Paper

Uncertainty Structure and µ-Synthesis of

a Magnetic Suspension System
Member Toru Namerikawa (Kanazawa University)
Member Masayuki Fujita (Kanazawa University)

This paper deals with modeling, uncertain structure and µ-synthesis of a magnetic suspension system. The
dynamics of magnetic suspension systems are characterized by their instability and complexity of electro-
magnets, and they should be robustly stabilized in spite of model uncertainties. First we derive a nominal
design model of the plant under some assumption, then we investigate the gap between the real physical
system and the obtained nominal design model. This gap has complex structure which is expressed by the
structured uncertainties that includes linearization error, parametric uncertainties, and neglected dynamics.
Then we set the interconnection structure which contains the above structurally represented uncertainties.
Next we design a robust controller which achieves robust performance using the structured singular value
µ. Finally, we evaluate the proposed interconnection structure and verify robustness and performance of the
designed µ controller by several experiments.

Keywords: Magnetic Suspension Systems, Uncertain Model, Linear Fractional Transformation, Robust Control,
µ-Analysis and Synthesis

avoided.
1. Introduction There are so many results of robust control of mag-
netic suspension systems, but in these results, the above
Active magnetic suspension systems allow contact- uncertainties have been treated as exogenous distur-
free suspension. They do not suffer from friction nor bances and as unstructured uncertainties (6) (8) , how-
wear, and this is the most important advantage of these ever, both of the uncertainty descriptions lead to conser-
systems. This technology is now used for various indus- vative analytic results for robust stability/performance
trial purposes, and has already been applied to magnet- tests. The goal of robust control theory is not only to
ically levitated vehicles, magnetic bearings, etc. (1) (2) . get robust controllers but also to know the quantitative
Recent overviews, advances and applications in this field limitation of stability/performance of the controller.
can be seen in (3) (5) . In (9) (11) , parameter perturbations were considered,
Since an active controlled-type magnetic suspension and the model uncertainties were described structurally,
system is inherently unstable, feedback control is indis- this result succeeded to reduce the conservativeness of
pensable to stabilize the system. A conventional PID analysis. But the considered model perturbation is
controller is often employed as a feedback compensator, imaginary and they assumed that all model parame-
and this method often yields enough stability and per- ters have same several percent uncertainties, but this
formance, but owing to model uncertainties and changes assumption can not fit the real physical phenomenon.
of the operating points, the entire system sometimes be- Recently the state-space control theory of uncertain
comes unstable. system with Linear Fractional Transformation(LFT)
To avoid this problem, the approach taken here is an has almost been settled up for practical use (12) . LFT
application of robust control methodologies. It is well have come to play an important role in control system
known this is one of the effective control techniques for design, and provide a uniform framework for realization,
unstable systems. On the control of magnetic suspen- analysis and synthesis for uncertain systems. Now its
sion technology field, one of the most critical problems practical evaluation is extremely expected.
is a description of a complex behavior of the dynamics In this paper, we present the model and uncertainty
of electromagnets and their forces. description of a magnetic suspension system by using
The exact description of this behavior is almost hope- LFT, which contains less conservativeness for robust
less, and even if it should be achieved by infinite di- stability/performance analysis. Further we propose a
mensional nonlinear differential equations, the result- method to quantify the magnitude of uncertainties.
ing model is only effective for the simulations/analysis, Uncertainties we consider in this paper are the lin-
but can not be used for a control system design as it earization error of the electromagnetic force, unmod-
is. Then some approximations and assumptions must eled dynamics of electromagnet, and parametric un-
be employed, and consequently the gaps between the certainties. We structurally describe these three types
real physical system and the design model cannot be

1080 T.IEE Japan, Vol. 121-C, No.6, 2001

 -

Fig. 2. Nominal linear model for M.S.S.

