632 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 10, NO.

6, DECEMBER 2005

Feedback Linearization of Active Magnetic Bearings:

Current-Mode Implementation
Min Chen and Carl R. Knospe, Member, IEEE

Abstract—Feedback linearization is a promising approach to developing simple, yet accurate, multiple-pole AMB models,
the nonlinear control problem posed by active magnetic bearing which have a significant influence on the success of the feedback
systems. In this paper, feedback linearization is employed in com- linearization approach. For multiple pole actuators, the dynamic
bination with robust control techniques for the regulation of a
single axis test rig actuated by a multiple pole magnetic bear- (voltage-mode) linearization approach proposed by Lindlau
ing. To this end, a nonlinear polynomial model of the magnetic et al. [3], which employed both magnetic finite element analysis
actuator was developed based on its experimental calibration. and experimental calibration, was felt to be cumbersome in
The effect of the amplifier and measurement system dynamics practice. The static (current-mode) linearization approach has
on the feedback linearization performance, was also examined, been experimentally investigated on a three-pole AMB system
and compensation filters were developed. Finally, an uncertainty
framework was proposed for the linearized plant, and a robust con- in [2]. However, the linearization controller results in a limit
troller was designed via µ synthesis. Experimental results demon- cycle after the rotor is levitated, due to impact of the remaining
strate that the feedback-linearized active magnetic bearing system bearing nonlinearity on the system stability. In this paper, it is
can achieve stability and the specified performance over the en- demonstrated that linearization can be successfully performed
tire range of bearing clearance. The introduction of compensation using current-mode amplifiers and static calibration data. The
filters is shown to be essential to this result.
use of current-mode feedback linearization necessitates the
Index Terms—Active magnetic bearings (AMBs), current mode, introduction of a compensation filter which is introduced here
feedback linearization, micro-synthesis. in an ad hoc fashion.
In practice, feedback linearization will not completely
I. INTRODUCTION compensate the actuator nonlinearity due to uncertainty in
the bearing model. In this paper, an uncertainty struc-
CTIVE magnetic bearings (AMBs) are noncontact
A actuators that employ attractive electromagnetic forces
to constrain and often manipulate a suspended object, called
ture was proposed for the linearized plant nominal model
associated with current-mode control. The magnitude of
this uncertainty is estimated from experimental data. As a
a flotor. These devices have been of increasing interest to the consequence, the feedback linearization approach can be com-
manufacturing industry due to their high force capacity, rela- bined with linear robust control techniques to develop con-
tively large travel range, high speed capability, and potential for trollers for the nonlinear system that achieve closed-loop sta-
enhanced damping via active control. However, AMB systems bility and adequate performance in the presence of model
are highly nonlinear and open-loop unstable. Thus, when high uncertainty.
performance is desired, they present a challenging control In this study, feedback linearization is applied to a flexible,
problem. In recent years, the feedback linearization approaches single-axis system driven by an AMB actuator. Discussions will
have been widely discussed in AMB applications [1], [3]–[5]. be focused on the following implementation issues: 1) devel-
This approach utilizes a complete nonlinear description of an opment of a high-order parametric model of the AMB so as to
AMB system to yield a feedback law that transforms the plant fit the calibration data; 2) development of a lookup table to im-
to a linear one. Therefore, feedback linearized AMB systems plement the feedback linearization algorithm within the digital
should behave linearly over a much larger travel range than controller; 3) evaluation and compensation of the influence on
achievable using traditional Jacobian linearization [6]. the feedback linearization of the dynamics of AMB peripheral
The feedback linearization approaches have been studied devices, such as power amplifiers and antialiasing filters; and
in simple magnetic levitation systems consisting of one or 4) characterization of the residual uncertainty and development
two electromagnets [1], [3], [5]. However, for typically-used of a robust controller.
multiple-pole radial magnetic bearings, most investigations This paper is arranged as follows. In Section II, the con-
have focused on theoretical analysis and computer simula- cept of feedback linearization is briefly reviewed. Section
tion [7], [8]. Few experimental results [2] have been reported III presents the experiment. Section IV follows with an
in the literature. One reason for this situation is the difficulty of examination of AMB modeling and the nonlinear compen-
sation. The influence of the amplifier, and measurement
Manuscript received May 31, 2005; revised August 4, 2005. Recommended system dynamics on the system is examined in Section V.
by Guest Editor L. S. Stephens. In Section VI, a feedback linearization uncertainty frame-
The authors are with the Mechanical and Aerospace Engineering De- work is proposed, which is then used in robust controller
partment, University of Virginia, Charlottesville, VA 22903 USA (e-mail:
[email protected]; [email protected]). design in Section VII. Finally, conclusions are provided in
Digital Object Identifier 10.1109/TMECH.2005.859824 Section VIII.

