Optimal control based LQR-feedback linearisation

for magnetic levitation using Improved Spiral

Dynamic Algorithm
A. M. Benomair∗ , F.A.Bashir, M.O.T okhi
Department of Automatic Control and Systems Engineering
The University of Sheffield
Sheffield, UK
ambenomair1@sheffield.ac.uk

Abstract—Magnetic Levitation (Maglev) system make signif- mode control [7] [8],adaptive backstepping control [9], gain
icant contribution to industrial application due their reduce scheduled control [10] [11], and H∞ control [12]. In this paper
power consumption, increased power efficiency and reduced cost the exact input-state feedback linearisation control algorithm
of maintenance. Common applications include Maglev power
generation (e.g. wind turbine), Maglev trains and Medical Devices is used rather than using linear approximation of the dynamics,
(e.g. magnetically suspended artificial heart pump). This paper based on voltage control due to some advantages over current
proposes an optimal linear quadratic regulator using improved control. Then linear quadratic regulator (LQR) is proposed as
spiral dynamic algorithm to actively control a magnetic levitation an optimal controller to balance between the state and effort
system with full-state feedback linearisation. Simulation as are controls.
performed with nonlinear mathematical model of the magnetic
levitation system, and the results show that the proposed lineari-
sation and control strategy perform well. II. NON-LINEAR DYNAMIC MODEL OF THE AML

Index Terms—magnetic levitation system; Linear quadratic The magnetic levitation system serves to keep a small steel
regulator; Full state input-linearisation, Improved spiral dynamic ball in stable levitation at some steady-state operating position.
algorithm An electromagnet is used to produce forces to support the
ball’s weight (see Fig. 1). The electromagnetic forces are
I. I NTRODUCTION related to the electrical current passing through the electro-
Active magnetic levitation (AML) systems have been widely magnet coil.
used in medical applications [1], turbo-machinery [2], mag- Electromagnetic force produced by current can be calculated
netic levitation vehicles [3], micro robot actuation [4] and by Appling Kirchhoffs voltage law to the electric system loop
wind turbine [5]. Contactless support of objects continues
to be a fantasy for several centuries. The utilization of d L(x)i
v(t) = vR + vL = i . R + (1)
magnetic forces seems to be the ideal solution in many dt
situations to such a goal. Using magnetic forces to support
an object without any mechanical contact is constrained by
the laws of magnetism. Earnshaws theorem [6] states that
when the inverse-square-law forces govern several charged
particles, they can never be within a stable equilibrium. The
interaction between ferromagnetic objects and electromagnets
of either the active or passive type, is associated with an
unstable nature. This unstable behavior can be represented
by highly nonlinear differential equations. In the literature
many researches are based on linearised models around a
specific nominal operating point. In such case, the tracking
performance can rapidly deteriorate as the deviation from the
operation point increases. However, in order to guarantee a
local asymptotic stability over larger range and ensure the
good tracking, it is necessary to consider a nonlinear model
rather than a linearised one. Several techniques have been
proposed to control the nonlinear dynamics such as sliding Fig. 1: Active magnetic levitation ball system (AML)

978-1-4799-8701-6/15/$31.00 ©2015 IEEE. 558

 where v: Applied voltage, i: Coil current, R: Resistance of
the electromagnet and L(x) is coil inductance. The inductance
of the coil is non-linear function of ball position (x). The
approximation inductance can be written as

2c
L(x) = Lc + (2)
x
where Lc is the constant inductance of the electromagnetic Fig. 2: Simulation scheme for control of the nonlinear model
coil in absence of the ball and c is magnetic constant. of the Magnetic levitation using exact input-state feedback
Substituting for L(x) from equation (2) into equation (1) and linearisation
rearranging yields
 
