VALIDATION OF A ROTOR DYNAMICS MODEL WITH

VISCOELASTIC BEARING SUPPORTS

Hedzer Tillema 1 , Jeroen Wensing 2
1

Faculty of Mechanical Engineering, Applied Mechanics and Polymers group

University of Twente, Drienerlolaan 5, P.O.Box 217, 7500 AE Enschede, The Netherlands
2
SKF Engineering & Research Centre
Kelvinbaan 16, P.O.Box 2350, 3430 DT Nieuwegein, The Netherlands

SUMMARY: The demand for silent bearing applications has started the research to asses the effect of a
viscoelastic damping layer in the vicinity of the bearing, for example between the bearing outer ring and the
housing. In order to predict the dynamic behaviour of rolling bearing applications a 3D nonlinear time dependent
computational model is developed. Both the vibration generation and the vibration transmission characteristics
of the rolling bearing are accounted for. Viscoelasticity is described both in the time and frequency domain by a
generalized Maxwell model. The modelling procedure is validated by means of experiments on a rotor dynamic
test rig. The results show that vibrations generated by the bearing-shaft arrangement can be isolated from the
housing by applying a viscoelastic bearing support.
KEYWORDS: rolling bearings, rotor dynamics, viscoelasticity, constrained layer damping, component mode
synthesis

1 INTRODUCTION
In recent years, an increasingly better performance is required concerning the vibrational and acoustical
behaviour of rolling bearings in the automotive and household appliance industry. In rotor dynamic applications,
such as electric motors and gearboxes, the bearings act both as vibration exciters and as vibration transmitters.
The vibration excitation of rolling bearings, however, has improved significantly due to improved manufacturing
technology. Therefore, for many applications the excitation of the bearings itself is negligible in comparison with
external excitation forces. In electric motors, for instance, often the main vibration source comes from the
(radial) electromagnetic forces due to a variable air gap between the stator and rotor. These vibrations are
transmitted both directly to the housing and indirectly from the rotor via the bearings to the housing, which
radiates noise. Obviously, in this case the rolling bearings act as vibration transmitters. However, the vibration
transfer path can be broken by means of a soft viscoelastic bearing support.
Within this study, the basic idea is to apply a thin viscoelastic layer between the bearing outer ring and the
surrounding structure such as a housing, as illustrated in Figure 1 for an electric motor.

Fig. 1a: Viscoelastic bearing support in an electric motor

Fig. 1b: Electric motor photograph

In this way, vibrations of the rotor/shaft can be isolated and damped resulting in a reduced noise radiation of the
housing. It should be noted that the static stiffness of the system will decrease when a relatively weak bearing
support is added. Therefore, a balance needs to be found between vibration isolation on one hand and static
support stiffness on the other.
With the increasing computer capacity available nowadays it is possible to model complex structures in great
detail using the Finite Element Method (FEM). In modelling rotor dynamic applications, however, the rolling
bearing is often still modelled as a single nonlinear spring. In these investigations, a computational tool is
developed that can predict the influence of the bearing and the viscoelastic support on the structural dynamic
behaviour of the application more accurately. The bearing model is presented in Section 2. Next, the equations of
motion for the viscoelastic support are derived in Section 3.1. Experimental data of a viscoelastic material are
fitted in the frequency domain using Maxwell elements (Section 3.2). In Section 3.3 a Component Mode
Synthesis (CMS) technique for the viscoelastic components is addressed. The model of a complete rotor
dynamic application was experimentally validated on a rolling bearing test rig, consisting of a flexible shaft
supported by two deep groove ball bearings mounted in flexible housings. The results are shown in Section 4.

