to problems of numerical error and excessive
computational load when implemented. Hence, the balanced controller reduction method described in Section 19.1.1 of 23 is
applied to reduce the controller order to around forty. It was found
that with this level approximation, there was very little degradation in the predicted closed loop performance.
Experimental Results
The reduced dimension controller designed in Section 4 was

implemented on a dSPACE DS1103 system as shown in Fig. 6.
The spectrum analyser is used to measure the closed loop
frequency response from the disturbance speaker input to the
microphone output. In order to implement each controller, it
was first discretized using the FOH method with a sample period
Fig. 10 LQG minimax controller
378 Vol. 125, JULY 2003
Fig. 11 Nyquist plot of the loop-gain
of 0.5103 seconds. The resulting discrete time controller

was then implemented on the dSPACE system with this sample
period.
5.1 Control of Low Frequency Modes. Experimental frequency response data was collected from 20500 Hz. It is not
practical to fit a model over the entire frequency range. A 40th
order model was fitted in the 20100 Hz frequency range. Figures
7 and 8 show the experimental response and the identified response for the disturbance and control speakers, respectively.
From the figures it can be seen that even a 40th order model is
unable to match the data exactly.
The uncertainty envelope W(s) and the magnitude of the 40th
order transfer function, for the identified models in Figs. 7 and 8,
derived using Yule-Walker method is shown in Fig. 9.
For the case under discussion the controller order is 80 and
balanced model reduction technique 23 was applied in order to
obtain a 40th order approximation to the 80th order controller. The
designed controller is shown in Fig. 10.
Fig. 12 Simulated closed loop and open loop response
JULY 2003, Vol. 125 379
Fig. 13 Experimental closed loop and open loop response
The predicted stability and robustness of the closed loop system

can be seen by the Nyquist plot based on the final reduced dimension controller transfer function and the measured plant
transfer function data P 2 ( j ). This Nyquist plot is shown in Fig.
11. The disturbance attenuation performance of the controller is
illustrated in Fig. 12.
Figure 13 shows the resulting measured frequency response for
the open loop and closed loop system.
method is used to demonstrate the controller performance between

250 450 Hz. The identified and experimental responses are
shown in Figs. 14 and 15. The choice of uncertainty envelope and
the experimental closed loop and open loop performance is shown
in Figs. 16 and 17, respectively. The frequency response outside
of the range shown in these figures remained unchanged. There is
a clear improvement in the damping of the resonant peaks in the
frequency range of interest.
5.2 Control of Higher Frequency Modes. The experimental results presented thus far were restricted to the lower frequency
range of 20100 Hz. In this section the minimax LQG control
5.3 Proper Choice of the Uncertainty Model. It is shown

in Section 3 that the minimax LQG control problem is equivalent
to the minimization of the H norm of the closed-loop transfer
Fig. 14 Experimental and identified control speaker response
380 Vol. 125, JULY 2003
Fig. 15 Experimental and identified disturbance speaker response
function between the noise input w

(t) and output (t) as defined
in Fig. 5. The output (t) consists of two terms, one of them is the
uncertainty output z(t) which is also known as performance output. To minimize the H norm the controller will attempt to keep
the output z(t) small. From this we realize that W(s) whose
output is z(t)) should be kept small for high authority control. In
obtaining models for acoustic enclosures it is seen that the modelling error is the largest at deep nulls. A proper placement of
Fig. 16 Uncertainty
P 2 j P 1 j
bound
actuators and sensors can be used to get rid of deep nulls. In other
words optimal actuator-sensor placement methods should use
modelling error as a measure of optimality.
The influence of the choice of W(s) on the controller can be
used to shape the controller. For example, the frequency band
where W( ) is large, the controller gain will be low. In the
closed-loop response in Fig. 13 it can be seen that there is an
increase in the gain near the deep null around 70 Hz. This is due
2 j
JULY 2003, Vol. 125 381
Fig. 17 Experimental open loop and closed loop response
to the extreme difficulty in finding a good match between the

identified model and the experimental data. To shape the closed
loop response at 70 Hz the choice of frequency weighting function
W(s) can be such that it is very conservative above 70 Hz as
shown in Fig. 18. The conservative choice of the envelope at
frequencies greater than 70 Hz ensures that the controller doesnt
have high gain in that region. The experimental response of the
controller designed with this method is shown in Fig. 19. From
the figure its clear that a proper selection of the weighting function helps in shaping the closed loop response.
Fig. 18 Uncertainty
P 2 j P 1 j
382 Vol. 125, JULY 2003
bound
Conclusions
This paper successfully demonstrates that feedback control can

be applied to very high order acoustical systems for active noise
control. Due to the practical difficulty in implementing very high
order controllers it is essential that a small frequency band be
selected for noise attenuation. The chosen minimax LQG control
method gives the flexibility to choose the frequency weighting
function W(s) such that the controller targets a specified frequency band. Overall the experimental results in this work are
2 j
Fig. 19 Experimental closed loop and open loop response
very impressive and further work in the selection of actuatorsensor placement to obtain a tighter fit between the model and
data and a choice of frequency weighting function can improve
the results significantly.
References
1 Elliott, S. J., 1999, Down with Noise, IEEE Spectrum, June, pp. 54 61.
2 Hu, J.-S., Yu, S.-H., and Hsieh, C.-S., 1998, Application of Model-Matching
Techniques to Feedforward Active Noise Control Design, IEEE Trans. Control Syst. Technol., 61, pp. 33 42.
3 Omoto, A., and Elliott, S. J., 1999, The Effect of Structured Uncertainty in
the Acoustic Plant on Multichannel Feedforward Control Systems, IEEE
Trans. Speech Audio Process., 72, pp. 204 212.
4 Pota, H. R., and Kelkar, A. G., 2001, Modelling and Control of Acoustic
Ducts, ASME J. Vibr. Acoust., 1231, pp. 210.
5 Kelkar, A. G., and Pota, H. R., 2000, Robust Broadband Control of Acoustic
Duct, Proceedings of the 39th IEEE Conference on Decision and Control,
Sydney, Australia, pp. 4485 4490.
6 Hong, J., Akers, J. C., Venugopal, R., Lee, M.-N., Sparks, A. G., Washabaugh,
P. D., and Bernstein, D. S., 1996, Modeling, Identification, and Feedback
Control of Noise in an Acoustic Duct, IEEE Trans. Control Syst. Technol.,
43, pp. 283291.
7 Clark, R. L., and Cole, D. G., 1995, Active Damping of Enclosed Sound
Fields Through Direct Rate Feedback Control, J. Acoust. Soc. Am., 973, pp.
17101716.
8 Erwin, R. S., and Bernstein, D. S., 1997, Discrete-time H 2 /H Control of an
Acoustic Duct: Delta-domain Design and Experimental Results, Conference
on Decision and Control, IEEE, San Deigo, CA, pp. 281282.
9 Petersen, I. R., and Pota, H. R., 2000, Minimax LQG Optimal Control of an
Experimental Acoustic Duct, IEE Control 2000 Conference, Cambridge, UK.
10 Petersen, I. R., 2001, Multivariable Control of Noise in an Acoustic Duct,
Proceedings of the European Control Conference ECC2001, Porto, Portugal.
11 Pota, H. R., and Kelkar, A. G., 2001, On Perfect Acoustic Noise Cancelling
Control, Control and Intelligent Systems, 292, pp. 48 54.
12 Petersen, I. R., Ugrinovski, V., and Savkin, A. V., 2000, Robust Control Design
using H Methods, Springer-Verlag, London.
13 Ugrinovskii, V. A., and Petersen, I. R., 1998, Time-averaged Robust Control
of Stochastic Partially Observed Uncertain Systems, Proceedings of the IEEE
Conference on Decision and Control, IEEE, Tampa, FL.
14 Fang, B., Kelkar, A. G., and Joshi, S. M., 2002, Modelling and Control of
Acoustic-Structure Interaction in 3-D Enclosures, IEEE 2002 Conference on
Decision and Control, Las Vegas, Nevada.
15 Ding, W. P., and Chen, H. L., 2001, A Symmetrical Finite Element Model for
Structure-Acoustic Coupling of an Elastic, Thin-Walled Cavity, J. Sound
Vib., 2433, pp. 547559.
16 Marburg, S., Beer, H.-J., Gier, J., Hardtke, H.-J., Rennert, R., and Perret, F.,
2002, Experimental Verification of Structural-Acoustic Modelling and Design Optimization, J. Sound Vib., 2524, pp. 591 615.
17 Banks, H. T., Brown, D. E., Smith, R. C., Metcalf, V. L., Wang, Y., and Silcox,
R. J., 1994, Noise Control in a 3-D Structural Acoustic System: Numerical
and Experimental Implementation of a PDE-Based Methodology, IEEE Proceedings of the 33rd Conference on Decision and Control, Lake Buena Vista,
FL, pp. 305310.
18 Nelson, P. A., and Elliott, S. J., 1992, Active Control of Sound, Academic
Press, London.
19 Elliott, S., 2001, Signal Processing for Active Control, Academic Press, Sydney.
20 McKelvey, T., Akcay, H., and Ljung, L., 1996, Subspace-based Multivariable
System Identification from Frequency Response Data, IEEE Trans. Autom.
Control, 417, pp. 960979.
21 Friedlander, B., and Porat, B., 1984, The Modified Yule-Walker Method of
ARMA Spectral Estimation, IEEE Trans. Aerosp. Electron. Syst., 202, pp.
158 173.
22 Petersen, I. R., Pota, H. R., and Jahromi, M. R. S., 2002, System Identification, Uncertainty Modelling and Actuator Placement in the Robust Control of
an Acoustic Duct, Proceedings of the Conference on Information, Decision,
and Control, Adelaide, Australia.
23 Zhou, K., Doyle, J. C., and Glover, K., 1996, Robust and Optimal Control,
Prentice Hall, NJ.
JULY 2003, Vol. 125 383
Active Vibration Suppression With

Time Delayed Feedback
Various active vibration suppression techniques, which use feedback control, are implemented on the structures. In real application, time delay can not be avoided especially in
the feedback line of the actively controlled systems. The effects of the delay have to be
thoroughly understood from the perspective of system stability and the performance of the
controlled system. Often used control laws are developed without taking the delay into
account. They fulfill the design requirements when free of delay. As unavoidable delay
appears, however, the performance of the control changes. This work addresses the stability analysis of such dynamics as the control law remains unchanged but carries the
effect of feedback time-delay, which can be varied. For this stability analysis along the
delay axis, we follow up a recent methodology of the authors, the Direct Method (DM),
which offers a unique and unprecedented treatment of a general class of linear time
invariant time delayed systems (LTI-TDS). We discuss the underlying features and the
highlights of the method briefly. Over an example vibration suppression setting we declare
the stability intervals of the dynamics in time delay space using the DM. Having assessed
the stability, we then look at the frequency response characteristics of the system as
performance indications. DOI: 10.1115/1.1569942
Rifat Sipahi
Nejat Olgac*
Mechanical Engineering Department,
University of Connecticut,
Storrs, CT 06269
Introduction
Active vibration suppression has long been implemented on

structures under oscillatory excitations in various ways 1,3,4. In
general, a partial or full state feedback control law is utilized in
order to achieve good suppression performance. Most generally
this is represented as a multi input-multi output MIMO state
space form of:
uF
xAxB
(1)
where u is the full state feedback, given by:

uKx
(2)
x ABxF
(3)
( nm ) are the system and

Here x( n1 ) is the state vector, A( nn ) , B
the control matrices, respectively, K( mn ) is the full state feedback
gain matrix, F( n1 ) is the excitation disturbance force acting on
K. Utilizing the full state feedback law Eq.
the structure, and BB
2, vibration suppression performance can easily be improved
within the frequency range of the excitation forces F.
It is very well known, however, that the control systems are
commonly subject to time delay, which converts Eq. 3 into a
retarded time delayed system RTDS which is given as:
xAxBx t F
(4)
where is the time delay influencing the system dynamics.

It is critical to understand that the time delay may appear due to
the inherent nature of the core dynamics 5 6 as well as the
feedback control 12,7 8. Regardless of the reason of the delay
in the dynamics
a the underlying mathematics i.e., retarded time delayed systems remains the same. The novel treatment presented in this text
is applied to vibration suppression problems, although it is generic
enough to be considered for a wide variety of TDSs.
b it may deteriorate the quality of vibration suppression and
in the worst case it can even drive the dynamics into instability.
*Author of correspondence. E-mail: olgac@engr.uconn.edu
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Feb. 2002;
Revised Nov. 2002. Associate Editor: J. Wickert.
384 Vol. 125, JULY 2003
In short, the dynamics of a typical active vibration suppression

system needs to be carefully evaluated with special consideration
of delay. The unforced dynamics of the system is:
xAxBx t
(5)
and it has to be studied from the stability perspective. Equation 5

represents a widely utilized class of linear time invariant retarded
time delayed systems LTI-RTDS. This general class of systems
has been extensively investigated for over four decades 919.
The deployment of a recently developed Direct Method DM
2,22 which yields unique observations on the system stability
and the ensuing vibration suppression performance of Eq. 4
form the main theme of the present text.
Until the Direct Method DM the stability posture of the LTITDS along the axis of time delay was not completely resolved.
That is, a complete set of values of delay which cause stable or
unstable behavior were not available through a systematic and
practical methodology. 18 gives an extensive review of the stateof-the art in this venue, including the techniques developed up to
2001. Most of the later methods convert the LTI-TDS into a distributed time delay equation via an integral transformation as explained in 20. They indicate that there are some additional dynamics appearing due to this transformation. Consequently, the
new dynamics invites additional characteristic roots, which impose adverse restrictions on the range of time delay. Other methods, which are mainly based on various Lyapunov-Krasovski
functional 18,2021, bring conservative solutions for the stable
regions of time delay. A recent publication 19 compares five
different approaches all developed in the 90s based on Lyapunov
and LMI linear matrix inequality techniques. It is rather intriguing to observe that even the best stability margin that these methods can provide is more than 50% conservative with respect to the
true margin which is detected by following a tedious Nyquist
method. In contrast, the DM 22 coincides with the Nyquist
results precisely, furthermore it is numerically very efficient.
These features make the Direct Method, extremely compelling.
Another shortfall of the existing methods is that, they all start
with a stable dynamics for nondelayed i.e., delay0 case. And
they try to assess the first positive delay value, which brings instability. This value forms the so-called, delay margin of the
system and the stability analysis stops at that point. However, as
realized by many investigators 2326, there can be more than
one such stable pocket in delay domain. This knowledge is extremely important, as one can imagine. If there are so called, the
secondary stability pockets one may opt to operate within one of
these, still performing a stable control despite the larger time delay in the picture. No existing methodology, however, could offer
a procedure to detect these additional pockets. The Direct Method
DM 22 suggests a unique construct ultimately yielding an explicit function of delay, which declares completely the stability
pockets. The end results clearly offer much stronger revelations
than simply stating the stability margin. The method is the first of
its kind as it handles the general n-dimensional LTI-RTDS with
commensurate time delays and uncovers its unique properties.
Some of these properties are realized in 23 for scalar dynamics
(n1), however the approach therein is quite restrictive to implement on a general LTI-RTDS i.e. n1). For these reasons, the
Direct Method is unique in the dynamic systems and controls
community, opening number of new directions for future research.
We wish to summarize the strong features of the Direct Method
here to better prepare the reader:
i. It does not have to start with a stable system for zero delay.
ii. It is exact, i.e. the limits of the stability pockets are precise.
They are not conservative results, which is a concern for the
peer methodologies 910,18 21,23.
iii. All stable pockets are declared exclusively, in one explicit
function of time delay.
In what follows we briefly review the procedure suggested by
the DM leaving the details to 22. The system in Eq. 5 has a
transcendental characteristic equation:
n
CE s, det sIABe s
a s e
j0
js
0 with 0
(6)
where a j (s)s are polynomials in s of degree n j for retarded

time delayed dynamics. Clearly CE(s, ) has infinitely many finite characteristic roots, which move on the complex plane as
varies. As indicated by the feature i above, this system could be
unstable when free of delay. The number of unstable roots for this
case, can be determined using, for instance, Rouths array. As a
matter of fact, these roots are nothing but the n eigenvalues of
AB and they can also be calculated numerically to display the
natural response characteristics of the system, for 0. The interesting outcome of this work could be that, if a system is unstable
for 0, it may return to stability as increases. This would be a
beneficial effect of the time delay on the dynamics.
As increases these n characteristic roots start moving, along
with infinitely many other roots because of the transcendentality
of Eq. 6, what we call the secondary roots. These secondary
roots are dormant, stable roots for 0 i.e. they are all in the
form of ai a finite and real, and they do not contribute to
the natural response of the system. But they play a very critical
role when 0, because they creep into the finite domain and
become influential. They may even destroy the stability of the
dynamics.
Obviously for certain values of 0 some of the infinitely
many characteristic roots cross the imaginary axis excluding
those dynamics that are stable independent of time delay, causing
a transition from stable to unstable half of the complex plane or
vice versa. These crossing frequencies ( i) and the values
generating them must all be detected exhaustively for a complete
stability analysis.
At this point, we state two very critical propositions from 22,
which eventually yield the three conclusions marked i, ii, iii
above. We will recite them here for clarity, leaving their proofs
to 2,22.
Proposition I: The time delayed system in Eq. 5 can have
only a finite number of imaginary characteristic roots call these
ck i, k1, . . . m), for all values of R . Note that, the
subscript c refers to the crossing of the roots over the imagiJournal of Vibration and Acoustics
nary axis. The corresponding values are k , k1, . . . m,

