www.elsevier.

com/locate/oceaneng
Authors Accepted Manuscript
Adaptive output feedback control based on DRFNN
for AUV
Li-Jun Zhang, Xue Qi, Yong-Jie Pang
PII: S0029-8018(09)00084-5
DOI: doi:10.1016/j.oceaneng.2009.03.011
Reference: OE1636
To appear in: Ocean Engineering
Received date: 19 February 2009
Accepted date: 29 March 2009
Cite this article as: Li-Jun Zhang, Xue Qi and Yong-Jie Pang, Adaptive out-
put feedback control based on DRFNN for AUV, Ocean Engineering (2009),
doi:10.1016/j.oceaneng.2009.03.011
This is a PDF le of an unedited manuscript that has been accepted for publication. As
a service to our customers we are providing this early version of the manuscript. The
manuscript will undergo copyediting, typesetting, and review of the resulting galley proof
before it is published in its nal citable form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply
to the journal pertain.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Adaptive output feedback control based on DRFNN for AUV
Li-Jun Zhang
a
, Xue Qi
b
, Yong-Jie Pang
c
a,b
College of Automation, Harbin Engineering University, Harbin Heilongjiang 150001,China
c
College of Shipbuilding Engineering, Harbin Engineering University, Harbin Heilongjiang 150001,China
Abstract
The tracking control problem of AUV in six degrees of freedom (DOF) is addressed in this paper. In general, the
velocities of the vehicles are very difficult to be accurately measured, which causes full state feedback scheme is
not feasible. Hence, an adaptive output feedback controller based on dynamic recurrent fuzzy neural network
(DRFNN) is proposed, in which the location information is only needed for controller design. The DRFNN is used
to online estimate the dynamic uncertain nonlinear mapping. Compared to the conventional neural network,
DRFNN can clearly improve the tracking performance of AUV due to its less inputs and stronger memory features.
The restricting condition for the estimation of the external disturbances and network's approximation errors, which
is often given in the existing literatures, is broken in this paper. The stability analysis is given by Lyapunov
theorem. Simulations illustrate the effectiveness of the proposed control scheme.
Keywords: Autonomous underwater vehicles; Adaptive control; Output feedback; Dynamic recrrent fuzzy neural network
Introduction
Due to the highly nonlinear dynamics of marine vehicles and the significant environment
disturbances, it is a great challenge to control marine vehicles, such as ships, submarines, remotely
operated vehicles and autonomous underwater vehicles. In AUV systems, linear translational

A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










velocities such as surge, sway and heave velocities are usually difficult to be obtained from
location measurements. And these measurements are usually corrupted by noises, so velocities
obtained by numerical location differentiation are not feasible.
Some adaptive control algorithms for AUVs tracking and dynamic positioning problems are
proposed in [1-5], where all of the given control algorithms are needed to know both locations and
velocities in 6 DOF, which cause more difficulties in practical operations. In order to solve the
problem of unmeasurable velocities, paper [6] provides an observer-based neural network adaptive
controller, in which the velocities are obtained by state observer. In [7], a full state observer is
designed to estimate the velocities of AUV moving in shallow water area, and an output feedback
controller is provided by using backstepping techniques to complete the tracking mission.
However, backstepping techniques are required to know the accurate modeling of AUV, which is
too much difficult for AUV with nonlinear structure and lots of uncertain parameters. In addition,
backstepping techniques also rise complexity of controllers structure.
In order to deal with the uncertain nonlinear parts in the AUV's dynamics, many researchers
concentrated their interests on the applications of neural networks to the AUV's control problems.
In [8], [9], and [19], a single hidden layered neural network is used to approximate the smooth
uncertainties of the vehicles' dynamics, where inputs of the networks are all the states of AUV.
This probably increases the number of hidden neurons and weighting parameters, and further slow
down the response velocity of the system. DRFNN is a dynamic recurrent multilayered
connectionist network with the memory elements and feedback loop. Compared to the traditional
fuzzy neural network (TFNN), DRFNN not only constructs a set of fuzzy rules and realizes fuzzy
inference, but also adds feedback connections to TFNN. Since a recurrent neuron has an internal
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










