ISA Transactions 47 (2008) 101–112

www.elsevier.com/locate/isatrans

Modeling and control of a flexible rotor system with

AMB-based sustentation
I. Arredondo ∗ , J. Jugo, V. Etxebarria
D. Electricidad y Electrónica, F. Ciencia y Tecnologı́a (U.P.V./E.H.U.), Spain

Received 7 November 2006; accepted 25 April 2007

Available online 4 September 2007

Abstract

In this work the modeling and basic control design process of a rotary flexible spindle hovered by Active Magnetic Bearings (AMB) whose good
capabilities for machine-tool industry extensively treated in the literature is presented. The modeling takes into account the three main behavioral
characteristics of such magnetically-levitated rotor: the rigid dynamics, the flexible dynamics and the rotating unbalanced motion. Besides, the
gyroscopic coupling is also studied proving that in this case, its effects are not significant and can be neglected. Using this model, a stabilizing
controller based on symmetry properties is successfully designed for the system and a complete experimental analysis of its performance is carried
out. Also, the predictions of the model are compared with the actual measured experimental results on a laboratory set-up based on the MBC500
Rotor Dynamics. Afterwards, a brief study about some nonlinear behavior observed in the system and its effect over the system stability at the
critical speed is included.
c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

Keywords: Magnetic bearings; Modeling; Control; Nonlinear system; Symmetry

1. Introduction loops to operate and take full advantage of their unique

In the last few years Active Magnetic Bearings (AMB) have
attracted great research interest due to their unique charac-
teristics and advantages with respect to conventional passive
bearings, particularly for high-performance mechatronic sys-
tems [1–3]. On the one hand, magnetic (without contact) bear-
ings are virtually friction-free elements, which allow to build
electromechanical devices with much less mechanical wear, no
need of lubrication and low maintenance requirements. These
characteristics make them specially suitable for high-speed ap-
plications, [4,5]. On the other hand, due to their active nature,
AMB are not only pure supporting elements but controlling ele-
ments as well, which can be used for active vibration damping,
mechanical impedance control or precision positioning control,
among other possibilities, [6–8].
However, since magnetic levitation systems are inherently
unstable, AMB require very well designed and tuned control
∗ Corresponding address: Fac. Ciencia y Tecnologia, Dpto. Electricidad y
Electronica, Campus de Leioa UPV/EHU, Barrio Sarriena s/n, 48940 Leioa,
Spain. Tel.: +34 946013235; fax: +34 946013071.
E-mail addresses: inigo.arredondo@ehu.es (I. Arredondo),
josu.jugo@ehu.es (J. Jugo), victor.etxebarria@ehu.es (V. Etxebarria).
potential. Moreover, AMB are much more complex than
their passive counterparts, and rather precise knowledge
of their mechanical, electrical and magnetic properties are
needed to make an appropriate robust stabilizing control
design of the device. Thus, an accurate model of the
whole system, comprising all the interacting parts, should be
developed to conveniently take into account the whole device
dynamics. This implies that it has to include the mechanical
flexibilities and resonances of the supported shaft, the nonlinear
electrical characteristics of the required power amplifiers and
the nonlinear spatial distribution of the generated magnetic
forces [9–11]. This is compulsory, both to be able to design
a reliable robust control for the system [12,13], and to test its
open- and closed-loop behavior in simulation, before trying
on a real experimental set-up, thus minimizing experimental
failure and avoiding potentially dangerous situations inherent
in high-speed machinery.
In this paper, the complete modeling (including rigid,
flexible and unbalanced dynamics and a gyroscopic analysis)
and control design process for a laboratory AMB set-up
based on the MBC500 Rotor Dynamics from LaunchPoint

