ARTICLE IN PRESS

Control Engineering Practice 15 (2007) 715725

www.elsevier.com/locate/conengprac

Neural network augmented identication of underwater vehicle models

Pepijn W.J. van de Vena,, Tor A. Johansenb, Asgeir J. Srensenc,
Colin Flanagana, Daniel Toala
a
Department of Electronic and Computer Engineering, University of Limerick, Limerick, Ireland
Department of Engineering Cybernetics, Norwegian University of Science and Technology, Trondheim, Norway
c
Department of Marine Technology, Norwegian University of Science and Technology, Trondheim, Norway

Received 22 September 2004; accepted 16 November 2005

Available online 4 January 2006

Abstract
In this article the use of neural networks in the identication of models for underwater vehicles is discussed. Rather than using a neural
network in parallel with the known model to account for unmodelled phenomena in a model wide fashion, knowledge regarding the
various parts of the model is used to apply neural networks for those parts of the model that are most uncertain. As an example, the
damping of an underwater vehicle is identied using neural networks. The performance of the neural network based model is
demonstrated in simulations using the neural networks in a feed forward controller. The advantages of online learning are shown in case
of noise impaired measurements and changing dynamics due to a change in toolskid.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Autonomous vehicles; Backpropagation; Marine systems; Neural networks; Nonlinear systems; System identication

1. Introduction
In recent years highly sophisticated nonlinear control
schemes for marine vehicles have been developed and
implemented. Although modelling of marine vehicles is
widely addressed, several parameters still pose uncertainties. This is due to the absence of accurate models to
describe the highly dynamic nature of these hydrodynamic
parameters. Of prime importance in this context is the
dependence of many hydrodynamic parameters and
coefcients on varying velocity regimes, proximity to the
sea bed, sea surface and other structures, just to mention a
few. At present, models are normally only valid for a
limited region of operational conditions (Fossen, 2002).
Certain model parameters can be determined analytically. Other parameters, however, will need to be determined using numerical methods or identied using (scaled)
Corresponding author. Tel.: +353 61 234230; fax: +353 61 202572.

E-mail addresses: pepijn.vandeven@ul.ie (P.W.J. van de Ven),

tor.arne.johansen@itk.ntnu.no (T.A. Johansen),
asgeir.sorensen@marin.ntnu.no (A.J. Srensen), colin.anagan@ul.ie
(C. Flanagan), daniel.toal@ul.ie (D. Toal).
0967-0661/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2005.11.004

model or full scale tests. Both methods can be time

consuming and expensive. In numerical calculations using
dedicated hydrodynamic software (Faltinsen, 1990), the
vehicle may be divided up into small sections and two
dimensional added mass contributions are calculated for
those sections. Consecutively, an integration over the
whole body yields the three dimensional added mass
parameters. In order to apply this method, which is called
strip theory, the user is required to provide a detailed
description of the vehicle in the form of a CAD drawing.
This part of the modelling process alone can take up
considerable time and, moreover, slender body theory must
be assumed.
For bluff bodies other methods must be used. The added
mass parameters can be measured using, e.g., a towing tank
or free decay tests (Ross, Fossen, & Johansen, 2004). Up to
date there are, to the authors knowledge, no methods
available to perform those tests for all coupled six degrees
of freedom simultaneously.
Additionally, it should be kept in mind that no means of
online updating of parameters is available from either
method. This possibly even affects the analytically derived
values of parameters that are normally assumed to be

