International Journal of Dynamics and Control

https://doi.org/10.1007/s40435-020-00629-8

Observer integrated backstepping control for a ball and plate system

Jie Ma1 · Hao Tao1 · Jingwen Huang1

Received: 22 February 2020 / Revised: 23 March 2020 / Accepted: 30 March 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
An observer integrated backstepping control is designed for a ball and plate system with the characters of cascade and model
uncertainties. The model of the system is written in the cascaded strict feedback form. The uncertainties of the model and
derivatives of the virtual controls are estimated by a linear extended state observer and a tracking differentiator, respectively.
The convergence of the derivative estimation is studied. The actual control and virtual controls are designed based on the
cascaded model with the Lyapunov theory. The stability of the closed-loop system is consequently proven. Simulations and
experimental results demonstrate the excellent tracking performance of the control design.

Keywords Ball and plate system · Backstepping · Linear extended state observer (LESO) · Tracking differentiator (TD)

1 Introduction the backstepping control algorithm with adaptive fuzzy con-

trol [6, 7] or neural networks [8, 9]. However, these methods
The ball and plate system is a typical benchmark for con- require knowing full-state output information. For the ball
trol theory research because of its characteristics of cascaded and plate system, only position information is available due
structure, under-actuated and strong-coupling. Some exam- to structure restriction. In addition, heavy noise is usually
ples of its application can be found today in robotics, associated with measurements, which may also deteriorate
aerospace or even medicine. Scholars have proposed con- the performance of these controllers.
trol schemes to solve the nonlinear, strong coupling and The Active Disturbance Rejection Control (ADRC),
uncertainty problems of the ball and plate system and have which was proposed by Han [10], can not only observe the
improved the control performance of the ball and plate sys- derivatives of position, but also provide real-time estimation
tem to a certain extent. These methods include: sliding mode of the uncertain model. The extended versions of the method
control [1], LQ control [2], fuzzy control [3], cascade PID such as Tracking Derivative (TD) [11] and Linear Extended
[4], neural networks-based PID [5], etc. State Observer (LESO) [12] have been used to estimate
Backstepping is a classic method to solve the control derivatives and deal with the mismatched disturbances. These
design problem of cascaded under-actuated systems by trans- methods have been combined with backstepping to estimate
forming the control design of the entire system into several unknown models and to suppress the flexible dynamics and
lower-order subsystems. At each step, a Lyapunov function parameters in power systems, unmanned aerial vehicle flight
is used to recursively obtain the progressive stability con- systems, and chemical processes [13–16].
trol law of the system, which effectively ensures the overall To the best of our knowledge, the works mentioned above
stability of the cascade system. However, the traditional back- assume that the TD is fast enough to ensure the tracking
stepping depends on the mathematical model of the system, stability. Moreover, the controller-observer separation prin-
and requires the system model in the form of strict feed- ciple is the main theory used for stability proof. However,
back structure. This requirement limits the wide application as is known, the controller-observer separation principle is
of the backstepping. In order to solve the model inaccuracy only applicable to linear systems and may not hold for the
problem of backstepping control, some scholars combined nonlinear ball and plate system. Furthermore, the parame-
ters for TD to be fast enough to track the reference should
B Jingwen Huang be addressed in detail so that the whole system stability is
[email protected]
guaranteed.
1 Beijing University of Chemical Technology, Beijing 100029,
China

123
 J. Ma et al.

Fig. 1 Schematic diagram of the

ball and plate system in
x-direction

In this paper, the linear ESO is integrated with the back- where 1 and 2 denote the unknown nonlinear parts of the
stepping method in a cascaded interaction mode so that the system and the disturbances, u is the control and represents
entire stability of the nonlinear system is guaranteed. The the input voltage V m of the servo motor.
parameters for TD to be fast enough to track the virtual con- Introduce the state variables as
troller is discussed in detail so that the tracking stability is
guaranteed. x1  x; x2  ẋ
This paper is organized as follows: The system structure x3  θ ; x4  θ̇ (2)
and the conversion to the strict feedback form are presented
in Sect. 2. Section 3 presents the ADRC based estimation of Then the system model (1) can be represented by the fol-
the uncertain model and derivatives, the cascaded interaction lowing cascaded feedback model:
control design procedure and the theoretical results. Exper-
imental setup and results are presented in Sect. 4. Section 5 ẋ1  x2 + f 1 , ẋ2  x3 + f 2
contains the conclusions. ẋ3  x4 + f 3 , ẋ4  ηu + f 4 (3)