These assumptions are not very strong and suitable

around the steady state.
Under these assumptions, we derive the following
three equations, which show an equation of the motion
of the iron ball, electromagnetic force and equation of
an electric circuit of the electromagnet, respectively.

d2 x(t)
M = M g − f (t), · · · · · · · · · · · · · · · · · · · · (1)
Fig. 1. Magnetic Suspension System dt2

2
I + i(t)
of uncertainties by using real/ complex bounded num- f (t) = k , · · · · · · · · · · · · · · · (2)
bers/matrices. For robustness analysis, we employ the X + x(t) + x0
mixed structured singular value(mixed µ) test (13) to re-
di(t)
duce conservativeness. L + R(I + i(t)) = E + e(t), · · · · · · · · · · · · (3)
Finally, based on an uncertainty model, we derived a dt
µ controller. At the end of our paper, we evaluate the where M is the mass of the iron ball, X is the steady
proposed model and uncertainty structure by several gap between the electromagnet(EM) and the iron ball,
experimental results concerned to the robust stability x(t) is the deviation from X, I is the steady current, i(t)
and robust performance. is the deviation from I, E is the steady voltage, e(t) is
2. Magnetic Suspension System and its Model the deviation from E, f (t) is the electromagnetic force,
k and x0 are coefficients of f (t) which are determined
In this section, we introduce the most basic magnetic
by experiments, L is an inductance of the EM, and R
suspension system which has only one degree of free-
is a resistance of the EM.
dom. Then we derive an ideal mathematical model of
In the case we apply the linear control theory with
the system based on physical laws and several assump-
respect to this system, the problem is that the equation
tions.
of the electromagnetic force (2) is nonlinear concerning
2.1 Construction of the system Consider the
x(t) and i(t). Here we utilize the standard linearization
elecromagnetic suspension system shown schematically
approach based on the Taylor series expansion around
in Fig.1. An electromagnet is located at the top of the
the operating point.
experimental system.
The control problem is to levitate the iron ball sta-
bly utilizing the electromagnetic force. The ball’s mass
2
I
M is 1.04kg, and steady state gap X is 5mm. Note f (t) := k − Kx x(t) + Ki i(t),
X + x0
that this simple electromagnetic suspension system is
unstable without feedback control. A standard induc- 2kI 2 2kI
Kx := , Ki := . · · · · · (4)
tion probe of eddy current type gap sensor is placed (X + x0 )3 (X + x0 )2
below the ball to detect the distance between the iron
ball and the electromagnet. Moreover, the steady state equations are given by
 2
2.2 Mathematical Model First, we introduce I
M g = k X+x 0
and RI = E, then according to equa-
following assumptions (1) (2) (6) in order to derive a nom-
inal model of this system by physical laws. tions (1), (3), (4) and these two steady state equations,
[A1] Magnetic flux density and magnetic field do not the nominal transfer function of the magnetic suspen-
have any hysteresis, and they are not saturated. sion system is easily derived as
[A2] There is no leakage of flux in the magnetic cir- −Ki
cuit. Gnom (s) := . · · · · · · · · · (5)
(M s2 − Kx )(Ls + R)
[A3] Magnetic permeability of the electromagnet is
infinite. This equation shows the system is unstable and oscil-
[A4] Eddy currents in the magnetic pole can be ne- latory. Further, in Fig.2 the nominal block diagram of
glected. the magnetic suspension system is shown, and it shows
[A5] Coil inductance is constant around the operating the structure of the plant. The positive feedback from
point, and any electromotive force due to a motion x to ẍ through Kx makes the system unstable.
of the iron ball can be neglected. The nominal model parameters are given in Table 1.

C 121 6 13 1081
 Table 1. Model Parameters

Parameter Nominal Value Unit

M 1.04 [kg]
X 5.00 ×10−3 [m]
I 0.789 [A]
k 1.71 ×10−4 [N m2 /A2 ]
x0 -1.80 ×10−3 [m]
Kx 6.27 ×103 [N/m]
Ki 25.7 [N/A]
L 0.859 [H]
R 24.76 [Ω]