feedback loop, it captures the dynamic response of a system, thus the number of inputs of DRFNN
are less than that of TFNN, this can simplify the network structure and speed up the response
velocity. DRFNN has been used to approximate the nonlinearity of the dynamics systems in
[10-12]. In this paper, we will use DRFNN to estimate the unknown parts of AUVs dynamic
modeling.
In practice, network's reconstruction error is not avoided in estimating the unkown nonlinear
part of dynamics, and the external disturbance will always exist, therefore, robustness has become
one of the most important issues in the control problems. Many researchers have studied the
applications of robust control methodology to the motion control of underwater vehicles in
[13-15]. In these literatures, the bounds of uncertainties are assumed to be known in advance. This
is a somewhat strong restricting condition in many practical applications because of the high
nonlinear and unpredictable operating environments of the vehicles. This kind of restrictions will
be relaxed in this paper.
Motivated by the results of the above mentioned papers, an adaptive output feedback controller
based on DRFNN for AUV is proposed to solve the trajectory tracking problem. The proposed
controller does not depend on the accurate modeling of AUV. The structure of the controller is
divided into three parts: dynamic compensator, adaptive control item and robust control item.
Dynamic compensator is used to stabilize the linear part of tracking error dynamics. Adaptive
control item is based on DRFNN which is used to approach the nonlinear dynamics of AUV.
According to the dynamic recurrenct features of DRFNN, the issues of the unmeasurable
velocities of AUV is solved without increasing the number of network nodes, which differs from
conventional neural networks. Robust control item is provided to compensate the unbounded
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










estimated error of DRFNN and external disturbances. Motivated by [20], an observer is designed
to estimate the derivative of tracking error which will be used in the designed adaptive updating
laws. The stability analysis will be given by Lyapunov theorem. In order to demonstrate the
effectiveness of proposed control scheme, certain simulation studies on an AUV are presented.
The tracking performance of the horizontal location, the depth location and the yaw angle are
considered in the simulation studies.
The remainder of this paper is organized as follows: AUV's dynamics equation is derived in
Section 2. In Section 3, a DRFNN-adaptive output feedback controller for trajectory tracking is
presented. Section 4 contains the stability analysis. In order to demonstrate the effectiveness of the
proposed control scheme, certain simulation studies are presented in Section 5. Finally, we make a
brief conclusion on the paper in Section 6.
2. Problem formulation
The mathematical model of an AUV in 6 DOF can be described as
( )
( ) ( ) ( )
=
+ + + + =
=

d
J
M C D g
yout
q q v
v v v v v q t t
q
(1)
where the positive definite inertia matrix = +
RB A
M M M includes the inertia
RB
M of the
vehicle as a rigid body and the added inertia
A
M due to the acceleration of the wave. The matrix
( )
6 6
e C R v groups the Coriolis and centripetal forces, this matrix is skewsymmetrical. The
hydrodynamic damping term
( )
6 6
e D R v takes into account the dissipation of energy due to the
friction exerted by the fluid surrounding AUV. The vector ( )
6
e g R q is the combined
gravitational and buoyancy forces in the body-fixed frame.
d
t is the external disturbances.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










| | =
T
x y z q | u denotes the vehicle location and orientation in the earth-fixed frame.
| | =
T
u v w p q r v is the vector of vehicles velocities expressed in the body-fixed frame.
6
eR t is the input of the system and the vector of the forces and moments on AUV induced by
the thrusters and fins. q is the output of the system.
( ) J q is the kinematic transformation
matrix expressing the transformation from the body-fixed frame to earth-fixed frame.
( )
( )
( )
1
2
0
0
(
=
(

J
J
J
q
q
q
( )
1
+ + (
(
= + +
(
(

c c s c c s s s s c c s
J s c c c s s s c s s s c
s c s c c
u u u | | | u
q u u | u | u |
u u | u |
( )
2
1
0
0
(
(
=
(
(

s t c t
J c s
s c c c
| u | u
q | |
| u | u
where ( ) ( ) sin , cos = = s c and ( ) tan = t .
For a given desired trajectory
d
q which satisfies
| | , ,
T
d d d
q q q is bounded, there exists a
maping
( )
1
2
=
=
d
d
J
q q
q v q
that transforms system (1) into the tracking error system
( ) ( ) ( )
1 2
1
2 1 2 1
1
,