where
2 System structure
f 1  0, f 2  −x3 + ẍ
The ball and plate system makes use of two same rotary f 3  0, f 4  −ηu + θ̈ (4)
servo base units made by Quanser. The plate is symmetrical.
The dynamics along each axis is therefore assumed to be the f i contains ball and plate system dynamic model and
same. As shown in Fig. 1, the motor drives the servo angle θ unknown disturbance is the pseudo-control coefficient. In the
to further control the displacement x of the ball through the following sections, the function f i will be estimated through
movement of the angle α of the beam. The system is a typical a reduced-order linear extended state observer in every step
cascaded, under-actuated system. of the backstepping design. The equivalence proof of the con-
From Newton’s second law, we can state that the dynamic trol between the pseudo-control coefficient and the original
equations for the displacement of the ball and the output angle controller will be given.
of the DC motor can be written in the following general form:

ẍ  a11 x + a12 θ + a13 ẋ + a14 θ̇ + b1 u + 1

θ̈  a21 x + a22 θ + a23 ẋ + a24 θ̇ + b2 u + 2 (1)

123
Observer integrated backstepping control for a ball and plate system

3 Backstepping control where L is the observer gain vector.

We use the cascaded feedback model (3) of the ball and plate Theorem 1 Assuming that the derivatives of the system mod-
system to design backstepping controls. We first present an els are bounded, the model estimation error obtained by the
adaptive disturbance rejection linear extended observer to observer (LESO) is bounded.
deal with model uncertainties. Some details of the tracking
derivative (TD) for estimating the derivative of the virtual Proof Define the state estimation error as e(t)  s(t) − ξ (t).
controls are presented, which is responsible for avoiding the According to (8) and (9), we have
‘differential explosion’ of the backstepping control design.
The guideline of choosing the parameter for TD is discussed. ė(t)  Ae e(t) + Eh(x, t) (10)
The virtual and actual controls are designed with the Lya-  
punov method. The stability of the closed-loop system is −β1 1
consequently proven. where Ae ( A − LC) 
−β2 0
By selecting a proper L  [β1 β2 ] and setting the roots of
3.1 Linear extended observer its characteristic equation s 2 +β1 s +β2 on the left half of the s-
plane, under the assumption that h(x, t) f˙i (x, t) is bounded,
The backstepping treats each state variable x i as a control in the state estimation error is also bounded as suggested by
the cascade system and creates a virtual control design x id for Eq. (10). 
the state. The overall control will reduce the tracking error
between the virtual controller x id and each state variable x i Note that he unknown model estimate can be expressed

defined as: as: fˆi (x, t)  ξi2  h i (x, t)dt, and the estimation error is
bounded.
z i  xi − xid , i  1, 2, 3, 4 (5)
3.2 Derivative estimation
Especially, when i  1, the tracking target is the given
reference denoted as x ref , which means x1d  xr e f Derivatives of various terms in the traditional backstepping
Introducing the tracking error (5) into Eq. (3), we have: control design for high-order systems are needed and often
lead to the so-called “differential explosion” involving many
ẋi  z i+1 + x(i+1)d + f i , i  1, 2, 3 terms in the control. When the system model is inaccurate
ẋ4  ηu + f 4 (6) or unknown, we cannot have analytical expressions of the
derivatives. In this paper, TD is used to estimate the deriva-
Each subsystem can be further expressed in the following tive of the virtual control in the backstepping design. The
extended form: following closed-loop system of a minimum time optimal
bang-bang control is used to achieve fast estimation.
ẋi  z i+1 + x(i+1)d +xi+4
ẋi+4  h i ẏi1  yi2
 