ARTICLE IN PRESS
716

P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

constant, as the physical characteristics of the craft may

change from mission to mission. As an illustration of this,
one might think of a remotely operated vehicle (ROV) or
an autonomous underwater vehicle (AUV) whose dynamics have been identied. However, due to a change in
toolskid both mass and geometrical characteristics of the
vehicle will change, thus changing the crafts (hydro)dynamic behaviour.
To overcome the above mentioned problems, neural
networks can be used as they offer a means of parameter
identication without the necessity of detailed model
knowledge and with the possibility of online updates. As
a result they can identify the parameters of interest over the
full region of operation and can adapt for changing
circumstances. Neural networks have several advantages
over other nonlinear identication methods. They can
represent a function with high accuracy in a smooth
fashion without the necessity of extensive amounts of
memory. The latter ability to compress data drastically
requires long training times, but in ofine training this is
generally not an issue. Model evaluation, on the contrary,
is fast. Furthermore, neural networks have an aptitude for
dealing with noisy training signals while still obtaining
accurate results. The application of neural networks for
control and modelling (Narendra & Parthasarathy, 1990)
has been given considerable attention in recent years. In
Van de Ven, Flanagan, and Toal (2003) three classes of the
application of neural networks to the task of modelling and
control of underwater vehicles were identied: (i) combined
control and learning, CCL, (ii) separate control and learning,
SCL and (iii) augmented control, AC. In the rst class,
CCL, a neural network is used in series with the craft.
While controlling the craft it learns to do so better and
better. Examples of this approach can be found in
Akkizidis and Roberts (1998), Farrell, Goldenthal, and
Govindarajan (1990), Guo, Chiu, and Wang (1995), Kim
and Yuh (2001), Labonte (2002), Seube (1991), Venugopal,
Sudhakar, and Pandya (1992), Wang, Lee, and Yuh
(1999a, 1999b), Wang, Lee, and Yuh (2000), Wang and
Lee (2002), Yuh (1990), Yuh and Lakshmi (1993) and Yuh
(1994). In SCL, neural networks are trained outside the
loop. Once satisfactory performance has been obtained
they are used in the control loop, either as a process model
or directly as a controller. The application of SCL is
described in Fujii and Ura (1990), Ishii, Ura, and Fujii
(1994), Ishii, Fujii, and Ura (1995), Ishii and Ura (2000)
and Ura, Fujii, Nose, and Kuroda (1990). The third group,
AC, contains all strategies that apply neural networks to
augment the performance of conventional methods in some
way. Examples of this strategy can be found in Campa,
Sharma, Calise, and Innocenti (2000), Kodogiannis,
Lisboa, and Lucas (1996), Li, Lee, and Lee (2002), Pollini,
Innocenti, and Nasuti (1997) and Yamamoto (1995).
All three approaches have their vices and virtues. For
CCL the fact that learning is performed with the latest
data, and thus circumstances, is a clear advantage.
However, nding an inverse model of the craft, which is

necessary in this strategy, might be hard or even

impossible. This becomes even more of an issue when
considering that control and identication both take place
in the same loop. The time for the identication process is
thus limited by the time between control commands. In
SCL this problem is circumvented by identifying the
necessary parameters outside the control loop. An initial
(possibly conventional) controller, or a previously synthesised controller, can be used during this learning process,
thus allowing longer time for learning. This, however, often
leads to a more complicated, and thus more expensive
architecture. As both these strategies exclusively make use
of neural networks, stability is a major issue. This stability
issue can be alleviated by applying AC. In AC neural
networks are used to improve the performance of a
controller, whereas minimum, but stable, control performance is guaranteed by a conventional linear or nonlinear
feedback controller.
No matter what strategy is used, in the literature
describing CCL, SCL or AC architectures, the neural
networks are always used in parallel with a known part of
the craft dynamics or controller (Campa et al., 2000; Li et
al., 2002; Pollini et al., 1997; Yamamoto, 1995). Or, if no
knowledge is assumed regarding model or controller, the
network is used to represent the whole model or controller
(Comoglio & Pandya, 1992; Ishii & Ura, 2000; Kodogiannis et al., 1996; Venugopal et al., 1992). Rather than using a
neural network in such an overall approximating fashion,
in this article the use of neural networks to model specic
parameters in the model is proposed. Although this
approach is not new, see e.g. Psichogios and Ungar
(1992) and Thompson and Kramer (1994), little attention
has previously been paid to it in the underwater vehicle
literature.
Before delving into the identication of the underwater
vehicle dynamics using neural networks, in Section 2 a brief
overview of marine craft dynamics is given. The (hydro)
dynamical model is introduced and the parameters
playing a role are discussed. Then, in Section 3, the use
of neural networks for identication is proposed. It is
argued why they should not be used in an overall
approximating fashion, but should rather be used to
approximate certain parameters in the model. To illustrate
this divide and conquer approach, a method to identify
the hydrodynamic damping with neural networks is
presented. With the presented identication method
simulations are performed in Section 4. The resulting
parameters are compared to parameters identied using
a least squares identication method in Section 5. Finally,
in Section 6 the ndings are summarised and conclusions
are drawn.
2. ROV kinematics, dynamics and hydrodynamics
In Fossen (2002) and Srensen and Ronss (2000) it was
shown that the nonlinear dynamic equations of motion of a
marine vehicle in six degrees of freedom can be expressed in

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

vector notation as
M_m Cmm Dmm gg s,

(1)

with the kinematic equation

g_ Jgm.