2 BEARING MODEL
2.1 Contact model
The dynamic behaviour of a single rolling bearing contact, i.e. between a rolling element and the raceway of the
inner ring or the outer ring, is represented as a spring-damper model. This model describes the nonlinear stiffness
of a dry contact (Hertz), whereas due to the lubrication a constant damping value is introduced. As a result, the
contact force is given by:

Fc =

3
2

+ c &

(1)

where is the contact indentation. The deflection coefficient is determined by the material and geometric
properties of the elastic bodies. The viscous damping coefficient c is determined by the lubricant properties. It
should be noted that more advanced structural bearing models were developed previously, in which the
elastohydrodynamic lubricated (EHL) contact is described in detail (ref. Wensing [1]). With the help of transient
dynamic calculations the constitutive behaviour of a single EHL contact could be predicted beforehand (ref.
Wijnant [2]).

2.2 Bearing model

In order to construct the equations of motion of the different bearing components, a Lagrangian approach is
adopted. In these investigations, the inner ring and outer ring of the bearing are modelled as rigid bodies.
Moreover, it is assumed that the rolling elements are massless. In that case, the inner and outer ball-raceway
contact models can be combined to a single spring-damper model. As a consequence, the resultant deflection
coefficient is expressed as:
3 / 2

1
1
(2)
res = 2 / 3 + 2 / 3

o
i

where i and o are the deflection coefficients for the inner an outer raceway contact, respectively. The cage is
only taken into account by assuming a constant circumferential angle between two consecutive rolling elements.
It is noted that in special cases the flexibility of the bearing outer ring can play a significant role in the overall
dynamic behaviour of the application. For that purpose a specific model was created by Wensing [1], in which
the bearing outer ring is modelled as a FEM component.
The constitutive relations of the contact model are used to describe the interaction between the structural
components of the bearing. The equations of motion for the inner and the outer ring yield:
Z

[mir ]{q ir } +

j =1
Z

[mor ]{q or } +

j =1

j
= { f ir }
{q ir }

(3a)

j
= { f or }
{q or }

(3b)

where Fcj, j, denote the force and indentation of the combined inner and outer ball-raceway contact,
respectively. The degrees of freedom for the inner and outer ring are stored in the vectors {qir} and {qor}. The set
of nonlinear equations is solved with the Newmark time integration method in combination with a modified
Newton Raphson process.

3 MODELLING OF VISCOELASTICITY
3.1 Time domain model

Dynamic excitation from the bearing is caused by transient mechanisms, which can only be described in the time
domain. To be able to predict the influence of a viscoelastic component on this transient behaviour the
constitutive relation for linear viscoelasticity and the equations of motion of a viscoelastic component need to be
derived as a function of time as well. In case of a Maxwell representation the constitutive equation for linear
viscoelastic material reads:
t t '
N
t

(t ) =
G0 +
Gi e i & (t ' )dt '
(4)

i =1

Here G0 is the instantaneous modulus of the material whereas Gi and i are the relaxation strength and relaxation
time corresponding to the i-th Maxwell element. The generalized Maxwell model consists of a number of
parallel spring-dashpot systems and a parallel spring. In general, time dependent behaviour for viscoelastic
materials is only present in shear direction. Therefore, the compression modulus is assumed constant and is
defined by a single frequency independent parameter K. Advantage of the Maxwell model is that the dynamic
behaviour can be described efficiently both in the time and frequency domain. The latter is important in
particular as experimental data for viscoelastic materials are usually given as a function of frequency.
Let us assume a generalized Maxwell model with an additional parallel spring and parallel dashpot. The
constitutive equation can then be expressed as:

(t ) =

1
K kk (t ) + 2G 0 e xy (t ) + 2 0 e& xy (t ) +
3

i =1

2Gi exp(

t 't
)e& xy (t ' )dt '
i

(5)

where kk is the hydrostatic (compressive) strain and exy denotes the deviatoric (shear) part of the strain vector.
The damping coefficient of the parallel dashpot is represented by 0. The convolution integral represents a sum
of the present strain rate and the historical strain rates of the material. Assuming a piecewise linear rate of
deformation e& xy , the integral can be solved and the constitutive equation can be discretized with time step t =

tn+1-tn according to Baaijens [3]:

nxy+1 (t ) =

1 n +1
n +1
n +1
n +1
K kk (t ) + 2G 0 e xy
+
H in +1
(t ) + 2 0 e& xy
(t ) + 2 i e& xy
3
i =1
i =1