=
= +
=

d
f J M
hout

t t

(2)
where
( ) ( ) ( )
1
1 2 1 1 2
, ( )

= + +

d
f J J q ( ) ( ) ( ) ( )
1 1 1
1 1 2 1 2
( ( ) ( )

+ +
d d
J M C J J q q
( ) ( ) ( ) ( )
1 1
1 2 1 2 1
( ) ( ) )

+ +
d d
D J J g q q
d
q
is a nonlinear uncertain function in which ( ) C v v and ( ) D v v contain velocities interaction in
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










6 DOF.
1
is the output tracking error.
The control objective of this paper is to design controller t only using the output tracking
error
1
such that the solutions of system (2) and the other signals are all uniformly ultimately
bounded (UUB).
3. Controller design
From the tracking error system (2), if ( )
1 2
, f and
d
t are exactly known, the controller
will be designed as follows
( ) ( ) ( ) ( )
1 1
1 1 1 2 1
,

= +
d
MJ u f J M t t
(3)
where
1
u is chosen as the output of a dynamic compensator to stabilize the following system at
point ( ) 0, 0

1 2
2 1
=
=

(4)
The structure of the dynamic compensator is shown in Fig. 1, i.e.
1
1 1
= +
= +

c d c d
d c d
A B
u C D


(5)
where e
c
R , , , ,
d d d d
A B C D are designed parameters such that the closed-loop system

1 2
2 1
1
=
= +
= +

d c d
c d c d
C D
A B



(6)
is asymptotically stable (in [18]).
However, in fact, ( )
1 2
, f is not known exactly. We might want to use a neural network to
approximate ( )
1 2
, f in controller (3) which compensates for its effects. This will be
introduced in following part.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










3.1 DRFNN architecture
In this paper, DRFNN is used to compensate the unknown nonlinear function ( )
1 2
, f online.
A schematic diagram of DRFNNs stucture with n inputs, ( ) 1, , = "
i
m i n term nodes for each
input variable, single output is shown in Fig.2.
Layer 1 accepts input variables and each node for each input variable. The nodes in this layer
only transmit input variables
1
, , ( =

"
T
N N N
n
x x x to the next layer directly.
Layer 2 represents the terms of the respective linguistic variables and is used to calculate
Gaussian membership value, i.e.

2
exp
(
| |
(
= |
|
(
\ .

N
i ij j
i
ij
x x

o
(7)
where
ij
x and
ij
o are the center and width of the Gaussian function respectively.
Nodes at layer 3 represent fuzzy rules. Layer 3 forms the fuzzy rule base. Links before layer 3
represent the preconditions of the rules, and the links after layer 3 represent the consequences of
the rule nodes, i.e.
( ) ( )
1
,
1 =
= +
_
m
a
ja c j
j
s a w k z o (8)
( ) ( )
,
1 =
c j j
k k o o (9)
( ) ( ) ( )
( ) ( ) 0.5
1
1

= =
+
a
s a
k f s a
e

o (10)
At time k , every rule node
( )
a
k o not only contains ( ) ( ) ( ) ( ) ( )
1
1 1
= "
n
i i a N N
n n
z k x k x k , but
also contains recurrent rule nodes at time 1 k .
Layer 4 realizes normalized unit for rules at layer 3, i.e.
1 =
=
_
a
a m
i
i
o
o
o
1, 2, , = " a m (11)
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










where m denotes rule number.
Layer 5 is the output layer. Each node is for actual output to the pumped out this system. The
links between layer 4 and layer 5 are connected by the weighting value
a
N
y , i.e.
( )
1
1
=
=
= =
_
_
m
a
N a
N a
N m
a
a
y
y F x
o
o
(12)
Define the parameter vector of DRFNN as follows
1
(

T
T T T T
N
x w y | o
Since ( )
1 2
, f is a function about
1
and
2
.
2
is not measurable directly, so it can't be
the input vector of DRFNN. However, there exists derivative relationship between
1
and
2
.
For the reason that DRFNN has memory elements and inner feedback loop which can reflect
dynamic relationship between
1
and
2
, choose
1
as the input vector is reasonable.
Due to the approximation capacities of neural networks for nonlinear mappings, we have
following lemma.
Lemma 1 ([16]). Given
*
0 > c , there exists a set of bounded parameter vector | , such that
( )
1 2
, f can be approximated over a compact set O by a DRFNN with the input vector
1
,
i.e.
( ) ( )
1 2 1
, = + f F | c
*
< c c (13)
for all ( )
1 2
, eO
Here, we replace ( )
1 2
, f in (3) with
2
u , i.e.