(2)

The matrix Jg converts velocity in a body xed frame, m,

to velocity in an earth xed frame, g_ . The interpretation of
the used symbols is as follows:
g
m
m_
M
Cm
Dm
gg

position and orientation of the vehicle in the

Earth-xed frame
linear and angular velocity of the vehicle in the
body-xed frame
linear and angular acceleration of the vehicle in the
body-xed frame
inertia matrix including added mass
matrix consisting of Coriolis and centripetal terms
matrix consisting of damping or drag terms
vector of restoring forces and moments due to
gravity and buoyancy
vector of control forces and moments

717

Accurate calculation of these phenomena is difcult.

Hence, often the damping is approximated by a diagonal
matrix containing the linear and quadratic damping terms
according to
Dm  diagfX u ; Y v ; Z w ; K p ; M q ; N r g
 diagfX ujuj juj; Y vjvj jvj; Zwjwj jwj; K pjpj jpj,
M qjqj jqj; N rjrj jrjg,

where X u , Y v , Z w , K p , M q and N r are the linear terms, and

X ujuj , Y vjvj , Zwjwj , K pjpj , M qjqj , N rjrj the quadratic terms of
the damping in six degrees of freedom. Although (7) is
a good approximation for decoupled motion, for manoeuvres involving movements along and about several
body axes at a time, such simple models might prove to be
insufcient.
3. System identication using articial neural networks

where u; v; w; p; q and r are the crafts velocity components in six degrees of freedom in a body-xed frame and uc
and vc are the velocity components of the surrounding
water in a horizontal plane. For brevity, and without loss
of generality, in this article the current velocity is assumed
to be zero. A detailed derivation of the nonlinear equations
of motion can be found in Fossen (2002). Below a small
summary of the model is given.
In the matrix M two inertial components are accounted
for,

In this paper neural networks will be used to identify the

nonlinear and coupled damping in an underwater vehicle.
Neural networks can be applied both as control plant
models and as controllers. A clear advantage of using a
neural network is that strong assumptions regarding the
damping, as discussed in Section 2, are not made. Higher
order terms and coupling between various degrees of
freedom can be taken into account by the neural network.
Normally the neural network is used in parallel with
conventional models or controllers in a switching or
output-blending fashion. In this case, however, the neural
network makes no, or only partial, use of the available a
priori knowledge. Due to looking at the neural network as
some nonlinear mapping between the plants input data
and output data, knowledge regarding the dependence
between parameters is lost. This might lead to an
unnecessarily complicated function to be learned by the
neural network. To illustrate this the process dynamics of
Eq. (1) are expressed in state-space form. First Eq. (1) is reordered:

M MRB MA .

m_ M1 s  Cmm  Dmm  gg.

In (1) it is assumed that no water current is present.

Introducing the latter with velocity m c results in the added
mass contribution to the Coriolis matrix and the damping
matrix to be a function of the relative velocity, m r , dened as
mr m  mc u v w p q rT  uc vc 0 0 0 0T ,

(3)

(4)

The rigid body inertial matrix, MRB , represents the mass

and inertia terms. Added mass is accounted for by the
matrix MA .
For the matrix Cm, a similar discourse can be held.
Both the Coriolis and the centripetal forces are functions of
the rigid body mass and added mass and the velocity, m:
Cm CRB m CA m.

(5)

CRB m accounts for the rigid body while CA m accounts for

the added mass.
In the damping matrix, Dm, four terms are combined:
Dm DP DS m DW DM m,

(6)

where DP is the potential damping (relevant when

operating in the wave zone), DS m the linear and quadratic
skin friction, DW the wave drift damping and DM m the
damping due to vortex shedding.

(8)

Taking the state vector to be m and the inputs to the system

as st and gt, (8) can be written in state-space form:
m_ Umt; st; gt,
y Wmt,

(9)

with Umt; st; gt the right hand side of (8) and Wmt
simply mt.
Fig. 1 shows the corresponding block diagram. Assuming that one has partial knowledge regarding the function
U, a neural network can be used to model the unknown
part of the system in parallel with the known part of the
system (in an all-in-one fashion) as depicted in Fig. 2. In
^ where UM
this approach one assumes U UM U
^ to the unknown
corresponds to the known part of U and U
part approximated by the neural network. If the same
assumption is made for the matrices M1 , Cm, Dm and

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

718

.