yi  C xi (7) yi2 |yi2 |
ẏi2  −r sign yi1 − xid + (11)
 2r
where xi+4  h i (x, t)dt f i (x, t), i  1, 2, 3, 4.
The state equation corresponding to the extended model where yi1 is the output of the system that is supposed to
(7) can be expressed as: track the target x id , for example, yi2 is then an estimate of
the derivative ẋid by virtue of the first line of Eq. (11), and
ṡ(t)  As(t) + BU (t) + Eh(x, t) r is the parameter that determines the response speed of the
y(t)  Cs(t) (8) tracking differentiator.
In order to ensure that the TD achieves excellent track-
where s  [xi xi+4 ]T , A  [01; 00]T , B  [10]T , E  ing in a short time, we examine the relationship between
[01]T , C  [10], U (t)  z i + x(i+1)d , when i  4, U(t)  ηu. the backstepping control speed and the tracking convergence
Equation (8) can be used to design a linear extended state time.
observer (LESO) as: In Eq. (11), we drop the index i for now and denote yi  ȳ.
The solution of Eq. (11) can be obtained as:
ξ̇ (t)  Aξ (t) + BU (t) + L(y(t) − ŷ(t))
1 2
ŷCξ (t) (9) ȳ1  ȳ + c+ (u  +1) (12)
2r 2

123
 J. Ma et al.

When c− − c+ > 0, on the path a → c, we have

ȳ2 (t)  r t + ȳ12 (18)

1
t1  ( ȳ22 − ȳ12 ) (19)
r

t 1 is the time the system takes to move from a to c. Simi-

larly, on the path c → b,we have

ȳ2 (t)  −r t + ȳ22 (20)

1
t2  ( ȳ22 − ȳ32 ) (21)
r

The total time for the system to move from point a to reach
point b is
Fig. 2 Phase trajectory and switching curves of the closed-loop
1 1
minimum-time optimal control system tab  t1 + t2  ( ȳ22 − ȳ12 ) + ( ȳ22 − ȳ32 ) (22)
r r

1 2 For the case when point a is to the right of point b, the

ȳ1  − ȳ + c− (u  −1) (13) required time t ab moving from a to b is
2r 2
1 1
where c+ and c− are integration constants. tab  t1 + t2  ( ȳ12 − ȳ22 ) + ( ȳ32 − ȳ22 ) (23)
Assume that the system initially starts from point a on the r r
phase plane shown in Fig. 2, the end position is at point b on where
the right. According to the minimum time optimal control,
⎧
any point in the phase plane can reach the target only once. ⎨ ȳ21  1 (c+ + c− )
Then, the path from the initial position to the terminal posi- 2 (24)
tion in the shortest time must be a → c → b where point c is ⎩
ȳ22  − r (c+ − c− )
the intersection of two switching curves as indicated in the
Fig. 2. The total time for the system to move from point a to It is clear that the time t ab is inversely proportional to r
reach point b can be calculated by the following methods: and is a function of the locations a and b. The differences of
At point a ( ȳ11 , ȳ12 ), u  + 1. We have the locations a and b represent the estimation error xid and
ẋid if the TD is used to track xid . Assume that we have suffi-
1 2
c+  ȳ11 − ȳ (14) cient data to estimate the upper bounds of various estimation
2r 12 errors. Then, we can find a real number S > 0 such that
At point b ( ȳ31 , ȳ32 ), u  − 1. We have S
tab ≤ (25)
− 1 2 r
c  ȳ31 + ȳ (15)
2r 32
Let the backstepping control sampling time be dt. We can
At point c ( ȳ21 , ȳ22 ), we must have select the value of r such that tab ≤ S/r < dt. It follows that
the derivative estimate ẋˆid by using Eq. (11) can be made
⎧ 1 2
⎪ to converge fast enough not to introduce time delay to the
⎨ ȳ21  ȳ + c+
2r 22 (16) control.
⎪
⎩ ȳ  − 1 ȳ 2 + c−
21
2r 22 3.3 Lyapunov design
Therefore,
Consider the Lyapunov function in terms of the tracking error
⎧ of each subsystem as:
⎨ ȳ21  1 (c+ + c− )
2 (17) 1
⎩ Vi  z i2 (26)
ȳ22  r (c− − c+ ) 2

123
Observer integrated backstepping control for a ball and plate system

Differentiating Eq. (26), and making use of Eqs. (5) and

(6), we have

V̇i  z i ż i z i (ẋi − ẋid )z i (z i+1 + x(i+1)d + f i − ẋid ), i  1, 2, 3

V̇4  z 4 ż 4 z 4 (ẋ4 − ẋ4d )  z 4 (ηu + f 4 − ẋ4d ) (27)