( )
( )
2 1 1 2

, = = u F f | (14)
where
1

(
=

T
T T T T
N
x w y | o is the estimation of | .
1 6
eR .
With equations (13) and (14), neural network's reconstruction error is written as follows:
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










( )
1 2 2
, f u ( )
( ) 1 1

= + F F | | c
( )
( )
1
2

=
c
= + +
c

D
T
F
| |
|
| | c
|
(15)
where Taylor expansion is used.

| | |
( )
2
2
lim 0

D
t
|
|
Equation (15) will be used in the immediate adaptive control design and stability analysis.
3.2 Robust controller
In literatures [13-15], the neural network's reconstruction error and external disturbances are all
bounded by known constants. This is a strong restricting condition for AUV dynamics, and it will
be broken in following part.
Assumption 1. Assume
( ) ( )
1
1 2 2 1
,

s
d d d
f u J M c t
(16)
with
d
c is an unknown constant and
d
is a known bounded nonnegative function for
( )
1 2
, eO . Clearly, Assumption 1 relaxes the restricting conditions in [13-15], which is very
valuable for the control design of AUV since its complex nonlinearity.
Here, we replace ( )
1
1

d
J M t in (3) with
3
u . The robust controller
3
u satisfies

3
=
d d
u c (17)

d
c is the estimation of
d
c defined in (16).
The controller in (3) is rewritten as
( )( )
1
1 1 2 3

= + MJ u u u t (18)
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










3.3 Observer and adaptive laws design
In order to obtain
2
u and
3
u , the adaptive laws of parameters

| and
d
c will need the error
vector
| |
1 2
T
c
in which
2
is not acquirable directly. So the observer will be first
designed to estimate the state
2
, and then the adaptive laws of

| and
d
c will be derived from
Lyapunov-based method.
Define
| |
1 2

T
c

0 1 0
0
0
(
(
(
(

d d
d d
A D C
B A
| | 0 1 0
T
b
From (2), (4) and (5), the error dynamics can be written as
( ) ( ) ( )
1
1 2 2 3 1
,

= + +

d
A b f u u J M t (19)
Define a new vector
| |
1

T
c
g
where g is measurable. With this definition, the error dynamics can be rewritten as
( ) ( ) ( )
1
1 2 2 3 1
,

= + +
=

d
A b f u u J M
g C
t

(20)
where
1 0 0
0 0 1
(
=
(

C
Clearly ( ) , A C is observable. With the measurable vector g as the input signal, the stucture
of a linear observer is shown as following
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t

= +
=

A Kg
g C

(21)
where = g g g is estimated error, K is a gain matrix and is chosen such that A KC is
asymptotically stable.
Define

A A KC
Then, the observer error dynamics can be written as
( ) ( ) ( )
1
1 2 2 3 1
,

= +

d
E AE b f u u J M t (22)
By (22) , we can obtain the estimate of
2
.
Based on Lyapunov theorem, the adaptive law of the parameter vector for DRFNN can be
designed as follows:
( )
( )
1
0

2
=
( c
= (
c
(

T
F
Pb
| |
| |
|
| | |
|
(23)
in which
0
| may be used to include a best guess of some | , where | is an ideal parameter
vector to estimate ( )
1 2
, f . 0 >
|
,
|

is adaptive gain matrix. P is the positive definite

solution of the Lyapunov equation
+ =
T
A P PA Q (24)
for some , 0 = >
T
P P Q .
In the similar idea,
d
c is updated by the following adaptive law:
( )
0

[ 2 ] =

d d
T
d c d c d d
c Pb c c (25)
Ior some 0 >
d
c
,
0 d
c is a best guess of
d
c , and
d
c is an ideal quantity.
d
c
is an adaptive
gain.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