3. Structured Uncertainties
The derived nominal model (5) and/or Fig. 2 with
nominal model parameters works fairly well around the
steady state operational point. However, if the state of
the system deviates from the nominal operating point, Fig. 3. Current-Force Curve and Gap-Force Curve
the model no longer suitably describes the physical sys-
tem.
We treat this gap as a model uncertainty, and we where Ki0 and Kx0 are nominal values, ki and kx
make a new extended model, which is a set of plant are uncertainty weights determined from slopes of the
models, that is constructed with the nominal model and dash-dotted lines.
3.2 Parametric Uncertainty The first request
model uncertainties. This set of models can cover the
relatively wider behaviors of the real plant, but still not for the system is robust stability against unexpected
globally. This set is an extension of the nominal model. exogenous force disturbances. Another general demand
The following three items are well known to be the most in practical use of the magnetic suspension system is a
general and serious uncertainties (14) : flexible change of the mass of suspended objects. These
• Linearization Error two specifications can be described by a parametric per-
• Parametric Uncertainty turbation of a mass of the iron ball M . We describe it
• Unmodeled Dynamics as
We discuss these uncertainties in the following sec- 1 1
tions, and include them into the set. = , |δM | ≤ 1, · · · · · · · · · · · · · (8)
M M0 + kM δM
3.1 Linearization Error There should be model
uncertainties caused by linearization of the electromag- where M0 is the nominal value, and kM is an uncer-
netic force, which was generated by the approximation tainty weight.
from equation (2) to (4). 3.3 Unmodeled Dynamics In this section, we
In Fig.3, Current-Force (i − f ) curve for a gap width discuss the dynamics of electromagnets. Nominally it is
1
X=5.0mm is plotted in the upper figure, and Gap-Force expressed by a transfer function GEM (s) := Ls+R . It is
(x−f ) curve at current I=1.15A is depicted in the lower well known that an inductance L and a resistance R of
figure, where a symbol “◦” denotes measured experi- the electromagnet have frequency varying and gap(x)-
mental data at each point, and solid curved lines show dependent characteristics. Further, these parameters
the determined Current-Force and Gap-Force curve, re- are very sensitive to be measured. Nominal values of L
spectively. These curves are determined from the least and R are determined as averages of five measurements
squares approximation laid on the equation (2). Two under the condition f = 10Hz and X = 5mm. Figure
dashed straight lines indicate tangents of each curve 4 shows the experimental data of GEM (s), where the
at the operating points. These inclinations of tangents solid curved line indicates the nominal frequency re-
are employed as Ki and Kx from equation (4), respec- sponse which is located in the center of a band, dashed
tively. The four dash-dot straight lines are sectors of lines show upper and lower bounds.
the linearization errors, which we will use them as sec- The transfer functions of the electromagnet are dis-
tor bounds in the following. tributed in a frequency dependent belt. Furthermore, if
Fig.3 shows that the perturbations between tangents the frequency of the input signal changes, this belt be-
and curves become bigger if the operating points move comes broad. We describe this belt as an unstructured
from the original points. These errors were caused by uncertainty of the following:
linearization. Here we employ sector bounds to account
for the linearization error, and describe Ki and Kx as 1
GEM (s) := + wi (s)∆i (s),
L0 s + R0
|∆i (jw)| ≤ 1. · · · · · · · · · · · · · · · · (9)
Ki := Ki0 + ki δi , |δi | ≤ 1, · · · · · · · · · · · · · · · · (6)
Kx := Kx0 + kx δx , |δx | ≤ 1, · · · · · · · · · · · · · · (7) Where L0 and R0 , are the nominal values of L and

1082 T.IEE Japan, Vol. 121-C, No.6, 2001

 -

4.1 Quantification of Uncertainties The ob-

tained set of models G̃(s) gives only a structure of the
uncertainties, not a quantity of each uncertainty.
In the previous research, tuning of the design param-
eters was depended on experimental/ simulated trial
and error. Development of a systematic tuning method
for the design parameters is now expected. Further,
the physical limit of allowable perturbations for robust
stability/performance was not clear. We quantify the
amount of uncertainty based on a change of the oper-
ating point.
4.1.1 Change of the Operating Point Our
approach taken here is to determine the set of the plant
models based on the change of the operating point.
The operating point of this system is characterized by
a steady state gap {X | Xmin ≤ X ≤ Xmax }. As
a design specification for real applications, the range
Fig. 4. Frequency responses of GEM (s) {X | Xmin ≤ X ≤ Xmax } is expected to be wider, but
this change of the operating point X causes perturba-
tions of Ki and Kx , hence the allowable range of the
operating point {X | Xmin ≤ X ≤ Xmax } is limited.
After the several iterations of controller design and
control experiments, we have finally chosen the allow-
able range of the operation point as follows, where µ-
synthesis based on D − K iteration and the mixed µ-
analysis were employed[13,15].