c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

0019-0578/$ - see front matter

doi:10.1016/j.isatra.2007.04.004
102 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
Fig. 1. MBC500 Rotor Dynamics scheme.
Technologies, Goleta, CA [14], is presented. Other related
works, such as [15] and [16], utilize similar devices such as
MBC500 and the MBC500 Turbo. However the MBC500 Rotor
Dynamics is an enhanced device which reaches higher speeds
and has a more complicated design in its rotor.
A mathematical model, predicting most of the effects
actually measured on the device, is presented in the paper.
By using this model a simple stabilizing controller is
successfully designed and then the closed-loop model behavior
is compared with the actual device. The proposed control
design methodology exploits the symmetry properties of the
machine, allowing to split the nxn MIMO system into n SISO
subsystems, thus making the control design process much
more easier. A related work, which however does not use the
symmetry properties, is developed by Huang et al. in [17],
where also a good review about decentralized control is shown.
As a second step the system is thoroughly tested and its
mechanical response is measured by means of a number of
sensors and laser-based noncontact measurement system. The
experimental data is then compared with the model predictions
and a satisfactory agreement between measurements and
simulations is obtained. The symmetry condition again plays
a relevant role in this comparison, allowing to separate each
individual feedback loop dynamics. In recent works [18]
and [19] similar analysis are performed but without using the
symmetry properties.
Finally, a brief study of the nonlinear properties observed in
the system (inherent to the magnetic levitation technology) is
carried out, and possible effects over the whole system stability
are outlined.
The article is organized as follows. In Section 2 the device
linear and nonlinear models are obtained by means of the
superposition of the rigid dynamics, the flexible behavior and
the unbalanced motion, including the reasons to neglect the
gyroscopic coupling in the rotation speed range of operation.
Then, the controller design based on symmetry properties is
performed. In the next section, using the obtained controller,
the model simulations and the experimental results are analyzed
and contrasted. Finally, conclusions are presented in Section 5.
2. Model of an AMB-based rotating device
The two most common ways of developing a model of this
type of devices are either by means of a pure Finite Element
Fig. 2. Pole configuration.
Analysis (FEA) or as is done in this article, by a constructive
method which is more clear from physical perspective. Both of
them provide reliable results.
The system to be modeled is a set-up based on the MBC500
Rotor Dynamics which is a laboratory device specially designed
for research purposes. It is composed of two AMBs and a
rotor which includes an air turbine drive, allowing speeds up
to 22 000 rpm. As is schematically shown in Fig. 1, the shaft
position is measured by Hall effect sensors and the currents
causing the forces in the bearings to maintain the hovering state,
are driven by voltage amplifiers. Thus, the system inputs are
the voltages given to the amplifiers, and the outputs are the
voltages provided by the position sensors. Also, as is explained
later, analog filters have been developed and connected to
the outputs to avoid noise at high frequencies. Finally, the
scheme is completed by closing the loop designing a stabilizing
controller, thus appropriately feeding back the system response
(Section 4).
Since the poles of the electromagnetic bearings are
symmetrically distributed (Fig. 2), and taking into account the
differential configuration which is used, the forces caused by
them over the horizontal and vertical directions (respectively
the x- and y-axis) are considered decoupled, so in the first
instance, the modeling can be performed independently for
each direction. On the other hand, when the shaft is rotating
the gyroscopic effect leads to a coupling between x and y.
 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112 103

Fig. 3. Rigid rotor configuration.

However, such effects which should be introduced in the model

 
0 0
(Section 2.5), can be neglected in this system because of the  1 1  
 F1
 
mechanical dimensions of the shaft. Moreover, the gravity +
 m m (2)
effect can be separately included for the vertical direction (as  0 0  F2

 1 L 
1 L

is presented in Section 2.3). These facts, (i.e. symmetric pole

− −l −l
configuration, gravity separate inclusion and gyroscopic effect I0 2 I0 2
neglectfulness), provide the chance of making the whole study I0 = I x x = I yy being the moment of inertia of the shaft with
only for one of the axis, being similar for the other one. Hence, respect to rotation about an axis in the x- or y-directions.
the modeling for the x-axis is divided into three different parts The relationship between the variables of the state-space
whose superposition describes the device behavior for this vector and the measured position in the Hall sensors and
direction. Then, the same dynamics for the y-direction and the bearing points are respectively,
gravity effects are introduced to complete the whole model.  
L
 
  x0 − − l2 sin θ
2.1. Rigid body model X rig1  2 
=    (3)
X rig2  L 
x0 + − l2 sin θ
The model of the rigid part is based on the geometry of 2
the rotor [20]. The rigid rotor motion can be described by
 
L
 
the linear translation of its mass centre (M.C.), x0 and the   x0 − − l sin θ
xrig1  2 .

rotary coordinate θ representing the angle between the principal xrig2
=
 
L
  (4)
inertial axis and the z-axis. This motion is driven by the x0 + − l sin θ
2
magnetic forces F1 and F2 generated by the bearings at a
distance l and its rigid part is described by variables xrig1 and However, as important as this rigid behavior is the flexible
xrig2 in bearing 1 and 2 positions, respectively. On the other dynamics which is discussed next.
hand, the Hall effect sensors, which measure the position of
the shaft provoked by the rigid dynamics X rig1 and X rig2 , are 2.2. Flexible mode model
located at a distance l2 (Fig. 3). Since the shaft is not a rigid structure, it is necessary
Applying Newton’s laws, it is easy to yield the rigid to include in the model the effects of the flexibilities. The
dynamics: resonances that arise because of this flexible phenomenon are
crucial because they are excited when the shaft is rotating at
F1 + F2 = m ẍ0
these frequencies. This leads to stability problems and also,
(1)
   
L L since the gap between bearings and rotor is very small, the
− − l F1 + − l F2 = I0 θ̈
2 2 mechanical resonances can easily cause destructive collisions.
The technique used to reach the flexible part of the model is
which can be rewritten under state-space form as:
     the well-known Finite Element Analysis (FEA). Some relevant
ẋ0 0 1 0 0 x0 points over the rotor are chosen to apply this technique as is
ẍ0  0 0 0 0 ẋ0 
 = shown in Fig. 4.
 θ̇  0 0 0 1  θ 
 