(6)

where
t i (1 e t / i ) t / n
i
+e
H in +1 = 2 i G i (1 e t / i )1
Hi
(7)

Here the tensor H in +1 is the history tensor containing the strain rates from the past (t < tn+1). Additionally, a
stiffness and a damping part can be recognized. It should be noted that the equation is written in a convenient,
linearized form. Therefore, the computation time remains relatively small. In matrix notation equation 6 can be
written as:
{} n +1 = [ E ]{} n +1 + [ D]{&} n +1 + {H } n +1
(8)
i = i Gi

t i (1 e t / i )
t

where [E] is the elastic material matrix and [D] the viscous material matrix. The equations of motion are derived
using the procedure of weighting and Galerkin's method. A vector {x} is introduced, containing the generalized
coordinates. Using a standard finite element approach, the equations of motion for a viscoelastic component
yield:

[ M ve ]{&x&} n +1 + [C ve ]{x&} n +1 + [ K ve ]{x} n +1 = {F } [ H ve ]{h} n +1

(9)

where [Mve] is the mass matrix, [Cve] the damping matrix and [Kve] the stiffness matrix of the viscoelastic layer.
The vector {F} contains the externally applied forces and [Hve]{h}n+1 the history forces. For a description of the
different matrices the reader is referred to Tillema [4]. The resulting set of equations is solved by means of the
Newmark time integration method.
3.2 Frequency domain model

Viscoelastic material data are obtained using a Dynamic Mechanical Analyzer (DMA), which measures the
stiffness and loss factor of a material specimen for a specific frequency and temperature range. In general, the
working frequency range of the DMA is limited from approximately 0.01 to 200 Hz, whereas the temperature
ranges from about -145 to 600 0C. In order to determine the shear modulus and accompanying loss factor for a
wide frequency range at a certain reference temperature, the time-temperature superposition (TTS) principle is
applied (ref. Ferry [5]). This principle says that the modulus of a material at time t and temperature T is
equivalent to the modulus at the reduced time and a reference temperature T0. It is widely used to extract the
material properties over a wide frequency range.
The dynamic shear modulus for viscoelastic materials can be written in complex form, reading:

G * () = G ' () + jG ' ' () = G ' () (1 + j())

(10)

where G'(), G''() and () denote the materials' storage modulus, the loss (or dissipation) modulus and loss
factor, respectively. denotes the angular frequency. For the generalized Maxwell model with N elements
including a parallel spring and dashpot, the storage and loss moduli are expressed as:
N

G ' () = G o +

( i ) 2

1 + ( )
i =1

G
2 i

G ' ' () = o +

1 + ( )
i =1

Gi

(11)

Note that the dissipation modulus of the parallel dashpot is linear with the frequency. Experimental data can be
fitted fairly easily using a number of Maxwell elements. In Figure 2 the storage modulus and loss factor of a soft
rubber are shown. Five Maxwell elements in combination with a parallel spring and dashpot were used to obtain
this result.

Fig. 2: Maxwell curvefit of the storage modulus and loss factor of a viscoelastic material

3.3 CMS reduction of viscoelastic components

For an efficient transient analysis of a bearing application, each FEM component in the system, such as the shaft
and the housing, is reduced individually by means of a Component Mode Synthesis (CMS) technique. In order to
deal with complex geometries and complete flexibility for a viscoelastic component one must resort to a finite
element approach as well. To enable an easy coupling of adjacent components, the CMS technique must
therefore also be applicable to a viscoelastic component. The CMS technique is based on the Ritz method, which
describes the displacements of a structure as a series of global shape functions. The number of degrees of

freedom of a system can be reduced significantly by choosing suitable shape functions in the frequency range of
interest. With CMS a set of shape functions or component modes consists of a static and a dynamic mode set.
These modes are required to describe both the static and dynamic solution (and if applicable, rigid body motions)
of the structure. Using the transformation from FEM coordinates {x} to the generalized coordinates {p}, the
component modes are stored column-wise in the transformation matrix []. Consequently, the reduced mass and
stiffness matrix are obtained by equation 12.
[m] = [ ]T [ M ][ ]