{X | 3.8 mm ≤ X ≤ 6.2 mm} · · · · · · · · · · · · · (10)

4.1.2 Perturbation of Ki and Kx Any change

of the operating point X also causes the perturbations
of the parameters Ki and Kx . We utilize Fig.3 to deter-
mine the bound of the perturbations of Ki and Kx . A
magnetic force f and a gap X and a current I are related
each other, and are written in (2). This equation (2) is
Fig. 5. Set of Plant Models with Uncertainties: transformed to (4) using the standard linearization ap-
G̃(s) proach based on the Taylor series expansion around the
operating point.
The experimental data corresponding to (2) is Fig.3.
R, respectively, and wi (s) is an uncertainty weight. The A symbol “◦” denotes measured experimental data at
magnitude of the weight wi (s) is determined as half of each point, and solid curved lines show the deter-
the width of the belt in Fig.4. mined Current-Force and Gap-Force curve, respectively.
3.4 Set of Plant Models: G̃(s) Up to now, These curves are determined from the least squares ap-
three types of uncertainties have been considered. Now proximation laid on the equation (2). In Fig.3, Current-
in this section, we use them to construct a model of Force (i − f ) curve for a gap width X=5.0mm is plotted
this plant. Actually, we make a set of plant models in the upper figure, and Gap-Force (x − f ) curve at cur-
G̃(s) in LFT form (12) which includes the above nomi- rent I=1.15A is depicted in the lower figure.
nal model, linearization error, parametric uncertainties, The following is a procedure to determine the bounds
and unmodeled dynamics. Using the LFT representa- of perturbations of Ki and Kx .
tion, then the block diagram model of the system is ( 1 ) First, decide the operating point X, which has
replaced from Fig.2 to Fig.5. The state space repre- a one-to-one correspondence to I, hence the I is
sentation of G̃(s) is easily derived, so in this paper it fixed. The operating point in the horizontal axes
is omitted. Now we have extended the nominal model in Figs 3 (upper and lower) have been fixed.
and obtained a set of plant models by LFT represen- ( 2 ) Calculate tangents of each curve at the oper-
tation which is expected to express the relatively wider ating points. These inclinations of tangents are
behaviors of the real plant. employed as Ki and Kx from equation (4), re-
4. µ-Analysis and Synthesis spectively. In Fig. 3, two dashed straight lines
indicate tangents.
In this section, we design a robust controller for the ( 3 ) Change the operating point, and continue
set of models G̃(s) step(1) and step(2).

C 121 6 13 1083
 In this case, parameters Ki and Kx perturb as 14.1 ≤ Table 2. Quantity of uncertainties
Ki ≤ 37.3 and 5.38×103 ≤ Kx ≤ 7.16×103 (3.8 ≤ X ≤ Value Value
ki 11.6 kx 8.90 × 102
6.2). Then we describe Ki and Kx as below. 1.28×10−3 (s+3.20)(s+900)
kM 7.25 × 10−2 wi (s) (s+25.8)(s+31.4)

Ki = 25.7 + 11.6 · δi , |δi | ≤ 1, · · · · · · · · · · · (11)

Kx = 6.27 × 103 + 8.90 × 102 · δx ,
|δx | ≤ 1. · · · · · · · · · · · · · · · · · · · · · · · (12)

4.1.3 Dynamical Uncertainties Uncertainties

in the dynamics of electromagnet wi (s) is also consid-
ered in this section. wi (s) is not only an uncertainty
of the dynamics of electromagnet but an important de-
sign parameter. After several design iterations, the final
bound of wi (s) and GEM (s) has been chosen to obtain
an appropriate robustness as below.