From the aforementioned analysis, taking into account the
θ̈ 0 0 0 0 θ̇ main elastic modes of the shaft, the first and the second, the
104 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112

Fig. 4. Detailed mechanical structure of the rotor showing FEA points.

expressions involving the flexible variables x f i are (5) and (6), 2.4. Linked model
respectively:
Superposing the rigid, flexible and unbalanced motion
1.14 × 103 x f 1
    
ẋ f 1 0 effects, and introducing the gravity (g), the whole mechanical
=
ẋ f 2 −1.14 × 103 0 x model is achieved. In this article the gravity is treated like an
  f 2
2.8 × 103 2.8 × 103 F1 external force in order to maintain the same model for the x-

+ (5) and y-axis. Note that the effect of the gravity supposes a bended
0 0 F2
beam in the y-axis direction, even in static position.
0 −0.1121 × 10−4 x f 1
    
X flex1 Then, the mechanical model is,
=
X flex2 0 −0.1121 × 10−4 x f 2
x̄˙ f x
    
Aflexx 04x4 04x4 04x4 x̄ f x
2.58 × 103 x f 3
    
ẋ f 3 0 x̄˙ f y   04x4 Aflex y 04x4 04x4  x̄ f y 
=   = 
ẋ f 4 −2.58 × 103 0 x  
  f 4  x̄˙ x   04x4 04x4 Arigx 04x4   x̄ x 
−2.9 × 103 2.9 × 103 F1

+ (6) x̄˙ y 04x4 04x4 04x4 Arig y x̄ y
0 0 F2 
Bflexx 04x2
 
Fx1
0 0.405 × 10−5  04x2 Bflex y   Fx 
     
X flex1 xf3 +   2
= −5  Brigx 04x2   Fy1 
X flex2 0 −0.405 × 10 xf4
04x2 Brig y Fy2
where F1 and F2 are the forces produced by the bearings  0 
10x2 0
and X flex1 and X flex2 are the positions measured in the Hall 10x1

 −e y ex  
effect sensor points (Section 2.1), corresponding to the flexible   01x1 
I I x z  sin ωt
 
dynamics. + ω2  yz +  03x1  (8)
 
 02x2  cos ωt

Finally, the modeling is now completed taking into account e e 
 g 
x y
the unbalanced motion caused by the rotating shaft. 01x1
−I x z I yz
2.3. Rotating unbalanced model
   
X1   x̄ f x
  = Cflexx 02x4 Crigx 02x4 x̄ f y 
 X 2  
The main effect caused by this unbalance is the introduction  Y1  02x4 Cflex y 02x4 Crig y  x̄ x 
of a centrifugal force Fct fi which is proportional to the square
Y2 x̄ y
of the rotational velocity ω, and a matrix composed by cross
inertia tensor elements I x z , I yz and the distance between the where A∗ , B∗ and C∗ are the matrices of the rigid and
M.C. and the principal axis of inertia ex , e y as is shown in flexible dynamics in state-space representation and 0i x j are null
(7) [2]. matrices with dimensions i x j. Schematically,
sin ωt
     
Fct f x x −me y mex 
x̄˙ = A x̄ + B F̄ + ω2 U + ḡ (9)
 sin ωt . cos ωt

 Fct f θx 
 = ω2  I yz Ix z 

(7)
me y  cos ωt

 Fct f y y   mex ȳ = C x̄
Fct f θ y −I x z −I yz
where x̄ is the state-space vector including the rigid and
Since the constants ex , e y , I x z and I yz are difficult flexible variables, A, B and C the matrices in state-space
to calculate, they have been determined according to the representation, F̄ the force signal, ω the rotation frequency,
experimental results measured on the device. Also, as is seen in U the unbalanced matrix from (7), ḡ the gravity vector and
(7) this unbalanced motion effect only appears when the beam ȳ the measured positions in Hall sensors points, all with the
is rotating (ω 6= 0). appropriate dimensions.
 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112 105

Up to this point, the inputs to the system are the forces in the
bearings, the nonlinear component of the unbalanced dynamics
(ω2 sin ωt, ω2 cos ωt) and the gravity, with the outputs being
the positions measured at the location of the Hall sensors.
The experimental device has also two extra (eddy current)
position sensors (Fig. 8) whose readings can also be inserted
by following the same steps utilized to obtain (3).
Once the mechanical model is determined, the electronics
and electromagnetic elements are added to reach the whole
model.
Firstly, the electromagnetic force is introduced, taking into
account the differential mode configuration of the bearings (10),
(i i + i 0 )2 (i i − i 0 )2
Fi = K − K with i = 1, 2 (10)
(xi − x g )2 (xi + x g )2
Linking all these expressions a complete linear model is
obtained (16).
˙x̄ = Alin x̄ + Blin V̄control + ω2 U sin ωt + ḡ
 