0, . . . which are distributed periodically by 2 / ck . This
proposition implies that the infinitely many characteristic roots of
Eq. 6 can only move across the imaginary axis through these m
pinholes ( ck i,k1, . . . m).
Proposition II: As transits through one of the k , k
1, . . . m, 0, . . . values, the respective root can pass
through the pinhole ck i only in one direction. That is, the root
tendency RT which is defined as
ck
RT
sgn Re
ds
d
s ck i
k
k1, . . . m
0, . . .
(7)
is invariant of the particular delay k p2 / ck , p

0, . . . which generates the crossing ck i.
What this proposition conceptually implies is that all of the m
pinholes for the possible passages of the characteristic roots
through the imaginary axis are guarded by one-way check
valves.
In Section 2, we introduce the steps of DM, and two propositions, as well as the novel stability analysis, the Direct Method,
are followed by an application on a realistic vibration control
problem, in Section 3.
The Direct Method
For completeness of the treatment, we wish to review the Direct

Method DM briefly here. The stability assessment of the system
5 is needed in the domain of the only free parameter in 5.
DM starts with detecting all the crossing frequencies ck (k
1, . . . m) as described in Proposition I and the corresponding
time delays
k k0
ck
0, . . .
(8)
which create these crossings.

The following structured steps are performed next for the DM
stability analysis:
1 All imaginary roots and generator s are assembled as:
ck , k ,
0, . . .
k1, . . . m,
2 The set of common root tendencies RT at each one of these

m crossing frequencies are evaluated, RT ck . As per Proposition
II, this is invariant of the value of k , k1, . . . m.
3 All k , k1, . . . m, 0, . . . are tabulated in an array, in ascending order, along with the ck s and RT ck s.
4 The number of unstable roots (NU) is determined for 0
non-delayed case. Call this number NU(0).
5 Take the next k in ascending order and check RT ck . It is
obvious that, for RT1, NU increases by 2, and for RT
1, NU decreases by 2. Repeat this step for the following s
until the particular target value of delay is reached.
6 Identify those regions in where NU( )0 as stable and
others as unstable. Also, state the NU( ) as the number of unstable roots in this interval of .
This procedure, indeed, lends itself to a novel explicit expression of NU( ), owing to the Propositions I and II:
NU NU 0
k1
k0
U , k0 RT k
k
(9)
where NU(0) is the number of unstable roots when 0,

U( , k0 )A step function in with the step taking place at k0
JULY 2003, Vol. 125 385
Table 1 The stability table
RT
NU
0
0.3071
0.5441
1.1444
1.9766
1.9818
2.0028
2.0052
7.5032
4.3864
7.5032
4.3864
7.5032
3.0446
2.98
1
1
1
1
1
1
1
0
2
4
6
8
10
8
6
3 Active Vibration Suppression Using Time Delayed

Full State Feedback
0 k0
Lets take a realistic vibration suppression setting given in Fig.

1 from 27 where the mass m 11 is excited harmonically and the
response of m 12 , x 12 , is desired to be minimized. This model is a
simplified display of two high-rise buildings (m 11 and m 21) interconnected via two passageways m 12 and m 22 . Two actuators
deploy the feedback control, which is determined using pole
placement techniques.
We alter the system parameters from the original example in
order to make the numerical results compact enough to include
here:
for k0 , ck 0
k0 , ck 0
(x)Ceiling function of x, returns the smallest integer greater

than or equal to x
The expression NU( ) of Eq. 9 requires the knowledge of
four things, all of which are known:
i NU(0)
ii k0 , k1, . . . m, the smallest 0 values corresponding to
each one of the ck s
iii k 2 / ck , k1, . . . m
iv RT k RT ck , k1, . . . m
The stable regions of i.e. the pockets are those which display NU( )0. That is s R ,NU( )0 . This completes
uF
xAxB
rad/s
the Direct Method of stability for LTI-RTDS. Now, we demonstrate how to deploy the DM for the active vibration suppression
applications by way of example.
Fig. 1 Dynamic model
U , k0
m 110.2,
c 1 2.2,
k 102,
k 114,
30
11
10
11
13.33
12.67
12.67
11
10
20
11
40 12.67 13.33
where x x 11 ,x 11 ,x 12 ,x 12 ,x 21 ,x 21 ,x 22 ,x 22 T is the state vector

and F are self
with x i j representing the motion of m i j , A, B
evident terms, f f 0 sin( t), and the full state feedback control is
uKx. Ultimately the solution to this equation in Laplace domain
is
x s sIAB 1 F s
(10)
k 122,
m 210.2,
m 220.15 kg
c 2 1.9 Ns/m
k 204,
k 212,
k 222 N/m
The governing equation of this controlled system is obtained as:
26.67 12.67
m 120.15,
88
6.67
6.67
0
f t
0
0
0
0
0
0
81
82
14.20 2.48 8.24 2.69 11.41 3.12 13.05 4.55

8.25 2.52 10.37 4.45 10.17 2.65 5.33 2.19
28
such that the controlled systems characteristic roots, i.e.

det sIAB 0
K, F(s)L F(t) , F(s)L f (t)

where BB
The characteristic roots of the uncontrolled structure is calculated from
are 2, 4, 5, 6, 13i, 3i. This is achieved using

MATLAB/Control Toolbox. When a delay appears in the feedback
line; i.e. u(t)Kx(t ), the characteristic equation of the system becomes
det sIA 0
det sIABe s 0
and they are 24.02, 20.69, 0.72, 1.79, 0.0186.44i,

0.0384.10i.
We select
386 Vol. 125, JULY 2003
(11)
which is identical to that given in Eq. 6. As we expand Eq. 11

it takes the form of:
a 0 s a 1 s e s a 2 s e 2 s 0
(12)
Fig. 2 Number of unstable roots NU plot in domain
where a p (s) is a polynomial in s of degree 8p, p0, 1, 2.

Notice that the largest commensurate delay is 2 in Eq. 12 because of the specific formation of A and B matrices in this case.
The stability analysis follows as described in Section 2.
All imaginary roots are calculated using the methodology presented in 22 and they are tabulated along with the corresponding
delays and root tendencies Proposition II below.
c1 4.3864
c2 7.5032
c3 3.0446
c4 2.98

2p
c1
2p
2k 0.307
c2
2p
3k 2.003
c3
2p
4k 2.005
c4
1k 0.544
from

RT 1 1
RT 2 1
RT 3 1
RT 4 1
p0, . . .
(13)
Fig. 3 Stability posture of the system
That is, only 4 sets of k s each with dimension of infinity

cause the roots pass through the pinholes of c1 , c2 , c3 , c4
in the direction given by root tendencies for 1 from stable to
unstable complex plane, and for 1 vice versa.
Continuing the listings of Eq. 13 one can simply form the
Table 1, which we call the stability table, in ascending order of
k , k1, . . . m, 0, . . . .
The corresponding NU( ) variation is like in Fig. 2. The stability posture of the system is clear: its stable only for the time
delay 0307 ms. Afterwards it never returns to stability. What
this implies in practice is that the delay on the feedback can only
be within this range for stability. Beyond 307 ms, the given feedback structure renders instability by enforcing unstable righthand-side roots numbered by NU( ), as per Fig 2. A numerical
simulation study validates this finding as given in Fig. 3.
The end result of this effort is not the declaration of stable or
unstable regions only, but also what happens to vibration suppression in the stable regions when the delay exists. We take 250
ms knowing that it is harmless from stability perspective and
study the frequency response of the mass m 12 under f 0 sin(t)
harmonic excitation on m 11 . When Eq. 10 is solved for x 12 it
results in:
x 12 s
16.78 s5 186.52 s4 887.89 s3 2930.75 s2 6516.5 s6155.36
8
F s
s 25 s7 272 s6 1724 s5 7172 s4 20516 s3 39520 s2 46000 s24000
(14)
Fig. 4 Frequency responses of the dynamics
JULY 2003, Vol. 125 387
for non-delayed feedback uKx(t) and

x 12 s p 0 s p 1 s e s p 2 s e 2 s
F s
q 0 s q 1 s e s q 2 s e 2 s
(15)
for delayed feedback uKx(t ), where p i (s), q i (s) are selfevident polynomials of s.
The frequency response features of this system are given in Fig.
4 for various cases. It is clear that the feedback controlled case Eq.
14 registers substantial improvement over the uncontrolled case
for 2.7 rad/s. And the delayed case Eq. 15 is bringing further
improvement except in the frequency range of 7.18.1 rad/s.
In any event, it offers much better vibration suppression than the
uncontrolled passive case.
It is obvious that, Fig. 4 has meaning if and only if the system
is stable, which can only be assured using the Direct Method
presented. Therefore the novel method plays a very critical role in
this operation.
Conclusions
A general active vibration suppression problem under delayed

feedback is investigated in this paper. It is well known that for
such control systems, time delay plays a very important role on
the vibration suppression quality and more importantly on the
stability of the system. In this phase, we need to fully understand
the stability picture of the dynamics under time delay influence,
first. For this we deploy a unique procedure, the Direct Method.
The method offers certain structured steps for the stability analysis
and uncovers very interesting features of linear time invariant
time delayed systems LTI-TDS, which were not recognized earlier. Using the DM, one can determine the exact boundaries of
stable and unstable regions in time delay domain. It is then
possible to investigate the efficiency of a vibration suppression
application, maintaining the stability of the system under time
delay effects. And we include a comparison of controlled vibration suppression for various frequency zones between the time
delayed and non-delayed cases.
References
1 Olgac, N., 1995, Delayed Resonators as Active Dynamic Absorbers, United
States Patent 5,431,261.
2 Olgac, N., Special Lecture at the 3rd IFAC Workshop on TDS 2001, http://
www.siue.edu/ifacdelay/
3 Seto, K., and Furuishi, Y., 1991, A Study on Active Dynamic Absorber,
ASME Proceedings, Paper DE, Vol. 38, pp. 263270.
4 Seto, K., and Yamashita, S., 1991, Simultaneous Optimum Design Method
for Multiple Dynamic Absorbers to Control Multiple Resonance Peaks, SAE
Transactions, 100, pp. 14811489.
388 Vol. 125, JULY 2003
5 Olgac, N., and Hosek, M., 1998, A New Perspective and Analysis for Regenerative Machine Tool Chatter, Int. J. Mach. Tools Manuf., 387, pp. 783
798.
6 Tlusty, J, 1985, Machine Dynamics, R. I. King, ed., Handbook of High
Speed Machining Technology, Chapman and Hall, New York.
7 Olgac, N., and Holm-Hansen, B., 1994, A Novel Active Vibration Absorption
Technique: Delayed Resonator, J. Sound Vib., 176, pp. 93104.
8 Olgac, N., Elmali, H., Hosek, M., and Renzulli, M., 1997, Active Vibration
Control of Disturbed Systems Using Delayed Resonator with Acceleration
Feedback, ASME J. Dyn. Syst., Meas., Control, 119, pp. 380388.
9 Chen, J., Gu, G., and Nett, C. N., 1994, A New Method for Computing Delay
Margins for Stability of Linear Delay Systems, Syst. Control Lett., 26, pp.
107117.
10 Chen, J., 1995, On Computing the Maximal Delay Intervals for Stability of
Linear Delay Systems, IEEE Trans. Autom. Control, 406, pp. 10871092.
11 Hale, J. K., and Verduyn Lunel, S. M., 2001, Effects of Small Delays on
Stability and Control, Operator Theory; Advances and Applications, 122, pp.
275301.
12 Hale, J. K., and Verduyn Lunel, S. M., 2001, Strong Stabilization of Neutral
Functional Differential Equations, IMA J. Math. Control Inform., 19, pp.
119.
13 Hale, J. K., and Verduyn Lunel, S. M., 1993, Introduction to Functional Differential Equations, Springer-Verlag.
14 Hale, J. K., Infante, E. F., and Tsen, F.-S. P., 1985, Stability in Linear Delay
Equations, J. Math. Anal. Appl., 105, pp. 533555.
15 Hertz, D., Jury, E. I., and Zeheb, E., 1984, Simplified Analytic Stability Test
for Systems with Commensurate Time Delays, IEE Proc., 1311, PtD, pp.
5256.
16 Jalili, N., and Olgac, N., 1999, Multiple Delayed Resonator Vibration Absorber for MDOF Mechanical Structures, J. Sound Vib., 2234, pp. 567585.
17 Kolmanovski, V. B., and Nosov, V. R., 1986, Stability of Functional Differential Equations, Academic Press, London, Great Britain.
18 Niculescu, S-I., 2001, Delay Effects on Stability, Springer-Verlag.
19 Zhang, J., Knospe, C. R., and Tsiotras, P., 2001, Stability of Time-delay
Systems: Equivalence Between Lyapunov and Scaled Small-gain Conditions,
IEEE Trans. Autom. Control, 463, pp. 482 486.
20 Gu, K., and Niculescu, S.-I., 2000, Additional Dynamics in Transformed
Time-Delay Systems, IEEE Trans. Autom. Control, 453, pp. 572575.
21 Park, P., 1999, A Delay-Dependent Stability Criterion for Systems with Uncertain Time-Invariant Delays, IEEE Trans. Autom. Control, 44, pp. 876
877.
22 Olgac, N., and Sipahi, R., 2002, An Exact Method for the Stability Analysis
of Time Delayed LTI Systems, IEEE Trans. Autom. Control, 475, pp. 793
797.
23 Cooke, K. L., and van den Driessche, P., 1986, On Zeroes of Some Transcendental Equations, Funkcialaj Ekvacioj, 29, pp. 7790.
24 Thowsen, A., 1981, The Routh-Hurwitz Method for Stability Determination
of Linear Differential-Difference Systems, Int. J. Control, 335, pp. 991
995.
25 Thowsen, A., 1981, An Analytic Stability Test for a Class of Time-Delay
Systems, IEEE Trans. Autom. Control, 263, pp. 735736.
26 Thowsen, A., 1982, Delay-independent Asymptotic Stability of Linear Systems, IEE Proc., 29, pp. 7375.
27 Seto, K., 1995, Structural Modeling and Vibration Control. Internal Report,
Nihon University. Department of Mechanical Engineering, College of Science
and Technology, Nihon University, 1-8-14 Kanda Surugadai Chiyoda-ku, Tokyo, 101-8308, Japan.
Gaetan Kerschen
e-mail: g.kerschen@ulg.ac.be
Jean-Claude Golinval
Vibrations & Identification des Structures,
Department of Aerospace,
Mechanics and Materials,
University of Lie`ge,
Chemin des Chevreuils 1 (B52),
B-4000 Liege, Belgium
Francois M. Hemez
Engineering Science & Applications Division,
ESA-WR, Mail Stop P946,
Los Alamos National Laboratory,
Los Alamos, New Mexico 87545
e-mail: hemez@lanl.gov
Bayesian Model Screening for the