4. Stability analysis
Define

d d d
c c c
Choose Lyapunov function candidate as

1 1
1 1
2 2

= + + +


d
T T T
d c d
V P P c c
|
| | (26)
where

P is the positive definite solution of the Lyapunov equation

+ =

T
A P PA Q
in which
, 0 = >

T
P P Q
Here we assume that

( )
2
min
> +

Q Pb Pb (27)

2
2 >
d
c d
(28)
where
( )
min

Q is the minimum eigenvalue of matrix

Q .
Theorem 1. For system (1) and (2), we take observer dynamics (21), the controller (5), (14), (17)
and (18), adaptive updated laws (23) and (25), then all the states in the closed-loop system and the
control signals are uniformly ultimately bounded.
Proof. Consider the Lyapunov function candidate
1 1
1 1
2 2

= + + +


d
T T T
d c d
V P P c c
|
| |
The derivative of V along (19) and (22) will be
( )
1 2 2 3

2 ( , = + +

T T T
V Q Q Pb f u u ( )
1
1
)

d
J M t
( ) ( )
1 2
2 ( , +

T
Pb Pb f
( )
1
2 3 1
)

+ +
d
u u J M t
1 1
+

d
T
d c d
c c
|
| |
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










= +

T T
V Q Q
( )
( ) ( )
1
2 1
3 1

2 ( )

=
c
+ + +
c

D
T T
d
F
Pb u J M
| |
|
| | c t
|

( ) ( ) ( ) ( )
1
1 2 2 3 1
2 ,

+ + +

T
d
Pb Pb f u u J M t
1 1
+

d
T
d c d
c c
|
| |
From (23), it follows
( ) ( )
( )
2 1
3 1

= + + +

D
T T T
d
V Q Q Pb u J M | c t

( ) ( ) ( ) ( )
1
1 2 2 3 1
2 ,

+ +

T
d
Pb Pb f u u J M t
( )
1
0

+


d
T
d c d
c c
|
| | |
( ) ( )
2
2
min min
s +

V Q Q
( )
1

2
=
| | c
+ |
|
c
\ .

T T
d d d d
F
Pb c c
| |
|
|
|
( )
2 ( ) + +

T
d d d
Pb Pb c c
( )
1
0

+


d
T
d c d
c c
|
| | |
( ) ( )
2
2
min min
s

V Q Q

2 +
T
d d
Pb c
( )
1

2
=
c
c

T T
F
Pb
| |
|
|
|
2 ( ) +


T
d d
Pb Pb c
( )
1
0

+


d
T
d c d
c c
|
| | | (29)
Put (17) and (25) into (29), we have
( ) ( )
2
2
min min
s +

V Q Q

2
c

T
F
Pb |
|
2 ( ) +


T
d d
Pb Pb c
( )
( )
0 0

d
T
c d d d
c c c
|
| | |
Since
2 2
2 2
2 ( ) + s + +


T
d d d d
Pb Pb c Pb Pb c
( ) ( )
0 0

= +

T T
| |
| | | | | | |
2
2
0
2 2
s +

| |

| | |
( ) ( )
0 0
= +
d d
c d d d c d d d d
c c c c c c c
2
2
2 2
s +
d d
c c
d d do
c c c

we find
( ) ( )
2
2
min min
s +

V Q Q

2
c
+
c

T
F
Pb |
|
2 2 2
2 2
2
+ + +


d d
Pb Pb c
|

|

2 2
2
0 0
2 2 2
+
d d
c c
d d d
c c c
|

| |
Since
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










2
2
2

2
2
c
c c
+ s
c

T
T
F
Pb
F
Pb
|
|

|
| |
|
we have
( ) ( )
2 2
2
min min
[ ] s +

V Q Q Pb Pb
2 2
2
| |
+
|
\ .