(16)
cos ωt
V̄sensor = Clin x̄
where V̄control is the voltage control signal and V̄sensor the Hall
sensors measured voltage. Note that, since one of the inputs
to the system is the nonlinear component of the unbalanced
dynamics (ω2 sin ωt, ω2 cos ωt) the model is linear but, in
addition, it maintains certain nonlinear nature. Moreover,
this nonlinearity loses importance when working at constant
rotating speeds. In this way, it could be rewritten as:

V̄control


0  ω sin ωt 
where K = 2.8 × 10−7 N m2 /A2 is a geometric constant  2 
x̄˙ = Alin x̄ + Blin ω2 cos ωt  (17)
depending on the bearing, i 0 is a bias current and x g =
0.0004 m is the effective distance between the bearing and the ḡ
rotor. V̄sensor = Clin x̄.
On the other hand, the magnetic bearings in the MBC500
Rotor Dynamics are driven by a set of amplifiers whose One important characteristic of the resultant model is
dynamics is described by (11) the symmetry in the matrix when using transfer function
representation. This symmetry can be exploited in order to
0.25
i controli = A/V × Vcontroli simplify the design and analysis process (Section 3) [21]. In
(1 + 2.2 × 10−4 s) fact, the system can be written as follows for one of its axes
with i = 1, 2. (11) when ω = 0.
Vsensor1 (s) A(s) B(s) Vcontrol1 (s)
    
In addition, the Hall sensors provide a voltage which measures
= . (18)
the position following the expression (12). Vsensor2 (s) B(s) A(s) Vcontrol2 (s)
Vsensori = 100 00V /m X i ± V offset with i = 1, 2. (12) The nonrotating condition imposed is a simplification to
facilitate the control design, but as is stated in [2], the control
In this way, the input forces are changed into input voltages
designed for zero rotational speed may also be used at high
and the position outputs into position voltages, maintaining
rotational speed, even though it will not be “optimal”.
the nonlinear component of the unbalanced dynamics and the
The whole model has been linked and simulated in Scilab
gravity as inputs.
and Matlab/Simulink environments, in order to obtain the
Finally, (13) shows the filter mathematical model, which is
results which are presented in the following sections. These
introduced in the model with the aim of minimizing the noise
simulations are also being compared with the experimental
effect.
results on the real device to validate the model.
4000
Vfilti = Vsensori with i = 1, 2. (13) 2.5. Analysis of the gyroscopic effect
s + 4000
For controller design reasons, it is convenient to obtain a Owing to the rotation dynamics of the body and neglecting
linearized version of the model around the operation point (0, the effects above the z-axis, an x–y-axis coupling is generated.
0). Hence, using Taylor’s expansion (10) changes to (14): To include this coupling and supposing an x–y symmetric body
Fi = K x xi + K in i i with i = 1, 2 (14) (I x x = I yy = I0 ), the Eq. (1) must be extended to the y-axis
and the following dynamics representing the gyroscopic effect
4K i 02 4K i 0
with K x = and K in = . by means of (19) must be added [22],
x g3 x g2
Moreover, supposed small changes of the θ angle can be ωIzz ẋ τ
       
I0 0 ẍ 0
considered sin θ ≈ θ. Therefore, (3) can be related to the state- + = x . (19)
0 I0 ÿ −ωIzz 0 ẏ τy
space vector of (2) through (15)1 :
  Since the rotor that is being modeled has two flywheels
L
  
x0 (Fig. 4), it is a good approximation to suppose that the main
  1 0 − − l 2 0
X rig1 2  ẋ0  coupling is provoked by them. Hence, four force inputs and
θ .

=
      (15) another four outputs are added to the model in their positions
X rig2 L
1 0 − l2 0 θ̇ to include this effect, extending (9)–(20).
2
˙x̄ = A x̄ + B 0 F̄ 0 + ω2 U sin ωt + ḡ
 
1 This expression is the same for [x , x ]t (4), changing l by l. (20)
1 2 2 cos ωt
106 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112

Fig. 5. Eigenvalues evolution. Fig. 6. Zoom on eigenvalues evolution of 220 Hz resonance.

ȳ 0 = C 0 x̄
with ȳ 0 = [ ȳ, ȳfly ]t , F̄ 0 = [ F̄, F̄fly ]t , B 0 = [B, Bfly ] and
C 0 = [C, Cfly ]t .
The gyroscopic coupling is modeled as F̄fly = G ȳ˙ fly =
GCfly x̄˙
ωIzz /I0
 
0
with G = ,
−ωIzz /I0 0
and then,
sin ωt
 
x̄˙ = A x̄ + B F̄ + Bfly GCfly x̄˙ + ω2 U + ḡ
cos ωt
sin ωt
   
= [I − Bfly GCfly ]−1 A x̄ + B F̄ + ω2 U + ḡ
cos ωt
ȳ = C x̄ (21)
to conclude that the new state-space A and B matrices are,2 Fig. 7. Zoom on eigenvalues evolution of 450 Hz resonance.

