Design ofROV Adaptive Sliding Mode Control System for Underwater Vehicle

Based on RBF Neural Network

Wei Chen1, Shilin Hu2,Qingyu Wei3

I. School of Electronics and Information Jiangsu University of Science and Technology,Zhenjiang,

China,212000,E-mail: [email protected]

2.School of Electronics and Information Jiangsu University of Science and Technology,Zhenjiang,

China,212000,E-mail:h [email protected]

3.School of Electronics and Information Jiangsu University of Science and Technology,Zhenjiang,

China,212000,E-mail: I99030044@stu. just.edu.cn

Abstract: The dynamic positioning control of ROY near the water surface under wave disturbance is still a challenging
2021 33rd Chinese Control and Decision Conference (CCDC) | 978-1-6654-4089-9/21/$31.00 ©2021 IEEE | DOI: 10.1109/CCDC52312.2021.9602771

problem. The principle of sliding mode control and the method of approximating unknown function by RBF neural

network are studied. The adaptive sliding mode controller of RBF neural network is designed. The stability and

convergence of the proposed algorithm are deduced and verified, and compared with the simulation results of traditional

adaptive sliding mode control methods.The simulation results show that the ROV's trajectory tracking effect is good in the

wave disturbance environment. The experimental results prove the effectiveness of the method and achieved satisfactory

performance.

Key Words: ROY, Control System, Sliding Mode Controller, RBF neural network

change of disturbance[4]. Traditional sliding mode

1 INTRODUCTION
control usually has chattering problems, which will cause
The development of the marine field has become the wear of the executive structure. When some scholars use
focus of strategic competition among countries in the neural networks to design control systems, which have
world. The dynamic positioning control of underwater not considered the problem of excessive calculations
robots is a key technology that can perform the specified caused by the estimation of unknown functions and
tasks well. This is also a hot topic for domestic and neural network control[5]. The above methods overcome
foreign research[ I]. a few related control problems caused by uncertain
In the current research results on the control method of hydrodynamic coefficients[6-7]. Disturbance of water
underwater robots, many controllers designed according flow environment could change the dynamic
to the nominal model have not analyzed the influence of performance of the control system. The disturbance
external interference and parameter uncertainty on the causes the uncertainty of ROY dynamics model and
control performance[2-3]. When the PID control affects the tracking accuracy of ROY along the desired
algorithm has time-varying and complex working trajectory of the target. There is no better solution yet.
environment for the underwater operation robot system, To solve the above problems, this paper adopts a
the parameters of PID control will change with the dynamic positioning method of RBF neural network
adaptive sliding mode control. First, the RBF neural
This project is supported by National Natural Science Foundation network is used to approximate the wave disturbances
of China (Grant no.51809128).This research is supported by Industrial existing in the operation process. Then the iterative
Foresight Project of Zhenjiang City (GY2018018) . function of the neural network weights and the sliding
mode controller are designed for adaptive online

978-1-6654-4089-9/21/$31.00 @2021 IEEE 2976

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.
 adjustment. Finally, the simulation results under the ideal environment interference, and regard them as the system
environment and wave disturbance environment verify uncertainty term F(4, v) :
the efficiency of this method. This control system F(;;,v)=J(;;,v)+d(;;,v) (3)
overcomes the shortcomings of insufficient positioning According to the nonlinear approximation ability of the
accuracy and unstable resistance to external disturbances. RBF neural network, there is a theoretically optimal
The experimental results and data fully prove that the weight such that:
ROY adaptive sliding mode control system of the
underwater robot based on the RBF neural network has (4)
better performance than the traditional dynamic
In the formula, H(V4)=exp(-(V4-m) 2 /a2 and H
positioning control.
are the output of the hidden layer, and the trajectory
2 RESEARCH ON SLIDING MODE tracking error 4 is the input of the hidden layer. m and
CONTROL BY RBF NEURAL NETWORK a are obtained by K clustering method . c is the
APPROACHING UNKNOWN FUNCTION network approximation error.

