NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

PID Controller Design for a Magnetic Levitation System using an Intelligent

Optimisation Algorithm

Nasir Ahmed Al-awad 1 Ayam Mohsen Abbass 2 Majid Eddan Dubkh 3

[email protected] [email protected] [email protected]

1 ,2
Computer Engineering Department, 3 Electrical Engineering Department
1 ,2 ,3
Faculty of Engineering, Mustansiriyah University
Baghdad, Iraq.

Abstract – Magnetic Levitation System (MLS) parameters can be tuned using different controlling techniques. Magnetic levitation
systems are used to levitate objects by using attraction force or repulsive force between magnetic force and ferromagnetic material.
The levitation of an object is possible using a control system to help stabilize the magnetic force. This paper investigates a
controlling technique for a MLS. The numerical model of the system shows that it has high non-linearity and inherent instability.
Different Controlling Methods were applied. A Proportional-Integral-Derivative (PID) controller is tuned using Neural Network
(NN) and Genetic Algorithm (GA) due to their ability to control nonlinear plants. The results show better performance of the MLS
when NN and GA are used, compared to other classical methods.

Keywords - MLS model, PID controller, LQR controller, IMC controller, NN controller, GA controller, Matlab/Simulink.

I. INTRODUCTION function with any precision [8]. A feedback error learning

PID controller with NN is presented in [9] to control an
MLS is very useful system that can be connected to MLS. A single multiplicative neuron model is proposed to
many applications, such as, fast transportation, control an MLS, which derived from the computation
magnetised bearing systems, vibration disengagement, model of a single neuron. Particle Swarm Optimisation is
levitation of wind control, and combination Energy used for offline training of network weights and biases
Materials preparing in magnetised levitation furnaces [1]. [10]. Different combinations of NN and nonlinear
MLS is an electromagnetic gadget, which suspends method are presented in [11]; Lyapunov method is
ferromagnetic components using the rule of considered to insure stability. The authors in [12] suggest
electromagnetism. MLS has the capacity to work in Back-propagation NN to improve the behaviour of the
vacuum condition. Due to the steady need for levitation, MLS and obtain the required response.
an MLS is subjected to constant change in its parameters, A method based on NN and GA to control MLS is
and consequently the numerical model becomes highly presented in [13], GA is used to adjust the NN
non-linear and unstable. Despite the fact that magnetic parameters, frequently updates the control signals, where
levitation is nonlinear behaviour and it is described by they are computed using the back propagation technique.
nonlinear differential equation, mostly design approaches In [14] GA is used to tune a PID controller for MLS. The
are based on linear model [2]. results claim that the response of the system of GA-PID is
Non-linearity occurs due to electromechanical better than traditional PID controller. An Integral-Tilted-
dynamics [3]. Different methods are proposed to control Derivative (I-TD) controller for MLS is proposed in [15],
the MLS in the literature, like PID, Linear-Quadratic and compared with traditional Tilted-Integral-Derivative
Regulator, and Internal Model Control. However, all these (TID) controller, the parameters of controller are
methods still suffer from the problems of tuning and optimised by using GA, the results show that the I-TD
optimisation for stability, overshoot and settling time. In controller is better than TID controller.
spite of the fact that PID control is a capable procedure The remainder of the paper is organised as follows:
for dealing with of non-linear plant yet. In this way, Section 2 presents the dynamic model of the MLS.
utilizing an established controller isn't appropriate for Section 3 explains the design of classical and modern
nonlinear control application [4, 5, 6]. NN –and GA based controllers for MLS. Section 4 presents the intelligent
controllers are proposed to control non-linear system in algorithms which considered to enhance the response of
[7]. the system. Section 5 addresses the results and discussion,
The NN controller has the ability of self-learning to with a comparison between the proposed and traditional
cope with complicated environments and requirements of controllers based on simulation results. Finally, Section 6
multi-objectives control, and approach to non-linear contains the main conclusions.