Agyro = [I − Bfly GCfly ]−1 A (22)

−1
Bgyro = [I − Bfly GCfly ] B.
Regarding the whole model of the previous section and
closing the loop with the controller presented in Section 3,
the evolution of the eigenvalues of the flexible modes with the
rotating speed is shown in Fig. 5 and in more detail in Figs. 6
and 7.
The eigenvalues which appear in these figures represent the
forward and backward whirls, which are synchronous rotations
in the same and opposite directions around the fixed coordinate
system respectively. Although they do not exclusively arise
due to the gyroscopic effect, the presented evolution of the
eigenvalues is characteristic when it is taken into account [22].
Even so, as is displayed in Fig. 5 the maximum dispersion of
the resonances is 2.5 Hz (0.5%), therefore, it is concluded that
the gyroscopic effect can be neglected in the working rotating
speed range.
2 Note that the inputs and outputs are the same as that of (9).
2.6. Nonlinear model
(23)
It is worth mentioning here, that the system is obviously
nonlinear, so the linearizations made in Section 2.1 are merely
approximations around the hovering state.
The main source of nonlinearity is the magnetic force, which
change quickly outside the equilibrium position (10).
Despite the simulations of this article having been carried
out without taking into account these nonlinearities in general,
the experimental results on the real device are in agreement
with the simulation while the nonlinear effects are not very
strong. This condition is well preserved inside the useful range
of rotating speeds, where the actuators are capable of inducing
forces strong enough so as to maintain the shaft within working
limits. Outside these limit rotating speeds, the linearity is lost
and sudden nonlinear effects appear. Some results describing
the importance of these nonlinear effects, are presented in
Section 4.3.3. Notice that following the steps of the previous
subsections, it is easy to yield the nonlinear model, too.
 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112 107

Fig. 8. MBC500 Rotor Dynamics-based testbed.

2 2
3. Control design based on symmetry = 1 + A0 − B 0 + 2A0 = 0
= (1 + (A0 + B 0 ))(1 + (A0 − B 0 )) = 0
The first step to analyze the suitability of the model is
to levitate the shaft closing the loop with an appropriate = (1 + (C + D)(A + B))(1 + (C − D)(A − B)) = 0. (26)
control. Given that AMB systems are characterized by a large
In conclusion, the MIMO system can be split into two
bandwidth, they have very fast dynamics, and therefore, a
different control design SISO problems, simplifying the overall
Digital Signal Processor (DSP) has been chosen to implement
control procedure
the control. In this case, the DSP DS1003 board manufactured
by dSPACE GmbH (Paderborn, Germany) driven by a PC in the 1 + C1 (A + B) = 0 (27)
Matlab/Simulink environment is used. The experimental set- 1 + C2 (A − B) = 0 (28)
up, including the MBC500 Rotor Dynamics and the real-time
control platform is shown in Fig. 8. leading to the MIMO system controller
To obtain this controller, and with the aim of making the
C1 + C2
design process easier, only one axis of the linearized model is C + D = C1 =C
considered, leading to the model (18) (Section 2.4), where in ⇒ C −2 C (29)
C − D = C2 1 2
this case, = D.
2
−0.94061(s − 1.261 × 109 )(s + 169.4)(s − 169.4)
A= In this article, it is assumed for the sake of simplicity, that
(s + 4545)(s + 277.8)(s − 277.8)(s + 133.5)(s − 133.5)
D = 0. For this case, it results in C1 = C2 = C.
(s 2 + 13.8s + 9.27 × 105 )(s 2 + 49.51s + 6.802 × 106 ) Physically, this mode separation can be interpreted as the
(24)
(s 2 + 27.91s + 1.866 × 106 )(s 2 + 57.07s + 8.111 × 106 ) superposition of the translational and conical modes which can
0.73335(s − 1.257 × 109 )(s − 46.22)(s + 46.22) describe any motion of the shaft (Fig. 9).
B= It is important to note that this procedure is independent of
(s + 4545)(s + 277.8)(s − 277.8)(s + 133.5)(s − 133.5)
the control design technique used for the two SISO equivalent
(s 2 + 3907s + 7.501 × 106 )(s 2 − 3820s + 7.647 × 106 )
. (25) control design problems, so, separate and different control
(s 2 + 27.91s + 1.866 × 106 )(s 2 + 57.07s + 8.111 × 106 ) methodologies can be applied in each case, depending on the
As it can be observed in (18), the matrix of transfer functions required closed-loop specifications. In this paper, the root locus
fulfills a symmetry condition. This property allows to divide the technique is directly utilized.
MIMO problem into two SISOs, thus making the control design The stability problem requires at least a proportional
easier [21]. h i derivative (PD) controller because this is the minimum required
Considering the symmetric controller G c = CD CD the by almost all AMB systems by nature [2]. Therefore, (30) is
closed-loop characteristic equation results in: inserted to close the loop.