Identification of Nonlinear
Mechanical Structures
The development of techniques for identification and updating of nonlinear mechanical
structures has received increasing attention in recent years. In practical situations, there
is not necessarily a priori knowledge about the nonlinearity. This suggests the need for
strategies that allow inference of useful information from the data. The present study
proposes an algorithm based on a Bayesian inference approach for giving insight into the
form of the nonlinearity. A family of parametric models is defined to represent the nonlinear response of a system and the selection algorithm estimates the likelihood that each
member of the family is appropriate. The (unknown) probability density function of the
family of models is explored using a simple variant of the Markov Chain Monte Carlo
sampling technique. This technique offers the advantage that the nature of the underlying
statistical distribution need not be assumed a priori. Enough samples are drawn to guarantee that the empirical distribution approximates the true but unknown distribution to
the desired level of accuracy. It provides an indication of which models are the most
appropriate to represent the nonlinearity and their respective goodness-of-fit to the data.
The methodology is illustrated using two examples, one of which comes from experimental
data. DOI: 10.1115/1.1569947
Introduction
The importance of diagnosing, identifying and modelling nonlinearity has been recognized for a long time, e.g., for the design
of shock absorbers and engine mounts. The identification of nonlinear systems began in 1979 with the introduction of the restoring
force surface RFS method by Masri and Caughey 1. An
equivalent method, referred to as force-state mapping, was proposed independently by Crawley, Aubert and ODonnel 2,3.
Since then, numerous methods were proposed. It is not our intention to review all the methods available but rather to cite the most
popular techniques that have been considered during the last
twenty years.
The first application of the Hilbert transform was made in the
frequency domain 4. The time-domain Hilbert transform was
also utilized to solve an inverse problem 5,6. The use of the
Volterra series in the field of structural dynamics began in the late
1980s 7. NARMAX models consist of polynomials that include
various linear and nonlinear terms combining the inputs, outputs
and past errors and were introduced by Leontaritis and Billings
8,9. Another area of signal processing that has gained importance in studying nonlinear systems deals with higher-order spectra 10,11. These are a natural extension of the ordinary linear
spectral analysis. For a detailed description of all these techniques, the reader is referred to reference 12.
The development of frequency response function-based approaches has received increasing attention in recent years. The
reverse path technique has been proposed by Rice and Fitzpatrick
13 and applied to simulated and experimental data 14,15. The
conditioned reverse path formulation 16 extends the application
of the reverse path algorithm to systems characterized by nonlinearities away from the location of the applied force. This method
exploits the spectral conditioning techniques introduced by Bendat 17. A related series of papers by Adams and Allemang also
develop the frequency response function-based approaches
18,19.
Finally, it is worth pointing out that there has also been a
Revised January 2003. Associate Editor: M. I. Friswell.
growth in interest in a particular class of identification techniques

based on a finite element model and referred to as finite element
model updating techniques 2022.
The problem of variable selection is one of the common issues
in the field of identification of nonlinear systems. The purpose is
to model the relationship between the response variable of interest
and a subset of predictor variables, possibly with interactions between these latter variables. Generally speaking, there is uncertainty about which subset to use.
A possible means of determining which variables should be
included in the model is through least-squares parameter estimation and the use of the significance factor 12. Cumulative and
multiple coherence functions may also be used in conjunction
with the conditioned reverse path formulation 23. The present
study investigates an inference technique based on the Bayesian
definition of probabilityas opposed to the frequentists point-ofviewfor identifying promising subsets of predictors 24 27.
While the frequentist interpretation defines probability strictly as
the number of occurrences of an event among a collective of
possibilities, the Bayesian approach defines probability as the subjective opinion of the analyst or expert. To stress the difference
between the two approaches, consider the simple question What
is the probability of life on the planet Mars? Such question
makes no sense in the frequentist framework because observations
can obviously not be obtained from a collective of planets similar
to the planet Mars. Similarly and even though we might not always be aware of it, many problems occur in structural dynamics
that require probability to be defined in terms of our a priori
knowledge of the phenomenon studied. The identification of the
form of a model is one such problem addressed in this work.
The procedure developed in this work exploits priorsthat
is, a probability structure that reflects the analysts a priori opinion
about the phenomenon investigatedon the variables of the regression model in order to give the list of all visited models together with their relative posterior probabilities. Models are visited according to their goodness-of-fit to the data, which, in the
Bayesian framework, represents the likelihood of predicting the
observed response. This implies that models well fitted to the
datathat is, more likely modelsare visited more often. The
marginal probabilities of inclusion of single variables are also
JULY 2003, Vol. 125 389
computed. To avoid the overwhelming burden of calculating the

posterior probabilities of all models, a Gibbs sampler is considered to perform an efficient stochastic search of the model space.
It is emphasized that the main difficulty of this inference problem
is that the multi-dimensional probability density function PDF of
the family of models must be sampled. However, this PDF is
unknown, making random walk techniques such as the Markov
Chain Monte Carlo MCMC sampling the only possible choice.
Gibbs sampling has been proposed as a computationally attractive
alternative to MCMC, yet, it can explore an unknown PDF
24,26.
While the principle of Bayesian inference has previously been
applied to various problems in structural dynamics e.g., References 28,29, no attempt has been made, to the best of the authors knowledge, to adapt the Bayes updating rule to the screening of model form during nonlinear system identification. After a
brief discussion of model fitting in Section 2, the Bayesian screening algorithm for model selection is outlined in Section 3. The
methodology is illustrated using two examples. Section 4 discusses a numerical simulation intended at demonstrating the overall performance of the screening method. The second example
involves experimental data sets collected during the European
COST-F3 program Section 5. The numerical predictability of the
identified model is finally assessed in Section 6.
Model Fitting
Model fitting generally refers to the calibration of model coefficients given a sequence of points (t k ;y k ) in the design space.
It is assumed that a model is available:
yM ;t
(1)
where t denotes the input variables, y denotes the output variables

and denotes the models coefficients.
For clarity, the discussion will assume that the model form is
polynomial-like. Nevertheless, nothing prevents the Bayesian
model screening proposed in Section 3 to be applied to other
functional forms. Fractional models could be considered, for example, to fit the poles and zeros of frequency response functions.
Exponential models could be considered to represent the decay of
propagating waves as a function of time or distance. Artificial
neural networks are increasingly used in a variety of applications
in structural dynamics because they can, depending on their form,
approximate any non-linear function 30.
Another notion that must be clarified before proceeding with
the discussion is the notion of effect. The model shown in Eq.
1 depends on inputs t where t does not necessarily refer to time.
Functions of the input variables t can be defined that will be
referred to as effects and denoted by x in the following. Such
functions can assume any form, linear or non-linear. For example,
the 2-input, 1-output nonlinear model:
y0.3t 1 2.0 sin t 2 1.5e t 1 t 2
(2)
can equivalently be defined through the three effects x 1 t 1 , x 2

sin(t2) and x 3 e t 1 t 2 as:
y0.3x 1 2.0x 2 1.5x 3
(3)
While the input variables t 1 and t 2 might be independent, note

that the effects x 1 , x 2 and x 3 are neither independent nor uncorrelated. The Bayesian model screening discussed in Section 3
does not require the effects to be independent or uncorrelated.
With the definition of effects x that can be functions of the input
variables t, the polynomial-like model can be simply represented
as:
y
k1 . . . m
x k k x T
(4)
The commonly encountered method of fitting the coefficients

is to define an objective function that represents the prediction
390 Vol. 125, JULY 2003
error and minimize it. The most straightforward choice is to adopt

the Euclidean norm of the prediction error e k y k x Tk :
J
k1 . . . N
e Tk e k e T e
(5)
in which case the best, linear, unbiased estimator of the coefficients is provided by:
X T X 1 X T y
(6)
where the column-vector y collects N observations and the N rows

by m columns matrix X evaluates the m effects for each of the N
observations:
y1
y2
]
yN
;X
x 1,1
x 1,2
x 2,1
x 2,2
x N,1
x N,2
x 1,m
x 2,m
x N,m
(7)
Clearly, other objective functions yield different estimators. The

generalization of the objective function 5 is commonly referred
to as the generalized least-squares GLS estimator 31. Weighting matrices are introduced and a regularization term penalizes
solutions too distant from the user-defined starting point o . Eqs.
8 and 9 show the GLS objective and the corresponding GLS
estimator, respectively:
T 1
J e T W 1
ee e o W bb o
(8)
1 1 T 1
X T W 1
X W ee y
ee XW bb
(9)
In general, weighting matrices are chosen arbitrarily or based

on experience, for example, to weight the importance of some
observations more than others. When covariance matrices are
used, the GLS estimator becomes similar to the Bayesian estimator. Rigorously speaking, other factors should appear in the definition of the Bayesian objective function. Because these additional factors are constant, however, the same estimator as the one
shown in Eq. 9 is obtained. An important benefit of Bayesian
inference is that it provides a posterior estimate of the covariance
matrix:
T 1
1
bbposterior W 1
W
bb X W ee X
(10)
Correlation coefficients of the posterior covariance matrix 10

provide insight into the quality of the estimator. Reference 32
discusses a shock propagation application where significant posterior correlation is obtained between coefficients that have no
physical reason to be correlated. The authors conclude that the
form of the model is inappropriate. They further demonstrate that
it is indeed the case when improved goodness-of-fit and posterior
correlation indicators are obtained with a different model.
With the exception of investigating the posterior correlation,
however, no practical tool is available to select the appropriate
form of a nonlinear model, which is the process we refer to a
model screening. Model formfor example, replacing a linear
contribution by a cubic stiffnessis usually selected based on
experience or empirical observation. Sometimes, several choices
seem equally likely and the analyst has to go through the painstaking process of fitting each model and assessing their goodnessof-fit. Because it is based on the concept of goodness-of-fit, such
approach leads to over-fitting.
Another subtle but important issue is to estimate the posterior
probability of a particular model as opposed to simply relying on
the goodness-of-fit. By definition, the posterior probability is conditioned on the evidence availablethat is, experimental observations. Posterior probability and goodness-of-fit complement each
other because the former indicates if the analysts prior opinion of
the form of the model is consistent with the evidence. In Section
3, a practical tool is proposed for model screening based on the
concept of posterior probability.
Bayesian Model Screening
In the previous section, the state of the practice in model fitting

has been briefly overviewed. Polynomial models have been considered for simplicity. It is emphasized that the Bayesian model
screening technique proposed here applies to any model, no matter which functional form it takes. Essentially, the only two general assumptions made are as follows. First, a model y
M ( ;x) must be available. As previously mentioned, the effects
x can be linear or nonlinear functions of the input variables t.
Second, an inference is available for calibrating the coefficients .
The inference procedure is usually referred to as best-fitting
with polynomials and training with neural networks.
Model screening consists in identifying the most probable models based on a family of models defined by the user and reference
data that the models predictions must reproduce with the highest
possible fidelity. It is emphasized that model screening does not
necessarily identify the best model but rather ranks potential models according to their posterior probability of occurrence.
The procedure starts by, first, defining a family of models. This
is achieved by defining various effects x i and how these effects
are allowed to interact to form the population of potential models.
Figure 1 illustrates the concept of a family of models by showing
two effects x 1 and x 2 that interact with each other. The modelforming rule illustrated in Fig. 1 is that linear and quadratic interactions are allowed between the effects x 1 and x 2 . The horizontal
plane represents the family of all potential models that must be
explored. The vertical dimension represents the likelihood that a
particular model is appropriate to represent the data. It is this
notion of likelihood that will be employed to guide the search for
the most appropriate models. Figure 1 illustrates a hypothetical
situation where the model shown with a star symbol, y o
1 x 2 2 x 21 3 x 2 x 21 , is the maximum likelihood model.
The second step of the procedure is to assign the prior probability of occurrence of each effect x i . The priors can reflect
empirical observations, experience or the analysts knowledge of
the system investigated. In the application discussed in Section 5,
for example, no specific knowledge of the system can be used to
guide a pertinent choice of priors. Probabilities of occurrence are
therefore set to a uniform 25% level for all effects.
The next step is to let the Bayesian screening method find the
most appropriate models among all possible combinations of effects. To identify the most probable models a measure of
goodness-of-fit to the reference data must be defined. This can be
assessed using a conventional root mean square RMS error between data and predictions. Assuming Gaussian distributions, the
RMS error becomes proportional to the likelihood function
L(y ) that estimates the likelihood that the model is appropriate
given the available data:
L y
k1 . . . N
y k x Tk 2
Fig. 1 Concept of family of models
posteriors of the nth iteration become the priors of the (n1)th

iteration. All models visited are kept in memory and, once enough
samples have been drawn, the probability of occurrence of each
model is estimated by the frequency of occurrencethat is, the
ratio between the number of times each model is visited and the
total number of models visited. The iterative procedure is summarized in Fig. 2.
In summary, Bayesian model screening provides the probabilities of occurrence of the most appropriate members of a user-
(11)
Note that the likelihood function 11 is similar to Eq. 5 previously discussed. Other functions can be used, in particular the
Bayesian objective 8, as well as the many objective functions
commonly used in test-analysis correlation and model updating
33.
Once the likelihood of a particular model has been estimated,
the posterior probabilities of the models effects can be updated
according to the Bayes Theorem that states that the posterior probability PDF( y) is equal to the likelihood function L(y ) multiplied by the prior probability PDF( ) and divided by the probability of the data PDF(y):
PDF y
L y PDF
PDF y
(12)
The probability of the observed data PDF(y) is generally kept

constant and omitted in the updating Eq. 12. Because the procedure is iterative in nature, the Bayes update 12 is repeated and
Fig. 2 Simplified flow chart of the Bayesian model screening

algorithm
JULY 2003, Vol. 125 391
The main difference between the two is that the Gibbs algorithm
samples one direction of the design space at a time, which makes
for simpler numerical implementation. Figure 4 illustrates the difference between MCMC and Gibbs sampling. It pictures two random walks from the lower left corner (x0;y0) to the upper
right corner (x1;y1). A constraint is enforced that prevents
the 30 points drawn in both sequences from being repeated and
from moving backwards. Pentagram symbols show a sequence of
Gibbs samples while hexagram symbols picture a realization of
the MCMC chain. In the former case, the solution is advanced in
one direction at a time whereas the MCMC chain randomly advances the solution in the two dimensions simultaneously.
4
Fig. 3 Concept of random walk optimization
Numerical Application
The first application presented is extremely simple and aims at