DOI 10.5013/IJSSST.a.19.01.06 6.1 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

II. MLS DYNAMIC MODEL F(i, x) 2K

K i  Fi (i 0 , x 0 )   2i 0
i iio, x  x o, x0
MLS model [16] relies upon the ball kinematics and
electrodynamics conditions. The Dynamic Equation of the (5)
Control ball is given in (1), steel ball is moved upward or F(i, x) 2K 2
descending as indicated by compel connected on it by the K x  Fx (i 0 , x 0 )   3i 0 (6)
x ii0, x  x 0, x0
attractive power F and gravity drive mg, as appeared in
figure(1). Applying Newton's second law of development Ki and K x are the stiffness coefficients of the
vertical way.
magnetic force to current and air gap at the equilibrium
point respectively. From equations (4), (5), (6), (3) and
(1) for the whole system, the (F) is:-
F(i, x)  K i i  K x x  Fi (i 0 , x 0 ) (7)

d2x
m  K i (i - x 0 )  K x (x - x 0 ) (8)
dt 2
The voltage equation of the electromagnetic coil is
given by:
U ( t )  R i ( t )  L(d i / d t ) (9)
Where (L) is the static inductance between the ball and
attractive field. In system demonstrating, the input is the
control current of the electromagnet, the effect of
inductance isn't considered here. Expect the power
enhancer yield current is entirely straight with input
voltage immediately. The system can be portrayed by
Figure1. MLS Configuration [3]. following condition, mg = - K(i 0 - x 0 ) 2 , subsequent to
taking Laplace change and putting limit condition ,the
system open mathematical model is:-
d2 x
m  F(i, x) + mg (1)
dt 2 d2x 2Ki 0 2Ki 2
Where, (x) is the air hole between the ball and the m  K i (i - i 0 )  K x (x - x 0 )  i  30 x (10)
dt 2 x0 x
magnet, (m) is the mass of steel ball, (i) is current in
electromagnet and force is generated and can be given 2ki 0 2ki 02
x (s)s 2  i(s) - x (s) (11)
using following equation, (g) is acceleration speed of mx 02 mx 30
gravity. When the control current goes over the winding, x (s ) -1
 (12)
electromagnetic power is created and can be given using i(s) As 2 - B
the following condition:
Where A  i 0 , B  i 0 characterize the input
F(i,x)=K(i/x) 2 (2) 2g x 0

Where K is constant, depending on physical parameters. variable as the input voltage of the power amplifier Uin ,
F(i,x) is the attraction power, the current (i) and air hole output variable as the output voltage Uout, the System
control object model can be described as:-
(x) are nonlinearly related. For the controller
k
configuration, the system can be linearized using linear ( s )
Uout(s) ks x(s) ka
hypothesis, first at balance point (i0, x0). Now Taylor’s G(s)    (13)
Uin(s) kai x(s) As2 - B
expansion of (2), ignore higher order terms, then
F(i,x)  F(i0 , x0 )  Fi (i0 , x0 )(i- i0 )  Fx (i0 , x0 )(x- x0 ) (3) Then the system state variables are x 1  Uout ,
In which F(i0 , x 0 ) is the magnetic force equal to the x 2  Uout and the system state equations are as:-
ball gravity force, when the air gap is x 0 and the current     0 1  x   0 
 x 1    2g  1  2gK s  U in (14)
is i0 .     0   x    
 2
x 2   x 0   i0Ka 
F(i 0 , x 0 )  mg (4)

