International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041

Available online at www.sciencedirect.com

ScienceDirect

Original article

Design of Course-Keeping Controller for a Ship Based on Backstepping and

Neural Networksյ
Qiang ZHANG a, Na JIANG b*, Yancai HU c, Dewei PAN d
a
Navigation College, Shandong Jiaotong University, China, Zq20060054@163.com
b*
Dept.of Basic Teaching, Shandong Jiaotong University, China, 14209912@qq.com, Corresponding Author
c
Dept.of Logistics and Maritime Studies, Mokpo National Maritime University, South Korea, 727210362@qq.com
d
Navigation College, Shandong Jiaotong University, China, ludnaldo@qq.com

Abstract
Due to the existence of uncertainties and the unknown time variant environmental disturbances for ship course
nonlinear control system, the ship course adaptive neural network robust course-keeping controller is designed by
combining the backstepping technique. The neural networks (NNs) are employed for the compensating of the
nonlinear term of the nonlinear ship course-keeping control system. The designed adaptive laws are designed to
estimate the weights of NNs and the bounds of unknown environmental disturbances. The first order commander
are introduced to solve the problem of repeating differential operations in the traditional backstepping design
method, which let the designed controller easier to implement in navigation practice and structure simplicity.
Theoretically, it indicates that the proposed controller can track the setting course in arbitrary expected accuracy,
while keeping all control signals in the ship course control closed-loop system are uniformly ultimately bounded.
Finally, the training ship of Dalian Maritime University is taken for example; simulation results illustrated the
effectiveness and the robustness of the proposed controller.

Keywords: Ship course control; DSC; Backstepping; Adaptive neural network

Do not DELETEPG
յ(D

Copyright གྷ 2017, International Association of e-Navigation and Ocean Economy. Hosting by Elsevier B.V.
This article is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer review under responsibility of Korea Advanced Institute for International Association of e-Navigation and Ocean
Economy.

յ
This is a revised version presented at the 4th Ai-MAST held at Mokpo, Korea, December 12-14, 2016.
http://dx.doi.org/10.1016/j.enavi.2017.06.004
 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041 35

1. Introduction part existing in backstepping method, and in the

literature (Zhang, Zhang and Chen, 2015), the linear
Ship motion control is an important research field in
weakening of backstepping was presented and "Yulong"
traffic engineering. Its ultimate goal is to improve the
ship was taken as an example to simulate the ship course
level of automation and intelligence, and ensure the
keeping. The results showed that the method is effective.
safety, economy and comfort of the navigation (Zhang,
In practice, the unknown time varying ocean
2012). Ship motion has the characteristics of nonlinear,
environment of the ship was not considered in the
strong coupling and large inertia, and it is a kind of
literature (Zhang, Zhang and Chen, 2015). In order to
typical uncertain nonlinear system (Jia and Yang, 1999).
overcome the uncertainty constant interference problem
Ship course control research has become an important existing in the ship system, Guan, Zhang and Wang
direction in the field of ship motion control. At first, the (2009) put forward a design scheme of ship course
PID controller is widely adopted (Xia, 1993). However, controller using closed-loop gain shaping algorithm and
due to the poor robustness of PID to the external integral backstepping method. The simulation results
environment, PID cannot show its unique advantages. showed that the designed controller can eliminate the
With the development of adaptive control theory, static error caused by the sea wind, and have a better
various control strategies were used in the course control. ship course control effect. According to the wave
Nomoto model was used as the model of the ship motion environment disturbance existing in ship course control
control system in Yang, Yu and Jia (1999). The method system, Liu, Sui and Xiao, et al (2011) used a
of designing the robust PID-type autopilot was presented disturbance observer to estimate the disturbance of the
by using the Lyapunov stability theory. According to the environment, and combined with the backstepping
characteristics of nonlinear and ship operation control in sliding mode to design the keeping controller. The
the process of complex interference, Liu, Wu and Zou simulation results showed the effectiveness of the
(2000) designed a fuzzy and integral controller for ship control strategy. For the effective stabilization of drift
course. The controller can realize the ship course uncertainty in nonlinear ship course keeping system,
keeping in case of disturbances. Considering the Tian and Liu (2015) combined the Lyapunov energy
characteristics of the rudder in the ship course control function with the preset stabilization boundary with
system, Wu and Zhu (2009) presented the backstepping adaptive backtsepping method, and finally constructed
sliding mode design algorithm using state vector the robust control with preset stabilization boundary.
linearization of feedback linearization theory. And it Although the disturbance influence of the external
designed a ship course-keeping backstepping controller. environment was considered, the uncertainty of ship
In the literature (Liu, Wu and Zou, 2000; Wu and Zhu, model was ignored. Li, Bai and Xiao (2014) designed a
2009), the control algorithms were designed based on new type of ship course sliding mode controller
the linear model of ship. However the nonlinearity of the combined extended state observer with sliding mode
ship model was ignored. Based on the Norbin nonlinear technology for the nonlinear ship motion mathematical
ship model, the nonlinear heading controller was model. While the model of uncertainty and external
designed in paper (Wu, Jia and Zhang, 2002 ) by using disturbance were considered, but with the controller
the backstepping method, and the simulation results appears obvious chattering, and this will not be
showed that the method could achieve a better course conducive to the rudder control of the implementation
keeping effect. The control strategy based on RBF for the directive.
neural network and backstepping was designed for ship
Li, Bai and Xiao (2014) constructed the Lyapunov
course nonlinear control system in paper (Xia and Luan,
function in backstepping method. It is favored by
2015). Finally, the control designing strategies were
researchers for its systematic advantage of design
used in the ship's course keeping control system.
process. However, the repeated derivation of the virtual
Considering the characteristics of the rudder, Wang, Liu
control law in the backstepping design increased the
and Li (2016) proposed an adaptive control design
complexity of the controller. The derivation of the
method based on the neural network and applied to the
virtual control law cannot be obtained when considering
control of the ship course. It is difficult to eliminate the
model uncertainty. In order to solve the problem of
practical problems of the output energy and nonlinear
36 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041