d
c
d d
c

2 2
0 0
2 2
+
d
c
d d
c c
|

| |

2
2

2
c
c
+
T
F
Pb
|

s + V (30)
where and are positive constants defined by
( )
( )
( )
( )
( )
2
min
min 2
max max
min{ , , 2 }
+
=

d d
c c d
Q Pb Pb
Q
P P

2
2
2 2
0 0

2
2 2
c
c
= + +
d
T
c
d d
F
Pb
c c
|
|

|
| |

Inequality (30) satisfies

( ) 0 0

| |
s s +
|
\ .
t
V V e

(31)
When t , V is bounded by

. Moreover, if
1
sV , then

1
1
lim

s
t

(32)
That is, ,

,

| and
d
c are uniformly ultimately bounded. ,
5. Simulation analyses
To illusrate that the approach is applicable to the AUV systems, simulation study is carried out
based on omni-directional intelligent navigator [17]. Due to the complicated motion system of
AUV, make the following assumption to simplify the problem: if AUV is moving with a changing
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










yaw angle, its underwater depth will keep in constant; If AUV is moving with a changing depth,
its yaw angle will keep in constant. In this part, we will show the trajectory tracking of surge,
sway, yaw angle and depth respectively. The AUV dynamics is described by equation (1). The
desired velocities and positions are generated by reference trajectory generators. The initial
parameters of DRFNN are random. In the simulation studies, the external disturbance is taken as
10 ( ) =
d
rand t v with ( ) | | 1,1 e rand , and the corresponding restricting functions are chosen
as 1 =
d
. Dynamic compensator parameters are designed as
5, 1, 30, 8 = = = =
d d d d
A B C D
Observer gain matrix
1 1
0.2 5
0.1 1
(
(
=
(
(

K
Other design parameters are selected as
1, 0.02, = =
| |

0
2, 0.25, 10 = = =
d d
c c d
c
In the following figures, solid lines are the trajectory of AUV, and dot lines are the desired
trajectory.
Fig.3 and Fig.4 show the trajectory tracking of surge with DRFNN and with TFNN respectively.
The input variables of TFNN contain the measurable state and the estimated states of the AUV
system. This will let the estimated error into the neural network, and then increase the output error
of the TFNN. So its trajectory tracking performance is worse than that of the trajectory tracking
with DRFNN.
Fig.5 and Fig.6 show the trajectory tracking of sway with DRFNN and tracking error
respectively.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.7 and Fig.8 show the trajectory tracking of depth and yaw angle with DRFNN respectively.
At the beginn ing, the output of the controller is almost provided by dynamic compensator. With
the learning of DRFNN parameters, after a few minutes the output of the controller is almost
provided by DRFNN. From these simulation figures, we can see that excellent tracking of position
is obtained by the controller proposed.
6 Conclusion
For the tracking control of AUV, an adaptive output feedback controller based on DRFNN was
proposed and shown to track the desired trajectory in the earth-fixed frame. DRFNN was chosen
to compensate the nonlinear uncertain part of the AUV dynamics online. Based on the advantage
of DRFNN, the measurable state variable is only chosen as the input of the neural network, it
decreased the output error of the neural network. This was helpful for improving the accuracy of
trajectory tracking. A focus of this paper was taken on an attempt to break the traditional
restricting condition, which is usually added to the AUV's external disturbances. Here, DRFNN's
error and the external disturbances was bounded by a unknown constant, and certain adaptation
scheme for this constant was introduced through Lyapunov-based method. A linear observer was
proposed to estimate the tracking error which was used in updating adaptive and robust laws. The
stability of the controller was analysed by Lyapunov theorem. Simulation results showed that the
output feedback controller performs well with stability and robustness.
References
[1] Jian-Cheng Yu, Ai-Qun Zhang, Xiao-Hui Wang, Li-Juan Su, Direct adaptive control of underwater vehicles
based on fuzzy neural networks, ACTA Automatica Sinica. 33 (2007) 840-846.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