1083-4435/$20.00 © 2005 IEEE

CHEN AND KNOSPE: FEEDBACK LINEARIZATION OF ACTIVE MAGNETIC BEARINGS: CURRENT-MODE IMPLEMENTATION 633

Fig. 1. Feedback linearized AMB system.

II. FEEDBACK LINEARIZATION

For a laminated magnetic bearing driven by a trans-
conductance (current-mode) amplifier, the electromagnetic
force f can be expressed as a static nonlinear function of the
controlled coil current Ip and the flotor displacement from the
center position x: Fig. 2. Photograph of the AMB test rig.
F = f (Ip , x). (1)
In current-mode feedback linearization, the controlled current for the application. The air gap of the bearing is approximately
Ip is chosen as the inverse function of (1) with respect to a 250 µm on each side. An eddy current transducer is used as a
reference force Fc , and is parameterized by the displacement x position sensor to measure the journal motion (hence air gap). A
dSPACE DS1103 digital control system is used for regulation
Ip = f −1 (Fc , x) (2) of the system. The control goal is to minimize the vibration of
the assesment platform in response to exogenous forces applied
so as to cancel the nonlinear function f (Ip , x) as illustrated in
to it through active control of the actuator platform. Please note
Fig. 1.
that the measurement of the assessment platform’s vibration is
Eddy currents are not included in the model (1) as the ac-
not made available to the controller.
tuator is composed of laminations. Hysteresis is present but
less significant than other sources of nonlinearity due to the air
IV. AMB MODELING
gap reluctance. Furthermore, this feedback linearization scheme
does not depend on the flotor dynamics. The resulting linear sys- A. AMB Calibration
tem is valid across the entire range of flotor displacements and
An accurate nonlinear model is necessary for feedback lin-
coil currents. Since the system from Fc to x behaves linearly, the
earization. Due to the complexity of the electromagnetic fields,
control design problem is simplified to a linear one. Of course,
it is difficult to obtain accurate analytical models for AMBs.
this approach ignores the amplifier dynamics. As shall be shown
However, as indicated by (1), the force-current-gap relationship
in the sequel, this presents some difficulties in application. An
can be determined from appropriate experimental tests on the
alternative to the feedback linearization approach that avoids
AMB to avoid the disadvantages of analytical models.
this issue is to employ a voltage-mode amplifier to drive the
For calibration, a stiff spring was used to pull the actuator
actuator and perform dynamic feedback linearization, as done
platform along (or opposite to) the axis of intended motion. A
in [3]. This may prove to be difficult for multiple pole AMBs,
proportional–integral–derivative (PID) controller was designed
however, due to the electrodynamic coupling between axes, as
to stabilize the actuator platform at any specified air gap. The
the magnetic circuits are not independent. We leave this avenue
spring force, air gap, and bearing current were measured by
for a future investigation and examine here how current-mode
a load cell, eddy current sensor, and current monitors, respec-
linearization may be effectively employed.
tively. The calibration was carried out in the following manner.
With a known spring force applied and a specified position
III. AMB TEST RIG
offset on the platform, the PID controller regulated the AMBs
Fig. 2 shows the magnetic bearing experiment employed in current to generate the proper electromagnetic force to balance
this study. The test rig consists of two platforms, the assess- the external force acting on the platform
ment platform and the actuator platform, connected by a leaf
F = −Fspring − kx (3)
spring, and constrained to single axis rectilinear motion by flex-
ures. The actuator platform is rigidly connected to the bearing where k is the static stiffness of the actuator platform flex-
journal, which is made of cobalt iron laminations. This journal ures, and Fspring is the measured spring force obtained from
can be driven by the AMB stator which is mounted above the the load cell. After transients have decayed, the calculated elec-
actuator platform. The magnetic bearing stator used is a con- tromagnetic force, the specified air gap, and the corresponding
ventional eight pole radial design. Two groups of coils, driven measured input current together are one set of static calibra-
by two identical 150 V pulse width modulation (PWM) power tion data. The AMB was calibrated at different positions over
amplifiers, generate opposing attractive forces. A bias current of the entire clearance, and at different perturbation coil currents
1 A is applied to the coils to guarantee sufficient force slew rate ranging from −1 A to 1 A. During the calibration, the AMB was
634 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 10, NO. 6, DECEMBER 2005