For a given ROY expected motion trajectory [771'772 y , In actual work, the neural network approximates the
the trajectory tracking error4 = [41,42 ] can be defined. unknown vector F(!;, v) as follows:

!;1=1J-1J1 A A A

F(4, v) = WF H(V F 4) (5)

(1)
!;1 and (2 are respectively the trajectory error and the In the formula, F (4 , v ) is the estimated output of the
W F and V F are the estimated
A

trajectory error rate of change. Before and after the wave neural network.
disturbance, the ROY heading and vertical dynamics weights.
models change greatly, which does not make it possible to 3 DESIGN OF ADAPTIVE SLIDING MODE
obtain a more accurate ROY dynamics model in the CONTROLLER
actual ROY operation. Therefore, this paper converts the
According to the underwater robot model, the following
dynamic model into a nonlinear uncertain system in the
state equation of the nonlinear uncertain system is
form of state space equations. Through this conversion,
established:
the uncertain part of the ROY model and the water flow
environment interference can be formed into a set of two {x1 = Xz
(6)
unknown items as shown in equation (2). It is convenient x2 = f (x , t) +Bu+ d(t)
for the next step of online approximation through neural The attitude information of the ROY for underwater

network[8]. operations is selected as the state vector of the system,

x E Rn ; u is the input vector of the system, u E Rm ;
A nonlinear uncertain system describing the state space
d(t) is the external random disturbance; f(x,t) is the
can be established as follows:
nonlinear function, and B is the constant value
matrix.Transform formula (6) into:
x = f(x,t) +Bu+ d(t) (7)
. -
42 = f C4, V )+Bu+d(4,v)-17d (2) Define xd as the desired attitude, e as the attitude error,
and e as the rate of change of attitude error:
RBF neural network approximates the model uncertainty
and water flow interference online. In actual operations,
{� = �d - �
(8)
the ROY dynamics model uncertainty term f(4, v) and e = xd -x

water flow interference d(4, v) in equation (2) are Design sliding mode switching surface:

unknown uncertain terms. For this reason, this paper uses s = ce +e (9)

the neural network method to approximate the unknown Derivation of the sliding mode surface designed in this
items f(4, v) and d(4, v) . Combine the model paper, and substituting equations (7) and (8) into the

uncertainty term of equation (2) with the flow result:

2021 33rd Chinese Control and Decision Conference (CCDC) 2977

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.
 s= ce + e = ce +id - f(x,t) - Bu - d(t) 4 SYSTEM STABILITY ANALYSIS
=- K·sgn(s) - d(t) (10)
According to the sliding mode controller of equations(lO)
Design sliding mode control rate:
and (11), select the Lyapunov function:

u= iCce +id - f(x,t) - d(t) + K·sgn(s)) (11) (18)

The ROY for underwater operations is susceptible to
Derivation of the above formula:
external disturbances such as water currents and waves in
the actual working process. This unknown function is
L1 = ss = s(ce +id - f(x,t) - Bu - d(t))
= s(- K·sgn(s) - d(t)) (19)
denoted as F(x) [9-12]. Since the RBF neural network
In the formula, d(t) is the external disturbance, satisfying
has the characteristics of approximating any nonlinear
the conditions of ld(t)I � D and K;::: D, then the above
function, the network algorithm for approximating
formula is:
F(x) is:

{ _

2b 2
2
h·J = exp ( llx-cill )
(12)
L1 = s(- K·sgn(s) - d(t)) �- Klsl - s ·d(t)

According to L1 ;::: 0 and L1 � 0 , the sliding mode

� 0
(20)