[2] Yoerger D R, Slotine J J E, Robust trajectory control of underwater vehicles J, IEEE Journal of
Oceanic Engineering 10 (1985) 462-470.
[3] Fossen T I, Sagatun S I, Adaptive control of nonlinear systems: a case study of underwater robotic
systems J, Journal of Robotic Systems 8 (1991) 393-412.
[4] Corradini M L, Orlando G, A discrete adaptive variable structure controller for MIMO systems and
its application to an underwater ROV J, IEEE Transactions on Control Systems Technology 5 (1997)
349-359.
[5] Antonelli G, Chiaverini S, Sarkar N, et al, Adaptive control of an autonomous underwater vehicle:
experimental results on ODIN J, IEEE Transactions on Control Systems Technology 9 (2001)
756-765.
[6] Jian-Shu Gao, Zhi-Wei Xing, Hong-Bo Zhang, Observer-based neural network adaptive control of
underwater vehicles, ROBOT 26 (2004) 515-518.
[7] Shu-Yong Liu, Dan-Wei Wang, Eng Kee Poh, Chin Swee Chia, Nonlinear output feedback
controller design for tracking control of ODIN in wave disturbance condition, OCEANS. (2005)
1808-1810.
[8] Ji-Hong Li, Pan Mook Lee, A neural network adaptive controller design for free-pitch-angle diving
behavior of an autonomous underwater vehicle, Robotics and Autonomous System 52 (2005)
132-147.
[9] J.H.Li, P.M.Lee, S.J.Lee, Neural net based nonlinear adaptive control for autonomous underwater
vehicle, Proceedings of the IEEE International Conferenceon Robotics and Automation, Washington
DC, USA, (2002) 1075-1080.
[10]You-Wang Zhang, Direct adaptive dynamic recurrent fuzzy neural network control and its
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










application. Information and Control 37(3), 269-274.
[11]Liu-Ping Mao, Yao-Nan Wang, Wei Sun, Yu-Xin Dai, An adaptive control using recurrent fuzzy
neural network, ACTA Electronica Sinica 34 (2008) 2285-2287.
[12]You-Wang Zhang, Rong-Zhu Wang, Design of indirect adaptive dynamic recurrent fuzzy neural
network controller, J.CENT.SOUTH UNIV.(NATURAL SCIENCE) 35 (2004) 253-257.
[13]P.Kiriazov, E.Kreuzer, F.C.Pinto, Robust feedback stabilization of underwater robotic vehicle, Rob.
Autonom. Syst.21 (1997) 415-423.
[14]P.M.Lee, S.W.Hong, Y.K.Lim, C.M.Lee, B.H.Jeon, J.W.Park, Discrete-time quasi-sliding mode
control of an autonomous underwater vehicle, IEEEJ. Ocean. Eng.24 (1999) 388-395.
[15]A.J.Healey, D.Lienard, Multivariable sliding mode control for autonomous diving and steering of
unmanned underwater vehicles, IEEE J. Ocean. Eng.18 (1993) 327-339.
[16]Naira Hovakimyan, Flavio Nardi, Nakwan Kim, Adaptive output feedback control of uncertain
nonlinear systems using single-hidden-layer neural networks, IEEE Transacyions on neural networks
13 (2002) 1420-1431.
[17]Gianluca Antonelli, Stefano Chiaverini, Nilanjan Sarkar, Michael West, Adaptive control of an
autonomous underwater vehicle: experimental results on ODIN, IEEE TRANSACTIONS ON
CONTROL SYSTEMS TECHNOLOGY 9 (2001) 756-765.
[18]Frederick M Brasch, JR., James B.Pearson, Pole placement using dynamic compensators, IEEE
Trans. Automat.Contr.15 (1970) 34-43.
[19]A.J.Calise, N.Hovakimyan, M.Idan, Adaptive output feedback control of nonlinear systems using
neural networks, Automatica 8 (2001) 1201-1211.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










[20]N.Hovakimyan, F.Nardi, A.Calise, A novel observer based adaptive output feedback approach for
control of uncertain systems, IEEE Trans. Automat. Contr. 147 (2002) 1310-1314.
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.1. Controller architecture
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.2. DRFNN architecture
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.3. Trajectory tracking of surge with DRFNN
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.4. Trajectory tracking of surge with TFNN
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.5. Trajectory tracking of sway with DRFNN
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t











Fig.6. Trajectory tracking error of sway with DRFNN
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t










Fig.7. Trajectory tracking of yaw angle with DRFNN
Figure
A
c
c
e
p
t
e
d

m
a
n
u
s
c
r
i
p
t











Fig.8. Trajectory tracking of depth with DRFNN
Figure