The linearizing current was calculated for each journal dis-

placement from −250 µm to 250 µm, with 5 µm increments,
and each force from −1000 N to 1000 N, with 10 N increments.
Then these results were stored in a 2-D look-up-table (LUT).
For inputs not in the table, the LUT uses either interpolating or
extrapolating techniques to perform linear mapping from inputs
Fc and x to output Ip . Compared to the traditional inverse func-
tion computation methods, the LUT has better computational
efficiency and is easier to implement in the dSPACE digital
controllers.

V. EFFECT OF SYSTEM DYNAMICS

Ideally, the LUT generates the proper current to cancel the
AMB nonlinearity independent of the frequency of the desired
force signal. Nevertheless, in practice, the position input signal
Fig. 3. AMB calibration results. to the LUT is not dynamically equivalent to the true position
signal, since the latter is colored by the eddy current sensor
and antialiasing filter dynamics. Furthermore, the current signal
degaussed frequently to reduce the effect of hysteresis. Fig. 3 commanded by the LUT is not equivalent to the actual coil
shows the calibration results, which confirms the nonlinearity current signal. Rather, the latter is a colored version of the
of the AMB. command signal due to the power amplifier and D/A converter
dynamics. These dynamics can degrade the performance of the
B. Parametric AMB Model feedback linearization.
A multivariable Taylor series expansion of (1) around Ip,0 = In order to mitigate this problem, high bandwidth sensors
0 and x0 = 0 yields and amplifiers should be used in the system. However, even
with good devices, it may be necessary to employ carefully de-
f (Ip , x) = a0 + a1 Ip + a2 x + a3 Ip2 + a4 Ip x signed corrective filters, connected in series with these devices,
to compensate their dynamics. These filters should make the
+ a5 x2 + a6 Ip3 + a7 Ip2 x + · · · (4) input/output signals of the LUT very close to the corresponding
signals in the AMB system over the control bandwidth em-
where a0 , a1 , . . . are the coefficients in the Taylor expansion. ployed.
These values can be determined by a least-squares fit of the cal- An alternative approach to this would be to include these dy-
ibration results. Note that the traditional Jacobian linearization namics into the feedback linearization itself (i.e., a dynamics
uses only the first three terms in (4), while the order of the gen- feedback linearization as opposed to the static approach taken
eral nonlinear model may be chosen as desired so as to achieve here-see [3], for example). However, this more technically rig-
an accurate representation over the entire operating range. orous approach may prove to be very difficult to carry out,
It was found that inclusion of terms up to 5th order resulted considering the complexity of the actuator dynamic behavior,
in a highly accurate representation of the data. A comparison as discussed previously.
of the errors for this model and that of the standard Jacobian Fig. 5 shows the feedback linearized system block diagram
linearization is shown in Fig. 4. with two supplementary digital filters. Filter I is used to com-
pensate for the dynamics of the eddy current sensor, antialiasing
C. Nonlinear Compensator filter, and A/D converter. Filter II is used to compensate for
For a static feedback linearization, a nonlinear compensator the dynamics of the D/A converter and PWM amplifiers.
was sought that yields AMB current inputs such that the applied These filters were designed in the following fashion. First, the
electromagnetic force F equals the reference force provided to frequency response of the eddy current position sensor was
the nonlinear compensator Fc at the measured bearing air gap found to be a constant gain over the bandwidth of interest by a
x. Therefore, the proper current Ip should be the solution of the sine sweep test employing the magnetic bearing as a shaker and
equation a high bandwidth accelerometer as a reference. This result is
in agreement with manufacturer’s claims regarding this device.
Fc − f (Ip , x) = 0. (5) The A/D converter introduces a delay, which in this case is
small enough to be ignored. Therefore, only the antialiasing
Substituting (4) into (5), the current Ip can be solved. Al- filter’s dynamics need to be compensated by Filter I. A second
though there are multiple solutions due to the high-order of order lead-lag filter was then developed by adjustment of
the AMB model, the appropriate solution may easily be chosen filter coefficients so as to compensate for the phase lag of the
by satisfying two conditions: Ip ∈ R and |Ip | ≤ 1 A (the bias antialiasing filter without changing the gain of the filter up to
current). 1000 Hz. For Filter II, the frequency response from the LUT
CHEN AND KNOSPE: FEEDBACK LINEARIZATION OF ACTIVE MAGNETIC BEARINGS: CURRENT-MODE IMPLEMENTATION 635