illustrating the overall performance of the model screening procedure. Consider an output variable y defined by the following
input-output model:
y2 sin 2t 3 cos t 1.5 sin 3t cos 2t
(13)
where t is an input variable that varies from zero to fifty with

increments of t0.05. It is assumed that the model form shown
in Eq. 13 is unknown. Instead, observations y k y(kt), for k
0 . . . 100, are obtained and the problem consists in identifying
the numerical model that best matches the observed data. It is
emphasized that, in this numerical simulation, no actual experiment is performed. The continuous solution 13 is shown in Fig.
5 with a solid line. The hexagram symbols represent the discrete
samples assumed to be collected.
Next, consider a set of candidate predictors:
Fig. 4 Illustration of MCMC and Gibbs random walk sampling

strategies
defined family of models, their goodness-of-fit indicators and the

posterior probabilities PDF( y) of effects involved in the most
likely models.
To do so, however, the unknown posterior probability function
must be sampled. The problem of exploring an unknown PDF is
solved with the Markov Chain Monte Carlo algorithm. The
MCMC sampling is advantageous in this situation because it can
sample any distribution, whether it is Gaussian or not. The
MCMC sampling can be viewed conceptually as an optimization
solver that performs a random walk through the optimization
space. This concept is illustrated in Fig. 3 where points in the
optimization space are sequentially visited. More appropriate solutions are guaranteed more frequent visits because the acceptance
criterion of a given solution is based on its likelihood function.
Each candidate point in the design spacehere, the design
space is the horizontal plane of potential models illustrated in
Figs. 1 and 3is accepted or rejected based on its value of the
likelihood function 11 and a Chi-square test. This particular acceptance criterion implies that inappropriate models have a small
chance of being accepted just like appropriate models have a
small chance of rejection. If rejected, a new point is randomly
selected in the neighborhood of the last accepted point. The sequence of points accepted is stored to estimate, once the process
has been completed, the probability of occurrence of each model.
The sampling procedure used in this work is the Gibbs sampling, the simplest of the many variants of the MCMC algorithm.
392 Vol. 125, JULY 2003
x 1 sin t
x 2 cos t
x 3 sin 2t
x 4 cos 2t
x 5 sin 3t
x 6 cos 3t
(14)
In addition to the six predictors of Eq. 14, six other predictors

labeled x 7 , x 8 , x 9 , x 10 , x 11 and x 12 are defined as random functions. It can be observed that, if the functional form of the output
variable y were known, it could be written as:
y3x 2 2x 3 1.5x 4 x 5
(15)
Fig. 5 Simulated non-linear function 13
Table 1 Top five models and number of appearances

Model
Effects
Posterior Probability
1
2
3
4
5
x 2 , x 3 , x 4x 5
x 2 , x 3 , x 4 x 5 , x 10
x 2 , x 3 , x 4x 5 , x 3x 4
x 2 , x 3 , x 4 x 5 , x 2 x 10
x 2 , x 3 , x 4x 5 , x 2x 5
52.0%
3.0%
2.0%
2.0%
2.0%
Clearly, y does not depend on predictors x 1 , x 6 , x 7 , x 8 , x 9 , x 10 ,

x 11 and x 12 . The objective of model screening is to identify the
model form 15. Equivalently, it can be stated that the objective
of model screening is to identify the linear effects x 2 , x 3 and the
linear interaction effect x 4 x 5 from all the potential combinations
defined by the family of models considered.
The family of models defined for this illustration is composed
of the linear models that include the twelve linear effects x i and
the linear interaction models, defined as the previous models augmented with the 66 interaction effects x i x j . The total number of
different effects x i and x i x j with twelve predictors is therefore
equal to 78. The total number of different models that can be
defined belonging to this family by combining the 78 effects is in
excess of 3.02 1023 models, a number that approaches the number of atoms in the known Universe. Clearly, exploring such a
large number of combinations without focusing on the models of
highest likelihood would not be feasible.
The procedure described in the foregoing section is applied to
the data using 50 samples dedicated to the initialization of the
Gibbs sampler and 100 samples for the computation. Initializing
the Markov chain is referred to as burn-in and guarantees that
the remainder of the chain is not biased due to a particular choice
of starting point. The samples drawn during burn-in are disregarded and only the 100 samples drawn during the optimization
itself are kept to estimate the final probability of occurrence of
each model in the family. The top five models are listed in Table
1. It can be observed that the best model in terms of posterior
model probability is the actual model. The mean-square error for
the top five models is about 0.003%. This means that it is not
necessary to include other terms than the ones present in the best
model.
Figure 6 represents the marginal posterior probability of each
effect being in a particular model. The prior probabilitiesthat
reflect the prior knowledgeare set to 25% for each linear effect
x i ; 10% for the interaction effects x i x j if one of the parent effect
x i or x j is selected in the model; and 1% only for the interaction
effects x i x j when neither x i nor x j are considered in the model.
Fig. 6 Marginal posterior probability of each effect included in

the family of models
These uniform probabilities reflect the fact that little is known

about the form of the model before starting the analysis. It can be
observed that effects 2, 3 and 43, namely x 2 , x 3 and x 4 x 5 , are
associated with a probability of 100% while the other effects may
be ignored because their posterior probabilities are reduced to
insignificant levels.
In conclusion, the Bayesian model screening clearly suggests a
model that includes the three effects x 2 , x 3 and x 4 x 5 . The identified coefficients corresponding to these effects are equal to 2.99,
2.02 and 1.52, respectively, and they are in good agreement with
the actual coefficients shown in Eq. 15. The algorithm is implemented as interpreted Matlab functions and it performs the
analysis in a few seconds of CPU time with a typical desktop
personal computer.
Experimental Application
In this Section, Bayesian model screening is applied to the

problem of identifying the form of a nonlinear model using real,
experimental data. The analyzed data sets are chosen from those
proposed by the VTT Technical Research Center of Finland within
the framework of the European COST action F3 working group
on Identification of Nonlinear Systems 34.
The structure investigated consists of wire rope isolators
mounted between the load mass and the base mass, as shown in
Fig. 7. The load mass acts like a free inertial mass. The motion
and forces experienced by the isolators are measured. In particular, the acceleration responses x 2 and x 1b of the load mass and
bottom plate, the applied force f and the relative displacement x 12
between the top and bottom plates are measured. The excitation
produced by an electro-dynamic shaker corresponds to a white
noise sequence, low-pass filtered at 400 Hertz. What makes this
system interesting for identification is that the attenuation of the
vibration across the interface is difficult to characterize because
the mechanics of the isolators is unknown to a large extent. Significant nonlinear dynamics are expected due to the geometrical
nonlinearitypre-loading in the wire rope isolators changes with
the load mass.
Fig. 7 Wire rope isolators
JULY 2003, Vol. 125 393
Table 2 Testing matrix of the VTT benchmark

Forcing Level
Level
Level
Level
Level
1
2
3
4
0.5
2.0
4.0
8.0
Mass 1 2.2 kg
volt
volt
volt
volt
Test
Test
Test
Test
1
2
3
4
Mass 2 5.8 kg
Test 5
f x,x
Four excitation levels are considered ranging from 0.5 Volt up

to 8.0 Volt. A nominal series of four tests are performed with a
load mass of 2.2 kg. A fifth test is also carried out with the heavier
load mass of 5.8 kg. Table 2 defines the testing matrix from which
data sets have been collected.
Reference 35 discusses the identification of the VTT benchmark structure using the RFS. The main idea behind the RFS
method is briefly overviewed to explain the system identification
approach and the reader is referred to Reference 35 for more
details.
The derivation of the main equations of the RFS method starts
by writing Newtons second law for the load mass m 2 , which
yields:
m 2 x 2 f NL x 2 x 1b ,x 2 x 1b 0
(16)
where f NL denotes the nonlinear internal force. Clearly, the force

f NL is unknown but it can be ascertained, as shown in Eq. 16,
that its value depends on the displacement and velocity of the load
mass relative to those of the bottom plate. Introducing the relative
displacement, x 12x 2 x 1b , Eq. 16 becomes:
f NL x 12 ,x 12 m 2 x 1b m 2 x 12
The next step of the RFS method is to describe the restoring

force by means of a mathematical model. This is achieved through
model fitting such as described in Section 2. The generic form of
models sought is usually given by:
(17)
Equation 17 can be viewed as describing the response of a

SDOF system subjected to a base acceleration. Because the acceleration signals shown in the right-hand side of Eq. 17 are measured and the mass is known, it is possible to compute the restoring force f NL at each instant from Eq. 17.
The value of the restoring force is shown in Fig. 8 in the four
cases where the load mass is equal to 2.2 kg Tests 1 4. At low
excitation level, the systems behavior is predominantly linear because the restoring force varies linearly with the displacement, as
can be observed for the 0.5 Volt and 2.0 Volt levels. As the excitation level is increased, a softening stiffness nonlinearity appears,
as can be observed from the 4.0 Volt and 8.0 Volt levels.
i0 j0
i jx
i j
(18)
To resolve the problem of order determination, which refers to the

identification of the most appropriate dimensions m, n in Eq. 18,
an over-determined system of linear equations is formed with the
available restoring force data. The singular value decomposition is
then used to select the appropriate order. Reference 35 details
the identification procedure and shows that the final model includes a linear stiffness term, a viscous damping term and a nonlinear stiffness contribution:
f NL x 12 ,x 12 k l x 12c l x 12k nl x 12 sign x 12
(19)
where the coefficients k l , c l , k nl and identified with the RFS

method and singular value decomposition are listed in Table 3.
These results are used in the remainder as the reference through
which the performance of the Bayesian model screening is
assessed.
The final model features a mean square error MSE equal to
2.11%, which indicates an excellent correlation to test data. The
MSE indicator is a normalized metric that measures the goodnessof-fit between model predictions and physical observations. It is
defined as:
MSE
100
N 2y k1 . . . N
y k x Tk 2
(20)
where, to comply with notations introduced in Section 2, y k represents the available restoring force data and y is the standard
deviation of data y k . The vector collects the coefficients k l , c l
and k nl , assuming that the exponent is known and equal to 1.5,
and the vector x k collects the corresponding effects in Eq. 19.
The exercise of identifying the most appropriate model form is
now repeated with the Bayesian model screening. First, three effects are defined in agreement with Eq. 19. They are the linear
stiffness x 12 , linear damping x 12 and nonlinear stiffness
x 12 sign(x 12). The corresponding coefficients are denoted by k l ,
c l and k nl , as before. Second, model-forming rules are defined
which are that main effects and linear interactions between the
main effects are allowed. This means that a total of six effects
leading to sixty two different model forms are allowed. Such combinatorial complexity is trivial compared to the example discussed
in Section 4. The complexity here stems from the fact that real
data sets are analyzed with all the risk of erroneous identification
caused by noisy measurements and signal conditioning issues.
Because the exponent is unknown, the Bayesian model
screening is repeated for several assumed values of . The value
that leads to the smallest MSE is retained. Repeating model
screening could become CPU-time intensive if long MCMC
chains are requested for each analysis. For this application, an
initial chain of length 50 is dedicated to burn-in and a chain of
length 300 is requested for the optimization. It has been verified
that requesting more samples does not improve the quality of the
final results. Figure 9 shows the evolution of the MSE as a function of the exponent . The minimum value is obtained for
Table 3 RFS identification of Eq. 19

Coefficient
Fig. 8 Estimation of the restoring force at the four levels 0.5

Volt, 2.0 Volt, 4.0 Volt and 8.0 Volt
394 Vol. 125, JULY 2003
kl
cl
k nl
Value
6
1.09 10
183.44
8.52 107
1.5
Units
N/m
N.sec/m
N/m1.5
Unitless
Table 5 Bayesian identification of Eq. 19.

Coefficient
kl
cl
k nl
Difference*
Value
6
1.12 10 N/m
198.19 N.sec/m
9.07 107 N/m1.5
1.5
2.75%
8.04%
6.46%
0.00%
* Difference relative to coefficients in Table 3.
Fig. 9 Evolution of the MSE as a function of the non-linearity

exponent
1.5, the same value as the one identified in Reference 35 with

the RFS method and singular value decomposition. It may seem
paradoxical that the MSE greatly increases for 1, i.e., for a
linear model while its value remains low in the neighborhood of
1 e.g., 0.99). The reason is that there is still a slight
curvature for values of different from 1 that can be enhanced by
taking high values of the corresponding nonlinear parameter k nl .
Table 4 and Figure 10 display the top five models and the
marginal posterior probability of each effect, respectively. The
mean square error for each of the top five models is around
2.37%, very similar to the RFS results previously reported. Such
low MSE values indicate that the agreement with experimental
data meets the expected level of accuracy. The most likely model
Table 4 Top five models and number of appearances

Model
Effects
1
2
3
4
5
x 12 ,x 12 , x 12 1.5 sign(x 12)

Model 1x 12 x 12 1.5 sign(x 12)
Model 1x 12 x 12 1.5 sign(x 12)
Model 1x 12x 12
Model 1x 12 x 12 1.5 sign(x 12)
x 12 x 12 1.5 sign(x 12)
Posterior
Probability
86.0%
7.0%
3.0%
2.6%
0.3%
Fig. 10 Marginal posterior probability of each effect included

in the family of models
in Table 4 includes only the main effects x 12 , x 12 and

x 12 1.5sign(x 12) and appears 86% of the time in the Markov chain.
The main effects, labeled 13 in Figure 10, have a posterior
probability of 100% while interaction effects, labeled 4 6, may
be considered negligible because their posterior probability is below 10%. For this application, the prior probabilities were set to a
uniform 20% for the main effects; 10% for an interaction effect
when at least one of the parent main effect is selected in the
model; and 1% only for an interaction effect when none of the
parents are selected. The increase in probability for effects 13 in
Figure 10 and the reduction for effects 4 6 are therefore significant. From these results it can be concluded that a suitable model
for the restoring force is given by Eq. 19 with an exponent equal
to 1.5.
The coefficients k l , c l , k nl and identified with the Bayesian
model screening are listed in Table 5. The last column in Table 5
compares the identification results to those of the RFS method in
Reference 35. To calibrate the models coefficients, the Bayesian
model screening currently relies on the least-squares estimator 6
even though other solvers could be implemented. Although the
true solution is unknown, it can be stated that both methods
provide consistent results because the maximum difference is less
than 10%.
The small differences witnessed between the RFS identification
and Bayesian model screening may be attributed to the different
data sets used. Referring to Table 2, the RFS identification was
conducted using the five combinations of input levels and load
masses Tests 15. The Bayesian model screening is restricted to
four of the five cases, as discussed in Section 6, to provide a
validation of the models predictive accuracy.
Validation of the Identified Model
It was pointed out previously that only four of the five data sets
are considered during the identification. These are Tests 1, 2, 4
and 5. The remaining data set Test 3, 2.2 kg load mass, 4.0 Volt
level is exploited to assess the predictive accuracy of the identified model.
Although it might not yet be the state of the practice in structural system identification, many authors, among whom we cite a
recent discussion of model validation in Reference 36, have emphasized that identified models should be independently validated.
It essentially means that independent experiments or data sets
should be used for model screening and parametric calibration, on
one hand, and model validation and predictive accuracy assessment, on the other hand. The predictive accuracy of a model cannot be objectively assessed over the operational range of interest
as long as the independence between training data and validation
data is not met.
Here, data sets collected during Tests 1, 2, 4 and 5 are used for
model screening and system identification while the data collected
during Test 4 are used for model validation. Essentially, Eq. 19
is evaluated with the coefficients of Table 5 to predict the restoring force. Displacement and velocity time series in Eq. 19 are
estimated from numerical integration of the measured acceleration
signals. The evolution of the predicted restoring force versus time
is then compared to the true restoring force measured during
Test 3. The true restoring force is estimated directly from acceleration measurements, as shown in Eq. 17. In Fig. 11, the two
JULY 2003, Vol. 125 395
References
Fig. 11 Comparison between measured and predicted restoring forces 4 Volt level
time series are compared. It can be observed that the measured