1 + C A + DB CB + DA
 k(z − 0.9)
det C1 (z) = with T = 1/12 000. (30)
CB + DA 1 + C A + DB z
0
B0
 
1+ A Since the MagLev system is rather fast, the sampling
= det =0
B0 1 + A0 frequency has been fixed to the highest attainable value with
108 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
Fig. 9. Translational and conical modes decomposition.
Fig. 10. Root locus of C1 (A + B).
Fig. 11. Root locus of C1 (A − B).
the used DSP controller board, taking into account that DSP
computes not only the stabilizing controller, but is also in
charge of the rotary speed control and the active control, if
any. On the other hand, it has been experimentally proved that
the stability is not maintained for sampling frequencies under
9 kHz.
Using the same controller for (27) and (28) with a gain of
k = 6.5 the stability is reached for both the SISO problems
(Figs. 10 and 11). This gain is selected to maintain the
stability, and it is the same in the theoretical and experimental
schemes. In this case, the utilized criterion is getting the best
experimental behavior, selecting the gain to be near the middle
of the stability margin and trying not to amplify the noise. In
this way, the complete MIMO system stability is achieved.
This controller has been experimentally tested in the
MBC500 Rotor Dynamics with satisfactory results which
are discussed in the remaining sections. Moreover, although
the stability range obtained in simulation is larger than
the actual experimental range, the effects introduced by the
digital implementation and the noise, make the results become
qualitatively similar.
Since the main objective of this paper is to prove the validity
of the described procedure, only a basic controller is designed,
but with the mode separation control methodology based on the
symmetry presented, other higher level control methods could
be applied leading to a better behavior.
4. Simulation and experimental results
The previously designed controller is extended also for the
y-axis leading to (31), with which it is possible to analyze
the suitability of the model since the controllers stabilize the
closed-loop system.
G c (z)
 6.5(z − 0.9) 
0 0 0
 z 
 6.5(z − 0.9) 
 0 0 0 
z .
 
= (31)