1
GEM (s) :=
0.859s + 24.8
1.28 × 10−3 (s + 3.20)(s + 900)
+ ∆i (s),
(s + 25.8)(s + 31.4)
|∆i (jw)| ≤ 1. · · · · · · · · · · · · (13) Fig. 7. Frequency response of the weight wi (s)

In this wi (s), the parametric uncertainty of L and R

as 0.782 ≤ L ≤ 0.936 (9% perturbation) and 24.5 ≤
R ≤ 25.0 (1% perturbation) are involved. In addition
to these parametric perturbation, unmodeled dynamics
in the high frequency range is also included in wi (s).
Frequency response of the final set of electromagnetic
dynamics is shown in Fig. 6.

Fig. 8. Interconnection Structure

4.2.1 Interconnection structure We construct

an interconnection structure by LFT representation
in Fig.8, where Wperf is a performance specification
and also is a weight for a sensitivity function S :
Fig. 6. Final transfer function GEM (s) with un- = (I + Gnom K)−1 .
certainty wi (s) For the disturbance attenuation and the tracking for
reference signal, the controller is expected to have in-
From the above discussion in these three subsections, tegral property. In order to achieve this specification,
the final quantity of uncertainties are selected in Table Wperf was chosen as the following function.
2, where 7% perturbation of a mass M is considered.
Final frequency response of wi (s) is shown in Fig.7. 100
4.2 Control System Design Utilizing the struc- Wperf (s) = . · · · · · · · · · · · · · · · · · · · · (14)
1 + s/0.1
tured singular value µ (13) (15) (16) , we design the con-
troller which achieves robust performance against vari- 4.2.2 Control Problem Next, for the robust
ous types of uncertainties. performance synthesis, we define the block structure ∆

1084 T.IEE Japan, Vol. 121-C, No.6, 2001

 -

as follows.
∆ := {diag[δi , δx , δM , ∆i , ∆perf ] : δi , δx , δM ∈ R,
∆i , ∆perf ∈ C}. · · · · · · · · · · · · · · · · · · · · · (15)
It is well known that the structured singular value
µ∆ (M ) is defined for matrices M ∈ Cn×n with the
block structure ∆ as

µ∆ (M )
1
:= −
(16)
min{σ (∆) : ∆ ∈ ∆, det(I − M ∆) = 0}

unless no ∆ ∈ ∆ makes (I − M ∆) singular, in which

case µ∆ (M ) := 0. Then, the control problem is to find
the controller K(s) which achieves the following robust
performance condition, where P (s) is the generalized Fig. 9. Final Controller K̂(s)
plant which is expressed by the gray rectangle box in
Fig.8.
less than 1, the closed-loop system with uncertainties
sup µ∆ [Fl (P (jω), K(jω))(jω)] < 1. · · · · · · · (17) achieves the robust performance condition.
ω∈R
This result shows K̂(s) guarantees robust perfor-
We apply the standard D − K iteration (16) to find mance against uncertainties caused by a change of op-
the sub-optimal µ controller for the system. We thus erating point { X | 3.8 ≤ X ≤ 6.2}(10).
iteratively solve the following problem:

sup inf {σ̄(D(jω)Fl (P, K)D−1 (jω))} < 1. (18)

ω∈R D(ω)

The block structure (15) is used to calculate the

mixed µ value (17), but the real-valued blocks in (15)
are replaced to complex-valued ones in the calculating
and fitting process of the scaling matrix D.
4.2.3 Robust µ Controller After the 3rd iter-
ation, we obtained a controller K(s), where the supre-
mum of µ∆ [Fl (P, K)] is 0.9766. Final scaling matrix
D(s) has 12 states, then K(s) has 30 states. We em-
ployed the Hankel norm approximation technique to cal-
culate the reduced order system of K(s). Final balanced
controller K̂(s) is as follows, and its bode diagram is
shown in Fig.9. The supremum of the µ∆ [Fl (P, K̂)] is
also 0.9766.
3.27 × 1010 × (s + 486 ± 885i) Fig. 10. µ∆ [Fl (P, K̂)] and σ̄[DFl (P, K̂)D −1 ]
K̂(s) =
(s + 1740)(s + 949 ± 1320i)
(s + 389 ± 626i)(s + 335)(s + 79.1)
× 5. Experimental Evaluation
(s + 472 ± 794i)(s + 391 ± 599i)
In order to evaluate the proposed set of plant models
(s + 29.5)(s + 14.7)(s + 4.86)
× G̃(s), we implement the obtained controller K̂(s) via
(s + 348)(s + 8.16)(s + 2.66) a digital control system, and carried out experiments.
(s + 2.63)(s + 0.175)(s + 0.114) The sampling period of the controller is 95µs, and a well
× (19)
(s + 0.210)(s + 0.127)(s + 0.0778) known Tustin transform was employed for discretiza-
tion. All experimental results of time response of the
Calculated upper and lower bounds of µ∆ [Fl (P, K̂)] iron ball position are shown in Fig.11, 12, 13.
with the controller K̂(s) are shown in Fig.10, where 5.1 Evaluation of Nominal Performance Step
the two solid lines respectively show upper and lower response of the position x of the iron ball at X =5mm
bounds of µ and the dashed line shows the maximum (nominal steady gap) is shown in Fig.11, which indi-
singular value of (D(jω)Fl (P (jω), K̂(jω))(jω)D−1 (jω)). cates the stable levitation with the controller K̂(s) at
Since the peak value of the upper bound of µ is the nominal steady gap X = 5.0 mm. The magnitude