∂i −R 2c i ∂x 1
= i+ + v (3) ⎡ ⎤
∂t L L x 2 ∂t L ⎡ ⎤ z2
ż1 ⎢ ⎥
The ball is suspended due to balancing between the force ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ z3 ⎥
of gravity and electromagnetic force. Applying Newtons 3rd ⎢ ż2 ⎥ = ⎢ ⎥ (7)
⎢ ⎥ ⎢ ⎥
law of motion, the dynamic form of the mechanical system ⎣ ⎦ ⎢   ⎥
⎣ 2c x3 x˙3 x23 x˙1 ⎦
can be written as ż3 − − 3
m x21 x1
By Substituting ẋ1 and ẋ3 form Eq. (5) into Eq. (7), the
fnet =fg − fem (4a) resultant Formula can be written as
 2
i ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
m ẍ = m g + c (4b) ż1 z2 z2
x ⎣ ż2 ⎦ = ⎣ z3 ⎦ = ⎣ ⎦
z3 (8)
Define the states x1 = x (position), x2 = V (velocity), x3 = ż3 w f (x) + b(x)u
i (current) and the system input u = v (applied voltage). The
where
nonlinear state space model of AML system can be expressed
as ⎡ ⎤    
x2 2c 2c 1 x2 x23 R x23
⎡ ⎤ f (x) = 1− +

ẋ1 ⎢ ⎥ m L x1 x13 L x31

⎢  2 ⎥
⎢ ⎥ ⎢ c x3 ⎥ (9)
⎢ ⎥ ⎢ g − ⎥  
⎢ ẋ2 ⎥ = ⎢ ⎥ (5) 2c x3
⎢ ⎥ ⎢ m x1 ⎥ b(x) = −
⎣ ⎦ ⎢ ⎥ mL x21
⎢   ⎥
ẋ3 ⎣ R 2c x2 x3 1 ⎦
− x3 + + u The nonlinearities can be cancelled by the control law of
L L x21 L the from
1
III. FEEDBACK LINEARIZATION SCHEME u= [w − f (x)] (10)
b(x)
The central idea of feedback linearisation approach is to where w is an equivalent input (synthetic control) to be
cancel the nonlinearity in a nonlinear system so that the designed, leading to a linear input-state relation of the form
closed-loop dynamics is in a (fully or partly) linear form.
Feedback linearisation is achieved by exact feedback and ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
ż1 0 1 0 z1 0
exact state transformations (see Fig. 2), rather than by linear ⎣ ż2 ⎦ = ⎣ 0 0 1 ⎦ ⎣ z2 ⎦ + ⎣ 0 ⎦ w (11)
approximation of the dynamics. Define the transformed states,
ż3 0 0 0 z3 1
for which the equivalent model is in a simpler form, as
⎡ ⎤
⎡ ⎤ x1 IV. CONTROL DESIGN
z1 ⎢ ⎥
⎢ ⎥ ⎢ ⎥ The transformed system using feedback linearisation can be
⎢ ⎥ ⎢ x2 ⎥ expressed in state variable form as
⎢ z2 ⎥ = ⎢ ⎥ (6)
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎢  2 ⎥
⎣ c x3 ⎦ ż(t) = A z(t) + B w(t) (12)
z3 g−
m x1
Before proceeding to design the controller, the new
To ensure that transformation is invertible, the system states transformed system needs be verified to be controllable.
are restricted to the region x1 > 0 and x3 > 0. In the new Satisfaction of this property means that the state of the
coordinates the system equations become system can be driven anywhere in finite time as desired.