Especially when i  1, xid  xr e f . The total Lyapunov

function of the system is equal to the sum of the Lyapunov
functions of the subsystems. To make the system stable, we
consider the following virtual and actual controls which make
each subsystem stable as well as the entire system stable.

x(i+1)d  −ci z i − fˆi +ẋˆid − z i+1 − z i , i  1, 2, 3

u  (−c4 z 4 − fˆ4 + ẋˆ4d )/η (28)

where the estimated function fˆi and the estimated derivative

of the virtual control ẋˆid are used. According to Theorem 1,
the estimates are bounded and can be made very small. For
this reason, in the stability of proof, we assume that the esti-
mation errors are negligible. Consequently, we can show that

V̇i  − (ci +1)z i2 , i = 1, 2, 3

Fig. 3 The plate and ball system
V̇4  − c4 z i2 (29)

Hence, when ci > −1, i  1, 2, 3 and c4 > 0 ,the closed- present the results of numerical simulations in X direction.
loop system is stable in the Lyapunov sense. Since the experimental setup is not perfect, a weak coupling
between the dynamics in X and Y directions exists.
3.4 Control equivalence
4.1 Numerical simulations
In [17], it has been shown that under the assumption that
the sign of the control coefficient b2 and its upper bound are The equations in X direction including the servo motor in
known, when |η|  |b2 |, the control design with the substi- the loop can be found by using the Lagrange method, and are
tuted control coefficient η converges to the control designed given by
with the known coefficient b2 . ⎡ ⎤
x2
⎡ ⎢ ⎤ ⎥
ẋ1⎢ 2m b grar m rb2 ⎥
⎢ ⎥ ⎢ ⎢   sin x3 + ξ (y) ⎥
4 Simulation and experimental results ⎢ ẋ2 ⎥ ⎢ L plate m b rb + Jb
2 ⎥
⎥
Ẋ  ⎢ ⎥⎢
⎢ ẋ ⎥ ⎢ ⎥
⎣ 3 ⎦ ⎢ x4 ⎥
Numerical simulations and experiments of the ball and plate ⎥
⎢ ⎥
system are carried out to demonstrate the performance of the ẋ4 ⎣ Beq,v Am ⎦
− x 4 + u
proposed control design. The hardware system used in the ηg K g2 Jm + Jl ηg K g2 Jm + Jl
experiment is made by Quanser. As shown in Fig. 3, the sys- (30)
tem consists of a ball, a smooth plate, two servo motors in the
X and Y directions, and an overhead camera. The position where x 1 is the displacement of the ball, x 2 is the velocity of
of the ball is measured with the overhead camera. The ball the ball, x 3 is the servo angle displacement, x 4 is the servo
displacement and speed are controlled indirectly by the servo angle speed. ξ (y) is the disturbance induced by the movement
motors adjusting the plate angles to track the desired trajec- of Y direction. The control variable u represents the input
tory. The component names and system parameter definitions voltage V m .
are listed in Tables 1 and 2 in the appendix. It is assumed that the dynamics of each axis is the same.
When the X and Y axes are perfectly perpendicular to each Therefore, only X direction is simulated in the numerical
other, the dynamics of the system in these two directions are simulation. In the simulation, the square trajectory with
decoupled and independent. For this reason, we shall only amplitude of 3 cm and a sinusoidal trajectory with period

123
 J. Ma et al.

4
x
2 of 40 s are considered as the reference. The control parame-
0
0 10 20 30 40 50 60 70 80
ters are taken as c1  0.08, c2  9, c3  120000, c4  0.06, η
1
 10000000. The tracking speed parameter r for estimating
θ (deg)

0
the derivative of each stage of the virtual control is taken to
−1
10
0 10 20 30 40 50 60 70 80
be 3.6 × 105 , 8 × 1010 and 8 × 1014 . The LESO observer
gains are β1  0.8 and β2  1.2.
V m (V )