and predicted forces agree to the point where the difference between signals is not visible.
The MSE, that assesses the overall fidelity between measurements and model predictions, is equal to 0.77% when the restoring
force of Test 3 is predicted based on the identified Eq. 19. Based
on the low MSE values obtained during model identification
2.37% and validation 0.77%, it can be stated that the prediction accuracy seems to be within 97%. This means that a prediction should be within 3% of a measurement, should a hypothetical
experiment be conducted. Clearly, the main drawback of our assessment of predictive accuracy is that it is based on a single
analysis. Other validation data sets would be required to reach a
more quantitative statement of accuracy. Nevertheless, this analysis leads to the conclusion that a reliable identification has been
performed over the operational range of interest, that is, within the
ranges of 2.2-to-5.8 kg load mass and 0.5-to-8.0 Volt excitation
level.
Conclusion
Model selection is one of the common issues in the field of

identification of nonlinear systems. A Bayesian inference approach for giving insight into the form of the nonlinearity has
been proposed in this paper. The key advantage of the method is
that a collection of potential models together with their posterior
probability is obtained instead of the single best model. It allows
for more flexibility in deciding the most appropriate model of the
non-linearity. In addition, the marginal posterior probability of
each effect being in a particular model may also be evaluated.
The results obtained using two illustrative examples enable us
to confirm the performance and the utility of the proposed technique. We believe that Bayesian model screening will become an
important addition to the structural dynamicists toolbox.
Acknowledgments
The authors acknowledge the VTT Technical Research Center
of Finland for sharing the Dynamic properties of resilient
mounts benchmark data. The authors are very grateful to Mandy
Cundy, technical staff member at Los Alamos National Laboratory, for helping them with the numerical implementation. Dr.
Kerschen is supported by a grant from the Belgian National Fund
for Scientific Research FNRS, which is gratefully acknowledged.
396 Vol. 125, JULY 2003
1 Masri, S. F., and Caughey, T. K., 1979, A Non-parametric Identification Technique for Non-linear Dynamic Problems, ASME J. Appl. Mech., 46, pp.
433 447.
2 Crawley, E. F., and Aubert, A. C., 1986, Identification of Nonlinear Structural
Elements by Force-state Mapping, AIAA J., 24, pp. 155162.
3 Crawley, E. F., and ODonnel, 1986, Identification of Nonlinear System Parameters in Joints using the Force-state Mapping Technique, AIAA Paper
86 1013, pp. 659 667.
4 Simon, M., and Tomlinson, G. R., 1984, Use of the Hilbert Transform in
Modal Analysis of Linear and Non-linear Structures, J. Sound Vib., 96, pp.
421 436.
5 Feldman, M., 1994, Nonlinear System Vibration Analysis using the Hilbert
Transform-I. Free Vibration Analysis Method FREEVIB, Mech. Syst. Signal Process., 8, pp. 119127.
6 Feldman, M., 1994, Nonlinear System Vibration Analysis using the Hilbert
Transform-II. Forced Vibration Analysis Method FORCEVIB, Mech. Syst.
Signal Process., 8, pp. 309318.
7 Gifford, S. J., and Tomlinson, G. R., 1989, Recent Advances in the Application of Functional Series to Non-linear Structures, J. Sound Vib., 135, pp.
289317.
8 Leontaritis, I. J., and Billings, S. A., 1985, Input-Output Parametric Models
for Non-linear Systems: Part IDeterministic Non-linear Systems, Int. J.
Control, 41, pp. 303328.
9 Leontaritis, I. J., and Billings, S. A., 1985, Input-Output Parametric Models
for Non-linear Systems: Part IIStochastic Non-linear Systems, Int. J. Control, 41, pp. 329344.
10 Bendat, J. S., and Piersol, A. G., 1980, Engineering Applications of Correlation and Spectral Analysis, John Wiley & Sons, New-York U.S.A..
11 Choi, D., Chang, J., Stearman, R. O., and Powers, E. J., 1984, Bispectral
Interaction of Nonlinear Mode Interactions, Proceedings of the 2nd International Modal Analysis Conference, Orlando, FL, pp. 602 609.
12 Worden, K., and Tomlinson, G. R., 2001. Non-linearity in Structural Dynamics: Detection, Identification and Modelling, Institute of Physics Publishing,
Bristol, Philadelphia PA.
13 Rice, H. J., and Fitzpatrick, J. A., 1991, A Procedure for the Identification of
Linear and Non-linear Multi-degree-of-freedom Systems, J. Sound Vib., 149,
pp. 397 411.
14 Rice, H. J., and Fitzpatrick, J. A., 1991, The Measurement of Non-linear
Damping in Single-degree-of-freedom Systems, ASME J. Vibr. Acoust., 113,
pp. 132140.
15 Esmonde, H., Fitzpatrick, J. A., Rice, H. J., and Axisa, F., 1992, Modelling
and Identification of Non-linear Squeeze Film Dynamics, J. Fluids Struct., 6,
pp. 223248.
16 Richards, C. M., and Singh, R., 1998, Identification of Multi-Degree-ofFreedom Non-linear Systems Under Random Excitations by the Reverse-Path
Spectral Method, J. Sound Vib., 213, pp. 673708.
17 Bendat, J. S., 1990, Nonlinear System Analysis and Identification from Random Data, John Wiley & Sons, New-York U.S.A..
18 Adams, D. E., and Allemang, R. J., 1999, Characterization of Nonlinear
Vibrating Systems using Internal Feeback and Frequency Response Modulation, ASME J. Vibr. Acoust., 121, pp. 495500.
19 Adams, D. E., and Allemang, R. J., 2000, A Frequency Domain Method for
Estimating the Parameters of a Non-Linear Structural Dynamic Model through
Feedback, Mech. Syst. Signal Process., 14, pp. 637 656.
20 Hemez, F. M., and Doebling, S. W., 2001, Review and Assessment of Model
Updating for Non-linear, Transient Dynamics, Mech. Syst. Signal Process.,
15, pp. 4574.
21 Lenaerts, V., Kerschen, G., and Golinval, J.-C., 2001, Proper Orthogonal
Decomposition for Model Updating of Non-linear Mechanical Systems,
Mech. Syst. Signal Process., 15, pp. 31 43.
22 Dippery, K. D., and Smith, S. W., 1998, An Optimal Control Approach to
Nonlinear System Identification, Proceedings of the 16th International Modal
Analysis Conference, Santa Barbara CA, pp. 637 643.
23 Bendat, J. S., 1986, Random Data: Analysis and Measurement Procedures,
John Wiley-Interscience, New York NY, second edition.
24 Chipman, H., Hamada, M., and Wu, C. F. J., 1997, A Bayesian Variable
Selection Approach for Analyzing Designed Experiments With Complex Aliasing, Technometrics, 39, pp. 372381.
25 Wu, C. F. J., and Hamada, M., 2000, Experiments: Planning, Analysis, and
Parameter Design Optimization, Wiley, New-York, NY.
26 George, E. I., and McCulloch, R. E., 1993, Variable Selection via Gibbs
Sampling, Journal of the American Statistical Society, 88, pp. 881 889.
27 Carlin, B. P., and Chib, S., 1995, Bayesian Model Choice via Markov Chain
Monte Carlo, Journal of the Royal Statistical Society Series B, 77, pp. 473
484.
28 Kulczycki, P., 2001, An Algorithm for Bayes Parameter Estimation, ASME
J. Dyn. Syst., Meas., Control, 123, pp. 611 614.
29 Yuen, K. V., Katafygiotis, L. S., Papadimitriou, C., and Mickleborough, N. C.,
2001, Optimal Sensor Placement Methodology for Identification With Unmeasured Excitation, ASME J. Dyn. Syst., Meas., Control, 123, pp. 677 686.
30 Cybenko, G., 1989, Approximation by Superpositions of a Sigmoidal Function, Mathematics of Control, Signals, and Systems, 2, pp. 303314.
31 Walter, E., and Pronzato, L., 1997, Identification of Parametric Models From
Experimental Data, Springer-Verlag, Berlin Germany.
32 Hasselman, T. K., Anderson, M. C., and Wenshui, G., 1998, Principal Components Analysis for Nonlinear Model Correlation, Updating and Uncertainty
Evaluation, Proceedings of the 16th International Modal Analysis Conference, pp. 664 651.
33 Mottershead, J. E., and Friswell, M. I., 1993, Model Updating in Structural
Dynamics: A Survey, J. Sound Vib., 1622, pp. 347375.
34 http://www.ulg.ac.be/ltas-vis/costf3/costf3.html, portal Web site of the
COST-F3 action.
35 Kerschen, G., Lenaerts, V., Marchesiello, S., and Fasana, A., 2001, A Frequency Domain vs. a Time Domain Identification Technique for Non-linear
Parameters Applied to Wire Rope Isolators, ASME J. Dyn. Syst., Meas.,
Control, 123, pp. 645 650.
36 Doebling, S. W., 2002, Structural Dynamics Model Validation: Pushing the
Envelope, Proceedings of the International Conference on Structural Dynamics ModellingTest, Analysis, Correlation and Validation, Madeira Island
Portugal.
JULY 2003, Vol. 125 397
Journal of
Vibration
and Acoustics
Technical Briefs
Closed-Form Exact Solution to H

Optimization of Dynamic
Vibration Absorbers Application to
Different Transfer Functions
and Damping Systems
standard model of the DVA. In 1928, Ormondroyd and Den Hartog 2 pointed out the damping of the DVA has an optimum value
so as to minimize the maximum amplitude response of the primary system. Such optimization criterion is now known as H
optimization. In 1932, Hahnkamm 3 derived the optimum tuning
parameter opt of the DVA using special points in which all re-
Toshihiko Asami
Mem. ASME
Professor, Department of Mechanical Engineering,
Himeji Institute of Technology, 2167 Shosha, Himeji,
Hyogo 671-2201, Japan
e-mail: asami@mech.eng.himeji-tech.ac.jp
Osamu Nishihara
Mem. ASME
Associate Professor, Department of Systems Science,
Kyoto University, Yoshida-Honmachi, Sakyo-ku, Kyoto
606-8501, Japan
e-mail: nishihara@i.kyoto-u.ac.jp
H optimization of the dynamic vibration absorbers is a classical

optimization problem, and has been already solved more than 50
years ago. It is a well-known solution, but we know this solution is
only an approximate one. Recently, one of the authors has proposed a new method for attaining the H optimization of the
absorber in linear systems. The new method enables us to obtain
the exact algebraic solution of the H optimization problem of the
absorber. In this paper, we first apply this method to the design
optimization of a viscous damped (Voigt type) absorber and a
hysteretic damped absorber attached to undamped primary systems. For each absorber, six different transfer functions are taken
here as performance indices to vibration suppression or isolation.
As a result, we found the closed-form exact solutions to all transfer functions. The solutions obtained here are then compared with
those of the approximate ones. Finally, we present the closed-form
exact solutions to the hysteretic damped absorber attached to
damped primary systems. DOI: 10.1115/1.1569514
Fig. 1 Systems with viscous damped DVA Voigt type DVA a

Force excitation system b Motion excitation system
Table 1 Transfer functions of the systems shown in Figs. 1
and 2
Force excitation system
shown in Fig. 1a
1f
2f
3f
Motion excitation system shown in Fig. 1b

Absolute response
x1
x1
xst
f 0 /k 1
x1
x1
1xst
f 0 /k 1
x1
21xst
x1
f 0 /k 1
Relative response
x1
x0
x 1
x1
1x 0
x0
x 1
21 x 0
x1
x0
21 y 1
1 y1
2
x 0
x0
1y 1 1 y 1
x 0
x0
y1
x0
Introduction
The dynamic vibration absorber DVA, first invented by Frahm

1 in 1909, is a passive vibration control device. As shown in Fig.
1, it essentially consists of a mass, a spring and a dashpot, which
is attached to a primary system subjected to an exciting force or
motion. This type of DVA is called the Voigt type DVA and it is a
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received April 2000;
Revised Nov. 2002. Associate Editor: R. L. Clark.
398 Vol. 125, JULY 2003
Fig. 2 Systems with hysteretic damped DVA a Force excitation system b Motion excitation system
Table 2 Expressions for the transfer functions and definition of the symbols of the viscous
damped systems shown in Fig. 1
No.
Transfer function
N f
D
x1
f 0 /k1
1f
x1
x0
N a
D
y1
x0
N r
D
1 k 1 /m 1
natural frequency
of primary system
2 k 2 /m 2
x1
y 1 x 1 x 0
Expressions
N f ()( 2 2 ) 2 (2 2 ) 2
N a () 1(2 1 ) 2 ( 2 2 ) 2 (2 2 ) 2
N r () (1 ) 2 2 2 4 (1 ) 2 (2 2 ) 2 4
D() (1 2 )( 2 2 ) 2 2 4 1 2 2 2
(2) 2 ( 2 2 ) 1 1(1 ) 2 2 2
/ 1
forced frequency ratio
m 2 /m 1
mass ratio
natural frequency
of absorber
2 /1
natural frequency
ratio
absolute displacement
of primary system
c1
1
2m11
damping ratio
of primary system
relative displacement
between primary system
and foundation
c2
2
2m22
damping ratio
of absorber
sponse curves pass through independent of the absorber damping

c 2 or 2 when the primary system has no damping (c 1 or 1
equals zero. Finally, Brock 4 derived the optimum absorber
damping 2opt in 1946. And now, we all know these optimum
tuning and damping parameters through the textbook written by
Den Hartog 5. Their method utilizing the special points of the
curves is now called the fixed-points method.
Although the fixed-points method gives us a simple and convenient way for solving the H optimization problem of the DVA, it
is only an approximate method. Recently, we proposed a new
method for finding the exact solution to the H optimization 6.
Using this method, we solved a standard H optimization problem, that is, the minimization of the compliance transfer function
of the system with Voigt type DVA. In this paper, we will apply
this method to different transfer functions listed in Table 1. We
also apply this method to optimizing another DVA system shown
in Fig. 2. This type of DVA is called the hysteretic damped DVA.
The principal difference between the viscous and hysteretic damping is that for the viscous damping the energy dissipated per cycle
depends linearly on the frequency of vibration; whereas for the
hysteretic damping it is independent of the frequency.
Table 1 shows the transfer functions taken here as performance
indices to vibration suppression or isolation: No. 1 f through No.
3 f are transfer functions in the force excitation systems shown in
Figs. 1a and 2a, and No. 1 through No. 6 are those in the
motion excitation systems shown in Figs. 1b and 2b. Transfer
functions No. 1 through No. 3 represent absolute responses of the
primary system and No. 4 through No. 6 represent relative responses between the primary system and the base. In this table,
dot over a symbol indicates the derivative with respect to time.
Other symbols are explained in Table 2.
The exact solutions obtained here are compared with the approximate ones reported by Korenev and Reznikov 7.
Fig. 3 Frequency response curves of the undamped primary

system with a viscous damped absorber graphical representation of transfer function No. 1 a General view of the curves
b Close up of the curves near the points P and Q
Fixed-Points Approach as an Approximate Method
Figure 3a shows some frequency response curves of the undamped primary system ( 1 0) with an optimally tuned DVA
JULY 2003, Vol. 125 399
Table 3 Approximate solutions to the H optimization in the design of viscous damped absorber attached to undamped primary
system derived from the fixed-points method
No.
1
Transfer
function
Optimum tuning opt

x1
x0
N a
D
x 1
x1
1x 0
x0
x 1
21 x 0
x1
x0
21 y 1
1 y1
2
x 0
x0

1y 1 1 y 1
x 0
x0
y1
x0
N r
D
Optimum damping ratios P and Q
1
1
1
1
2
2
1
1
2
2
1
1
8 1
1
4 2
2
2
2 1
4 1 2
8 1
3 Approximate and Exact Solutions to the H Optimization