[k ] = [ ]T [ K ][ ]

(12)

The conventional Component Mode Synthesis works most efficient for undamped, slightly damped or
proportionally damped systems. In case of a viscoelastic component we assumed damping only in shear direction
implying that it is a non-proportionally damped system. However, if the dynamic behaviour of a viscoelastic
component can be described efficiently with the use of undamped modes, the reduced matrices can be obtained
conform equation 13:
[c ve ] = [ ]T [C ve ][ ]
(13a)
[m ve ] = [ ]T [ M ve ][ ]
[k ve ] = [ ]T [ K ve ][ ]

[hve ] = [ ]T [ H ve ][ ]

(13b)

Note that the reduced history matrix can be obtained because the vector {h} is a linear combination of the
generalized degrees of freedom. It is clear that the system mass and stiffness matrix can be calculated in a FEM
package directly. The system damping and history matrix, however, are not proportional to the mass or stiffness
matrix, but proportional to the deviatoric part of the stiffness matrix [Kve()]dev. The deviatoric part of the
stiffness matrix [Kve()]dev can be obtained in a FEM package by setting the compression modulus K to zero or
close to zero in order to avoid numerical problems. Hence, the reduced damping and history matrices can be
calculated using the original set of undamped modes:
N

0 +
[c ve ] = [ ] [C ve ][ ] =
T

[hve ] = [ ]T [ H ve ][ ] =

j =1

G0

[ ]T [ K ve ] dev [ ]

1
[ ]T [ K ve ] dev [ ]
2G 0

(14)
(15)

In these investigations a viscoelastic layer was considered having a ring shaped geometry with a relatively small
wall thickness. The inner and outer surface of the layer are modelled as two rigid interfaces both having 3
translational and 3 rotational degrees of freedom. No dynamic effects over the thickness of the layer are
expected. As a consequence, the three typical dynamic modes of a viscoelastic layer are an axial, a tilting and a
radial mode, such as shown in Figure 3.

Fig. 3a: axial mode

Fig. 3b: tilt mode

Fig. 3c: radial mode

4 EXPERIMENTAL VALIDATION
4.1 Rotor dynamic test rig

In these investigations a specially designed test rig is analysed. It was developed within SKF ERC for testing the
vibration behaviour of low-noise rolling bearings under operating conditions similar to that given in real-life
applications. An overview of the test rig is shown in Figure 4.

Fig. 4: Schematic overview of the rotor dynamic test rig

The test rig consists of the following components:

A very stiff and heavy base-frame supported in a soft suspension.

Exchangeable housings rigidly bolted to the base-frame.
A long, flexible rotor supported by two rolling element bearings in the housings.
A preloading tool used for axial preloading of the bearings.
An external frequency controlled electric motor to drive the test rig. A highly flexible coupling between
the motor and the test rig ensures that motor vibrations are not transferred to the test rig rotor.

The test rig is designed to analyse bearing related parameters under operating conditions. It can also be used to
study the vibration transmission characteristics of viscoelastically supported bearings. The main purpose of this
setup is to demonstrate that vibrations of the shaft-bearing arrangement can be isolated from the housing through
the viscoelastic layer. The vibration excitation in this setup is mainly generated by the imbalance of the rotor and
to some extent by the deep groove ball bearings.