:
F(x) = W*Th(x) +E E �En
control system is asymptotically stable[13].
In the formula, W* 1s the ideal weight, and E is the
For the RBF neural network adaptive sliding mode
approximation error.
controller designed by equations (15) and (16), select the
Define the output of the RBF neural network:
Lyapunov function:
F(x) = wrh( x) (13)
1 1 - ­
In the formula, wr is the estimated weight. When W = L =-s 2 +-yWrW (21)
2 2 2
W* - W, then:
Then derivate the Lyapunov function, because the
F(x) - F(x) = W*Th(x) +E - wrh( x ) = wrh(x) +E
approximation error of the RBF neural network can be
(14)
limited to small enough, and substitute the adaptive law
Introducing the above-mentioned RBF neural network
into the formula:
idea of approximating the disturbance unknown function
L =- s(K·sgn(s) + E)
into the sliding mode controller, the sliding mode control 2
=-ES- Klsl
rate based on the neural network is:
�- Klsl � 0 (22)
u= i(ce +id - f(x,t) - F(x) - d(t) + K·sgn(s)) Since L ;::: 0 , L � 0 , and thus s and W are
2 2
bounded, when t --> oo is s --> 0, and thus e--> O, e--> 0,
(15)
so the controller system is asymptotically stable.
Substituting the sliding mode control rate into the sliding
mode surface derivative of equation (10), we get: 5 ROV DYNAMIC POSITIONING CONTROL
SIMULATION AND EXPERIMENT BASED ON
s= ce + e = ce +id - f(x,t) - F(x) - Bu - d(t)
RBF NEURAL NETWORK
= ce +id - f(x,t)
1 5.1 Simulation And Result Analysis
- F(x) - B· ( ce +id - f(x,t) - F(x) - d(t)
-

B
To verify the performance of the dynamic positioning
+ K·sgn(s)) - d(t)
controller designed in this paper. This article uses
= F(x) - F(x) - K·sgn(s)
MATLAB software to simulate the ROV's underwater
=- WTh(x) - K·sgn(s) - E (16) movement and analyze the simulation results.This article
Design system adaptive law: researchs the simulation of heading and vertical dynamic

_,__1 positioning. Analyzing the ROY motion model can

W =-- s·h (x ) (17)
y further simplify the above formula to a single degree of

In the formula, y is the adaptive parameter, y > 0. freedom. The obtained heading and vertical single degree
of freedom kinematics model is expressed as:

2978 2021 33rd Chinese Control and Decision Conference (CCDC)

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.
 m�� + d�� + d�1�1�l�I + g� = T� (23) velocity of water wave motion in the actual experiment
In the formula, m� is the coefficient of inertia. T� 1s may be slow, so it can be approximately considered that
the force generated by the thrust in the � degree of the water wave is stable and invariant, and the wave is in
freedom. g� is the component of the ROY vertical force a state of random motion. The target trajectory tjJ =
in the � degree of freedom. d� is the primary damping 12 cos (0.016rrt) is tracked in the bow direction. The
coefficient. d�l�I is the secondary damping coefficient. simulation results are shown in Figure 3. Tracking the

In order to verify the performance of the dynamic preset depth trajectory in the vertical direction. The target

positioning control method designed in this article, this trajectory expression is: H = 0.05 sin (0.026rrt) - 2. The

section applies the control method proposed in this article simulation results and error comparison are shown in

and the traditional adaptive sliding mode control method Figure 4.

to the ROY dynamic positioning control system under the

same working conditions. Comparative experimental Traditional sliding mode control tracking error
• RBF network adaptive sliding mode control tracking error
verification is carried out. 20

In an ideal situation, the target trajectory tjJ =

10
12 cos (0.016rrt) is tracked in the bow direction, and the
""
simulation results are shown in Figure!. Tracking the Ci 0
"'
c:

preset depth trajectory in the vertical direction, the target

-10
trajectory expression is: H = 0.05 sin (0.026rrt) - 2. The
simulation results are shown in Figure 2. -20

-�--
-30 �-
0 50 100 150 200 250 300
time/s

Fig 3. Heading trajectory tracking error

0.1 .----�--�-�
--Traditional sliding mode control tracking error
- • RBF network adaptive sliding mode control tracking error