repeated derivation virtual control law in the traditional where ¢,£ are nonlinear parameters of ship.
backstepping, Farrell, Polycarpou, and Sharma et al
Let x1 \ ˈ x2 \ ˈ u G ˈ and from the
(2009) and Dong, Farrell and Polycarpou, et al (2012)
introduced command filtering into backstepping method formula (1) and (2) we can obtain
and proposed a filtering backstepping method. In x1 x2 (3)
addition, when the nonlinear mathematical model of the
system is highly nonlinear, the application of K
x2 f ( x2 )  u  d (t ) (4)
backstepping is limited. T
Inspired by the above literature, this paper proposes an y x1 (5)
adaptive neural network control method based on the
instruction filtering technique for the nonlinear where y  R is the output of the system ˈ
mathematical model of ship course motion. The method 1
f ( x2 )  H ( x2 ) is the unknown nonlinear function
uses neural networks to approximate the unknown T
nonlinear of ship course mathematical model online. of the system, H ( x2 ) Dx2  Ex23 ˈu is the designed
Using a two order filter to obtain the derivation of the
control law.
virtual control law significantly simplifies the design
process of backstepping method, and avoids the Assumption 1. d (t ) is a bounded external environment
"computation explosion" problem. Considering disturbance
unknown time-varying environment disturbance, an
d (t ) d d (6)
adaptive law is designed to estimate the unknown time-
varying disturbance of the environment. Then the
where d is the upper bound of external environment
controller output of the equivalent rudder angle
disturbance.
counteracts the environmental disturbance, and makes
the actual output of the ship to track the desired course. Assumption 2. Expected reference course yd is

Symbol Description˖Rn the n-dimensional Euclidean smooth and bounded, and has two order continuous
bounded derivative. There exists a positive constant B
space, ˜ is norm of Euclidean, Omax (˜) and Omin (˜)
respectively represent the maximum and minimum
having – ^( y d `
, y d , yd ) : y d2  y d2  yd2 d B .

values of the matrix, diag(g) is a diagonal matrix, ˜ is

absolute value. 2.2 RBF Neural Networks
In this paper, we use the radial basis function neural
networks to approximate the unknown nonlinear terms
2. Problem description and preliminary knowledge
in the mathematical model (1) of the nonlinear ship
2.1 Problem description course control system. According to the paper (Wang
The relationship between rudder angle and course in X2016), it can be expressed as
the mathematical model of nonlinear ship course control H ( Z ) T *T S ( Z )  E ( Z ) (7)
system can be described as (Fossen, 1994):
where H (Z ) is the unknown nonlinear term to be
1 K
\  H (\ ) G  d (t ) (1) approximated, Z is the input vector, T * is a ideal
T T
weight, E ( Z ) is the approximation error having
where T and K is ship manoeuvrability index; G is the
rudder angle, \ is the heading angle, d (t ) is external E (Z ) d E * (8)
environment disturbance, H (\ ) is a nonlinear function
where E * is the bound of the approximation error.
related to \ , it can be approximated as
S ( Z ) is the radial basis function vector of neural
3
H (\ ) D\  E\ (2) networks, can be expressed as
 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041 37

ª (Z  P i )T (Z  P i ) º where k1 ! 0 is a design parameter.

