Modelling and Control of Ball-Plate System

Final Project Report

Mohammad Nokhbeh and Daniel Khashabi
Under the supervision of Dr.H.A.Talebi

Amirkabir University of Technology, 2011

Abstract
AbstractIn this report we will have a detailed look into controlling a
Ball-on-Plate system apparatus. We will first derive the exact non-linear
differential equations of Ball-on-Plate apparatus by the use of Lagrange-
Euler equations. At the rest we derive several useful linearised model for
our system. We use the linearised model to design compensator for the
system. We simulate the results using non-linear model of system and
evaluate the results by step responses and trajectory tracking.

Contents
1 Introduction 2

2 Review on Previous Works 2

3 Mathematical Modelling 3
3.1 Plant Values: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Motor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3 Model Linearisation . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3.1 Linearised Model of a Ball-on-Plate . . . . . . . . . . . . 7
3.3.2 State-space Model of System . . . . . . . . . . . . . . . . 8
3.3.3 Complete Linearisation of Ball-on-Plate and DC Motor . 9

4 Controller Design 10
4.1 Single Compensator in Open-Loop . . . . . . . . . . . . . . . . . 10
4.2 Compensator With Prefilter . . . . . . . . . . . . . . . . . . . . . 11
4.2.1 Multi-loop controller 1 . . . . . . . . . . . . . . . . . . . . 12
4.2.2 Multi-loop controller 2 . . . . . . . . . . . . . . . . . . . . 12

5 Result Evaluation 13
5.1 Non-linear Simulation . . . . . . . . . . . . . . . . . . . . . . . . 15
5.1.1 Trajectory Tracking Using Non-linear Simulator . . . . . 16

1
 5.1.2 Coulomb Friction . . . . . . . . . . . . . . . . . . . . . . . 16
5.2 Two-Dimensional Overshoot . . . . . . . . . . . . . . . . . . . . . 17

6 Conclusions 18

7 Acknowledgement 18

1 Introduction
Balancing systems are one of most challenging problems in Control field. There
are lots of platforms for this sake like cart-pole system(inverted pendulum),
ball-beam system, double and multiple inverted pendulums[1].
The ball-on-plate system is a promoted version of the traditional ball-on-beam
control problem. The problem consists of a plate which its deviation can be
manipulated in two perpendicular directions. The goal is to carry the ball
moving on the plate to a desired position, that is to control a freely rolling ball
on a specific position or moving on a trajectory on the plate. In more precise
way, one can do stabilization control in which the goal is to carry the ball from
one specific position and hold it in the desired position or do trajectory tracking
control in which the goal is to make the ball follow a predefined trajectory with
the least error and in a the minimum time.

2 Review on Previous Works

We can classify the previous works on ball-on-plate system into to general cate-
gories. Those based in mathematical modelling and those proposing model-free
controllers.
The [2] is devoted to a mechatronic design of the system which benefits a touch
sensor feedback. In [1] a Fuzzy Reinforcement Learning method is proposed.
The previous paper benefits a high accuracy stepping motor with 0.9 degree
size of steps which is reducible by the drivers down to 1/200 of step size, i.e.
4.5 103 . It uses precise incremental encoder (3600 ppr) which is coupled to
motors shaft. Altogether by taking the mechanical limitations, it benefits the
smallest measurable and applicable amount of rotation 0.1 degrees. It uses a 15
inches touch sensor that its output is a message packet being sent through RS-
232 serial communication with 19200 bps. Thus the fastest sampling rate of the
whole sensor block is 384 samples per second. This implies that the maximum
available time for decision making is 1/384
= 2.6103 seconds. The area of the
surface of the touch sensor on which pressure can be sensed 30.41 22.81cm.
Sensor resolution is 1900*1900 pixels so if the sensor sensitivity is uniformly
distributed on its sensitive area, then each pixel is assigned to an area approxi-
mately equals to 1.8*1.2 mm2 of the surface of the sensor.
In [3] a PID neural network controller(PIDNN) based on genetic algorithm(GA)
is used for control. In this approach GA is a applied in training weighting factor
of multilayer feed-forward neural network in order to overcome the disadvantage

2
 Figure 1: Side view of ball on plate system.

of back-propagation algorithm which falls into local minima. It is claimed that

the proposed method is robust, adaptive and strong control performance.
In [2] ball position measured by a touch pad rather than a camera. In the pre-
vious paper a linear quadratic state feedback regulator was designed after the
system was linearised around a few operating points. Accuracy of stabilization
5 mm. Steady state error in average was 18 mm in circle following while tra-
jectory velocity was less than 4.2 mm/s
In [4] a hierarchical fuzzy control scheme with a TS type fuzzy tracking con-
troller was proposed. In [5] sliding mode control with uncertain items observe
compensation, is used to control a ball-on-plate system.