3.1 Case 1: Viscous Damped Absorber. Table 3 shows the
solutions to the H optimization design of the viscous damped
absorber obtained from the fixed-points method. All solutions are
expressed by one parameter, mass ratio . As mentioned previously, the optimum tuning parameter opt described here is such
that the two fixed points P and Q have equal heights; the abscissas
and heights of them are shown in this table. They are approximate
solutions except for No. 5, since the two damping ratios P and Q
are different from each other. Of the double sign of the expressions the upper sign represents P and the lower Q . No. 5 is the
only exact solution obtained from the fixed-points method.
We propose the exact solution shown in Table 4. The procedure
for obtaining the solution is refer to 6. The expressions are more
complicated than those in Table 3, but most of them are easy to
calculate on the hand-held calculators. The calculation examples
for a mass ratio 0.1 are shown in Fig. 4, where the exact
solution is drawn with a solid line and the approximate one with a
dashed line. It may be observed from this figure that the two
43 85
2 1 2
1
2
1
4 8
3
8 1
whose mass is one-tenth of the main mass i.e., 0.1. Three

response curves corresponding to zero, optimum and infinite absorber damping 2 are drawn in this figure. If there is no primary
damping, then all response curves pass through two fixed points,
named P and Q, as shown in Fig. 3a. It has been observed that
there is a trade-off relationship between the heights of the fixed
points. The fixed-points method is essentially based on this observation. On the basis of this method, one can find the solution such
that first by a proper choice for the two fixed points P and Q are
adjusted to equal heights, and then by a proper choice for 2 the
response curve passes simultaneously with a horizontal tangent
through the two fixed points. However, the special values of 2 ,
say P and Q , for which the curve passes horizontally through
the points P and Q, respectively, are generally not equal to each
other see Fig. 3b. The average of them is usually adopted as an
optimum damping 2opt , but the two resonant amplitudes of the
curve are not exactly equal in this case, as shown in Fig. 3b.
Hence we recognize that, in general, we cannot reach the exact
solution to the H optimization by the fixed-points method.
400 Vol. 125, JULY 2003
2424 5 2 4 85
1
1
1
1
Abscissas P and Q
of the fixed points
4 1
2 1
Height of the
fixed points
2 1
resonant amplitudes are adjusted to equal heights in the exact

solution and they are smaller than the higher resonant amplitude
in the approximate solution.
3.2 Case 2: Hysteretic Damped Absorber. Similarly,
Tables 5 and 6 show the approximate and exact solutions of the
H optimization design of the hysteretic damped absorber. This
time, the expressions based on our method are as simple as those
of the fixed-points theory, except for No. 2 and 5. Figure 5 compares the approximate and exact solutions for the condition that
0.1. This figure reconfirms us in the view that the two resonant
amplitudes of the system are equal in heights in the exact solution.
In comparison with Figs. 4 and 5, it can be seen that the optimized
maximum amplitude responses of the system with hysteretic
damped absorber is greater than those of the system with viscous
damped absorber. This means that the hysteretic damped absorber
will give a poor performance than the viscous damped absorber
when they are equal in size.
Exact Solution to the Damped Primary System
Unlike the fixed-points method, our method is applicable to the

design of DVA attached to the damped primary system. However,
for the viscous damped systems shown in Fig. 1, we could not find
the algebraic solution for all transfer functions defined in Table 1.
On the other hand, for the hysteretic damped systems shown in
Fig. 2, we obtained the closed-form exact solution for the transfer
functions 1 f , 1 and 6. The solution is given in Table 7. We can
demonstrate that if the primary system has no damping or 1
0, then the expressions shown in Table 7 reduce to those in
Table 6. A numerical representation for typical primary damping
1 is given in Table 8. Figure 6 shows an example of the response
curves of the system with optimally tuned and damped DVA. It
may be observed that the two peak values of the curves shown in
solid lines are exactly the same; whereas those of the dashed lines
the solution to the undamped primary system are different. It is
interesting to note that the optimum tuning and damping parameters, opt and 2opt of the force excitation system are the same as
those of the motion excitation system compare the expressions of
No. 1 f and No. 1.
Table 4 Exact solutions to the H optimization in the design of viscous damped absorber attached to undamped primary system derived from our method
No.
1
Transfer
function
x1
x0
2
85
21 x 0

21 y 1
x 0
1
r1
r2
16
1
r3
r4
16
JULY 2003, Vol. 125 401
3 6480 27 2
x 1
1
85
2
167 6416 26 2
3
1
2r 1 16r 21 4r 1 2
6
p 1 q 1
p 1 q 1
2q 1 p 1
2q 1 p 1
4 q 0
4 q 0
1 2 1 2
1
2r 1
p 1 q 1
p 1 q 1
1
2
1
2 1
1
169 4 43
6
Maximum amplitude
89 4 43
89 2 169 128 43 3/2
2
2 1
1625 2 6416 26 2
3 3227
1
1r 21
p 0 262144196608 233472 2 598016 3 448320 4 120624 5

6371 6 486 3/2 6456 13 2 3/2 2 3240 2
1/3
p 1 p 1/3
87 512192 168 2 31 3
0 p 0
q 0 9672 13 2
q 1 12832 19 2
2 1
1r 1
2r 1 16r 21 4r 1 2
1
3 3227
1
2 1 2
6488 33 2
43 c 0
2
c 0 6416 26 2
1
1
1y 1
x 0
y1
x0
Optimum damping 2opt
2 1623 9 2 2 2 43
x 1
1x 0
1
1
Optimum tuning opt
3
8 1
89 4 43
1
1
3
2 1
89 2 169 128 43 3/2

3 3227
Fig. 4 Comparison between exact and approximate solutions

to the H optimization of the system with a viscous damped
DVA a Transfer function No. 1 b Transfer function No. 4
Fig. 5 Comparison between exact and approximate solutions

to the H optimization of the system with a hysteretic damped
DVA a Transfer function No. 1 b Transfer function No. 4
Table 5 Approximate solutions to the H optimization in the design of hysteretic damped absorber attached to undamped
primary system derived from the fixed-points method
No.
1
Transfer
function
Optimum tuning
opt

N a
D
x1
x0
x 1
x1
1x 0
x0
x 1
21 x 0
21 y 1
x 0
x1
x0
1 y1
2 x0

1y 1 1 y 1
x 0
x0
y1
x0
N r
D
1
1
1
1
1
1
2
2
1
1
2
2
1
1
1
1
402 Vol. 125, JULY 2003
2
2
Optimum loss factors P and Q
32 2 2
2450 30 2 5 3 2 55 2 85
2
1
2
3 23 2 2 1
6 2 2
2
3 4 8
8
32 2 2
2
Abscissas P and Q
of the fixed points
Height of the
fixed points
2
1
43 85
2 1 2
2
1
2 1
4 8
4 1
2 1
JULY 2003, Vol. 125 403
Coefficients
y1
x0
1 y 1
x0
x 0
21 y 1
21 x 0
x 1
x1
1x0
1 2
q1 1 22opt 2 1 22opt
2
2
q1 1 22opt2 1
2a
b b 2 4ac
3
2
1
q3
2 1
3
2
2a
b b 2 4ac
2 1
1
q3
4 1
2 2
Maximum amplitude
3 8
2 2
3 8
p 2
2p 1
2 1 2 12
3
p
2 0
c (c 0 q 3 c 1 q 2 c 2 qc 3 )
b2(b 0 q 3 b 1 q 2 b 2 qb 3 )
aa 0 q 3 a 1 q 2 a 2 qa 3
p1
p 2 2 1 2
p 1 2(1 2 )3p 0
p 0 2 2 (1 ) 2 1/3
Nomenclature
a 0 512094464 50688 2 281728 3 129492 4 106667 5 43693 6 10176 7 1024 8

a 1 6144156160 95616 2 913632 3 449296 4 667848 5 381560 6 146199 7 32064 8 3072 9
a 2 23552301824 397568 2 507584 3 776140 4 227683 5 727803 6 399091 7 156903 8 33600 9 3072 10
a 3 12288123904 191872 2 494176 3 548168 4 778822 5 643981 6 495675 7 193878 8 61309 9 11712 101024 11
b 0 102437376 36224 2 56976 3 7868 4 22581 5 18779 6 7040 7 1024 8
b 1 102462208 66432 2 250672 3 5564 4 41644 5 93648 6 66753 7 22656 8 3072 9
b 2 5120124928 104064 2 434608 3 32052 4 18799 5 28627 6 89049 7 72753 8 24192 9 3072 10
b 3 307249920 57472 2 102288 3 1140 4 29134 5 93055 6 72685 7 66494 8 31691 9 8576 101024 11
c 0 1715212288 172624 2 160188 3 161061 4 84723 5 24768 6 3072 7
c 1 2124825664 446624 2 517192 3 800496 4 602424 5 290889 6 78912 7 9216 8
c 2 77568132736 227504 2 651828 3 200997 4 713205 5 614421 6 314361 7 83520 8 9216 9
c 3 3(1305629888 96768 2 184128 3 286918 4 289257 5 224231 6 114406 7 42977 8 9792 9 1024 10)
1
1
2 2 2
4
q1 1 22opt 2 1 1 22opt
q1 1 22opt2
1
1
x1
x0
Optimum tuning opt
T.F.
No.
Table 6 Exact solutions to the H optimization in the design of hysteretic damped absorber attached to undamped primary system derived from our method
for 1 and 6
ae 0 e 1 e 23
be 3 3e 1 e 2
ce 1 e 2 3e 23
2 33 2 2 3 4 2 1 21 q 1 1 22opt
6 1
1 1 21
6 1 1 22opt
1 1
x1
x0
y1
x0
404 Vol. 125, JULY 2003
Expressions
x1
f 0 /k1
1f
for 1 f and 1
for 6
b b 2 4ac
2a
2 33 2 2 3 4 2 1 21 q 1
e 0 4 2 q 21 8 q 1 36 6 2 6 3 4 (1 )(3 2 ) 21 4(36 6 2 6 3 4 ) 2
16 2 (1 )(96 12 2 18 3 3 4 4 5 ) 21 32 4 (1 ) 2 (3 2 ) 41
e 1 (1 ) 2 q 1 2 (3 ) 2 4 2 (1 ) 21 q 1 2 (36 2 )4 2 (1 ) 21
e 2 2(1 ) q 21 q 1 318 6 2 6 3 4 2 2 (1 )(3 ) 21
2 (3 )(915 18 2 6 3 3 4 5 )4 2 (1 )(69 3 2 9 3 7 4 ) 21 8 4 (1 ) 2 (3 ) 41
e 3 12 (1 )(1 ) 2 1 q 1 2(33 2 2 3 )4 2 (1 ) 21
p 0 (3 ) 4 4(1 )(3 )(93 2 2 ) 21 4 2 (1 ) 2 41
6
p 1 (3 ) 3(1 )(3 ) 3 (99 21 2 5 3 ) 21 12 2 (1 ) 2 (453 2 2 3 ) 41 8 4 (1 ) 3 61
q 0 2 p 1 3(1 ) 2 1 3 2 (1 ) 21 (3 ) 3 8 2 21 3
2p0
q1 1/3 q 1/3
0
q0
6 1 1 21
Primary
damping 1
q 1 2 36 2 2 1 3 2 21
for 1 f
6 1
q 1 2 36 2 2 1 3 2 21
Maximum amplitude h max

Optimum tuning opt
No.
Transfer function
Table 7 Exact solutions to the optimization problem in the design of hysteretic damped absorber attached to damped primary system
Table 8 Numerical exact solutions for damped primary system

Transfer function is x 1 x 0 and mass ratio is equal to 0.1
0.00
0.01
0.02
0.05
0.10
0.15
0.20
0.50
Optimum
tuning opt
Optimum loss
factor 2opt
Maximum
amplitude
0.909091
0.907941
0.906801
0.903437
0.898000
0.892754
0.887677
0.859966
0.393700
0.395643
0.397566
0.403232
0.412355
0.421124
0.429584
0.475493
4.69042
4.51256
4.34740
3.91648
3.36195
2.94908
2.63249
1.68819
Fig. 6 Absolute displacement response of the damped primary system with optimally tuned and damped DVA
Concluding Remarks
The H optimization of a dynamic vibration absorber attached

to a linear primary system is a classical optimization problem.
Although the system is very simple, it is difficult to find the exact
solution. The conventional method, well-known as the fixedpoints method, is only an approximate one to solve the problem.
In 2002, we proposed a new method to find the exact solution to
the H optimization.
First in this paper, we applied this method to solve the H
optimization problem to different transfer functions and DVAs
viscous damped DVA and hysteretic damped DVA under the
condition that the primary system has no damping, and found
closed-form exact solutions for all transfer functions. Some of the
solutions are simple and others complicated, but we can handle
them by the hand-held calculators. And now, we can evaluate the
accuracy of the conventional solutions by comparing them with
our solutions. From the comparison of these solutions we see that
the difference between the exact and approximate solutions is negligibly small in the practical range of mass ratio is less than
about 0.2. Nevertheless, we believe that the conventional expressions shown in Tables 3 and 5 should be replaced by the new
expressions in Tables 4 and 6, respectively, because the latter are
exact solutions.
Next, we tried to solve the same optimization problem for a
generalized system where damping is present in the primary system, and found the closed-form exact solutions for the hysteretic
damped system.
The last unsolved problem is the H optimization of the viscous damped dynamic vibration absorber attached to damped primary systems. However, the algebraic approach is probably impossible to this system. We proposes simultaneous equations for
calculating numerically the exact solution 6.
References
1 Frahm, H., 1911, Device for Damping Vibrations of Bodies, U.S. Patent,
No. 989, 958, pp. 3576 3580.
2 Ormondroyd, J., and Den Hartog, J. P., 1928, The Theory of the Dynamic
Vibration Absorber, ASME J. Appl. Mech., 507, pp. 922.
3 Hahnkamm, E., 1932, Die Dampfung von Fundamentschwingungen bei veranderlicher Erregergrequenz, Ing. Arch., 4, pp. 192201, in German.
4 Brock, J. E., 1946, A note on the Damped Vibration Absorber, ASME J.
Appl. Mech., 134, p. A-284.
5 Den Hartog, J. P., 1956, Mechanical Vibrations, 4th ed., McGraw-Hill, New
York.
6 Nishihara, O., and Asami, T., 2002, Closed-Form Solutions to the Exact
Optimizations of Dynamic Vibration Absorbers Minimizations of the Maximum Amplitude Magnification Factors, ASME J. Vibr. Acoust., 1244, pp.
576 582.
7 Korenev, B. G., and Reznikov, L. M., 1993, Dynamic Vibration Absorbers, p.
11, John Wiley & Sons, New York.
Determining Stability Boundaries

Using Gyroscopic Eigenfunctions
Problem Formulation and Initial Analysis
Consider an equation of motion describing parametric excitation of the form