4.2 Structural modelling

The test rig is subdivided into different structural components, which are separately modelled and validated with
measurements. First, it was found that the stiff base-frame of the structure had no significant effect on the
dynamic behaviour of the system under normal operating conditions. Therefore, the flexible housings are
assumed fixed to the ground at the bolt positions.
Finite element models are built for the housings, the rotor and the viscoelastic bearing supports. The FEM
models of the housing and the rotor are validated with experimental modal analysis, showing that up to 5 kHz
the predicted eigenfrequencies of the components are within 10% deviation of the measurements. Subsequently,
the models are reduced conform the CMS technique for elastic components as discussed in Section 3.3.
The viscoelastic bearing supports consist of a steel inner and outer ring constrained by a thin viscoelastic layer.
These "sandwich" rings can easily be mounted between the bearing and the housing, such as illustrated in Figure
5. Due to their small size, i.e. with an outer diameter of 50 mm, it was not possible to perform an experimental
modal analysis on the sandwich rings. However, experiments on larger ring samples showed good comparison
with the predictions (see Tillema [4]). In accordance with the theory explained in Section 3.3, a FEM model is
constructed for the viscoelastic layer and reduced with the CMS approach. Note that all interfaces in the bearingsandwich ring-housing arrangement are assumed rigid. The measured viscoelastic material properties are fitted
with a sufficient number of Maxwell elements to simulate the frequency dependent constitutive behaviour from
10 Hz to 10 kHz.

Fig. 5: Viscoelastic bearing support mounted in the test rig

The deep groove ball bearings are modelled according to the theory discussed in Section 2. The deflection
coefficients i and o are calculated using the geometric and material properties of the bearing. To account for
the effect of lubrication, the bearing is given a constant contact damping coefficient. It was found, however, that
the contact damping is of minor importance in comparison with the viscoelastic damping. In the unloaded
situation a ball bearing has a radial clearance of several m. This radial clearance disappears under an externally
applied axial load, and therefore ensures full contact with the raceway for each rolling element. In case of the test
rig, an axial preload of 60 N is applied on the left bearing. For this purpose, the outer ring of this bearing can
move freely in axial direction.
The complete model of the rotor dynamic test rig is constructed by connecting the components at the interfaces.
The system DOFs are now stored in a vector {p}, containing the generalized coordinates. Whereas the number of
finite element DOFs is in the order of ten thousands, the number of generalized DOFs is in the order of
hundreds. As a result, a modal analysis of the application is performed in a couple of seconds, whereas a
transient dynamic calculation of 100.000 time steps can be performed in about 1 hour on an ordinary
workstation. Note that the displacements of the original FEM systems can be obtained by back-substitution of
the vector {p}.
4.3 Results

A modal analysis was performed on the non rotating, preloaded test rig, both with and without a viscoelastic
bearing support. In the analyses, the most important eigenmodes up to about 5 kHz were determined. In Figure 6
some important predicted and measured eigenfrequencies are shown for an assembly with and without
viscoelastic layer.

Fig. 6: Measured and predicted eigenfrequencies of the test rig

A good agreement is obtained between the prediction and measurement. Up to 5 kHz each eigenfrequency is
predicted within 10% accuracy. Note that the frequency dependent complex stiffness of the viscoelastic support
is accounted for in the modal analysis. In Figure 7 some important eigenmodes are shown, which are also excited
by the system under operating conditions.

Fig. 7a: eigenmode at 486 Hz

Fig. 7b: eigenmode at 525 Hz

Fig. 7c: eigenmode at 628 Hz

Fig. 7d: eigenmode at 675 Hz

Next, the velocity spectra of the axial vibrations of the housing are measured for increasing rotational speeds.
The measured spectra are collected in a matrix and are presented in contour plot form, a so-called Campbell
diagram. In such a diagram the resonance frequencies show up as vertical bands, whereas excitation frequencies
are characterized by straight (order) lines proportional to the speed. In Figure 8 the Campbell diagrams of the
axial vibrations of the end tube of the right housing are shown. For clarity the frequency range is limited from 0
to 1500 Hz.

