Vous êtes sur la page 1sur 5

Inverse Model Based Adaptive Control of Magnetic

Levitation System
Muhammad Shafiq and Sohail Akhtar
Department of Systems Engineering
College of Computer Science & Systems Engineering
King Fahd University of Petroleum and Minerals
Dhahran 31261, Saudi Arabia
Email: mshafiq@ccse.kfupm.edu.sa, sakhtar@ccse.kfupm.edu.sa

ABSTRACT well in the same zone. A linear controller for such a


This paper presents, an adaptive finite impulse re- system was designed and implemented in [1]. The non-
sponse (FIR) filter based controller used for the track- linear system equations were perturbed about an oper-
ing of a ferric ball under the influence of magnetic ating point, resulting in an approximate linear model
force. The adaptive filer is designed online as approx- which has an open-loop pole in the right-half plane.
imate inverse system. To stabilize the open-loop un- A phase-lead (linear) compensator is used to stabilize
stable and highly nonlinear magnetic levitation sys- the system for step responses of 1.5 mm around the
tem, PID controller is designed using polynomial ap- operating point.
proach. To improve the stability, an adaptive FIR fil-
ter is added along side the PID controller while the Recently, extensive work has been reported for
use of the proposed controller has improved tracking. controlling a Maglev system by taking the non-
Since adaptive FIR filters are inherently stable so the linearities of the system model into account. A com-
controller remains stable. Experimental results are in- parison of a sliding mode controller versus a linear
cluded to highlight the excellent position tracking per- phase-lead controller is carried out in [2]. The study
formance. has shown that the sliding mode controller has pro-
vided a better transient response than the linear con-
KEY WORDS troller. The current dynamics are neglected in the used
Adaptive Control, Magnetic Levitation, Adaptive FIR feedback controller and the ball’s motion was limited
Filters to a range of 1mm only. In [3], the authors have car-
ried out a comparative study of linear and non-linear
1 Introduction state-space controllers for Maglev system. A linear
state-feedback controller-observer, based on linear sys-
Magnetic levitation (Maglev) systems are attracting tem model; and a non-linear controller in conjunction
the attention of researchers as a mean of eliminat- with a non-linear velocity observer providing tracking
ing the Columbic friction due to mechanical contacts. and regulation was implemented. These experimen-
They are widely used in various fields such as fric- tal observations have shown that the two controllers
tionless bearings, high-speed Maglev passenger trains, have a comparable response to step inputs. However,
levitation of wind tunnel models, vibration isolation of feed-back linearization controller has provided signifi-
sensitive machinery, levitation of molten metal in in- cantly better trajectory tracking. A highly non-linear
duction furnaces, and levitation of metal slabs during model without premagentizing current in the coil has
manufacture [1-14]. Maglev systems are usually highly been used in [4]. A non-linear controller is proposed
non-linear and open-loop unstable. Therefore, it is al- to reduce the effect of the imbalance on the rotor dis-
ways a challenging job to construct a high performance placement without estimating the perturbation. The
feedback controller to control the position of a levi- results highlight the use of a non-linear controller. A
tated object. In recent years, many approaches have single degree of freedom (1-dof) magnetic suspension
been reported in literature based on linear and non- system has been designed and constructed in [5] to
linear system models for controlling this system. The compare the performance of linear and non-linear dig-
standard linear technique is quite straight forward. It ital control schemes. It is shown that due to a non-
is based on an approximate linear model found by per- linear relationship between force, current and air gap,
turbing the system dynamics about a desired operat- a non-linear controller will outperform a conventional
ing point. As this linear system model is valid only controller in an environment where large variations of
around a ‘small’ region of the operating point, the re- operating points are considered. Feedback lineariza-
sulting linear controller is only expected to function tion technique has been employed though it requires a
very accurate system model. approach. The proposed controller has shown a swift
Polyhedral Lyapunov functions and linear vari- response and a large operating range (100mm) to a