DOI 10.5013/IJSSST.a.19.01.06 6.2 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

x  III. MLS CONTROLLERS

Y  1 0 1   x 1 (15)
x 2  There are many controllers for MLS, can be classified
After substituting the real system model parameters between classical and modern approaches:-
[17, 18, 19], in equations (14) and (15), then
Y  x1 A. PID Controller
For unstable systems, the one level of opportunity
The sytem state conditions can be written as:
controller neglects to provide a smooth reference
.
X  AX  BU in , Y  CX following execution because of proportional and
derivative kick. With end goal to improve the overall
And the transfer function in equation (13) is: closed loop performance, it is basic to consider a Two
Degree of Freedom (2DOF) PID structure. A detailed
Y(s) 77.8421 examination on different 2DOF structures is illustrated in
G0 (s)   (16) [20].
Uin (s) 0.0311s 2 - 30.5250
In this paper, an endeavour has been made with the
Feed Forward (FF) 2DOF PID structures. Figure (3)
There is an open loop pole at the right of s-plane, by shows the Feed Forward type controller structure with PD
stability basis; stable system ought to have all the open controller in the feed forward loop and a PID controller in
poles on the left plane. Therefore the GML system is the closed-loop. The PID controller responds on error e(t)
unstable. See figure (2). and the PD controller works on the reference input r(t)
.The controller values for this structure are given in the
conditions below [21]:

1
C3 (s)  k p (1   d Df (s))  k p  k i  k d Df (s) (17)
i s
C4 (s)  kp (  dDf (s))  kp  kdDf (s) (18)

Figure2. Shows unstable transient response of the MLS.

Figure3. Feed forward structure of 2DOF PID [21].

In this structure, the numbers of parameters to be tuned kd


are K p , K i , K d . () And () are the controller weighting kp
parameters, ranging from 0 to 1, while D f (s) is the Where ( K d ) is the derivative gain. Using PID tune by
approximate derivative which given by: MATLAB can be calculated:-
s K p  9.07 , K i  32.4 and K d  0.0753 . Substitute, these
D f (s)  (19)
s  1

DOI 10.5013/IJSSST.a.19.01.06 6.3 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

values, in equation (17), and for C 4 controller (FF) 2DOF PID structure. The response requires more
let     0.6 , Figure (4) shows the SIMULINK block- tuning, to reduce the settling time, but this may increase
the overshoot.
Diagram, and Figure (5) shows the transient response of

Figure4. SIMULINK (FF) structure of 2DOF PID.

1 0 
R  1; and Q   
0 1 
Using MATLAB LQR statement, the optimal K is
equal to: K= [1.4664 1.0006]. Figure (6) shows closed
loop response with (LQR Controller).

Figure5. Closed Loop Response (2DOF PID Controller).

B. LQR Controller
LQR controller is intended to investigate the system in
closed-loop. The controller demonstrated enhanced
execution for various tracks [22]. To make the response of
the system stable feedback path was used. The LQR Figure6. Shows Closed Loop Response LQR Controller.
controller looks to limit the vitality show in the system,
and the control that is gotten by limiting this criterion is C. IMC
linear.The strategy is by finding a K for input in state The IMC is a notable procedure and a successful
factors to such an extent that it limits practical cost strategy for planning and actualizing robust controllers.
[23].Using equation (14) and let IMC structure is a contrasting option to the classic

DOI 10.5013/IJSSST.a.19.01.06 6.4 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

feedback structure .The input setup is appeared in Figure IV. THE INTELLIGENT ALGORITHMS FOR
(7). In the IMC plan, the controller, Gc(s), is construct ENHANCING THE RESPONSE OF THE SYSTEM
specifically with respect to the "great" some portion of the
system under control[24].The transfer function of NN and GA are used to optimise the attributes of the
controller Gc(s) is:- PID controller, to obtain the required response for the
(MLS).The two methods are explained in follows.
1
G o (s)  G p (s) -1 (20) A. NN Algorithm
(s  1) n
The NN processes the information of the system with
many properties as biological NN [25]. The intelligent
controller can estimate the behaviour of the system to
produce the desired response .The most familiar network
used is feed forward NN [26].
The NN is divided into three main layers:
1- The input layer: It takes the data to be processed
through the network. The required output of the MLS at a
time is input to the NN controller during different time.
2- The hidden Layer: In this layer by the activation
functions processing the data that take from input layer.
3-The output layer: Each neuron at the hidden layer is
connected with weights to all neurons at the output layer.
Figure7. The IMC feedback configuration [24]. The tan sigmoid (tansig) activation function is considered
here for the MLS is. The structure of NN is shown in
Figure (9).
The IMC detailing for the most part results about just
a single tuning parameter, the closed loop time constant
(  ) , the IMC filter factor. The filter order (n) is selected
large enough to achieve proper Gp(s). Choose n=2 and
 =0.01. Figure (8) shows the closed-loop response for
(IMC), the response is satisfactory, but at one second the
response is out of control, due to unstable poles at high
frequency. (  ) can be adjusted to obtain the bandwidth
and the stability required for the closed loop system.