559
  
Rank of the controllability matrix B AB A2 B (3×3)
B. Tracking control design using feedforward gain
= (number of states). Feedforward gain doesn’t affect the stability of the system
(which is completely determined by the eigenvalues of λ (A −
The controller makes use of the states of the transformed B K) but it has only an effect on the steady state solution.
system z and the reference input r to produce the command Hence a scalar feedforward gain Kr is introduced to track
for the process through its new command input w which can the reference signal and to clear off steady state error. The
be written as transformed closed-loop system equations can be written as

w(t) = −K z(t) + Kr r (13) ż(t) = (A − BK) z(t) + B Kr r(t) (19)

Particularly, the equilibrium point (i.e no change in the sys-
where K is the full-state feedback gain to be designed using tem dynamics) and steady-state output for the asymptotically
LQR, Kr is the feedforward gain and r is the reference input stable closed loop system are given as
(or command signal).
ze = −(A − B K)−1 B Kr r (20)
A. Full state feedback control (Linear Quadratic Regulation)
The stability of the new dynamics can be achieved using yr,e = C ze (21)
the linear state feedback control law that achieves the given Thus for zero steady state offset Kr should be chosen such
requirements. that yr,e = r (the desired output value). Solving equations (20)
and (21) for scalar gain Kr yields to
w(t) = −K z(t) (14) Kr = −1/(C(A − B K)−1 B) (22)
In this paper, linear quadratic regulation method is used to
determine full state-feedback control gain matrix K, seek to V. SIMULATIONS AND RESULTS
minimize a cost function J Simulations were carried out with the following values for
the system parameters: of M = 21.2e-3 kg, C = 8.248*1e-5
∞
N*m/Aˆ2 , R = 4.2 Ω, L = 0.02 H.
J(z) := min = z T Qz z(t) + wT Qw w(t) dt (15) LQR minimizes a performance index (i.e. quadratic cost
w(.) 0 function) which mainly has two penalty matrices: state weight-
ing matrix Qz and control weighting matrix Qw .Both of these
Subject to the constrains parameters have greatly influence the performance of LQR
controller being selected by designer. However, it is not a
ż(t) = A z(t) + B w(t) ∀z(0) = z0 (16) trivial task to tune those two matrices.
Initially the weights of the LQR are to be assumed such
where Qz (3 × 3) and Qw (1 × 1) are typically positive semi- as Qw =1 and Qz = C T ∗ C, so that the corresponding cost
definite and positive definite matrices respectively. Tuning function will pose equal importance on the control effort and
these parameters will balance the relative importance between the fast convergence of the state variables. This implies to
the control accuracy versus effort. The solution to the LQR start with Qw = 1 and Qz = diag( 1 0 0 ). Adjusting
problem is a linear control law of the form: these parameters intuitively by trial and error to maintain the
smallest possible error between the reference input and the
w(t) = −Q−1 T
w B P z(t) (17) system output for an acceptable control effort, led to the initial
results of system response and the control effort shown in Figs.
where P is a positive definite matrix which satisfies the 3 and 4.
algebraic Riccati equation The classic approaches are labour-intensive, time consuming
and do not guarantee the expected performance. Furthermore,
these techniques only aim to minimize the quadratic perfor-
P A + AT P − P B Q−1 T
w B P + Qz = 0 (18) mance index and do not consider other control objectives
such as minimizing the overshoot, rise time, settling time,
Thus,the LQR method is used for determining the optimal and steady state error.Based on aforementioned results, the
state-feedback control matrix K to allocate the eigenvalues searching range for both parameters are specified. The state
of the closed-loop system λ(A − BK) on the left half plane. weighting matrix and control weighting matrix are refined to
Since the design of the state feedback matrix K is based on be as Qz = diag( 9.31 ∗ 104 0.7287 0.7378 ) and Qw =
solving the regulator problem (no reference inputs), the stable 6.34 ∗ 10−4 using the improved spiral dynamic algorithm
closed-loop system will reach a steady state to a constant input (ISDA) [13] to obtain a satisfactory performance (see Figs.
reference signal but steady state errors may be present. 5 and 6).