3 Mathematical Modelling
In this section we are going to derive the motion equations of system. In these
part we assume following simplifications:

There is no slipping for ball.

The ball is completely symmetric and homogeneous.
All frictions are neglected.

The ball and plate are in contact all the time.

Here we derive dynamical equations of ball-on-plate system by the help of

Lagrangian. The following precise mathematical equations are based on [6][4][7]
([4, 6, 7]). The Euler-Lagrange equation of ball-plate system is as followings:
d T T V
+ = Qi (1)
dt qi qi qi

3
Where qi stands for i-direction coordinate, T is kinetic energy of the system, V
is potential energy of system and Q is composite force.
The system has 4 degree of freedom; two in ball motion and two in inclination
of plate. Here we assume the generalized coordinates of system to be (is) xb
and yb balls position on plate and and the inclination of the plate. It is
important to note that we assume the center of x y coordinates be at center
of plate. The kinetic energy of ball consists of its both rotational with respect
to its center of mass and translational energy:
1
1
mb x 2b + y 2b + Ib x 2 + y 2


Tb = (2)
2 2
Where mb is mass of the ball and I b is moment of inertia of the ball; x b and
x b are balls translational velocities along x-axis and y-axis; x and y are balls
rotational velocities along x-axis and y-axis. The following relations between
translational velocities and rotational velocities:

x b = rb y , y b = rb x (3)
In which rb denotes balls radius. By substituting equations 3 into equations
4 we will have:

   
1 2 2

Ib 2 2 1 Ib
x 2b + y b2



Tb = mb x b + y b + 2 x b + y b = mb + 2 (4)
2 rb 2 rb

The kinetic energy of the plate, by considering ball as a point mass which is
placed in (xb , yb ), , consists of the (ezafi) its rotational energy with respect to
its center of mass:

   
Tp = 1
2 (I p + Ib ) 2 + 2 + 12 mb xb + yb 2
    (5)
= 1
2 (I p + Ib ) 2 + 2 + 12 mb x2b 2 + 2xb y
b + yb2 2

Where and are plates angle of inclination along x-axis and y-axis, re-
spectively. Therefore and are plates rotational velocity. Here we can
calculate the kinetic energy of system as followings:
Where and are plates angle of inclination along x-axis and y-axis, re-
spectively. Therefore and are plates rotational velocity. Here we can
calculate the kinetic energy of system as followings: (paragraf tekrari)

T = Tb+ Tp   
= 21 mb + rIbb2 x 2b + y b2 + 21 (Ip + Ib ) 2 + 2


  (6)
b + yb2 2
+ 21 mb x2b 2 + 2xb y

The potential energy of the ball relative (the relative potential) to horizontal
plane in the center of the inclined plate can be calculated as:

4
 Vb = mb gh = mb g (xb sin + yb sin ) (7)
Here we can derive the systems equation by Lagrangian and equations 4-7:

L = Tb + Tp Vb (8)
We use L to derive systems equations:
T   L
= (Ip + Ia ) x + mb xb xb + yb , = mg cos (9)

T   L
= (Ip + Ia ) x + mb yb yb + xb , = mg cos (10)

 
T Ib L  
= mb + 2 xb , = mb xb + y b (11)
xb rb xb
 
T Ib L  
= mb + 2 yb , = mb xb + y b (12)
yb rb yb
We assume generalized toques as x and y which are exerted torques on the
plate. From Lagrange-Euler equation we can write:
d T L
dt
= (Ip + Ib ) + mb x2 + 2mb xb xb + mb xb yb