able structure stabilization controller is used in [6]. step command signal.
These functions have the property that the control This paper proposes an adaptive FIR filter based
and state constraints can be taken into account during control algorithm implemented on a laboratory scale
their construction. A success in using a chaos control Maglev system. Motivations of the use of adaptive FIR
method for stabilizing a Maglev system based on time filters is the fact that these filters are always stable and
delay auto-synchronization controller is reported in [7]. easy to implement. This means that the controller will
A linear quadratic Gaussian (LQG) design method always remain stable. The objective is accomplished
with additional design freedom for frequency shaping by estimating the plant parameters and the approxi-
based on inverse linear quadratic (ILQ) design method mate inverse system by normalized least mean square
is reported in [8]. This gives a lower order controller (NLMS) estimator [15]. The paper is organized as fol-
than the H∞ controller. The results have shown the lows: Section 2 is devoted to problem statement and
same tracking ability as that of the ILQ controller and the controller design. Section 3 gives the mathemati-
at the same time suppresses the spillover. A non-linear cal model of the Maglev system which is used for PID
model has been developed in [9]. Feedback lineariza- controller design. Experimental results are shown in
tion approach is used to control the ball position over Section 4, and finally the paper concludes with Section
long distances (in the range of 40mm). The robust- 5 which briefly discusses the obtained results.
ness of the scheme is validated by the variation of
mass and the evolution of resistance and inductance 2 Problem Statement
of electromagnet. A SISO second order non-linear dif-
ferential equation is considered to represent the system Let us describe a single-input single-output (SISO)
in [10]. In the presence of parameter uncertainties, an stable linear discrete-time plant by
adaptive robust non-linear controller in a back step-
ping manner is proposed. At first step, a PI controller A(q)y(t) = q −d B(q)u(t), (1)
is designed to stabilize the position error. Then an
where
adaptive robust non-linear controller consisting of an
adaptive feedback linearization control term and a ro- 
n 
m
A(q) = 1 + ai q −i , B(q) = bj q −j ,
bust non-linear damping term is designed to undo the
i=1 j=0
effect of parameter uncertainties. This combined ap-
proach has achieved a better performance as compared q −1 is a back shift operator, d represents a known time-
to the case where either is employed alone. Variable delay and b0 = 0. A(q) and B(q) are co-prime polyno-
structure control (VSC) for robust stabilization and mials of degrees n and m, and n ≥ m. It is supposed
disturbance rejection of a 1-dof Maglev system is ap- that parameters ai and bj are unknown. Further, de-
plied in [11]. A method called Reaching law method grees of B(q) and A(q) may be unknown. The plant
complemented by a sliding mode equivalence technique can be a minimum or non-minimum phase system.
is used. u(t) and y(t) are the measurable input and output,
A novel 2-sigma artificial neural network to esti- respectively. In these circumstances, the objective is
mate the bounds of modelling uncertainties resulting to synthesize a bounded control input u(t) using a two
from parameter variations, un-modelled dynamics and degrees of freedom adaptive controller based on adap-
non-deterministic processes in a dynamic plant has tive FIR filters such that the plant output y(t) tracks
been proposed in [12]. The application of this network desired bounded input sequence yd (t).
to robust control strategies such as sliding mode con-
trol enables intelligent, spatially-dependent controller 2.1 Controller Design
gains that retain robust performance without exces-
sive control activity. [13] gives a new model and un- Let us synthesize the control input u(t), from the fol-
certainty structure of a Maglev system by using lin- lowing equation
ear fractional transformations (LFT), which contains
less conservativeness for robust stability/performance u(t) = F̂ (q) [P (q)yd (t) − δy(t)] , (2)
analysis. A method for quantifying the magnitude
where F̂ (q) is a polynomial, which satisfies the follow-
of uncertainties is also developed. Later on a ro-
ing approximation
bust controller by µ-analysis and synthesis is designed,
which has shown robust performance. A combina- B(q)
tion of fuzzy-sliding-mode and state-feedback control q −d F̂ (q) ≈ q −L , (3)
A(q)
is used in [14]. This concoction has the advantage of
both methods: high-speed response of fuzzy-sliding- L is a positive integer and L ≥ d. We say that F̂ (q) is