560
Fig. 3: system response for different manual setting of state Fig. 5: Optimized controlled position using ISDA with step
weighting matrix input

Fig. 4: Control effort (input voltage) for different setting of Fig. 6: Optimized controlled position using ISDA with sinu-
control weighting matrix soidal input

VI. CONCLUSION [3] S.-M. Jang, Y.-S. Park, S.-Y. Sung, K.-B. Lee, H.-W. Cho, and D.-J.
You, “Dynamic characteristics of a linear induction motor for predicting
In this paper nonlinear control algorithm design has been operating performance of magnetic levitation vehicles based on electro-
presented for nonlinear simulation model of Active Magnetic magnetic field theory,” Magnetics, IEEE Transactions on, vol. 47, no. 10,
levitation (AML) using of input-state linearisation. The results pp. 3673–3676, 2011.
[4] M. Hagiwara, T. Kawahara, T. Iijima, Y. Yamanishi, and F. Arai, “High
show that the controller can stabilize the system in a larger speed microrobot actuation in a microfluidic chip by levitated structure
region which differs from conventional linearisation (small with riblet surface,” in Robotics and Automation (ICRA), 2012 IEEE
rage operation around the equilibrium point). Moreover, using International Conference on. IEEE, 2012, pp. 2517–2522.
[5] C. Aravind, R. Rajparthiban, R. Rajprasad, and Y. Wong, “A novel
LQR method can balance between minimizing the excursion magnetic levitation assisted vertical axis wind turbinedesign procedure
in state trajectories of a system while requiring minimum and analysis,” in Signal Processing and its Applications (CSPA), 2012
controller effort. However, full state feedback linearisation IEEE 8th International Colloquium on. IEEE, 2012, pp. 93–98.
needs all states to be measurable and also it cant grantee a [6] H. Bleuler, M. Cole, P. Keogh, R. Larsonneur, E. Maslen, Y. Okada,
G. Schweitzer, A. Traxler, G. Schweitzer, E. H. Maslen et al., Mag-
global stability. netic bearings: theory, design, and application to rotating machinery.
Springer Science & Business Media, 2009.
R EFERENCES [7] N. Al-Muthairi and M. Zribi, “Sliding mode control of a magnetic
levitation system,” Mathematical Problems in Engineering, vol. 2004,
[1] P. Allaire, E. Maslen, H. Kim, G. Bearnson, and D. Olsen, “Design of a no. 2, pp. 93–107, 2004.
magnetic bearing-supported prototype centrifugal artificial heart pump,” [8] J. Yang, S. Li, and X. Yu, “Sliding-mode control for systems with
Tribology transactions, vol. 39, no. 3, pp. 663–669, 1996. mismatched uncertainties via a disturbance observer,” in IECON 2011-
[2] R. Field and V. Iannello, “A reliable magnetic bearing system for 37th Annual Conference on IEEE Industrial Electronics Society. IEEE,
turbomachinery,” in Proceedings of the 6th ISMB, 1998, pp. 42–51. 2011, pp. 3988–3993.

561
 [9] C.-M. Huang, J.-Y. Yen, and M.-S. Chen, “Adaptive nonlinear control
of repulsive maglev suspension systems,” Control Engineering Practice,
vol. 8, no. 12, pp. 1357–1367, 2000.
[10] Y. C. Kim and K.-H. Kim, “Gain scheduled control of magnetic
suspension system,” in American Control Conference, 1994, vol. 3.
IEEE, 1994, pp. 3127–3131.
[11] M.-S. Kim, Y.-S. Byun, Y.-H. Lee, and K.-S. Lee, “The levitation
controller design of an electromagnetic suspension vehicle using gain
scheduled control,” in 5th WSEAS Int. Conf. on Circuits Systems,
Electrics, Control & Signal Processing, 2006.
[12] P. K. Sinha and A. N. Pechev, “Nonlinear h controllers for electromag-
netic suspension systems,” Automatic Control, IEEE Transactions on,
vol. 49, no. 4, pp. 563–568, 2004.
[13] A. Nasir and M. Tokhi, “An improved spiral dynamic optimization
algorithm with engineering application,” Systems, Man, and Cybernetics:
Systems, IEEE Transactions on, vol. PP, no. 99, 2015.

562