(13)
+ mb x b yb + mb xb y mb g cos = x
d T
dt L
= (Ip + Ib ) + mb yb2 + 2mb yb yb + mb xb yb
(14)
+ mb y b xb + mb yb x b mb g cos = y

   
d T
dt x b L
x b = mb + Ib
rb2
b mb xb + yb + mb g sin = 0
x (15)

   
d T
dt y b L
y b = mb + Ib
rb2
yb mb yb + xb + mb g sin = 0 (16)

So the non-linear differential equations for the ball-plate-system as follow-

ings:
 
Ib  
mb + 2 x b mb xb 2 + yb + mb g sin = 0 (17)
rb
 
Ib  
mb + 2 yb mb yb 2 + xb + mb g sin = 0 (18)
rb

+ 2mb xb x b + mb xb yb


x = Ip + Ib + mb x2b
(19)
+mx b yb + mb xb y b + mb gxb cos

y = Ip + Ib + mb yb2 + 2mb yb y b + mb yb xb



(20)
+mb x b yb + mb xb y b + mb gyb cos

5
 The equations (18)-(17) shows the relation between balls state and plates
state that is plates inclination. The equations (19)-(20) shows the effect of
external torque on the ball-on-plate system. It is obvious that working with such
complex equations is too hard. So in order to do Laplace analysis on systems
transfer function, we derive the linearised model of system in neighbourhood of
working state.

3.1 Plant Values:

In this report we use the values that are shown in following table that are based
on [4]:

Parameter Description Value

mb Mass of the ball 0.11kg
rb Radius of the ball 0.02m
Sp Dimension of the plate 1.0 1.0m2
vmax Maximum velocity of the ball 4mm/s
mp Mass of the plate 0.1kg
Ip Mass moment of inertia of the plate 0.5kg.m2
Ib Mass moment of inertia of the ball 1.76 105 kg.m2
g Gravitational acceleration 9.81

3.2 Motor Model

Here we assume (use or consider) an Armature-Controlled DC motor which its
block diagram is shown in figure 2. The motor uses armature current ia as the
control variable. The stator field can be established by a field coil current. The
detailed description of equations can be found in [8] and [9]. When we have a
constant current in the field coil, we have the following postulates:

Tm (s) = Km Ia (s)
V (s) = (Ra + La s)Ia (s) + Ea (s)
V (s) = Kb (s)
= Ia (s) = [V (s) Kb (s)] / [Ra + La s]
The absolute torque exerted on the load can be found by:

TL (s) = Tm (s) Td (s)

By letting Td (s) = 0 and considering GL (s) = TLL (s)

(s)
to the load torque-
velocity characteristic, one can find the input-output transfer function:

L (s) Km GL (s)
G(s) = V (s) = s[(Ra +sLa )+Km Kb GL (s)]

6
 Figure 2: Block diagram for an armature-controlled DC motor.

We can assume an approximate function for load characteristic as following :

1
TL (s) = Js+B Where J stands for load inertia and B for load damping. We
rewrite the motors transfer function in this way:
L (s) Km
G(s) = V (s) = s[(Ra +sLa )(Js+B)+Km Kb ]

We assume following typical values for DC motor[10]:

Parameter Description Value

Km Toque constant 6.876 104 N.m/A
Kb Back-emf constant 7.2 10 V /rpm or 1.9099 107 V.s/rad
5

Jm Rotor inertia 1 103 N.m.s2 /rad

B Rotor damping 108 N.m/(rad/s)
f Field time constant 1ms
Rotor time constant 100ms
R Armature resistance 3.9
L Armature inductance 12 106 H
Pmax Maximum output power 0.25hp
Ts Stall torque 2.44 N m
Vf l Rated DC supply voltage 1.5V

3.3 Model Linearisation

In this part we will derive several linearised model of system that we will use
them in next sections.

3.3.1 Linearised Model of a Ball-on-Plate

2 2
The approximate value for a solid balls moment of inertia is Iball = 5 mb rb .
Therefore equations (17) and (18) can be written as:

7
  
5  
mb b xb 2 + yb + g sin = 0
x (21)
7
 
5 
2


mb yb yb + xb + g sin = 0 (22)
7
We can linearise above equation by assuming:
Small angle of inclination for the plate(up to 5 ):  1 and  1
= sin ' , sin '
Slow rate of change for the plate:  0 and  0 = ' 0 , 2 ' 0,
2
' 0.