NN2

M-1

C+D
+

NN1
-

-1

MM

+ .


CM+DM
+

g()


+
gM ()
+

Fig. 1. State-space representation of the nonlinear dynamic equations of

motion.

NN





Fig. 3. State-space representation of the nonlinear dynamic equations

with several neural networks to model unknown parameters.

.


Fig. 2. State-space representation of the nonlinear dynamic equations

with neural network in parallel, to model unknown parameters.

gg, (8) can be written as

 Dm ppm m ppm  ggppm ,

M
1

^
^ fs  CM mm  Cmm
M
^
 gM g  g^ gg.
 DM mm  Dmm

m_ cpm M1 s  Cmcpm m cpm  ggcpm .

10

^ These terms will be

The last three lines of (10) represent U.
estimated by the neural network.
The proposed method in which various model parameters are identied with their own neural network is
shown in Fig. 3. The neural networks in Fig. 3 model
^ D;
^
NN 2 C

other contributors to the damping cannot be estimated

accurately. Therefore, it is assumed that only the damping
matrix Dm is unknown and hence, initially, not accounted
for. In this case the dynamic equation for the reference
system or process plant model (ppm) and the initial
approximate or control plant model (cpm), respectively,
can be written as
m_ ppm M1 s  Cmppm mppm

m_ M1
M fs  CM mm  DM mm  gM gg
^
^
 Dmm
 g^ gg
M1 fCmm

^ 1 ;
NN 1 M

NN3

NN 3 g^ g,

(11)

which is a considerably easier task, demonstrating that the

all-in-one neural network mapping is unnecessarily
complicated.
Another advantage of the approach depicted in Fig. 3 is
that use can be made of known structural properties during
the learning stage. Examples of features are symmetry of
the mass matrix, independence of the rigid body mass
matrix on the velocity, m, and known coupling between
certain degrees of freedom. To benet from the mentioned
advantages, the alternative conguration of the neural
networks, as shown in Fig. 3, might prove worthwhile and
is therefore used in this study.
In most practical cases, all parameters, apart from the
damping parameters, can be calculated with sufcient
accuracy. The contribution of potential damping can be
calculated accurately through the application of potential
theory. However, without towing tank experiments, the

12
(13)

With the ppm, a simulation is performed in which the craft

is controlled by a human operator through joystick
commands. During the simulation one-step-ahead predictions are computed from the cpm. The inuence of the
damping matrix can then be computed as follows. At every
time step one assumes both systems have the same state
vector: mppm m cpm m. Substituting m for mppm and mcpm in
Eqs. (12) and (13), the difference between Eqs. (12) and
(13) becomes the damping matrix Dm, multiplied by the
state vector m and the inverse of the mass matrix, M1 . The
product Dmm can thus be calculated as
Dmm M_mcpm  m_ ppm .

(14)

Eq. (14) will be used for training of the neural networks

and hence is the identication model.
The neural networks are initially trained ofine using the
LevenbergMarquardt algorithm. To prevent problems
during the learning stage, rather than using one neural
network to represent the damping, six neural networks are
used. This is done for two reasons. First of all it should be
noted that the six elements of the vector Dmm do not have
to be of the same order of magnitude. Especially the
elements pertaining to degrees of freedom that cannot be
actuated will in general have a small magnitude. As back
propagation learning relies on feeding back an error signal
which will generally be larger for larger output values, the
speed of learning can drastically change from output

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

element to output element. As a result parts of the neural

network will learn faster than other parts. This again might
result in overtraining of certain parts (i.e. the neural
network might learn to represent the noise in the training
data rather than the actual information). On the other
hand, parts of the neural network will be undertrained,
again resulting in poor performance. For ofine training
the above described problem can be circumvented by
normalising the data. However, for online training there is
no guarantee that this normalisation still holds. The second
reason is closely related to the rst reason: in general, it is
impossible to guarantee the same rate of learning for
different output nodes in a neural network. The straightforward way to tackle the mentioned problems is to apply
six neural networks, with each one output neuron, to the
identication problem.
If improvements are expected the neural networks are
further trained online, using a standard back propagation
algorithm with a momentum term.
4. Simulations
In this section the identication method described in
Section 3 will be applied to an underwater craft in a
simulation study. This open frame underwater craft,
named Tethra, is currently being developed within the
Mobile & Marine Robotics Group of the University of
Limerick (Molnar, Toal, Flanagan, & Hayes, 2004). To
demonstrate the advantages of neural networks for coupled
and nonlinear damping characteristics, the damping matrix
inhibits signicant off-diagonal elements and nonlinearities. In the simulations the following assumptions are
made:




Acceleration and velocity of the vehicle can be

measured.
Only the damping matrix is unknown.