Fig. 4. Force errors of the 5th order nonlinear (top) and the Jacobian linear (bottom) models.

output to amplifier current sensor output was measured via a

sine sweep test using a dynamic analyzer. A second order filter,
(4s2 + 9050s + 2.0 × 108 )/(s2 + 8020s + 2.2 × 108 ), was
then developed so as to achieve compensation of the amplifier’s
phase lag and the associated gain peak around 1000 Hz. Fig. 6
shows the frequency responses before and after using these two
filters. The comparison shows that both the magnitude variations
and the phase lags of the signal paths were reduced by using Fig. 5. Filtered system diagram.
these filters.
One method to verify the effectiveness of this compensation
is to examine the frequency response of the feedback linearized evaluate this, impact tests were performed. A comparison of
system. If perfect linearization were achieved, the linearized measured Gpf l (jω) with and without the filter compensation
system’s frequency response from reference input Fc (as shown along with the measured Gp (jω) is shown in Fig. 7. It is clear
in Fig. 5) to plant output x, Gpf l (jω), would be equal to the that the compensation has greatly improved the quality of the
open-loop plant frequency response from F to x, Gp (jω). To feedback linearization.
636 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 10, NO. 6, DECEMBER 2005

Fig. 6. Frequency responses comparison for sensor and amplifier signal paths before compensation (thin line) and after compensation (thick line).

Fig. 7. Open-loop frequency response comparison. Thick line: open-loop plant G p (jω); thin/dotted line: plant with feedback linearization G p f l (jω), before/after
using corrective filters.

VI. UNCERTAINTY REPRESENTATION compensator effectively depend on a feedback of the platform

Since feedback linearization renders the nonlinear AMB and displacement, it is natural to choose an uncertainty structure
amplifier static with respect to gain, the actuator platform’s consisting of displacement feedback to represent the remaining
model can be employed as the nominal model of the entire mismatch between the nominal system model and the feedback
feedback linearized system. As both the AMB and the nonlinear linearized system behavior. Fig. 8 shows this structure, where
CHEN AND KNOSPE: FEEDBACK LINEARIZATION OF ACTIVE MAGNETIC BEARINGS: CURRENT-MODE IMPLEMENTATION 637