5
x
b + g = 0 (23)
7
5
yb + g = 0 (24)
7
By linearising the above equations, we find two separate differential equations
for each of x and y axis. Note that we can use above linear differential equations
to estimate the ball-on-plate systems states xb , yb , x b , y b . By assuming (s)
and (s) as inputs to ball-on-plate system, we find the transfer functions:

Xb (s) g
Px (s) = = 5 2 (25)
(s) 7s

Yb (s) g
Py (s) = = 5 2 (26)
(s) 7s

3.3.2 State-space Model of System

One can write state equations of ball-on-plate system by considering Equa-
tions (17)-(20)[3]. As mentioned x and y are toques exerted on the plane(or
on the ball from the earth) in y-axis and x-axis direction respectively. By
defining state variables X = [x1 , x2 , x3 , x4 , x5 , x6 ]T = [xb , x b , , , T,
y , y , , ]
T T 2

b b
U = [ux , uy ] = [, ] and constant value B = m/ m + Ib /rb we write the
non-linear state equation x = f (x, u):

x1 x2 0 0

x2


B(x1 x24 + x4 x5 x8 g sin x3 )


0 0


x3


x4


0 0
 

x4
= 0
+ 1 0
ux

x5


x6


0 0
uy

x6


B(x5 x28 + x1 x4 x8 g sin x7 )


0 0

x7 x8 0 0
x8 0 0 1

8
 Figure 3: Block diagram of the Plant.

Because
   1 and
of the low velocity and acceleration of the plate rotation(||
 

   1) we can omit the coupling term in f (x, u) and divide the system into
to sub-systems and control them independently.

x1 x2 0
x2 B(x1 x24 g sin x3 )
+ 0 ux
 
x3 =

x4 0
x4 0 1

x5 x6 0
x6 B(x5 x28 g sin x7 )
+ 0 uy
 
x7 =

x8 0
x8 0 1

The implementation of the plant is shown in figure 3.

3.3.3 Complete Linearisation of Ball-on-Plate and DC Motor

We use MATLABs internal function to generate a linear model of our non-
linear model. First we consider (ezafi) are going to make a linear of model
which consists of Plant which is ball-on-plate and a DC motor. The goal of
this model is to consider the loading effect of ball and plate on the motor.
Because if have ball and plates that have considerable weight, they affect the
position/velocity of motor. The complete Simulink model of system is show in
figure (5).
The linearisation must be done in infinitesimal neighbourhood of equilibrium
point. The equilibrium point is where speed and its derivatives are zero. By
considering zero value for velocity and acceleration terms, we find the following
point as equilibrium:

9
 Figure 4: Block diagram of the whole system.

X = [xb , x b , , , T = [0, 0, 0, 0, 0, 0, 0, 0]T

yb , y b , , ]

MATLAB CODE:
DefineConstants();
[A,B,C,D]=linmod(PM,[0,0,0,0,0,0,0,0,0,0,0,0,0,0],[0,0])
sys = ss(A,B,C,D)
sys2=tf(sys)

The resulting output transfer functions are as followings:

X(s) 802.2
V1 (s) = s5 +3.25105 s4 15.09s4.905106
Y (s) 802.2
V2 (s) = s5 +3.25105 s4 15.09s4.905106

If we plot the root-locus plot of the above transfer function, we will get the
figure (6).

4 Controller Design
In this section we are going to design compensator for the transfer functions we
derived in previous parts. Here we propose several methods of compensation
and in

4.1 Single Compensator in Open-Loop

The first compensator consists of only a compensator in open-loop. Unfor-
tunately this system is unproper and DC gain of this controller is 100 (for
stabalizing), its settling time (Ts ) and overshoot is also unacceptable.

10
 Figure 5: Block diagram of the motors with non-linear load characteristics.

3461538.4615(s+0.01)2 (s+2)(s2 +10s+26)

C(s) = (s1.8)(s+100)

Gain Poles Zeros

100 2 5

Root-locus and step response of the first compensator are shown in figure (7).