Figure 9. Structure of NN [25].

The input to the hidden neurons can be expressed as

follows:
n
neti2  ∑ k 0
2
w ik * I1k (21)
n
net 2j  ∑ q 0
w 2jq * I1q (22)

Figure 8. Closed-loop response IMC.

Where: net i2 and net 2j are the inputs to the hidden
2
layer, w ik and w 2jq represent the hidden layer weights , I1k
and I1q are the output of the K th and q th inputs of the input

DOI 10.5013/IJSSST.a.19.01.06 6.5 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

layer, k and q represent the number of input layer controller. In this structure, the same parameters of the
neurons. PID controller are used, which are obtained from using
The control action (U) represents the output of the 2DOF PID controller. Figure (12) shows the response of
NN, which applied to the plant (MLS). In this structure, the system using the proposed NN controller.
the input signals to the NN are Z1, Z2 and Z3 from the
PID controller. S1 and S2 are fed from the NN. The block
diagram of the PID-NN controller for the MLS is shown
in Figure (10). The NN is used to control the MLS
directly. The NN controller has two inputs, the first input
is the feedback from the NN and second input is the
output of the PID controller .The optimized control action
is applied to the plant to obtain the required response with
small values of overshoot and settling time.
Figure (11) shows the Simulink model of the PID -
NN controller for the MLS. This NN is self-learning, Figure 10. A block diagram of PID - NN controller for MLS.
there is no need to modify the activation function or
neural block, if the reference input is changed. The NN
controller receives data from the output of same block
with the output of the PID controller as inputs to

Figure 11. Simulink model of PID-NN controller.

B. GA
GA follows the natural evolution for Drawin theory
"The survival of the fittest". In 1975 John Holland
proposed the GA for the first time In GA, the smallest
unit of data is the gene, the data which carried by a set of
genes is called individual, and population is a set of
individuals, synonyms are chromosome and individual
[27]. GA is used to choose the optimum parameters of the
PID controller to obtain the desired response of the
system. The process of optimisation is calculated by
maximisation of the fitness function (F) which is the
mean error between the current value of the system output
and the desired reference:
1 T
MSE 
 0 ∫
(e( t )) 2 dt (23)

Fitness  1 / MSE (24)

Figure 12.The Response of MLS with NN controller.

DOI 10.5013/IJSSST.a.19.01.06 6.6 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

A block diagram of PID-GA is shown in Figure (13), The optimisation is achieved in iterations form called
to apply GA, the chromosome elements (Kp, Kd, Ki) of generations, and creates a new set of chromosomes at
PID controller is set. each generation through crossover and mutation and the
best chromosomes are allowed to the next generation. In
this work GA parameters are chosen according to the trial
and error method as follows:

Population size=80; Crossover rate= 0.4;

Mutation rate=0.01; Maximum generation = 100.

Roulette wheel selection method is considered with
uniform mutation and arithmetic crossover. The Elitism
strategy is applied to keep the best solution over
generations, and the stopping condition is accomplished
when the maximum number of generation is reached.
Figure (14) shows the flowchart of the PID –GA system
Figure 13. A block diagram of PID-GA controller for MLS. design. Figure (15) shows the response of the MLS
system of PID-GA controller.

Figure 15. The Response of MLS with GA controller.