problem. This provides a systematic technique for the design of

controllers that meet H∞ performance objectives and guaran-
tees robustness to structured model uncertainty. Fig. 10 shows
the synthesis block diagram with its uncertainty structure, where
the assessment platform and actuator platform blocks contain
their corresponding nominal models; Ks is the stiffness of the
leaf spring; FE and xA are the exogenous force input and asses-
Fig. 8. Complex uncertainty representation.
ment platform’s displacement, respectively; ∆1 is the unit norm-
bounded uncertainty associated with the feedback linearization,
and W1 is its weight (i.e., ∆c = W1 ∆1 ).
The control goal is to minimize the dynamic compliance of
the assessment platform (i.e., from FE to xA ), with a reason-
able magnitude of control effort. Two fictional complex uncer-
tainties, ∆2 and ∆3 , are included in the synthesis framework,
along with their associated weights, W2 and W3 , so as to spec-
ify the system’s performance goals. If µ < 1 is achieved in
synthesis, then the controller will achieve an assessment plat-
form compliance less than W2 (jω)−1  at each frequency,
and the controller gain will be approximately constrained
by W3 (jω)−1 .
A µ controller was designed by D-K iterations using the
MATLAB µ toolbox. The resulting controller was 18th order,
and was reduced to 6th order via a balance-and-truncate algo-
rithm without a notable loss of robustness or performance. For
evaluation purposes, a position setpoint x0 was introduced to
adjust the bearing air gap, as shown in Fig. 10. An integral term
−50/s was connected in parallel with the µ controller to im-
prove the static performance of tracking of the setpoint x0 . The
Fig. 9. Uncertainty estimation. controller was implemented on the dSPACE digital controller
with a sampling rate of 20 kHz.
To verify the effectiveness of the feedback linearization over
the uncertainty (ie., unmodeled error) may be estimated as the entire clearance, seven setpoint journal locations were spec-
Gpf l (jω) − Gp (jω) ified, ranging from +/−180 µm from the center position, with
|∆c | ≈ . (6) an increment of 60 µm. The closed-loop assessment platform
Gpf l (jω) · Gp (jω)
compliances were measured at each setpoint value; these are
An experimental investigation was conducted to estimate the
compared with the open-loop compliance in the top plot of
proper magnitude of this uncertainty for synthesis. First, the
Fig. 11. The open-loop structure has a lightly damped resonance
position of the actuator platform was regulated to a setpoint
frequency at 65 Hz, while the closed-loop compliance is well
by a PID controller. Then, frequency responses Gpf l (jω) and
damped. The plot shows that the variation in the closed-loop
Gp (jω) were measured at this position through impact testing.
frequency response obtained at different locations is very small,
The differences between these two frequency responses were
and the closed-loop compliance in each case is below W2 −1 at
calculated via (6) at the measured frequencies. Such tests were
all frequencies, as specified, indicating that the robust perfor-
performed at many setpoints throughout the entire operating
mance goal was achieved. The importance of the supplementary
range. The magnitude of the complex uncertainty set was then
compensation filters in the feedback linearization can be seen
determined by finding a circle in the complex gain plane that
in the bottom plot of Fig. 11. Here, the closed-loop compliance
encloses nearly all of the 2500 calculated data points, this circle
at the center setpoint position was found with and without the
and the data shown in Fig. 9. Although the feedback lineariza-
compensation filters. As can be seen, the performance goal is
tion does not render the system as truly linear and the method
not met when the compensation filters are not employed in the
outlined for estimating the magnitude of uncertainty presup-
feedback linearization. In fact, the peak compliance without the
poses that it does (i.e., that the uncertainty is linear), it shall
compensation filters is over 50% higher than that achieved when
nevertheless be employed, as it is likely that the small and mild
they are employed.
residual nonlinearities present after feedback linearization will
While the frequency responses obtained from impact testing
not affect system stability.
indicate that the system’s small signal behavior is as designed
and is invariant to setpoint, there remains the question of the sys-
VII. ROBUST CONTROLLER DESIGN tem’s behavior in the face of large input signals-does it respond
With feedback linearization rendering the control problem as in the same manner as the linear plant with the µ controller? To
an essentially linear one, µ synthesis is applied to the regulation examine this, a band limited white noise input was applied as the
638 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 10, NO. 6, DECEMBER 2005

Fig. 10. Nominal linearized system with uncertainty and performance weights.

Fig. 12. Linear model (thick line) versus experimental journal displacement
(thin line) when excited by band limited noise.

VIII. CONCLUSION
Current-mode feedback linearization of active magnetic bear-
ings is considered where calibration data is employed as a ba-
sis for nonlinearity inversion. The approach is advocated, as it
provides highly effective linearization while only requiring rel-
atively simple static testing of the actuators. For multiple pole
actuators, the dynamic linearization approach outlined by Lind-
lau et al. which employed both magnetic finite element analysis
Fig. 11. Compliance frequency response. and testing, was felt to be cumbersome in practice. The use
of current-mode linearization necessitated the introduction of a
compensation filter, which was introduced in an ad hoc fash-
setpoint x0 , and the journal’s displacement was recorded. This ion herein. A compensation filter was also employed so as to
same input was applied to the nominal linear system model (as remove coloring of the position signal by the antialiasing filter.
shown in Fig. 10 without the uncertainty blocks). Fig. 12 shows A model of the residual uncertainty after feedback linearization
the result of this comparison. It can be seen that the experimental was developed using experimental data. A controller was then
response closely matches that of the closed-loop linear system developed via µ-synthesis for minimization of compliance at a
over a large range of displacements, providing further evidence location on the test structure that was neither collocated with
of the effectiveness of the feedback linearization. the actuator nor the sensor. Experimental results demonstrated
CHEN AND KNOSPE: FEEDBACK LINEARIZATION OF ACTIVE MAGNETIC BEARINGS: CURRENT-MODE IMPLEMENTATION 639