mode and small steady state error of state-feedback the L-delay approximate inverse system of the plant.
The polynomial F̂ (q) can be estimated using some The current through electromagnet is given by the re-
suitable estimator like the NLMS estimator, same as lation:
in adaptive equalization problem [15, 16]. The polyno-
mial F̂ (q) can be considered as an adaptive FIR filter. I = 0.15U + I0 (7)
δ is a constant such that |δ| < 1 and P (q) = 1 + δq −L .
where I0 is the nominal operating current at nominal
Using equations (1), (2) and (3), the following relation
ball position X0 .
can be shown easily
The variation in current is given by:
lim y(t) = yd (t − L). (4)
t→∞ ∆I = 0.15∆U (8)
This means the plant output will track the desired
output with some delay. As F̂ (q) is an adaptive FIR The ball’s displacement is governed by the following
filter and P (q) is an FIR filter, so the controller will equation:
remain stable. It can also be shown by using equations I2
(1), (2) and (3) that the characteristic polynomial is mẌ = mg − k (9)
X2
P (q) = 1 + δq −L and |δ| < 1, so the closed-loop will
remain stable. The block diagram of the proposed
control strategy is shown in Figure 1. where m is the mass of the ball, g is gravitational
constant, k is magnetic force constant, X is ball’s dis-
placement and I is the current through the magnetic
coil.
Linearizing above equation around nominal operating
point and replacing ∆I by equation 8 will yield:
I02 I0
m∆Ẍ = 2k ∆X − 0.3k 2 ∆U (10)
X03 X0
where constant k is obtained by setting the derivatives
of the non-linear model to zero.
Figure 1. NLMS based Inverse Controller Approximating the above system with a 2nd order
transfer function will give:
η
3 Magnetic Levitation System Mod- G(s) = (11)
s2 − ω02
elling 
with η ∼
= 3γ
I0 and ω0 ∼
= 20
X0
Figure 2 below shows the simplified diagram of Ma-
glev system (adapted from Feedback Instrument Ltd.
manual). 4 Real Time Implementation and Ex-
perimental Results
A laboratory scale Maglev system, manufactured by
Feedback Instrument Ltd., is used to evaluate the per-
formance of the proposed adaptive controller. An elec-
tromagnet applies a magnetic force which holds the
steel ball in air counter acting the force of gravity.
By varying the current through the electromagnet, the
steel ball can be forced to follow a predetermined tra-
Figure 2. Simplified Maglev System Diagram jectory. The metallic ball is manually introduced in
the magnetic field. An array of infrared transmitters
The photo-sensors measure the ball’s position. Corre-
and detectors, positioned such that the sphere inter-
sponding to the ball’s position from the electromagnet,
sects the infrared beam, is used to determine the ball’s
the sensor generates a voltage output (Vsensor ) obey-
position. The output of the ball position sensor is a
ing the following relation:
voltage signal in the range of ±2.6 volts. This corre-
Vsensor = −γ(X − X0 ) (5) sponds to a ball movement of 25mm in the magnetic
field. System models were built in Simulink. Matlab
where γ > 0 and X0 is the nominal operating point.
Real Time Windows (RTW) Toolbox is used to gen-
For X=X0 , sensor output will be zero, so the variation
erate C sources and executable programs, which can
in sensor voltage is given by:
be run in real time. A standard Pentium III IBM per-
∆Vsensor = −γ∆X (6) sonal computer is used and a sampling frequency of
1kHz is selected. Advantech PC 1711 data-acquisition at the crests and troughs of the reference sinusoidal
card with a preamplifier is employed to interface the signal as shown in Figure 6. To improve the tracking
measured and control signals.
As mentioned before, Maglev systems are open-
loop unstable, so a PID controller is used to stabilize
the system as shown in Figure 3. The above mentioned
2nd order system model is used to design the PID con-
troller using polynomial approach. The obtained con-
troller’s gains were kP =1.5, kI =3 and kD =0.2 for the
system closed-loop poles at −0.6 + 4i, −0.6 − 4i and Figure 5. AFIR addition to improve the stability
−2 in the s-plane. We have applied a sinusoidal de-
mand signal of amplitude 1.5 volts. An effort to ap-
ply a larger amplitude input signal resulted in system
collapse. Figure 4 shows the actual and desired ball
positions and the applied control signal. It is evident
from the response that the ball has started to follow
the desired trajectory with some ripples especially at
the extreme sides of the desired signal. It was also
observed that when the system parameters (resistance
and inductance) vary with electromagnet heating, the
performance of the PID controller was deteriorated.