−10
0 10 20 30 40 50 60 70 80 Figure 4 shows the simulation results of the ball tracking
Time (s)
the square trajectory. Figure 5a, b show the simulation results
Fig. 4 Simulation results of tracking a square trajectory of the ball tracking the sine and cosine trajectories, respec-
tively. In the figures, the black curve is the reference, and the
red dash line is the tracking response. It is obvious that the

5 5

0 0
x

y
−5 −5
0 10 20 30 40 50 0 10 20 30 40 50

20 20
θ (deg)

θ (deg)
0 0
−20 −20
0 10 20 30 40 50 0 10 20 30 40 50

20 20
V m (V )

V m (V )

0 0
−20 −20
0 10 20 30 40 50 0 10 20 30 40 50
Time (s) Time (s)
(a) (b)
Fig. 5 a Simulation results of tracking a sine trajectory. b Simulation results of tracking a cosine trajectory

5 5
0
x

0
−5
−5
0 10 20 30 40 50 0 10 20 30 40 50
100 100
θx (deg)

θy (deg)

0 0

−100 −100
0 10 20 30 40 50 0 10 20 30 40 50
100 100
V mx (V )

V my (V )

0 0

−100 −100
0 10 20 30 40 50 0 10 20 30 40 50
Time (s) Time (s)
(a) (b)
Fig. 6 a Experimental results of tracking a square trajectory in X direction; b Experimental results of the tracking a sinusoidal trajectory in Y
direction

123
Observer integrated backstepping control for a ball and plate system

5 5

0 0
x

y
−5 −5
0 10 20 30 40 50 0 10 20 30 40 50
100 100
θx (deg)

θy (deg)
0 0

−100 −100
0 10 20 30 40 50 0 10 20 30 40 50
100 100
V mx (V )

V my (V )
0 0

−100 −100
0 10 20 30 40 50 0 10 20 30 40 50
Time (s) Time (s)

(a) (b)
Fig. 7 a Experimental results of tracking a sine trajectory. b Experimental results of tracking a cosine trajectory

proposed control has good performance in tracking square 6

and sine trajectories.
4

2
4.2 Experimental results

0
y

The controls are implemented in Matlab/Simulink 2014,

which interacts with the hardware through a data acquisi-
tion card and a power amplifier. The experiment uses the −2
same control parameters in both X and Y directions as the
numerical simulations. However, the LESO observer gains
are changed to be β1  2.5 and β2  750. −4
The system responses are shown in Fig. 6. Figure 6a is
a response of tracking the square trajectory in X direction, −6
while tracking the sine trajectory in Y direction as shown −6 −4 −2 0 2 4 6
x
in Fig. 6b. The black curve is the target trajectory and the
red dash line is the tracking response. It is a challenging Fig. 8 Experimental results of tracking a circle
problem to control the ball to follow a smooth trajectory in
one direction and a non-smooth trajectory in the other.
We should note that the open-loop motions in X and
Y directions of the physical system are weakly coupled. 5 Conclusion
Once the ball goes through step changes when tracking the
square trajectory in X direction, the response in Y direction We have presented an observer integrated backstepping con-
is affected. However, as observed in Fig. 6a, although there trol for the cascaded ball and plate system with model
are step changes in X direction at 20 s and 40 s, but the cou- uncertainties. The LESO is embedded in each backstepping
pling effect on the response in Y direction is suppressed by design step, and the derivative of the virtual control is esti-
the control as indicated by Fig. 6b. mated by TD. The guideline to select the parameter for TD
Figure 7a is a response in X direction to track the sinu- for fast estimation of the derivative in order to avoid time
soidal trajectory, while the ball tracks the co-sinusoidal delay is presented. The overall stability of the closed-loop
trajectory in Y direction as shown in Fig. 7b. In the X–Y system is proven with the Lyapunov stability theory. The
plane, the responses form a closed circular trajectory as results of numerical simulations and physical experiments
shown in Fig. 8. show that the proposed control has the ability of decoupling