0.05

-15

-� �-�--�--�-�--�-__J
0 50 100 150 � 250 300
time/s
-0.05
Fig I. Heading trajectory tracking curve

-0.1

-0.15 �--�--�-----'
0 50 100 150 200 250 300
time/s

Fig 4. Vertical trajectory tracking error

It can be seen from the simulation curve , under ideal

conditions, the heading trajectory and vertical trajectory
-2.06
are simulated. Compared with the traditional adaptive
-2.08

-2.1 �-�--�--�--�-�--_J
sliding mode control method, the method of this paper is
0 50 100 150 200 250 300
time/s closer to the expected value. In the presence of wave

Fig 2. Vertical trajectory tracking curve

disturbance, the target trajectory is tracked in the heading
and vertical directions. The interference function of wave
The main interfering factors in the ROY operation
process are water currents and waves. However, the disturbance is approximated online by RBF neural
network. The errors of the control method are more lower

2021 33rd Chinese Control and Decision Conference (CCDC) 2979

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.
 than that of the traditional adaptive sliding mode control
method. The simulation results show that the effect of the
RBF dynamic positioning control method is significantly
better than the traditional adaptive sliding mode control
method.

5.2 Experiment And Result Analysis

The anti-interference test of water flow is carried out for

the wave disturbance problem. The experiment uses a
water pump to generate random water flow, and the
purpose is to test the ROV's anti-interference ability to
Fig 6. Robot in hover
water flow. When the joystick controls the rotation of the
In the case of wave disturbance during the experiment,
underwater robot, the tail thruster rotates forward and
the target trajectory is tracked in the bow direction, and
reverse to make the ROY rotate in place. At the same time,
the interference function of the wave disturbance is
the rotation of the ROY can be achieved through the
approximated online through the RBF neural network.
differential rotation of the two propulsion motors on both
The data collected by the computer can be seen in Table 1
sides. Experiments show that ROY can achieve in-situ
and Table 2. The mean square error, maximum error and
rotation, flexible steering, faster response, better stability,
average error of the control method in this paper are
and better bow steering effect. It can be obviously
0.2467°, 11.85° and 0.112° respectively. Compared with
observed in the experimental pictures that the ROY is
the traditional adaptive sliding mode control method, the
under the action of water flow in three directions. It can
error percentage is reduced by 41.75%, 46.55%, 58.52%.
be seen that the steering of the ROY is hardly disturbed
The vertical trajectory is tracked in the presence of
by the external environment, as shown in Figure 5.
disturbance. The mean square error of the control method
is 0.0057m, which is 54.76% lower than the traditional
adaptive sliding mode control method. The maximum
error is 0.047m, which is 54.37% lower than the
traditional adaptive sliding mode control method, and the
average error is 0.0013lm. Compared with the traditional
adaptive sliding mode control method, it reduces by
46.96%. The effect of the dynamic positioning control
method in this paper is obviously better than that of the
traditional adaptive sliding mode control method.

Tab. I Heading trajectory tracking error data comparison

testing method Mean Maximum average

square error!° value!°

error/0

Fig 5. Steering movement of underwater ROV Disturbing Method of 0.2467 11.85 0.112

As shown in Figure 6, when the ROY is hovering, the environment this article

water flow is applied to the ROY, and the water flow is heading Comparison 0.4235 22. 17 0.27

suddenly increased or decreased, or the direction is tracking method

suddenly changed. The experimental results prove that the Percentage 41.75% 46.55% 58.52%

ROY has a good trajectory tracking effect in a wave reduction

disturbance environment.

2980 2021 33rd Chinese Control and Decision Conference (CCDC)

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.
 Tab.2 Vertical track tracking error data comparison results[C]//OCEANS, 2011 IEEE-Spain. IEEE, 2019, 1-10.