Figure 6. Input output signals with AFIR addition


Figure 3. PID controller for system stabilization

ability of the system, filters F̂ (q) and P (q) are added


as shown in Figure 7. Adaptive filter, F̂ (q), is the ap-
proximate inverse of the system. A filer order of 32
is used. Plant output is used as the filter input and
the difference of the delayed signal applied to the sta-
bilized plant and output of the F̂ (q) is used for filter
adaptation. Filter P (q) is used for input scaling. The
system response is shown in Figure 8 along with the
control signal. It can be seen that, with the approach
proposed in this paper, the object in levitation follows
the reference set point very well on a long range of
travel (±2.6 volts, which is the full operating range of
the prototype system used in these experiments).

Figure 4. Input output signals using PID controller


In an effort to minimize the effect of parameter
deterioration due to heating effect and to increase the
stable range of the Meglev system, an AFIR filter is
added in parallel with the PID controller as shown by
the schematic in Figure 5. A 10th order adaptive filter
is employed. The error between reference and actual
trajectories is used as filter input as well as adapting
signal. The addition has increased the stable operat- Figure 7. Discrete-time IMC Configuration
ing range to ±2.2 volts. But still mismatch between
desired and actual ball positions is observed, notably
[5] D. L. Trumper, S. M. Olson and P. K. Sub-
rahmanyan, Linearizing control of magnetic sus-
pension systems, Trans. IEEE Control Systems
Technology, 5(4), 1997, 427-437.
[6] F. Blanchini and S. Carabelli, Robust stabi-
lization via computer generated Lyapunov func-
tions: An application to a magnetic levitation
system, Conference on Decision and Control,
Lake Buena Vista, FL, Dec. 1994, 1105-1106.
[7] B. Z. Kaplan and G. Sarafian, Control strategy
for stabilizing magnetic levitation , Electronic
Letters, 33(23), 6th Nov. 1997, 1960-1961.
[8] M. Sakai, Y. Kuroe, K. Nakashima and T. Fu-
Figure 8. Input output signals using inverse controller jii, Application of ILQ-based frequency-shaping
control to a magnetic levitation system, Pro-
ceedings of the 1998 IEEE Int. Conference on
Control Applications, Trieste, Italy, 1-4 Sep.
5 Conclusion 1998, 593-597.
In this paper an inverse model based adaptive control [9] A. E. Hajjaji and M. Ouladsine, Magnetic and
algorithm is proposed. Inherently stable NLMS filters nonlinear control of magnetic levitation systems,
are used as the approximate inverse model. Since a Trans. IEEE Industrial Electronics, 48(4), 2001,
bounded input is applied to the system and stable 831-838.
FIR filters are used with the stabilized plant, so the
controller output always remains bounded. The con- [10] Z. -J. Yang and M. Tateeishi, Adaptive robust
troller’s performance is verified on a highly non-linear nonlinear control of magnetic levitation system,
Maglev system. The results have shown excellent Automatica, 37, 2001, 1125-1131.
tracking of the sinusoidal reference trajectories over [11] M. M. H. Ikbal and M. M. Abdelfatah, Variable
the full operating range. It is also shown that addition structure control of a magnetic levitation sys-
of an adaptive FIR filter in parallel with the PID tem, Proceedings of the American Control Con-
controller has improved the system stability. ference, Arlington, VA, 25-27 Jun. 2001, 3725-
3730.
Acknowledgment
The authors would like to thank KFUPM for its sup- [12] D. B. Gregory, Intelligent bounds on modeling
port to carry out this research work. uncertainties: application to sliding mode con-
trol of a magnetic levitation system, Proceed-
ings of the 2001 IEEE International Conference
References
on Systems, Man and Cybernetics, 1, 2001, 81-
[1] T. Wong, Design of a magnetic levitation system 86.
- an undergraduate project, IEEE Trans. On Ed- [13] T. Namerikawa and M. Fujita, Uncertainty
ucation, 29, 1986, 196-200. structure and µ-synthesis of a magnetic levita-
[2] D. Cho, Y. Kato and D. Spilman, Experimental tion system , T. IEE Japan, 121-c, 2001, 1080-
comparison of sliding mode and classical con- 1087.
trollers in magnetic levitation systems, IEEE [14] S. M. Hosseini, A. R. Alfie, M. Farrokhi and M.
Control Systems Magazine, 13(1), Feb. 1993, 42- R. S. Motlagh, Fuzzy sliding state-feedback con-
48. troller of non-linear ball suspension sysetm, 15th
[3] W. Barie and J. Ckiasoson, Linear and nonlin- Teiennial World Congress, Barcelona, Spain,
ear state-space controllers for magnetic levita- 2002.
tion, International Journal of Systems Science, [15] S. Haykin, Adaptive Filter Theory, Pearson Ed-
27(1), 1996, 1153-1163. ucation, USA, 2002.
[4] A. Charara, J. D. Miras and B. Caron, Nonlin- [16] H. Babak, Ali H. Sayed and T. Kailath, H∞ Op-
ear control of magnetic levitation system with- timality of the LMS algorithms, IEEE Trans. on
out premagnetization, Trans. IEEE ControlSys- Signal Processing, 44(2), Feb. 1996, 267-280.
tem Technology, 4(5), 1996, 513-523.

Vous aimerez peut-être aussi