V. DISCUSSION

The responses of all presented controllers are

compared with respect to settling time, maximum
overshoot, and steady-state error, as shown in Table (1).
TABLE I. COMPARSION BETWEEN ALL CONTROLLERS

Controller Settling Max. Overshot Steady-

Types Time (sec) % State Error

PID 1.3 0 0
LQR 5.3 0 0
IMC 0.1 Infinity(after1sec) infinity
PID - NN 0.378 0.108% 0
PID - GA 0.009 0 0

It is noticed that GA provides better results than other

methods, but GA requires 45.74 seconds computational
time to tune the PID controller. The parameters of the PID
controller which are optimised by using the GA are:
Figure 14. The flowchart for the PID –GA system design.

DOI 10.5013/IJSSST.a.19.01.06 6.7 ISSN: 1473-804x online, 1473-8031 print

 NASIR AHMED AL-AWAD et al: PID CONTROLLER DESIGN FOR A MAGNETIC LEVITATION SYSTEM .

Kp = 155.7678, Kd = 28.9475, and Ki = 489.8021. [10] D. Hall and B. Samanta,” Nonlinear Control of a Magnetic
Levitation System Using Single Multiplicative Neuron
Models”, ASME International Mechanical Engineering Congress
When comparing the proposed method PID-GA, with and Exposition, San Diego, California, USA, November , 2013.
the same system equation. (16) in [1] and [19], using [11] J. Jesús ,et al. “Modeling and control with neural networks for a
intelligent controller (fuzzy logic controller), it is clear magnetic levitation system”, in Elsevier Neurocomputing,
vol.227, pp.113-121, 2017.
that the design specifications (Max. overshot, settling [12] Arun K. and Veena S. , “ANN approach for Magnetic Levitation
time, and rise time) are better. stabilization using gradient and Quasi Newton learning”, in IEEE
Students Conference on Engineering and Systems (SCES) ,
VI. CONCLUSIONS Allahabad, India ,2014.
[13] Z. Daghooghi ,et al.” A real-time control of maglev system using
neural networks and genetic algorithms”,in IEEE International
This paper presents a controlling method for the Conference on Industrial Technology (ICIT), Athens, Greece,
(MLS), where different approaches are proposed. The 2012.
proposed methods are PID, state-space approach, LQR [14] I. Ahmad, M. Shahzad and P. Palensky,” Optimal PID control of
Magnetic Levitation System using Genetic Algorithm”,in IEEE
and IMC, are measured, in every one of these techniques, Energy Conference (ENERGYCON), Cavtat, Croatia, 2014.
the transient plan details are not met, exceptionally in [15] D.Sain,S. Kumar and S. Kumar,” TID and I-TD controller design
settling time (response speed), as shown in Table (1). for magnetic levitation system using genetic algorithm”,in
The NN controller system for the MLS is proposed. Elsevier Perspectives in Science, Vol. 8, pp. 370-373, 2016.
[16] Googol Technology Ltd, GML series Magnetic Levitation System
The neural system structure is self-learning and User Manual and Experimental Book, 2007.
straightforward. The response of the MLS with the NN [17] Anupam Kumar, Vijay Kumar "Design and Implementation of
controller is better than the traditional PID, providing fast IT2FLC for Magnetic Levitation System”, Advances in Electrical
response, with small estimation of overshoot. There is no Engineering Systems (AEES) 116Vol. 1, No. 2, 2012.
[18] C.J. Munaro, M.R. Filho," Modeling and Observer-Based
need to compare the proposed methods with the IMC Nonlinear Control of a Magnetic Levitation System”, Proceedings
since the steady-state error and overshot are infinity after of the 2002 IEEE International Conference on Control
one second. The PID-GA method results are the best Applications,pp.162-167, September 2002.