[5]Tian Y, Zhao J Q, Li Y C. A sliding mode control-based on a

testing method Mean Maximum average
RBF neural network for deburring industry vehicleic
square error/m value/m
system[J]. International Journal of Advanced Vehicleic
error/m
Systems, 2020, 12(1 ): 1-10.
Disturbance Method of 0.0057 0.047 0.00131
[6]Panwar V. Wavelet neural network-based trajectory motion
environment this article
for Vehicle manipulators using fast terminal sliding mode
vertical Comparison 0.0126 0.103 0.00247
control[J]. Vehicleica, 2018, 26(6): 1-16.
tracking method
[7]Zhang M J, Chu, Z Z. Adaptive sliding mode control based
Percentage 54.76% 54.37% 46.96%
on local recurrent neural networks for underwater robot[J].
reduction
Ocean Engineering, 2018, 45: 56-62.

6 SUMMARY OF ARTICLE [8]Binglei Guan, Hailun Wang, Yan Zheng. Design of Dynamic

The ROY adaptive sliding mode control system of the Positioning System for Anti-interference Ships under

underwater robot based on RBF neural network is Different Sea Conditions[J]. 2020, 83(spI):897-90I.

researched in this paper. The design process of the [9]Arash Pourhashemi, Amin Ramezani, Mehdi Siahi.

adaptive sliding mode controller is described, and the idea Designing dynamic fractional terminal sliding mode

of RBF neural network approaching the unknown controller for a class of nonlinear system with uncertainties.

disturbance function is introduced into the sliding mode 2019, 13(2):197-225.

controller. The article uses stability analysis and proves [l O]Tian Y, Zhao J Q, Li Y C. A sliding mode control-based on a
the stability of the control system. Simulations are RBF neural network for deburring industry vehicleic
performed under ideal conditions and wave disturbances system[J]. International Journal of Advanced Vehicleic
respectively. In the experiment, the water pump generates Systems, 2019, 12(1): 1-19.
random water flow to test the anti-interference ability of
[ll]Panwar V. Wavelet neural network-based trajectory motion
the control system designed in this paper against external
for Vehicle manipulators using fast terminal sliding mode
disturbances. The experimental results show that the
control[J]. Vehicleica, 2019, 26(6): 1-16.
adaptive sliding mode control dynamic positioning
[12]Zhang M J, Chu, Z Z. Adaptive sliding mode control based
method of the RBF neural network is significantly better
on local recurrent neural networks for underwater robot[J].
than the traditional adaptive sliding mode control method.
Ocean Engineering, 2019, 45: 56-62.

REFERENCES [l3]Fourati H, Manamanni N, Afilal L, et al. A complementary

observer-based approach for the estimation of motion in

[I]Francesco Cappello, Subramanian Ramasamy, Roberto
rigid bodies using inertial and magnetic sensors[C].//Control
Sabatini. A low-cost and high performance navigation system
Applications, 2018 IEEE International Conference on. IEEE:
for small RPAS applications[C]. Aerospace Science and
2018, 422-427.
Technology, 2019, 58:97-103.

[2]Yurong Lin, Wei Zhang, Jianqiong Xiong. Specific force

integration algorithm with high accuracy for strapdown

inertial navigation system[J]. Aerospace Science and

Technology. 2018, 42:58-65.

[3]Tannuri E A, Agostinho A C, Morishita H M, et al. Dynamic

positioning systems: An experimental analysis of sliding

mode control[C]. Control engineering practice. 2019, 18(10):

1121-1132.

[4]Dukan F, Ludvigsen M, Sorensen A J. Dynamic positioning

system for a small size ROY with experimental

2021 33rd Chinese Control and Decision Conference (CCDC) 2981

Authorized licensed use limited to: ULAKBIM UASL ISTANBUL TEKNIK UNIV. Downloaded on December 29,2021 at 18:04:21 UTC from IEEE Xplore. Restrictions apply.