S (Z ) exp« » (9)
¬ 2K i2 ¼
Take the equation (13) and (15) into equation (12), and
where Pi is center valueˈKi is the width of the Gauss have

function. z1 z 2  y d  k1 s1  a 20 (16)

Considering the ship unknown time-varying Define compensated tracking error as
environment disturbance of the system of nonlinear ship
course control, the control goal of this paper is using the z2 z 2  s2 (17)
instruction filtering technique and backstepping method
The virtual control law is chosen as
to design a neural network adaptive course tracking
controller and make the actual output to track the desired a20  k1 z1  s2  y d (18)
course direction.
Take the equation (17) and (18) into equation (16), and
have
3. Controller Design and Stability Analysis z1 k1 z1  z 2 (19)
In this part, aiming at the nonlinear ship course control
Choose pre-selection Lyapunov function as
(1), the external environment disturbance will be
considered. Command filter technology and 1 2
V1 z1 (20)
backstepping are used to design course tracking 2
controller. The stability analysis of the closed-loop
Put equation (19) into derivation of (20)
control system and the design process are given below.
2
Step 1˖Define tracking error as V1 z1 z1 k1 z1  z1 z 2 (21)

Step 2˖Put equation(13) and (4) into derivation of

z1 y  yd (10)
(20) and get
where y is the actual output of the system, yd is K
expected reference course.
z2 f ( x2 )  u  d  a 2  s2 (22)
T
Define compensated tracking error as Choose
z1 z1  s1 (11) s2 k 2 s 2 (23)
Get the derivative of the equation (11) where k 2 ! 0 is a designed parameter.
z1 x2  y d  s1 (12) Take equation(23)into equation(22), and have

Define tracking error K

z2 f ( x2 )  u  d  a2  k2 s2 (24)
z2 x2  a2 T
(13)
Choose Lyapunov function as
Let a20 get through the first order low pass filter with the
1 2
time constant e2 ! 0 , obtain filter virtual controller a2 V2 V1  z2 (25)
2
e2a2  a2 a20 (14) Take equation(21) and(24) into the derivation of
equation(25), and have
where a20 is the input of virtue control of the system.
V2 V1  z2 z2
Choose
2 ª K º
s1 k1s1  a2  a20 (15) k1z1  z1z2  z2 « f (x2 )  u  d  a2  k2s2 » (26)
¬ T ¼
38 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041

where f ( x2 ) is an unknown nonlinear function of the z2

k1 z1  k 2 z 2  z 2 E ( Z )  z 2 dˆ tanh(
2 2
V2 a )
system, it cannot be directly used for the controller H
design. f ( x2 ) has to be approximated by using RBF
neural networks. d is completely unknown external k3 ~ T ˆ ~ z k ~ ~
 z2 d  T T  z 2 d tanh( 2 )  4 d (d  d )
disturbance, and the adaptive law is designed to adapt to r H U
the upper bound of disturbance.
(33)
*T
f ( x2 ) T S ( Z )  E ( Z ) (27)
According to Young's inequality, we get
where Z x 2 is input of RBF neural networks, 1 2 1 *2
z2 E ( Z ) d z2  E (34)
S( Z ) is RBF neural networks vector, E (Z ) is 2 2
approximation error. According to equation˄26˅and k3 ~T ˆ k ~ ~ k T k ~ k
˄27˅, design control law and adaptive law as
T T d  3 T TT  3 T * T * d  3 T TTˆ  3 T M2
r 2r 2r 2r 2r
(35)
Tª z º
u  k2 z2  a2  dˆ tanh( 2 )  z1  TˆT S (Z )» (28)
«
K¬ H ¼ ~ 1~ 1
dd d d 2  d 2 (36)
2 2
Now it is time to conduct the update laws for d̂ and Tˆ .
where T M is the bound of ideal estimation T * .