that robust performance independent of large translations of Carl R. Knospe (M’91) received the B.S. degree in
the journal (hence large variations in gap length) was achieved arospace engineering and the Ph.D. degree in me-
chanical and aerospace engineering from the Univer-
by the µ-controller with feedback linearization. Furthermore, sity of Virginia, Charlottesville, in 1984 and 1989,
experimental results demonstrate that the compensation filters respectively.
introduced are essential to this result. He joined the faculty of the Department of Me-
chanical and Aerospace Engineering at the University
of Virginia in 1989 as a Research Assistant Professor,
REFERENCES and was promoted to Associate Professor in 1997. His
work has focused on the development and application
[1] D. Trumper, “Linearizing control of magnetic suspension systems,” IEEE of control theory to a variety of mechatronic systems,
Trans. Contr. Syst. Technol., vol. 5, no. 4, pp. 427–437. with a focus on magnetic levitation and vibration control. Much of his recent
[2] S. Chen, S.-H. Chen, and S.-T. Yan, “Experimental validation of a current- work has focused on the analysis and control of infinite dimensional systems
controlled three pole magnetic rotor bearing system,” IEEE Trans. Magn., that arise in engineering application, primarily due to delays resulting from
vol. 41, no. 1, pp. 99–112, Jan. 2005. the movement of materials and information within. He is the author of over
[3] J. Lindlau and C. Knospe, “Feedback linearization of an active magnetic 40 archival papers. He has taught several short courses in North America and
bearing with voltage control,” IEEE Trans Contr. Syst. Technol., vol. 10, Europe, most recently “Practical Techniques for Control Engineering” at the
no. 1, pp. 21–31, Jan. 2005. American Control Conference.
[4] H. Gutierrez and P. Ro, “Parametric modeling and control of a long-range Dr. Knospe was Editor of IEEE TRANSACTIONS ON CONTROL SYSTEMS
actuator using magnetic servo levitaition,” IEEE Trans. Magn., vol. 34, TECHNOLOGY’s Special Issue on Magnetic Bearing Control. He has received
pt. 2, no. 5, pp. 3689–3695, Sep. 1998. numerous awards including Best Paper at both the 5th and 6th International Sym-
[5] S. Joo and J. Seo, “Design and analysis of the nonlinear feedback lin- posia on Magnetic Bearings (Kanazawa, Japan; Cambridge, MA). He is an Asso-
earizing control for an electromagnetic suspension system,” IEEE Trans. ciate Editor for the IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY
Contr. Syst. Technol., vol. 5, no. 1, pp. 135–144, Jan. 1997. and for IEEE Control Systems Magazine.
[6] P. Schroder, “On-line evolution of robust control systems: An industrial
active magnetic bearing application,” Contr. Eng. Practice, vol. 9, pp. 37–
49, 2001.
[7] L. Lin and T. Gau, “Feedback linearization and fuzzy control for coni-
cal magnetic bearings,” IEEE Trans. Contr. Syst. Technol., vol. 5, no. 4,
pp. 417–426, Jul. 1997.
[8] T. Minihan, “Large motion tracking control for thrust magnetic bearings
with fuzzy logic, sliding mode, and direct linearization,” J. Sound Vib.,
vol. 263, pp. 549–567, 2003.

Min Chen received the B.Eng. degree and the M.Eng.

degree in mechanical engineering from Tsinghua
University, Beijing, China, in 1996 and 1999, respec-
tively. He is currently working toward the Ph.D. de-
gree in the Mechanical and Aerospace Engineering
Department, University of Virginia, Charlottesville.
He has also been associated with the Advanced
Servo Group, Western Digital Corporation, San Jose,
CA, since 2004. His research interests include robust
control, signal processing, control of dynamic sys-
tems with delays, and servo control technology for
disk drives.