4.2 Compensator With Prefilter

The second compensator consists of a compensator and a prefilter. Unlike first
compensator this system is proper and DC gain and percent overshoot of this
controller is acceptable but this compensator have a Ts = 15s settling time.
5 4 3 2 6
C(s) = 0.4294s 139600s 323700s 705500s 1.20210 s798.5
s5 +9.971s4 +43.77s3 +111.2s2 +195.5s
F (s) = 0.4294s5 +139600s4 +323700s798.5
3 +705500s2 +1.202106 s+798.5

Gain Poles Zeros Prefilter poles

4.08 5 5 5
In order to find better compensator, we change the architecture add following
add a new controlling loop. The compensator at the rest of report are designed
to control plant and previous controller.

11
 Figure 6: Pole-zero map of systems transfer function.

4.2.1 Multi-loop controller 1

the following system is designed to make the settling time of second compensator
appropriate. this compensators aim is to get have overshoot but rapid timerise.
Gain Poles Zeros Prefilter poles
1.6 8 8 1

F (s) = 16.12
s+10
C(s) =
7.2961013 s8 +5.5971014 s7 +1.9061015 s6 +3.7291015 s5 +4.5861015 s4 +3.6131015 s3 +1.7881015 s2 +4.9171014 s+6.2511013
s8 +400s7 +7104 s6 +7106 s5 +4.375108 s4 +1.751010 s3 +4.3751011 s2 +6.2511012 s+3.9071013

4.2.2 Multi-loop controller 2

the following compensator is desinged to make the settling time of second com-
pensator appropriate. this compensators aim is to critical damp mode.
Gain Poles Zeros Prefilter poles
.6 8 8 1

F (s) = 26.66
s+10
C(s) =
2.7361013 s8 +2.0991014 s7 +7.1481014 s6 +1.3981015 s5 +1.721015 s4 +1.3551015 s3 +6.7061014 s2 +1.8441014 s+2.3441013
s8 +400s7 +7104 s6 +7106 s5 +4.375108 s4 +1.751010 s3 +4.3751011 s2 +6.2511012 s+3.9071013

12
 Figure 7: Root-locus and step response of the first compensator.

5 Result Evaluation
In this part we are going to evaluate the results we got from previous parts.
The main goal of this part part is to check the validity of the previous results
and whether they are valid for the case of main non-linear equations. For this
goal, we designed the comprehensive non-linear simulator to check the validity
of results.

13
Figure 8: Root-locus and step response of the second compensator.

Figure 9: Architecture of added controlling loop.

14
 Figure 10: Root-locus and step response of the third compensator.

5.1 Non-linear Simulation

In order to have a realistic evaluation of the designed compensator and its
operation we need to design a non-linear simulator by the use of primary non-
linear equations (17)-(20).

15
 Figure 11: Root-locus and step response of the forth compensator.

5.1.1 Trajectory Tracking Using Non-linear Simulator

The task of tracking a trajectory is one of important tasks that one controller
must be able to accomplish. In figure (14) the results of several tracking exper-
iments are shown.

5.1.2 Coulomb Friction

Despite the complex microscopic origins of Coulomb friction, we can model it
by simple models. One popular model for formulating the Coulomb friction is
as following[11] [12] :

16
Figure 12: Main block diagram for simulation in presence of designed controller.

Figure 13: The designed controllers and non-linear implementation of the sys-
tem.

Tf = tanh ( x)

Where and are constant values. For large values of the Coulomb friction
is similar to a Sign Function and for lower values of we get softer Coulomb
friction, as shown in figure (15).

5.2 Two-Dimensional Overshoot

As it is described above systems model is discretely described in each dimension
and there is no relation between two subsystem. Consider measuring of over-
shoot of system, overshoot is the excessive value which systems takes for speed,
but what is the relationship between the overshoot measured in each subsystem
and overshoot of system. The output signal will be the joint of both systems
and because both systems overshoot times are not equal, system will track an
unknown trajectory which is for sure in the rectangle of O1 O2 . Overshoot is
usually considered in building of the physical system, the systems bounds are
based on overshoots measured. Considering all above a new measurement of

17
overshoot can be considered which is maximum distance from the zero point in
the 2D path. Figure 16 2d-overshoot point (the point with maximum distance
of zero) of a joint system consisting of two standard 2nd order linear subsystems
are drawn. Assuming n1 = n2 this will be just in the corner of rectangle, but
when n1 6= n2 this point will be on one of the sides.