Tˆ rz 2 S ( Z )  k 3Tˆ (29)
According to equation(34)-(36) and (6) ˈ

z2 equation(33)can be written as

dˆ Uz 2 tanh( )  k 4 dˆ (30)
H 2 § 1· 2 1 2
V2 a d k1 z1  ¨ k 2  ¸ z 2  E *
~ © 2¹ 2
where T T *  Tˆ is error vector of estimation weightˈ
~
Tˆ is ideal estimation T * is estimation, d d  dˆ ˈ ª z º k ~ k
 « z 2  z 2 tanh( 2 )» d  3 T 2  3 T M2
¬ H ¼ 2r 2r
d̂ is estimation of d ˈ k 3 ! 0, k 4 ! 0 are designed
parameters. k4 ~ 2 k4 2
 d  d (37)
According to (26)ˈconstruct the augmented Lyapunov 2U 2U
function as The characteristic of hyperbolic tangent function: For
1 ~T ~ 1 ~ 2 H ! 0 ˈ a  R , have
V2 a V2  T T  d (31)
2r 2U a
0 d a  a tanh( ) d EH (38)
where r ! 0, U ! 0 are designed parameters. H

Get the derivation of the augmented Lyapunov function where a constant E meeting E e  ( E 1) is E 0.2785 .

1 ~ ~ 1 ~ ~ Considering equation (38)ˈequation (37) is turned

V2 a V2  T T T  d d
r U into
2 § 1· 2 k ~ ~ k ~ 1 2
2 ª K º V2 a d  k1 z1  ¨ k 2  ¸ z 2  3 T T T  4 d 2  E *
k1 z1  z1 z 2  z 2 « f ( x2 )  u  d  a 2  k 2 s 2 » © 2¹ 2r 2U 2
¬ T ¼
k k
1 ~  1 ~  0.2785Hd  3 T M2  4 d 2
 T T Tˆ  d dˆ (32) 2r 2U
r U
Take equation (38)-(30) into equation(32), and obtain d PV2 a  C (39)

­ § 1· ½ 1 2
where P min ®2k1 , 2¨ k 2  ¸ˈk 3 , k 4 ¾, C˙ E *
¯ © 2 ¹ ¿ 2
 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041 39

k3 2 k 4 2 initial weight is Tˆ 0 . The controller designed

 0.2785Hd  TM  d , and the designed
2r 2U parameters are chosen as k1 1.82 ˈ k2 3.21 ˈ k3 2 ˈ
parameters k1 ˈ k 2 ˈ k 3 ˈ k 4 meet the following k4 2ˈr 5 ˈ U 3.495 ˈ H 0.01 .
conditions.

­k1 ! 0 30 yd
° y
°°k 2 ! 1

heading (deg)
20
® 2 (40)
°k 3 ! 0
° 10
°¯k 4 ! 0
0
Obtain from equation(36)
-10
C C 0 200 400 600
0 d V2 a d  [V2 a (0)  ]e  Pt (41)
P P t(s)
Figure 1: Course of ship
C
So thatˈwhen t o f , V2 a converges to ˈ all the In figures of the simulationˈFig. 1 is the course of the
P
ship, we can see from the figure that the design of the
signals in the closed-loop system are bounded(Li J. F,
ship course tracking controller has a faster response
2013). If properly select the design parameters k1 ˈ k 2 ˈ
speed and better tracking ability; Fig. 2 is the course
k3 ˈ k 4 , r , U , e2 , it Can make the course tracking error tracking error of ship. It can be seen that the error is
arbitrarily small. almost zero in the tracking process which indicates that
the keeping effect is better.

4. Simulation Study 2

In order to verify the effectiveness of the designed 0

tracking error (deg)

controller, take Dalian Maritime University ocean -2

training "Yulong" ship as an example for simulation.
-4
The ship parameters as below ˖ Length between
perpendiculars (L) is 126m, Breadth (moulded) (B) is -6
20.8m, load draught is 8.0m, square coefficient is 0.681ˈ -8
speed is 7.72m/s. By these parameters, the mathematical
-10
model parameters of ship nonlinear motion can be 0 200 400 600
t(s)
calculated K 0.478 ˈ T 216 , D 1 ǃ
Figure 2: Course-keeping error of ship
E 30 .When selecting the desired course signal; a
mathematical model representing the actual performance 0
requirements is selected. (Li and Li, 2013)
rudder angle (deg)

\m (t )  0.19\ m (t )  0.015\ m (t ) 0.015\ r (t ) (42) -10

where\ m is ideal system performance of ship course, -20

\ r is input command signal.