Fig .8a: Campbell diagram test rig

Fig. 8b: Campbell diagram test rig with layer

It is observed that resonance frequencies appear in bands around 200 Hz and 550 Hz. Moreover, high vibration
amplitudes are observed for the test rig without layer for rotational speeds higher than 4300 rpm. The resonance
frequency at 640 Hz is accompanied by second order harmonics and side harmonics indicating nonlinear effects.
This resonance behaviour is not observed for the test rig with the viscoelastic bearing supports. Clearly,
vibrations of the shaft-bearing arrangement are isolated from the housing effectively.
The excitation frequencies that appear in these spectra are dominated by the rotational speed and its frequency
harmonics. This indicates that the imbalance of the shaft is the main source of excitation.

The acceleration spectrum of the axial response of the housing at a rotational speed of 5000 rpm is extracted for
comparison with the numerical results. For the sake of graphical clarity, a convoluted spectrum of the velocity
level is determined. A running mode analysis was simulated by applying a small imbalance of the shaft. The
measured and predicted spectra are depicted in Figure 9.

Fig. 9a: Measured velocity spectrum test rig

Fig. 9b: Predicted velocity spectrum test rig

From the figures a fairly good agreement is obtained between model and experiment. In both cases it is observed
that the vibration level in the frequency range from 400 to 700 Hz is dominant. Moreover, the vibration level is
significantly lower in case of a viscoelastic support, both for the experiment and the prediction. In fact, the
overall vibration level has reduced with about 5 dB. However, as it appears very difficult to determine the exact
excitation forces beforehand, the agreement will not be very accurate over the complete frequency range. The
general trends, however, can be predicted quite well and therefore the numerical analysis can serve as a valuable
design tool to optimize the viscoelastic bearing support.

5 CONCLUSIONS

A computational model is developed to analyse the dynamic behaviour of rolling bearing applications with
viscoelastic bearing supports. The bearing model is based on previous research and includes both the nonlinear
stiffness due to the bearing geometry and the effect of lubrication damping.
The constitutive relation for viscoelasticity is based on a generalized Maxwell representation. With the help of an
exponentially decaying history tensor, the equations of motion of a viscoelastic component can be written in a
convenient linearized form. With this method transient dynamic calculations can be performed relatively fast.
Moreover, the Maxwell description enables one to curvefit the complex stiffness of the material in the frequency
domain.
The present modelling approach allows for the usage of FEM for the viscoelastic layer, which is of crucial
importance if the component needs to have a complex geometry. With the linear set of equations for a
viscoelastic component the Component Mode Synthesis can be assessed. Keypoint in this approach is that the
damping and history matrix are proportional to the deviatoric part of the stiffness matrix.
The modelling procedure is validated with results obtained from a rotor dynamic test rig. The eigenfrequencies
of the non rotating system are well predicted. For an accurate prediction of the vibration level under rotating
conditions, however, the excitation forces need to be known more precisely. It is shown both numerically and
experimentally that vibrations of the shaft-bearing arrangement are isolated from the housing effectively by the
viscoelastic bearing support.

REFERENCES

1.
2.
3.
4.

5.

Wensing, J.A., On the dynamics of ball bearings, PhD thesis, University of Twente, Enschede, The
Netherlands, ISBN 90-36512298, 1998.
Wijnant Y.H., Contact dynamics in the field of elastohydrodynamic lubrication, PhD thesis, University of
Twente, Enschede, The Netherlands, ISBN 90-36512239, 1998.
Baaijens F.P.T., Applied computational mechanics, lecture notes, University of Eindhoven, The Netherlands
Tillema, H.G., Dynamic behaviour of ball bearing applications with constrained damping layers ,
International Modal Analysis Conference, Proceedings of the Conference held in Kissimee USA, 2001, pp.
130-136.
Ferry, J.D., Viscoelastic properties of polymers, John Wiley & Sons, New York, 1980.