C 121 6 13 1085
of the step-type disturbance is 22 N, which is twice as
much as steady state force. Since it is difficult to give
disturbance forces to the iron ball directly and repeat-
edly, we add pseudo-disturbance by applying voltage
signal to the control input signal. This figure shows the
nominal performance.
5.2 Evaluation of Robust Stability Next our
interest is robust stability of the closed-loop system.
Time responses of the controllers K̂(s) are shown in
Fig.12, which indicate the stable levitation at the steady
state gaps X = 1.3, 5.0, 8.7 mm.
• The robust stability against the perturbation
{ X | 1.3mm ≤ X ≤ 8.7mm} is achieved.
If we change the steady state gap X to less than
X = 1.3, or greater than X = 8.7, however, the sys-
tem disappointingly becomes unstable.
5.3 Evaluation of Robust Performance The
final evaluative item is our main control problem, “ro-
bust performance”. For the sake of verification of the ro-
Fig. 11. Step Response at X = 5mm
bust performance, we measured time responses against
a step-type external disturbance (22 N) at the steady
state gaps X = 3.8, 6.2 mm. Results are shown in
Fig.13. Apparently, the controller K̂(s) shows enough
performance comparing the response in Fig11. We have
confirmed that
• K̂(s) achieves the robust performance against
model perturbations caused by a change of oper-
ating point { X | 3.8mm ≤ X ≤ 6.2mm} (10).
When we change the steady state gap X to
{ X | 2.8mm ≤ X ≤ 7.2mm}, the system keep up the
almost same response, but if the steady state gap X
would be less than X = 2.8, or greater than X = 7.2,
the response suddenly deteriorates.
6. Conclusion
In this paper, we proposed the novel model and uncer-
tainty structure of magnetic suspension systems by us-
ing LFT, which contains less conservativeness for robust
stability/performance analysis. Further we proposed Fig. 12. Time Response at X = 1.3, 5.0, 8.7mm
one method to quantify the magnitude of uncertain-
ties. Uncertainties we considered in this paper are the
linearization error of the electromagnetic force, unmod-
eled dynamics of the electromagnet, and parametric un-
certainties. We structurally described these three types
of uncertainties by using real/ complex bounded num-
bers/matrices. Next, we designed a robust controller by
µ-analysis and synthesis which achieves robust perfor-
mance by using the structured singular value µ. Here,
we employ the mixed µ test to reduce conservativeness.
Finally, we evaluate the proposed model and uncer-
tainty structure by several experimental results con-
cerned to the nominal performance, robust stability and
robust performance.