among other methods. However, the PID-GA [19] Shailendra Chauhan1, HimanshuKumar,”Fuzzy Controller design
using LQR Fusion for Magnetic Levitation System,” IJSRET,
improvement is sensitive to the reference which used. Volume 3, Issue 1, April 2014.
PID-GA appears to offer the most encouraging result. [20] M. A. Johnson and M. H. Moradi, “PID Control,” Springer-
Verlag, London, 2005.
REFERENCES [21] Mituhiko Araki and Hidefumi Taguchi,” Two-Degree-of-Freedom
PID Controllers”, International Journal of Control, Automation,
and Systems Vol. 1, No. 4, December 2003
[1] Anupam Kumar, Vijay Kumar "Design and Implementation of [22] Alvaro Romero Acero1, Alejandro Marín Cano1 ,” Modeling and
IT2FLC for Magnetic Levitation System”, Advances in Electrical Simulation of LQR and LFSV Controllers in the Magnetic
Engineering Systems (AEES) 116Vol. 1, No. 2, 2012. Levitation System (MLS)”, 2014 III International Conference of
[2] Santanu Kumar Pradhan, Bidyadhar subudhi, " Nonlinear Control Engineering Mechatronics and Automation (CIIMA).
of a Magnetic Levitation System Using a New Input-Output [23] Engr. Sadaqat Ur Rehman1, Engr. Muhammad,” Linear Quadratic
Feedback Linearization ", IFAC-PapersOnLine 49-1 (2016) 332– Regulator controller for Magnetic Levitation System,” (IJCSSE),
336. Volume 4, Issue 1, January 2015.
[3] Anupam Kumar, Manoj Kumar Panda, and Vijay Kumar , "Design [24] Adrian-Vasile, Mircea Dulău,Stelian-Emilian Oltean,"IMC based
and Implementation of Interval Type-2 Single Input Fuzzy Logic PID Control of a Magnetic Levitation System", 9th International
Controller for Magnetic Levitation System",Proceedings of Conference Interdisciplinarity in Engineering, INTER-ENG 2015,
ICAdC, AISC 174, pp. 833–840., Springer India 2013. 8-9 October 2015, Tirgu-Mures, Romania ,ELSEVIER..
[4] R. Lakshman Kumar Reddy,” Different Controlling Methods And [25] P. Kumar and K.P.Singh,”Comparative Analysis of Air
PID Controller Design For Magnetic Levitation System”, Conditioning System Using PID and Neural Network Controller”,
IJAREEIE, Vol. 2, Issue 12, December 2013. in International Journal of Scientific and Research Publications,
[5] Kashif Ishaque,, Yasir Saleem, “Modeling and control of magnetic Vol. 3, Issue 8,2013.
levitation system via fuzzy logic controller”, IEEE,2011. [26] N. Safaa.” Improve the Performance of PID Controller by Two
[6] De-sheng, Liu Li,Wen-Sen,Chang,” Internal Model Control For Algorithms for Controlling the DC Servo Motor”, Baghdad
Magnetic Suspension System”, Proceedings of the Fourth Journal of Engineering, No. 1 ,Vol. 22, January ,2016.
International Conference on Machine Learning and Cybernetics, [27] G. Joshi,” Review of Genetic Algorithm: An Optimization
Guangzhou, 18-21 August 2005. Technique”, International Journal of Advanced Research in
[7] Shabeer Ali K P, Sanjay Sharma, Dr.Vijay Kumar, "Comparison Computer Science and Software Engineering”, Vol. 4, Issue 4,
Of Interval Type-2 Fuzzy Controller And Self Tuning Interval April ,2014.
Type-2 Fuzzy Controller For A Magnetic Levitation
System",IJERT, Vol. 2 Issue 6, June – 2013.
[8] S.Zhao and X.Gao “Neural network adaptive state feedback
control of a magnetic levitation system”,in IEEE Control and
Decision Conference (CCDC), Changsha, China,2014.
[9] M.Aliasghary,et al.” Magnetic Levitation Control Based-on
Neural Network and Feedback Error Learning Approach”,in
IEEE International Conference on Power and Energy (PEC ),
Johor Baharu, Malaysia ,December, 2008.

DOI 10.5013/IJSSST.a.19.01.06 6.8 ISSN: 1473-804x online, 1473-8031 print