As described in Section 3, training data are derived from a

simulation in which the craft is controlled by a human
operator. Use was made of the Matlab virtual reality
toolbox to provide an interface between human and
computer.
Fig. 4 gives an impression of the virtual-world interface.
Joystick commands are interpreted in the block called
Hand Control Unit. These joystick commands, which
represent excitations in four degrees of freedom, are then
converted to thruster commands in the block Control
Allocator. Time delays and saturation of the thrusters are
described by the block Propulsion System and the

Hand
Control
Unit

Control
Allocator

719

vehicle dynamics are described in the block ROV model.

Interfacing with the Virtual Reality toolbox is provided by
the block Virtual World Interface.
In the Matlab Virtual World environment the University
of Limerick test pool was drawn to obtain similar
circumstances in simulation and future testing. Fig. 5 shows
the vehicle in the pool at the beginning of a simulation.
After gathering the training data, the neural networks
are trained to represent the damping matrix. The neural
networks have 12 input neurons, 5 neurons in one hidden
layer and 1 linear output neuron. The hidden layer neurons
use a hyperbolic tangent activation function. Both the
hidden and the output layers use bias parameters. Rather
than only supplying the six degrees of freedom velocity to
the neural networks, the magnitude of the velocity vector in
six degrees of freedom was supplied as well, resulting in an
input layer width of 12 neurons. In this way, the a priori
knowledge that the damping is a function of the magnitude
of the velocity can be explicitly used in the neural network
identication.
Initially the neural networks are trained ofine with the
LevenbergMarquardt algorithm. This algorithm was
implemented in Matlab. The ofine training, with a batch
of N 18 000 data samples, is stopped after 50 cycles, after
which the mean square error between network output, yk,
and desired output, yd k:
E

N
X

yd k  yk2 ,

(15)

k1

is smaller than 0.1% of the average magnitude of the

desired output for all six degrees of freedom. Generalisation of the neural networks is then tested with a test set
consisting of 4500 samples, which yields comparable mean
square errors.

Fig. 5. The University of Limerick in-house vehicle in the virtual pool at

the start of a simulation.

Propulsion
System

ROV model

Fig. 4. Block diagram of the simulation environment in Matlab.

Virtual
World
Interface

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

720

4.1. Case Study I: drag estimation using noise free signals

D ()

M-1

.


+
+



gM ()

Fig. 6. Model of vehicle with the damping effects modelled by a neural

network.

Identication of the drag under ideal circumstances, i.e.

without sensor noise, is shown in Figs. 7a and b. Fig. 7a
shows the path followed in the horizontal plane while
Fig. 7b shows the depth prole versus time of the dive.
As the two trajectories are on top of each other it can be
concluded that identication of the damping with ideal
measurements is highly accurate. Fig. 7 shows that the
horizontal projection of the trajectory is not a circle. This is
due to the fact that neither in the reference feedforward
controller nor in the neural network feedforward controller, the thruster dynamics have been incorporated.
4.2. Case Study II: drag estimation under noisy conditions
To investigate the ability of the neural network to
perform estimation of the damping using noisy training

After the identication of Dmm in (14) has been

performed, the neural networks can be used in the model
as shown in Fig. 6. The control plant model now becomes
(16)

1
0.5

-0.5
-1
-1.5
-2
-2

-1.5

-1

-0.5

0.5

1.5

(a)

(17)

Eq. (17) is used as the feedforward control law. The

performance of this feedforward controller is evaluated by
comparing the resulting path to a reference path made with
a feedforward controller that does have full knowledge of
the vehicles dynamics. To probe the neural network
representation of the damping for cross-coupling the
followed trajectory incorporates movement in several
degrees of freedom at the same time. A dive from 1.5 m
depth to  3:5 m depth is performed while the craft is
following a circular path in the xy plane.
In the rst computer experiment, Section 4.1, the
damping matrix will be approximated ofine using velocity
and acceleration data assuming zero measurement noise.
Then, in Section 4.2, noise is added to the setup. Online
learning will be used to decrease the state prediction
error. Finally, in Section 4.3, the benecial inuence of
online learning on changing vehicle dynamics will be
demonstrated.