Acknowledgment
The authors wish to thank Mr. Florian Lösch and Prof.
Gerhard Schweitzer with ETH for helpful discussions
and suggestions. Fig. 13. Step Response at X = 3.8, 6.2mm
(Manuscript received July 3, 2000, revised December

1086 T.IEE Japan, Vol. 121-C, No.6, 2001

 -

7, 2000) Masayuki Fujita (Member) He received the B.E., M.E.

and Dr. of engineering degrees in electrical
engineering from Waseda University, in 1982,
References 1984 and 1987, respectively. From 1995 until
1992, he was with Kanazawa University. From
1992 until 1998, he was with the Japan Ad-
(1) P. K. Sinha, Electromagnetic Suspension: Dynamics and vanced Institute of Science and Technology.
Control, IEE, Peter Peregrinus, London, 1987.
In April 1998, he joined Kanazawa Univer-
(2) G. Schweitzer, et al., Active Magnetic Bearings, Hochschul-
sity, where he is presently a Professor in the
verlag AG an der ETH Zurich, 1994.
(3) P.E. Allaire and D.L. Trumper (Editor): Proceedings of
Department of Electrical and Electronic Engi-
the Sixth International Symposium on Magnetic Bearings, neering. From 1994 to 1995, he held a visiting position in the
Technomic Publishing Company, Inc., 1998. Technical University of Munich, Germany. His research interests
(4) Special Issue on Magnetic Bearings, JSME International include robust control and its applications to robotics and mecha-
Journal Series C, vol.40, no.4, pp.553-606, 1997. tronics.
(5) C.R. Knospe and E.G. Collins (Editor): Special Issue on
Magnetic Bearing Control, IEEE Trans. on Control Sys-
tems Technology vol.4, no.5, pp.481-608, 1996.
(6) M. Fujita, T. Namerikawa, et al., “µ-Synthesis of an Elec-
tromagnetic Suspension System,” IEEE Trans. Automatic
Control, vol.40, no.3, pp.530-536, 1995.
(7) F. Lösch, C. Gähler and R. Herzog, “µ-Synthesis Controller
Design for a 3 MW Pump Running in AMBs,” Proc. of the
Sixth Int. Symp. on Magnetic Bearings, pp. 415-428, 1998.
(8) K. Nonami and T. Ito: “µ Synthesis of a Flexible Rotor Mag-
netic Bearing Systems,” IEEE Trans. on Control Systems
Technology, vol.4, no.5, pp.503-512 1996.
(9) T. Sugie and M. Kawanishi, “Analysis and Design of Mag-
netic Levitation Systems Considering Physical Parameter
Perturbations,” Trans. of ISCIE, vol.8, no.2, pp.70-79,
1995.
(10) R. Fittro and C. Knospe, “µ Synthesis Control Design Ap-
plied to a High Speed Machining Spindle with Active Mag-
netic Bearings,” Proc. of the Sixth Int. Symp. on Magnetic
Bearings, pp.449-458, 1998.
(11) M. Hirata, K. Nonami and T. Ohno, “Robust Control of a
Magnetic Bearing System Using Constantly Scaled H∞ Con-
trol,” Proc. of the Sixth Int. Symp. on Magnetic Bearings,
pp.713-722, 1998.
(12) C. Beck, R. D’Andrea, F. Paganini and J.C. Doyle, “A State-
Space Theory of Uncertain Systems,” Proc. of 13th IFAC
World Congress, pp.291-296, 1996.
(13) P.M. Young, “Controller design with real parametric uncer-
tainty,” Int. J. of Control, vol.65, no.3, pp. 469-509, 1996.
(14) F. Paganini, “Sets and Constraints in the Analysis of Uncer-
tain Systems,” Ph.D. Thesis, California Institute of Tech-
nology, 1996.
(15) B.J.Balas, J.C.Doyle, K.Glover, A.Packard and R.Smith, µ-
Analysis and Synthesis Toolbox User’s Guide, MUSYN Inc.
and Math Works, 1996.
(16) K. Zhou and J.C. Doyle, Essentials of Robust Control,
Prentice-Hall Int. Inc., 1998.

Toru Namerikawa (Member) He received the B.E., M.E.

and Dr. of Engineering Degrees in electri-
cal and computer engineering from Kanazawa
University, Japan, in 1991, 1993 and 1997,
respectively. Since 1994, he has been with
Kanazawa University. He is currently an As-
sistant Professor in the Department of Elec-
trical and Electronic Engineering. From 2001,
he holds a visiting position in the University of
California, Santa Barbara, USA. His research
interests are robust control applications to electro-mechanical sys-
tems.

C 121 6 13 1087