M u ,tt Gu ,t Kup t M u ,tt G u ,t K u 0
where M, G, K, M , G , and K are linear, real, spatial differential or matrix operators, u(x,t) is the system response, x is the
spatial coordinate, t is time, 1 is a small parameter, p(t) is a
periodic function in time, and a comma indicates partial differentiation. The operators M, G, and K are associated with a complex
inner product , and enjoy the properties
M 0

in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Aug. 2002;
Revised Jan. 2003. Associate Editor: L. A. Bergman.
(3)
d n t n
(4)
can always be constrained such that 3
n1, . . .
d n,t n
n1, . . .
i nd n n
(5)
This last constraint 5 enables one to decouple the unperturbed

modal equations 3. Substituting Eq. 4 into 1, taking the inner
product with i m m and utilizing Eq. 5 gives a system of
modal equations in which the left hand side, unperturbed equations are a set of first order, decoupled, ordinary differential equations:
d m,t i m d m i m p t
The complex eigenfunctions of positive definite, linear, gyroscopic systems possess a number of properties that make them
advantageous and efficient in analytic methods. First and foremost, these eigenfunctions can be used to decouple system equations, thus providing exact modal solutions even for continuous
systems 13. These eigenfunctions can also be used in perturbation studies 4. Unfortunately, these eigenfunctions do not appear to offer meaningful advantages for numerical methods 57.
In this Tech Brief, we demonstrate that complex gyroscopic
eigenfunctions are particularly efficient for determining the stability boundaries of parametrically excited systems using the method
of multiple scales. This efficiency is derived from two properties:
decoupling and reduction of order. Our contribution here is one of
presentation and the examples; we make no claim of originality
for the method which has been used in previous parametric stability studies 8 10. Our aim here is to present the analysis in as
general and simple a manner as possible so that the efficiency of
the procedure is obvious and its application is accessible to those
unfamiliar with it. To that end, the procedure is used to quickly
analyze three parametrically excited systems. The efficiency of the
procedure is evident when our brief analyses are compared with
the alternative approaches given in the literature.
(2)
where nm is the Kronecker delta. The eigenexpansion
n1,2, . . .
Introduction
K * K
n m M n , m K n , m nm
By taking advantage of modal decoupling and reduction of order,

we derive a simplified procedure for applying the method of
multiple scales to determine the stability boundaries of parametrically excited, gyroscopic systems. The analytic advantages of the
procedure are illustrated with three examples.
DOI: 10.1115/1.1569944
G * G
where the * indicates the adjoint. In other words, the unperturbed

0 problem is a positive definite gyroscopic system. The
eigensolutions associated with the unperturbed system are
i n , n where, due to 2, all the eigenvalues are imaginary,
i.e., n is real. These solutions are numbered such that n1,
where the overbar
2,3 . . . and i n , n i n ,
n
denotes complex conjugation.
The adjoint eigenfunction is the eigenfunction itself; hence
eigensolutions can be normalized such that
u ,t
M * M
K0
Anthony A. Renshaw
Associate Professor, Dept. of Mechanical Engng.
Columbia University, New York, NY 10027
(1)
d n,t G nm
d n K nm
d n,tt M nm
(6)
M n , m , etc.
where M nm
The parametric instability boundaries are determined from Eq.
6 by applying the method of multiple scales. Define the time
scales T 0 t and T 1 t and let d n (t)D n0 (T 0 ,T 1 )
D n1 (T 0 ,T 1 ) . . . . Substitution into Eq. 6 gives the leading
order equation
D m0,T 0 i m D m0 0
(7)
D m0 A m0 T 1 e i m T 0
(8)
whose solution is
The unknown coefficient A m0 (T 1 ) is determined by eliminating

secular terms of the form e i m T 0 from the right hand side of the
order equation, given by
D m1,T 0 i m D m1 A m0,T 1 e i m T 0 i m p T 0
A
n
n0 e
inT0
nm
(9)
where
i n G nm
K nm
nm 2n M nm
(10)
After this step, the procedure varies from problem to problem; we

illustrate different approaches with the following examples.
JULY 2003, Vol. 125 405
Example 1. Classical Parametric Resonance

For the classical case p(t)cos(t). The order equations
become
D m1,T 0 i m D m1 A m0,T 1 e i m T 0
e iT 0
A
n
1
i e iT 0
2 m
n0 e
inT0
nm
(12)
k,n
A kn0 e i kn T 0 R kn r 0 R jm r 0 e i jk T 0
(13)
1
i A e i T 1 mn 0
2 n m0
(14)
(21)
The detuning parameter for a two mode secular combination of
terms on the right hand side is
jm kn jk
(22)
giving the equations

A jm0,T 1 i jm R 2jm r 0 A jm0
i jm R jm r 0 R kn r 0 e i T 1 A kn0 0
1
A m0,T 1 i m A n0 e i T 1 nm 0
2
(23)
A kn0,T 1 i kn R 2kn r 0 A kn0

i kn R kn r 0 R jm r 0 e i T 1 A jm0 0
(24)
Assuming solutions of the form
whose solution has the form

A m0 a m e T 1
(11)
for some n and m, where, by assumption, is real. Eliminating the

secular terms requires both
A n0,T 1
A jm0,T 1 e i jm T 0
i jm
When the first term on the right hand side is the only secular term,
only bounded solutions result. We therefore look for combinations
of two secular terms. Define the detuning parameter by
m n
D jm1,T 0 i jm D jm1
A n0 a n e i T 1
(15)
A jm0 a jm e T 1
A kn0 a kn e i T 1
(25)
and solving, we obtain unstable solutions whenever
Solving these gives

i
2 n m nm mn 1/2
2
(16)
The first order stability boundary occurs when changes from

purely imaginary to complex. In other words,
m n n m nm mn
(17)
when the factor under the square root is positive. If it is not

positive, the unstable region vanishes. Primary instability occurs
for m n . Combination resonance of the summation type
occur when n m 0; this is not possible if nm mn 0 and real,
i.e., when M and K are skew symmetric and G is symmetric.
Combination resonance of the difference type occur when n m
0; this is not possible when nm mn 0 and real, i.e., when M
and K are symmetric and G is skew symmetric.
Example 2. A Stationary Spring Loading a Rotating Circular Plate
For an axisymmetric rotating circular plate in the rotating frame
of reference 11, the eigensolutions of the unperturbed system are
doubly indexed with eigenvalues i km and eigenfunctions of the
form
km e ik R km r
(18)
where (r, ) are polar coordinates and k and m are integers and
m0. This describes a mode with k nodal diameters and m
nodal circles. We do not need to know the actual forms of M, G,
K, R km , and km in order to analyze stability; these are only
needed if explicit values of the frequencies and boundaries are
required.
Here we consider a perturbation comprising a stationary, concentrated spring transversely loading the rotating plate. In the rotating frame of reference, we have p1 and
K
1
rr 0 t
r
(19)
where is the rotation speed of the plate and r 0 is the radial

location of the spring. M G 0. The inner product is
jm K e ik R kn ,e i j R jm R kn r 0 R jm r 0 e i jk t
K kn
(20)
Equation 9 is therefore
406 Vol. 125, JULY 2003
jm R 2jm r 0 kn R 2kn r 0 2 4 jm R 2jm r 0 0 (26)
The most important instability produced by the spring is divergence instability of a mode with a fixed number of nodal diameters. For this case, we let jm 0 0, kn 0 , jk, and
R jm (r 0 )R kn (r 0 ). The conditions for instability are then
4 0 R 2jm r 0 0
(27)
In other words, for a combination resonance of two modes with

the same number of nodal diameters, the stability boundary is
demarcated by the two lines
0 j
and
0 2 0 R 2jm r 0 j
(28)
The system is unstable at rotation speeds immediately above the

speed at which the natural frequency divided by the number of
nodal diameters equals the rotation speed. This, of course, is a
well known result 11,12. Compare this analysis with 13.
We also note that when R jm (r 0 )0, the instability region vanishes. This result is derived in a quite different manner in 4,14.
Example 3. Washboarding of Circular Saws
Washboarding of industrial circular saws is a self-excited vibration of the saw that occurs at certain cutting speeds and leaves the
cut surface of the workpiece with a sinusoidally varying profile
rather than the desired flat surface. Tian and Hutton 15 present a
model for washboarding based on the classical regenerative force
model used to analyze machine tool chatter 16. Here we consider a simplified version of their model.
A circular saw has N evenly spaced teeth located on its periphery at rr 0 and at fixed angles q 2 q/N for q1, . . . ,N,
where these angles are measured in the rotating frame of reference
i.e., fixed to the saw. The saw rotates at angular speed so that
the time interval between teeth passing the same angular location
on the non-rotating workpiece is T2 /N. In an actual saw,
only an angular sector of the saw cuts the workpiece at any given
time. For our simplified analysis, we make the assumption that the
entire periphery of the saw cuts the workpiece at all times. While
this assumption is physically unrealistic, it gives results that are
similar to those derived from more realistic but complicated modeling.
In the rotating frame of reference, the regenerative cutting force
is given by p1 and
K u
q1
References
1
rr 0 q u r 0 , q ,t
r
u r 0 , q1 ,tT
(29)
where u(r, ,t) is the transverse displacement of the saw and

is the delta function. In other words, the transverse load produced
by a given tooth cutting the workpiece is proportional to the difference between its current position and the location of the previous tooth when it cut through the same angular location of the
workpiece.
For this model, Eq. 9 becomes
D jm,T 0 i jm D jm A jm0,T 1 e i jm T 0
N
i jm
q1 k,n
kn r 0 R jm r 0 e
i jk q
A kn0 e i jm T 0 1e 2 i k kn /N
(30)
We eliminate secular response of a single mode by setting

A jm0,T 1 i jm NR 2jm r 0 A jm0 1e 2 i j jm /N
(31)
whose solution is
A jm0 exp i jm T 1 R 2jm r 0 N 1exp 2 i j jm /N
(32)
Therefore, the response is stable only when
sgn jm sin 2 j jm /N 0
(33)
The response changes from stable to unstable as the disk rotation

speed increases through critical speed ( cr jm / j). Since circular saws rarely operate above their critical speed, the more important factor is whether or not the response is stable between 0
cr . As 0 , we obtain an infinite number of alternating stable and unstable zones of rotation speed for each mode, the
width of each zone shrinking as 0. This form of response is
similar to the one determined by Tian and Hutton 15 using numerical methods. Based on experimental observations, they focus
their study on the regions at which the tooth passing frequency
approximately equals a natural frequency of the system in the
stationary frame of reference, i.e.,
jm jN
Timothy A. Brungart
e-mail: tab@wt.arl.psu.edu
Eric T. Riggs
e-mail: etr@wt.arl.psu.edu
The Pennsylvania State University, Applied Research
Laboratory, PO Box 30, State College, PA 16804-0030
Concluding Comments
The fundamental simplification derived by using gyroscopic

eigenfunctions is the reduction of order in 6. For this first order
ordinary differential equation, we obtain only one potential secular solution. In the traditional analysis, the decoupled equation is
second order and two potential secular solutions must be investigated. In practice, one of these solutions is always stable and the
critical stability criteria are determined by the other solution. The
procedure presented here eliminates the need to investigate the
stable secular solution.
Positive definite, self-adjoint systems are a special case of positive definite gyroscopic systems, so the analysis described here
applies to these systems as well. In fact, because they are performed in the rotating frame of reference, Examples 2 and 3 are
self-adjoint.
The restriction to positive definite systems (M 0, K0) ensures that all the natural frequencies ( j ) are real. This makes it
trivial to determine which terms on the right hand side of 9 are
secular. For systems in which the natural frequencies are not all
real e.g., damped systems, the analysis would be identical except
that more care would be required to determine secular solutions.
Rotor Isolation for Vibration and

Noise Reduction
(34)
where is a detuning parameter. Substitution of Eq. 34 into 33

shows that 0 is always the location of a transition from stable
to unstable behavior, as shown in their results.
1 Meirovitch, L., 1975, A Modal Analysis for the Response of Linear Gyroscopic Systems, ASME J. Appl. Mech., 422, pp. 446 450.
2 Wickert, J. A., and Mote, C. D., Jr., 1990, Classical Vibration Analysis of
Axially Moving Continua, ASME J. Appl. Mech., 57, pp. 738 744.
3 Renshaw, A. A., 1997, Modal Decoupling of Systems Described by Three
Linear Operators, ASME J. Appl. Mech., 64, pp. 238 240.
4 Hryniv, R. O., Lancaster, P., and Renshaw, A. A., 1999, A Stability Criterion
for Parameter Dependent Gyroscopic Systems, ASME J. Appl. Mech., 66, pp.
660, 664.
5 Jha, R. K., and Parker, R. G., 2000, Spatial Discretization of Axially Moving
Media Vibration Problems, ASME J. Vibr. Acoust., 122, pp. 290294.
6 Lee, K.-Y., and Renshaw, A. A., 1999, Solution of the Moving Mass Problem
Using Complex Eigenfunction Expansions, ASME J. Appl. Mech., 67, pp.
823 827.
7 Lee, K.-Y., and Renshaw, A. A., 2002, A Numerical Comparison of Alternative Galerkin Methods for Eigenvalue Estimation, J. Sound Vib., 2532, pp.
359372.
8 Parker, R. G., and Lin, Y., 2001, Parametric Instability of Axially Moving
Media Subjected to Multifrequency Tension and Speed Fluctuations, ASME
J. Appl. Mech., 68, pp. 4957.
9 Mockensturm, E. M., Perkins, N. C., and Ulsoy, A. G., 1996, Stability and
Limit Cycles of Parametrically Excited, Axially Moving Strings, ASME J.
Vibr. Acoust., 118, pp. 346 351.
10 Lee, K.-Y., and Renshaw, A. A., 2002, Stability Analysis or Parametrically
Excited Systems Using Spectral Collocation, J. Sound Vib., to appear.
11 Iwan, W. D., and Moeller, T. L., 1976, The Stability of a Spinning Elastic
Disk with a Transverse Load System, ASME J. Appl. Mech., 43, pp. 485
490.
12 Renshaw, A. A., and Mote, Jr., C. D., 1992, Absence of One Nodal Diameter
Critical Speed Modes in an Axisymmetric Rotating Disk, ASME J. Appl.
Mech., 59, pp. 687 688.
13 Mote, Jr., C. D., 1970, Stability of Circular Plates Subjected to Moving
Loads, J. Franklin Inst., 2904, pp. 329344.
14 Renshaw, A. A., and Mote, Jr., C. D., 1996, Local Stability of Gyroscopic
Systems Near Vanishing Eigenvalues, ASME J. Appl. Mech., 63, pp. 116
120.
15 Tian, J. F., and Hutton, S. G., 2001, Cutting-Induced Vibration in Circular
Saws, J. Sound Vib., 2425, pp. 907922.
16 Stone, E., and Askari, A., 2002, Nonlinear Models of Chatter in Drilling
Processes, Dynamical Systems, 171, pp. 65 85.
A technique for reducing the vibration and noise from marine