-30
In the simulation, Initial conditions are selected as
x1 10$ ˈ x2 0 , we choose Neural Networks -40
0 200 400 600
T S ( Z ) containing 25 nodes, where center Pl evenly
*T t(s)
Figure 3: Rudder angular of ship
distributed in > 10, 10@ ˈwidth is Kl 4(l 1, l1 ) ˈ
40 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041

0.2 Weihai Science and Technology Development

Program titled “Development of Mariners E-learning
0.15 and CBT System Based on Maritime Satellite
Communication”.
••
•
•›•
•E

0.1

0.05 References
Dong, W., Farrell, J. A., Polycarpou, M. M., et al. Command
0 filtered adaptive backstepping, Control Systems Technology,
0 200 400 600
t(s) IEEE Transactions on, 20(3): 566-580.
Figure 4: Norm of Neural Networks weight vector Farrell, J. A. (2009), Polycarpou, M., Sharma, M., et al.
Command filtered backstepping, Automatic Control, IEEE
Fig. 3 shows that the rudder angle of the ship is
Transactions on, 54(6): 1391-1395.
reasonable; the trajectory of estimation of the norm of
NN ideal weight vector is shown in Fig. 4. The Guan, W., Zhang, X.K., Wang, X.P.(2009), Non - linear
robust control of ship course based on integral backstepping,
simulation results verify the effectiveness of the design
China navigation, 32(3):66-70..
of the course tracking controller based on the
Brooks, X. L., Yang, Y.S.(1999), Ship motion mathematical
instruction filtering technique.
model: modeling mechanism modeling and identification,
Dalian: Dalian Maritime University Press.

5. Conclusions Li, Y., Bai, X. E., Xiao, Y. J. (2014), Ship heading sliding
control based on new extended disturbance observer, Journal of
A ship course-keeping controller is designed based shanghai jiaotong university, 48 (12):1708-1713.
on the commander filtering technique and the
Liu, Q., Wu, X. H., Zou, Z. J. (2000), Ship course fuzzy
backstepping method using neural network to
control and simulation. Ship and Ocean Engineering, 1(1):6-9.
approximate the nonlinear ship course control system.
Liu, W., Sui, Q., Xiao, H., et al.(2011), Sliding backstepping
The adaptive law is designed for the complete
control for ship course with nonlinear disturbance observer,
unknown time-varying disturbance of the ship. The
Journal of Information and Computational Science, 8(16):
method has the advantages of using instruction 3809-3817.
filtering technology to avoid the explosion "problem"
Li, J. F., Li, T. S. (2013), Direct adaptive neural network
in the traditional backstepping method; Neural
tracking control considering input saturation, Journal of
Network is utilized to approximate nonlinear terms in Applied Sciences, 31(3):294-302.
the ship model. It overcomes the problem of singular
Tian, B. J., Liu, Z. J. (2015), Ship course robust Ȝ regulation
value of controller. Finally, simulation is carried out control. China Navigation, 38(2):21-24.
by MATLAB, and the effectiveness of the proposed
Brooks, Fossen, T. I. (1994). Guidance and Control of
control algorithm is verified.
Ocean Vehicles, New York: John Wiley and Sons.

Wang, X. C., Jiang, X. H., Zhang, J. (2002) Nonlinear ship

Submitted : Feb. 15, 2017 course keeping backstepping controller design, Control
Accepted : 25 May, 2017 Engineering, 9(5):63-65.

Wang, X., Liu, Z. J., Li, T. S. (2016), et al.. Adaptive neural

network control for nonlinear ship-based nonlinear systems,
Journal of Harbin Engineering University, 37(1): 123-131.
Acknowledgements
Wu, X., Gao, S., Zhu, Q. D.(2009), Design of ship heading
The authors would like to acknowledge the support control system base on backstepping and sliding, Control
from the Natural Science Foundation of Shandong Engineering, 1(S1):51-55.
Jiaotong University [NO.Z201401], [NO.Z201631] Xia, G. Q., Luan, T. T. (2015), Study of Ship Heading Control
[NO.Z201515]. This research was also a part of the using RBF Neural Network, International Journal of Control
 ZHANG, JIANG, HU, PAN / International Journal of e-Navigation and Maritime Economy 7 (2017) 034–041 41

and Automation, 8(10): 227-236.

Xia, H. S. (1993), Ship course keeping PID control, Journal of

Fuzhou University. 5 (1):216-221.

Yang, Y. S., YU, X. L., Jia, X. L.(1999), Ship course robust

PID auto-pilot design, Journal of Dalian Maritime University,
23(4):11-15.

Brooks, Zhang, X. K. (2012), Ship motion concise robust

control, Beijing: Science Press, China.

Zhang, X.K., Zhang, G. Q. and CHEN, X. J. (2015), A kind of

linear reduction of Backstepping algorithm based on nonlinear
feedback, Control and Decision, 30(9): 1641-1645.

There is no conflict of interests for all authors.