 0 0
6.5(z − 0.9)
0


 z 
 6.5(z − 0.9) 
0 0 0
z
In the following subsections, the simulated and experimental
results are presented and compared following this outline.
Firstly, the model behavior is stated studying the shaft dynamics
when it is out of the machine and inside it, influenced by the
bearings. In the second situation two analysis are performed:
nonrotating and rotating. Afterwards, the same study is carried
out with the experimental results. Finally, a comparison
between simulation and experimentation is performed.
4.1. Model behavior
The simulation of the model developed in Section 2 has been
carried out in the Scilab and Matlab environments.
Initially only the shaft is considered obtaining the frequency
response versus the velocity/force magnitude. The units are
chosen in this way because the measurements have been
performed by hitting the rotor with an impact hammer
(Section 4.2). In order to simulate the previous situation, a force
signal is introduced in one of the bearings, to take the position
of the shaft beside the sensor. So, Fig. 12 shows the resulting
graphics with two resonances, one at 220 Hz and the other one
at 450 Hz, preceded by two dips.
Regarding the shaft inside the device, since the model has
symmetric properties, the whole model nonrotating behavior
can be represented by the transfer functions A and B, (24) and
(25) (Section 3).
 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
Fig. 12. Frequency response in the sensor position.
Fig. 13. A and B frequency-response curves.
Fig. 14. Rotational motion of the shaft.
In this way, Fig. 13 shows the frequency response for
transfer functions A and B where two resonances at 220 and
450 Hz can be observed. Considering A, Fig. 13 shows two
sharp dips before the resonances. On the other hand, B has a
soft one near the second resonance.
To conclude, the simulated motion of the shaft when it is
rotating is shown in Fig. 14. One main resonance at 40 Hz and
an increase in amplitude from 200 Hz are noticed.
109
Fig. 15. External measuring system.
Fig. 16. Measured frequency response in the sensor position.
4.2. Experimental results
Following the previous outline, first, the nonrotating
response is measured. These responses are acquired with a
high-accuracy external measuring system, namely the IVS 300
Laser vibrometer, connected to a PC via the OR24 noise &
vibration analyzer [23] (Fig. 15).
The IVS 300 measures vibrational velocity according to the
heterodyne interferometer principle, which generates a high-
frequency carrier signal at the photo detector with the aid of
a Bragg cell [24]. Then, thanks to the OR24 analyzer, real-
time FRFs (Frequency Response Functions) can be obtained
anywhere on the structure, appropriately pointing the laser spot.
Moreover, the provided software allows to save and analyze the
measurements to calculate the modes of the structure.
The measurement procedure when the shaft is outside the
machine consists of hitting with an impact hammer in the
bearing position node of the structure when it is hung and
recording the response through the vibrometer in the sensor
position node. Nevertheless, the final recorded value is the
mean of several of these intermediate measurements. Fig. 16
shows the frequency response hitting in one bearing position
and measuring in the sensor below. Here, two resonances (180
and 370 Hz) and two dips are displayed.
When the shaft is in its operation placement inside
the machine but without rotating, aiming at comparing the
simulation and the experimental results of the model, the
transfer functions A and B are obtained by introducing
110 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
Fig. 17. A and B frequency response curves.
Fig. 18. Rotational behavior.
reference signals in phase and in counterphase and adding and
subtracting the IVS 300 measured outputs respectively [21].
In Fig. 17 two resonances at 180 and 370 Hz and a similar
shape to that shown in Fig. 13 can be observed.
Finally, the characteristics when the shaft is rotating inside
the machine are measured. In this case the tube-guard is
indispensable in order to guarantee a safe operation, and the
Hall effect and Eddy current sensors are used, instead of the
external measurement system. The results which are presented
in Fig. 18 show a main resonance at 40 Hz and an increasing
gain for high frequencies. Then, the system falls into an
instability near 400 Hz (24 000 rpm) and it is impossible to
reach higher frequencies with the implemented controller.
4.3. Comparison between the model and the experimental
behavior
The shapes of the simulated and measured curves, i.e the
position and relative magnitude of the resonances, are similar.
However, there are differences which can be explained taking
into account the effects discussed in the following.
4.3.1. Poles and zeros real location
The position of the model zeros and poles is in part led by the
flexible dynamics. This is studied by Finite Element Analysis
Fig. 19. Nonlinear stable behavior rotating below and at critical speed.
(FEA) for which it is essential to have a precise datasheet of
the rotor structure: material composition, lengths, thicknesses,
etc. Therefore, the imprecisions between the datasheet and the
real shaft cause a difference between the zeros and poles of the
model and the “true” zeros and poles of the device.
Thus, comparing Figs. 12 and 13 with Figs. 16 and 17
respectively, a better tuning of the model is possible changing
the poles and zeros suitably. On the contrary, it is remarkable
that there is agreement between the model and the experimental
results with reference to the shape of the magnitude curve both
for the shaft behavior and for the whole model. In this case,
the decreasing tendency of the frequency response of A and the
dips before the resonances of the simulation are in agreement
with the measurements. Moreover, the increasing tendency of
the first part of the frequency response of B and the sharp slope
of the last part of the same curve are also maintained in the
experimental results.
4.3.2. Nonsymmetry
Although the model has been built assuming that the
mechatronic device has a symmetrical structure, and this
assumption can be a good first approximation to reality, the
actual device has several sources of asymmetry, such as the
location of the bearings, the bearings themselves which are
not exactly matching pairs, the air turbine location, the gravity
effect and even the internal mechanical structure of the shaft
itself.
One experimental proof of this loss of symmetry is found in
Fig. 18, where nonsymmetrical readings along the shaft and the
x- and y-directions are displayed. However, the symmetrical
approximation is good enough to predict stability ranges and
the general behavior of the system when the model purpose is
the design of a suitable controller for the AMB device.
4.3.3. Nonlinearity
The nonlinearities play an important role in the real behavior
of the device. The most important one is introduced by the
magnetic force (10), whose linear approximation is only valid
 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
Fig. 20. Nonlinear unstable behavior rotating below and at critical speed.