123
 J. Ma et al.

and suppressing uncertainty while achieving excellent track- References

ing performance.
1. Bang H, Lee YS (2018) Implementation of a ball and plate control
Acknowledgement This work was supported by the National Natural system using sliding mode control. IEEE Access 6:32401–32408
Science Foundation of China through the Grants (11702016, 11972070) 2. Spacek L, Bobal V, Vojtesek J (2017) Digital control of ball and
and by Educational Research Grant (2018BHDJGY23) of Beijing Uni- plate model using LQ controller. In: International conference on
versity of Chemical Technology. process control, Štrbské Pleso, Slovakia
3. Hamid M, Elaheh L (2017) Designing a fuzzy type-2 model-based
robust controller for ball and beam system. Procedia Comput Sci
105:125–130
Appendix 4. Adiprasetya A, Wibowo AS (2016) Implementation of PID con-
troller and pre-filter to control non-linear ball and plate system. In:
See Tables 1 and 2. 2016 International conference on control, electronics, renewable
energy and communications (ICCEREC), 2016
5. Amin M, Ji-Chul R (2020) Neural network-based PID compensa-
tion for nonlinear systems: ball-on-plate example. Int J Dyn Control
Table 1 Names of the plate and ID Component 8:178–188
ball 6. Moezi SA, Zakeri E, Eghtesad M (2019) Optimal adaptive interval
1 Ball type-2 fuzzy fractional-order backstepping sliding mode control
2 SRV02 for X-Axis method for some classes of nonlinear systems. ISA Trans 93:23–39
7. Jinzhu P, Rickey D (2019) Adaptive fuzzy backstepping control
3 SRV02 for Y-Axis for a class of uncertain nonlinear strict-feedback systems based on
4 Plate dynamic surface control approach. Expert Syst Appl 120:239–252
5 Support arm 8. Pan Y, Sun T, Liu Y, Yu H (2017) Composite learning from adaptive
backstepping neural network control. Neural Netw 95:134–142
6 USB camera
9. Wu J, Li J, Chen W (2014) Semi-globally stable adaptive NN
backstepping control for uncertain MIMO systems with tracking
Table 2 The parameters of the plate and ball system accuracy known a priori. J Frankl Inst 351:5274–5309
10. Han J (2009) From PID to active disturbance rejection control.
Symbol Description Value IEEE Trans Ind Electron 56:900–906
11. Ibraheem IK, Abdul-Adheem WR (2016) On the improved nonlin-
mb /kg Mass of ball 0.003 kg ear tracking differentiator based nonlinear PID controller design.
rb /cm Radius of ball 1.96 cm Int J Adv Comput Sci Appl 7:234–241
12. Zhang S, Wang Q, Yang G, Zhang M (2019) Anti-disturbance back-
g/(m/s2 ) Gravitational acceleration 9.81 m/s2
stepping control for air-breathing hypersonic vehicles based on
La /cm Support arm length 25.4 cm extended state observer. ISA Trans 92:84–93
Lplate /cm Table length 27.5 cm 13. Liu Y, Ma J, Tu H (2008) Robust command filtered adaptive back-
Lc /cm Camera support height 69.5 cm stepping control for a quadrotor aircraft. J Control Sci Eng 1–9
14. Yang T, Li A, Sun X (2015) An auto disturbance rejection back-
Jb /m4 Moment of inertia of the ball 4.6217e−07 m4 stepping super maneuver nonlinear control law design method. J
Beq,v/ (N.m.s/rad) Equivalent viscous damping 0.0150 N.m.s/rad Harbin Inst Technol 47:46–50
coefficient 15. Wang YX, Zhou Z, Zhang B, Shao Z, Zhu XP (2017) Flying-
Am Actuator gain 0.3347 wing attitude control based on backstepping and active disturbance
rejection control. Flight Dyn 35:30–34
ηg Transmission efficiency 0.9000 16. Cheng Y, Chen ZQ, Sun MW, Sun QL (2017) Multivariable inverted
Kg Total gear ratio 70 decoupling active disturbance rejection control and its application
Jl /(kg m2 ) Load moment of inertia 0.0001 kg m2 to a distillation column process. Zidonghua Xuebao/acta Autom
Sin 43:1080–1088
Jm /(kg m2 ) Motor armature momentof 0.0021 kg m2 17. Huang J, Zhang T, Sun JQ (2019) Data-driven backstepping control
inertia of underactuated mechanical systems. J Dyn Syst Meas Control
rarm /cm Distance between SRV02 2.54 cm 141:091003–091009
output gear shaft and
coupled joint

123