propulsors is examined theoretically and demonstrated experimentally. A single layer of elastomer is incorporated into a rotor
to form a rotating vibration isolation mount. The mount isolates
the rotor hub and shaft from the unsteady forces generated on the
blades due to their interaction with circumferentially non-uniform,
time-averaged inflow and turbulence. This results in reduced vibration and noise. Nonrotating isolated rotor transmissibility is
well predicted with a simple lumped parameter rotor model for
frequencies below the first bending mode of the blades. Tests performed under operational (i.e., rotating) conditions have shown
that rotor isolation provides a reduction of up to 15 dB in transmissibility compared to an identical rotor without isolation.
DOI: 10.1115/1.1569948
Revised February 2003. Associate Editor: J. Wickert.
JULY 2003, Vol. 125 407
Introduction
Propulsors used in marine applications are a source of tonal and
broadband vibration and noise. Unsteady forces at blade rate and
harmonic frequencies are generated by the interaction of rotor
blades with mean inflow spatial nonuniformities, due to upstream
obstructions, such as control surfaces, stators and support struts.
Broadband unsteady forces are generated by the interaction of
rotor blades with turbulence, formed in the boundary layers and
wakes of upstream surfaces, that is ingested into the rotor. The
tonal and broadband hydrodynamically-induced rotor unsteady
forces are transmitted to the watercraft through the rotor hub and
drive shaft. This results in watercraft vibration and noise. Blake
1 provides an in-depth analysis of these vibration and noise generating mechanisms.
Vibration and noise control is achieved by reducing the unsteady forces generated at the source and/or by inhibiting the
transmission of the unsteady forces to the watercraft. An example
of the latter type of control is the use of resilient mounts that
isolate the watercraft from vibrating components such as power
generating equipment and pumps 2. When power generating
equipment and pumps are mounted independently, flexible shaft
couplings are often used to accommodate shaft misalignment and
provide isolation of component torsional vibrations 2. Similar
isolation techniques can be incorporated directly into the rotor for
vibration and noise control.
Approach
A vibration isolation mount incorporated into the rotor hub can
attenuate the unsteady forces generated on the rotor blades before
they are transmitted to the shaft and watercraft. Such a mount is
shown schematically in Fig. 1. A single layer of elastomer separates the outer blade ring from the inner hub. In the event that the
rotors axial or torsional load exceeds the mount design specifications or the elastomer fails, mating axial and circumferential stops
are machined into the rotor. The stops are not engaged under
normal operating loads since this would create a shorting path
around the mount.
Fig. 1 Schematic of isolated rotor
inder is driven through the base plate with the attached shaker 3.
The transmissibility can also be shown, using the transmissibility
theorem, to equal the absolute value of the complex ratio of the
force transmitted to the cylinder to the force driving the mass 3.
This is illustrated in Fig. 3 where the foundation represents the
cylinder, M is the mounted mass, V 1 and V 2 are the foundation
and mounted mass velocities, respectively, and F 1 and F 2 are the
forces acting on the foundation and mounted mass, respectively.
With a shaker driving the cylinder of the ECTF, the signals
from two phase and amplitude matched accelerometers, mounted
on opposite sides of the elastomer, were measured with a dynamic
signal analyzer. The accelerometer transfer functions were measured over a frequency range of 0 to 2 kHz in 2.5 Hz frequency
bands with 256 ensemble averages. Values of G and G for
mounts fabricated from Hapco Inc. Hapflex 581, 570, 560 and 540
elastomers were determined by fitting Eq. 1 to the measured
transmissibility. Hapflex 540 was found to be the softest of the
four elastomers considered and, as a result, to provide the greatest
isolation capability. Axial load/deflection curves were measured
on an Instron Corporation Series IX Automated Materials Testing
Quantification of Elastomer Properties. In order to build a

suitable vibration isolation mount, the dynamic mechanical properties of candidate elastomers must be known, but may not be
adequately provided by product vendors. Elastomer properties are
measured with the Elastomer Characterization Test Fixture
ECTF shown in Fig. 2. A mass, M, is suspended within a cylinder by casting and curing the candidate elastomer into a vibration
isolation mount similar to that utilized in the isolated rotor. A base
plate attaches the cylinder to a shaker. The elastomer dynamic
shear modulus, G , and loss factor, G , are determined by performing a least-squared error fit of the measured mount transmissibility, T, to the lumped parameter expression for transmissibility
given by Snowdon 3
kG
G j
T
,
Z
(1)
Fig. 2 Schematic of elastomer characterization test fixture

ECTF
where Z is the impedance of the mechanical system

Z
kG
kG
j M
,
G
(2)
and k is a mount stiffness-related parameter

k
2rM
.
r c r M
(3)
Here r is the radius of the suspended mass M or cylinder C,

and is the bond length of the elastomer.
The transmissibility is the absolute value of the complex velocity or acceleration ratio measured across the mount when the cyl408 Vol. 125, JULY 2003
Fig. 3 Vibration isolation mount transmissibility
the clearance derivative which is defined as the reduction in

efficiency points for a one percent increase in normalized
clearance 4.
Results
Fig. 4 Experimental set-up for static measurement of isolated

rotor transmissibility
System with the Hapflex 540 in the ECTF. The data indicated that
Hapflex 540 also provided suitable stiffness, therefore, Hapflex
540 was used for the construction of the isolated rotor.
In order to assess the static stiffness of the mount fabricated
from Hapflex 540, isolated rotor axial load/deflection curves were
also measured on an Instron Corporation Series IX Automated
Materials Testing System. From the curves, the mount stiffness
was determined to be approximately 1056 N/mm, 965 N/mm, and
863 N/mm at loading rates of 12.7 cm/min, 1.27 cm/min, and
0.127 cm/min, respectively. The mating stops machined into the
rotor are designed to engage when the thrust moves the blade row
forward approximately 0.32 cm at speeds just above the maximum speed of operation. Adequate clearance hub-to-blade tip is
available to permit such movement. Rotor isolation is not expected to affect the powering performance of the propulsor since
the blade tip clearance does not change significantly throughout
the rotors axial range of motion. An increase in blade tip clearance has an adverse effect on propulsor efficiency 4. The extent
of the performance degradation has been quantified in terms of
Isolated Rotor Static Transmissibility. The axial transmissibility of the mount embedded in the isolated rotor was measured
in both air and water in a manner similar to that employed with
the ECTF. The experimental set-up is shown in Fig. 4. A shaker
was mounted to the inner hub of the rotor and supported with
elastic shock chords from overhead. The outer blade ring is unconstrained in this set-up, as required by the transmissibility theorem. Two phase and amplitude matched accelerometers were
mounted opposite one-another across the elastomer and oriented
in the direction of the shaker-applied excitation. The measured
transmissibility, and transmissibility predicted from Eq. 1, with
G and G measured with the ECTF, are shown in Fig. 5. The
measured transmissibility for a nonisolated i.e., solid rotor is 1.0
which is also shown in Fig. 5. The frequency is nondimensionalized by the rotor blade passing frequency at the subject operating
condition (1BR).
The in-air measured transmissibility is within 3 dB of the
simple lumped parameter model Eq. 1 prediction up to a frequency of approximately 4BR. The lumped parameter model
tends to under-predict the transmissibility somewhat at higher frequencies. The transmissibility measured in water is within 3 dB of
the Eq. 1 prediction up to a frequency of approximately 1.7
BR. Equation 1 also under-predicts the in-water transmissibility at higher frequencies but to a greater extent than the in-air
results. Water loading is responsible for the reduction in mount
resonance frequency and corresponding modifications to the transmissibility for the in-water measurements compared to the in-air
measurements. It is reasonable to expect that the measured transmissibility will deviate from the lumped parameter-based prediction at frequencies where the rotor begins to exhibit distributed
system behavior. Structural modal analyses of the rotor, performed
in-air and in-water, indicate that the lowest order mode, the first
bending mode, occurs at frequencies of approximately 6.5BR in
air and 2.8BR in water. Thus, we attribute the deviations be-
Fig. 5 Comparison of statically-measured and predicted of isolated rotor

transmissibility
JULY 2003, Vol. 125 409
Fig. 6 UUV model unsteady forces test set-up
Fig. 7 Typical shaft unsteady thrust spectrum measured with and without
rotor isolation
Fig. 8 Comparison of operationally-measured and predicted isolated rotor

transmissibility
410 Vol. 125, JULY 2003
tween the measured and predicted transmissibility, at frequencies

above approximately 4BR in air and 1.7BR in water, to the
occurrence of the rotor blade bending modes.
Isolated Rotor Transmissibility Under Operational Conditions. The hydrodynamically-induced rotor unsteady forces
transmitted to the shaft were measured on an Unmanned Undersea
Vehicle UUV model. Lauchle et al. 5 give a detailed description of the 1.22 m Water Tunnel facility in which the measurements were performed. A schematic of the test set-up is shown in
Fig. 6.
The rotor was housed in a duct with a downstream stator and
was driven with a shaft attached to two 75 Hp electric motors in a
downstream-mounted dynamometer. An unsteady force sensor
was mounted in the shaft and used to measure the rotor-generated
unsteady axial force thrust. Unsteady thrust spectra were acquired with a dynamic signal analyzer from 0 to 1 kHz in frequency with 0.3125 Hz resolution. The spectra were averaged
over 256 ensembles. Typical results are shown in Fig. 7. The
isolation reduces the shaft unsteady thrust measured at the 1
BR and harmonic frequencies up to 10 dB in level, and the
measured broadband shaft unsteady thrust up to 15 dB in level.
The transmissibility measured under operational conditions is
shown in Fig. 8 and compared to the transmissibility measured
statically in water and to the Eq. 1 prediction. The transmissibility increase measured statically in water at the mounts resonance
frequency (0.6BR) is reduced approximately 15 dB under operational conditions. In-water static transmissibility measurements
performed with and without the duct were identical. This result is
consistent with calculations that indicate the duct does not affect
the transmissibility since its impedance is negligible compared to
the impedance of the mechanical system. The 15 dB transmissibility reduction at the mounts resonance frequency, measured under operational conditions, appears to be associated with the dynamics of the rotating dynamometer and shaft assembly since the
resonance amplification is approximately 15 dB when the isolated
rotor is mechanically excited on the nonrotating dynamometer
shaft. We speculate that the rotation of the shaft and dynamometer
assembly alters the phase relationship between the excitation force
and the axial motion of the rotor at resonance resulting in reduced
transmissibility.
At frequencies below approximately 0.4BR and between
0.9BR and 1.9BR, both the Eq. 1 prediction and measured
static transmissibility are within 3 dB of the transmissibility
measured under operational conditions. At frequencies above 1.9
BR, the Eq. 1 predictions are well below the statically and
operationally measured transmissibility since the lumped parameter assumption inherent to Eq. 1 is violated. Resonances associated with the dynamometer and shaft assembly add to the difficulty of interpreting comparisons of the measured static
transmissibility to the transmissibility measured under operational
conditions. Even so, the Fig. 8 transmissibility comparisons show
that the isolated rotor attenuates the unsteady forces transmitted to
the shaft up to approximately 15 dB in level compared to an
identical non-isolated rotor.
Conclusions
A vibration isolation mount has been incorporated into the hub
of a rotor used to propel a marine vehicle. The mount transmissibility is predicted reasonably well under static conditions with a
simple lumped parameter transmissibility model at frequencies
below the first bending mode of the rotor blades. Under operational conditions, the mount resonance is eliminated but otherwise
the measured transmissibility is comparable to the static measurements. The isolated rotor reduces the unsteady forces transmitted
to the shaft up to 15 dB in level compared to an identical nonisolated rotor.
Acknowledgment
This work was sponsored by the Office of Naval Research,
Code 333, and monitored by Dr. Kam Ng. Their support is gratefully acknowledged. The contributions of Dr. Jason T. Gomez and
Mr. Stephen J. Plunkett, of the Naval Undersea Warfare Center
Division Newport, in various aspects of this work, are acknowledged as well.
References
1 Blake, W. K., 1986, Mechanics of Flow-Induced Sound and Vibration, Volumes I and II, Academic Press, Inc., Orlando.
2 Skaistis, S., 1988, Noise Control of Hydraulic Machinery, Marcel Dekker,
New York.
3 Snowdon, J. C., 1968, Vibration and Shock in Damped Mechanical Systems,
John Wiley and Sons, New York.
4 Wisler, D. C., 1986, Advanced Compressor and Fan Systems, Fluid Dynamics of Turbomachinery, Gas Turbine Division, American Society of Mechanical Engineers, ASME Turbomachinery Institute, Ames, Iowa.
5 Lauchle, G. C., Billet, M. L., and Deutsch, S., 1989, High Reynolds Number
Liquid Flow Measurements, Lecture Notes in Engineering, 46, Frontiers in
Experimental Fluid Mechanics, M. Gad-el-Hak, ed., Springer-Verlag, Berlin,
pp. 95158.
JULY 2003, Vol. 125 411

Asme 125 (Vibration & Acoustic) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Asme 125 (Vibration & Acoustic) PDF

Uploaded by

Copyright:

Available Formats

Transactions Journal of Vibration

of the ASME and Acoustics

Published Quarterly by The American Society of Mechanical Engineers

VOLUME 125 NUMBER 3 JULY 2003

Spectrum of High-Frequency Acoustic Noise in Inviscid Liquid-Linear

Theory on Pitch Noise and Its Application

Investigation of the Sound Transmission into an Advanced Grid-Stiffened

Optimal Stiffener Design for Interior Sound Reduction Using a Topology

New Evaluation Method on Gear Dynamics Using Continuous and

A Stochastic Model for Simulation and Diagnostics of Rolling Element

Nonlinear Dynamic Analysis of a Rotor Shaft System With Viscoelastically

Nonlinear Parameter Estimation in Rotor-Bearing System Using Volterra

Effect of Thrust Magnetic Bearing on Stability and Bifurcation of a

Torsional Vibration Analysis of Complicated Multi-Branched Shafting

Study on the Dynamics of a Rotor in a Maneuvering Aircraft

Control of Self-Excited Vibration of a Rotor System With Active Gas

Vibration of Flex Circuits in Hard Disk Drives

Vibration Control of a Traveling Suspended System Using Wave

Dynamic Stiffness Formulation and Its Application for a Combined Beam

Effect of Damping on Asymmetric Systems

DESIGN ENGINEERING DIVISION

Contents continued on inside back cover

Volume 125, Number 3

Robust Control of a 2D Acoustic Enclosure

Active Vibration Suppression With Time Delayed Feedback

Bayesian Model Screening for the Identification of Nonlinear Mechanical Structures

Closed-Form Exact Solution to H Optimization of Dynamic Vibration Absorbers Application to Different

Determining Stability Boundaries Using Gyroscopic Eigenfunctions

Rotor Isolation for Vibration and Noise Reduction

The ASME Journal of Vibration and Acoustics is abstracted and

It is known that the acoustic noise spectrum has some slope in

with addition of the Taits equation of state for water

Contributed by the Technical Committee on Vibration and Sound for publication

Journal of Vibration and Acoustics

the Euler equations 1 and 2 can be written according to Cole

the motion equation

where B300 MPa3000 atm and n7.

The initial system of governing equations is the Navier-Stokes

the continuity equation (2)

It should be noted that Eq. 4 is not an equivalent to / t(c

Copyright 2003 by ASME

JULY 2003, Vol. 125 249

and neglecting p 0 as compared with B, the acoustic pressure is

where all nonlinear terms are collected in the right-hand part of

Assuming hA(r)e t , where A is a function of r only, the

is the complex function of the frequency, .

The constant C 2 can be determined from the condition that p

therefore, in the first approximation,

and Eq. 11 becomes

Since the enthalpy must remain limited, the constant C 1 has to

and, consequently, the solution of Eq. 15 is

h C 1 cos krC 2 sin kr

Now, differentiating Eq. 10 with respect to r, differentiating the

The square of the continuity equation is

whose general solution for k0) is

Therefore, the solution of Eq. 17 can be written in the form:

and this expression can be substituted into the motion equation 5

uC 1 cos krC 2 sin kr.

From Eq. 6, the radial derivative of the velocity is:

According to the formula 2.101 from the handbook of Kamke

where the coefficient K1/2 p 1 /k sin kr/r is introduced for

250 Vol. 125, JULY 2003

Transactions of the ASME