2
n1
H1 (s) = 2
s2 +21 n1 +n1
2
n2
H2 (s) = 2
s2 +22 n2 +n2

In figure 17 the 2d-overshoot of different systems with different 1 = 2 is

shown.It shows that systems which their natural
frequencies of their subsystems
are equal have the worse 2d-overshoot: 2(O1d ) (It is also proven with some
other experiments). For example when n1 = 1, n2 = 1, 1 = 1 = .6, O2d =
1.5482, but for n1 = 1, n2 = .5, 1 = 1 = .6, O2d = 1.4767.

6 Conclusions
In this report we derived complete equations of ball-on-plate system. We did
linearisation analysis to find the transfer function of the system. We designed
several compensator and at the rest made some analysis on the results. We tried
several trajectory tracking tasks. The final results showed that the designed
controllers had almost acceptable performances.

7 Acknowledgement
The authors would like to thank all their classmates, friends and masters that
helped on this project. They would like to thank Ehsan Omidi and Farokh
Atashzar for their illuminating discussion and ideas. They also knowledge
Dr.H.A.Talebi for his comprehensive and off course perfect instructions.

References
[1] N. Mohajerin, M. B. Menhaj, and A. Doustmohammadi, A reinforcement
learning fuzzy controller for the ball and plate system, in Proc. IEEE Int
Fuzzy Systems Conference, 2010, pp. 18.
[2] S. Awtar, C. Bernard, N. Boklund, A. Master, D. Ueda, and K. Craig,
Mechatronic design of a ball-on-plate balancing system, Mechatronics,
vol. 12, no. 2, pp. 217228, 2002.
[3] X. Dong, Z. Zhang, and C. Chen, Applying genetic algorithm to on-line
updated pid neural network controllers for ball and plate system, in Proc.

18
 Fourth Int Innovative Computing, Information and Control (ICICIC) Conf,
2009, pp. 751755.
[4] X. Fan, N. Zhang, and S. Teng, Trajectory planning and tracking of ball
and plate system using hierarchical fuzzy control scheme, Fuzzy Sets and
Systems, vol. 144, no. 2, pp. 297312, 2004.

[5] D. Liu, Y. Tian, and H. Duan, Ball and plate control system based on
sliding mode control with uncertain items observe compensation, in IEEE
International Conference on Intelligent Computing and Intelligent Systems,
ICIS09, vol. 2. IEEE, 2009, pp. 216221.

[6] K. kyu Lee, G. Btz, and D. Wollherr, Basketball robot: Ball-on-plate with
pure haptic information, in IEEE International Conference on Robotics
and Automation, May 2008.
[7] D. Yuan and Z. Zhang, Modelling and control scheme of the ball-plate
trajectory-tracking pneumatic system with a touch screen and a rotary
cylinder, Control Theory & Applications, IET, vol. 4, no. 4, pp. 573589,
2010.
[8] F. Golnaraghi and B. Kuo, Automatic control systems, vol. 2.
[9] R. Dorf and R. Bishop, Modern control systems, 11th ed. Pearson Printice
Hall, 2008.

[10] W. Bolton, Mechatronics: electronic control systems in mechanical and

electrical engineering. Prentice Hall, 2004.
[11] D. Haessig and B. Friedland, On the modeling and simulation of friction,
in American Control Conference. IEEE, 2009, pp. 12561261.

[12] C. S.J. and T. R.W., On the modeling of coulomb friction, UNKNOWN

JOURNAL OR CONFERENCE, 1998.

19
Figure 14: Results of trajectory tracking.

20
 Figure 15: Coulomb Friction Models.

Figure 16: 2-D overshoot place. This figure shows the point which has the
maximum distance from zero for different parameters of 1 and 2 . Each curve
is a class of systems which have one subsystem and the other subsystem differs.

21
Figure 17: 2-D overshoot. This figure shows the the maximum distance from
zero for different systems with 1 = 2 according to their subsystems natural
frequencies ratios.

22