around the central hovering state. This issue is particularly
important near the critical speeds, where the position deviates
from the central point and nonlinear effects such as bifurcations
appear, leading the system to abrupt dynamical changes.
Figs. 19 and 20 show the time response of the shaft position,
measured when the rotation frequency is being increased. As
is deduced from the comparison of both graphics, a growing
nonlinear behavior appears until the critical speed is reached,
when the nonlinearity effect leads the system into instability.
It should be noted here, that the amplitude of the response
cannot be the only cause of the destabilization at high
frequencies, since in Fig. 18 it is seen that this amplitude
is larger at, for instance 40 Hz. Also the nonlinear behavior
is reflected in Figs. 19 and 20, taking into account the
appearance of a spurious signal near 40 Hz frequency when the
shaft is rotating at high speeds. The main observed nonlinear
characteristics can be predicted through an harmonic domain
analysis as is described in [11].
5. Conclusions
In this article a modeling process and control for a rotary
shaft levitated through active magnetic bearings has been
presented. This kind of device is of great interest in the
machine-tool industry because of the inherent advantages
of noncontact magnetic levitation technologies in high-speed
mechatronics. The modeling has been developed by the
superposition of three main dynamical effects present in the
system: the rigid shaft dynamics, the flexible dynamics and the
unbalanced motion dynamics. Hence, first, applying Newton’s
laws the rigid part of the model has been calculated, then, the
flexible dynamics has been obtained by using FEA analysis
and, finally, the unbalanced motion characteristics have been
modeled. This procedure has been followed for obtaining both
a linear and a nonlinear model of the whole AMB system.
On the other hand, a study of the gyroscopic coupling is also
performed, concluding its negligibility.
Afterwards, taking into account the symmetric properties
which characterize the obtained model and in general, this
111
type of devices, a controller is designed. This symmetry allows
to perform a separation into translation and conical modes
decoupling the MIMO system into SISO subsystems. Thus,
appropriate controllers for each SISO subsystem are designed
and a stable levitation state is achieved, making possible a static
and dynamic analysis of the system in closed-loop.
It is concluded that, despite the complexity of the system, the
proposed model is accurate enough so that a suitable stabilizing
controller can be successfully designed for the real device.
Moreover, the simulations are able to predict the response of
the machine if they are compared with the experimental results
obtained using either the device provided sensors or an external
measurement system.
Finally, a brief study of the nonlinear behavior of the system
is included, showing the effect of the nonlinearities over the
system stability at critical rotary speeds.
Acknowledgments
The authors are very grateful to the CICYT, the University of
the Basque Country and the Basque Government for the support
of this work, through projects DPI2002-04155-C02-01, PTR95-
0897.OP.CT and 9/UPV00224.310-15254/2003, and the pre-
doctoral grant BF104.466, respectively.
References
[1] Brunet M. Practical applications of the active magnetic bearings to the
industrial world. In: Proceedings of the 1st international symposium on
magnetic bearings. 1988. p. 225–44.
[2] Schweitzer G, Bleuler H, Traxler A. Active magnetic bearings:
Basics, properties and applications of active magnetic bearings. vdf
Hochschulverlag AG an der ETH Zurich; 1994.
[3] Stephenson DA, Agapionu JS. Metal cutting theory and practice. New
York: Marcel Dekker; 1996.
[4] Siegwart R, Larsonneur R, Traxler A. Design and performance of a high
speed milling spindle in digitally controlled active magnetic bearings. In:
Proceedings of the 2nd international symposium on magnetic bearings.
1990. p. 197–204.
[5] Knospe CR. Active magnetic bearings for machining applications.
Control Engineering Practice 2007;15:307–13.
[6] Tamisier V, Carrère F, Font S. Synchronous unbalance cancella-
tion across critical speed using a closed-loop method. In: Pro-
ceedings of the 8th international symposium on magnetic bearings,
http://www.mech.ibaraki.ac.jp/˜ismb8/pocpdf/0047.pdf; 2002.
[7] Overschie PM, Jabben L, Schieveen JP, Spronck JW, Eijk J. Capacitive
sensing system design for a rotating AMB with sub-nanometer positioning
accuracy. In: Proceedings of the 10th international symposium on
magnetic bearings. 2006.
[8] Li L, Shinshi T, Iijima C, Zhang X, Shimokohbe A. Compensation of
rotor imbalance for precision rotation of a planar magnetic bearing rotor.
Precision Engineering 2003;27(2):140–50.
[9] Stainschaden N, Springer H. Some nonlinear effects of magnetic
bearings. In: Proceedings of the 1999 ASME design engineering technical
conferences. 1999.
[10] Ji JC, Yu L, Leung AYT. Bifurcation behavior of a rotor supported by
active magnetic bearings. Journal of Sound and Vibration 2000;235(1):
133–51.
[11] Jugo J, Lizarraga I, Arredondo I. Nonlinear analysis of an AMB system
using harmonic LTV models. In: IEEE international conference in control
applications. 2006.
112 I. Arredondo et al. / ISA Transactions 47 (2008) 101–112
[12] Ren M, Nonami K, Kubo A, Kameno H. Zero bias H∞ control of flexible
rotor magnetic bearing flywheel system with gyroscopic effect using
singular value decomposition. In: Proceedings of the 10th international
symposium on magnetic bearings. 2006.
[13] Houpis CH, Rasmussen SJ. Quantitative feedback theory. New York:
Marcel Dekker; 1999.
[14] The LaunchPoint website, http://www.launchpnt.com/; 2007.
[15] Ogonowski Z. Suboptimal model predictive control for MIMO active
suspension. Mechanics 2005;24(2):130–4.
[16] Petteys R, Parker G, Redmond J. Disturbance rejection control of
an electromagnetic bearing spindle. In: Proceedings of the ASME
international mechanical engineering congress and exposition. 2000.
p. 461–8.
[17] Huang HP, Jeng JC, Chiang CH, Pan W. A direct method for multi-
loop PI/PID controller design. Journal of Process Control 2003;13:
769–786.
[18] Nakajima Y, Sagane T, Watanabe T, Tajima H, Seto K. Modeling, motion
and vibration control for flexible rotor supported by active magnetic
bearings. In: Proceedings of the ASME international design engineering
technical conferences and computers and information in engineering
conference. 2005. p. 1163–71.
[19] Ahn HJ, Han DC. System modeling and robust control of an AMB
spindle: Part I modeling and validation for robust control. KSME
International Journal 2003;17(12):1844–54.
[20] MBC500 Magnetic Bearing System Operating Instructions. Launch Point,
Goleta, CA; 2002.
[21] Jugo J, Arredondo I, Etxebarria V. Analysis and control design of MIMO
systems based on symmetry properties. In: Proceedings of CDC-ECC’05.
2005.
[22] Lee CW. Vibration analysis of rotors. Dordrecht: Kluwer Academic
Publishers; 1993.
[23] OROS S.A. website, http://www.oros.com/; 2007.
[24] IVS 300 Laser vibrometer user manual, Polytec GmbH. Waldbronn,
Germany; 2002.