^
where Dmm
indicates that this is the neural network
approximation of Dmm.
It should be noted that, unlike the other blocks, the
^
block with caption Dmm
should not be interpreted as a
multiplication of the inputs and the blocks argument, but
^
rather as a mapping from the input, m, to an output, Dmm.
To test the accuracy of the identication process, the
trained neural networks are used in a feedforward
controller. From Eq. (16) it can be seen that if the
parameters M, C, g and m are known, the neural networks
representing Dmm can be used to calculate a thrust force, s,
such that a wanted acceleration, m_ is obtained:
^
s M_m cpm Cmm gg Dmm.

1.5

-1

-1.5

-2
z [m]

^
m_ cpm M1 s  Cmm  gg  Dmm,

-2.5

-3

-3.5
0
(b)

10

15

20

25

30

35

40

t [s]

Fig. 7. Path followed by ppm (solid line) and cpm (with damping
identied using neural networks) (dotted line): (a) horizontal projection;
(b) depth prole versus time.

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

2
1.5
1
0.5
Y

signals, typical noise is added to the measurements. For the

linear accelerations a zero mean noise sequence with an
amplitude of 0:5 m s2 is added. The angular accelerations
are summed with a noise component of 0:05 rad s2 . The
velocity measurements are assumed to be impeded with a
zero mean noise sequence, the amplitude of which is equal
to 0.1% of the magnitude of the velocity. It should be
noted that in the simulations no noise preltering is
assumed. This would make matters considerably easier
for the neural network. Performance of the neural networks is shown in Figs. 8a and b.
Clearly, the added noise affects the prediction abilities of
the model. The prediction of the right control force, s, can
be improved by continuing the learning process online. The
result of this experiment is shown in Figs. 9a and b,
demonstrating that online learning improves the predic-

721

0
-0.5
-1
-1.5
-2
-2

-1.5

-1

-0.5

0.5

1.5

(a)
-1

1.5
-1.5

-2
z [m]

0.5
0

-2.5

-0.5
-1

-3
-1.5
-3.5

-2
-2

-1.5

0.5

0.5

1.5

(b)

(a)

10

15

20

25

30

35

40

t [s]

Fig. 9. Path followed by vehicle with damping approximated from noisy

measurements and online updating (ppm solid line, cpm dotted line):
(a) horizontal projection; (b) depth prole versus time.

-1

-1.5

tions considerably. However, in this case the neural

network does not necessarily represent the damping any
more. Online learning results in a better local estimate of
the damping as the neural network is now trained with
larger weight on the latest data points. The global
estimation accuracy may suffer from this, unless a
considerable amount of the state space is visited and thus
used in the online learning process.

-2
z [m]

-2.5

-3

-3.5

4.3. Case Study III: drag estimation for time-varying

dynamics

-4
0
(b)

10

15

20

25

30

35

40

t [s]

Fig. 8. Path followed by vehicle with damping approximated from noisy

measurements (ppm solid line, cpm dotted line): (a) horizontal projection;
(b) depth prole versus time.

In this simulation the suitability of neural network based

identication methods for changing toolskids is demonstrated. The AUV from Case Study II is equipped with a
60 kg heavy torpedo shaped sidescan sonar. As a result
several matrices, which were previously assumed to be

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

known, will change drastically. However, changes in the

mass matrix, M (and thus the Coriolis matrix, Cm) and
the restoring forces, gg, are assumed to be known.
Changes in the damping matrix will be assumed to be
unknown and online learning is performed to adapt to
these changes. Figs. 10a and b show the trajectory for the
process plant model and the control plant model using the
neural network to model damping, but without online
learning.
Clearly, the change in damping, which amounts to
parameters changing up to a factor 10, has a detrimental
effect on the performance of the feedforward controller.
Again online learning can be applied to perform an
update of the neural network knowledge regarding the
damping. The trajectories resulting from such an online
update while controlling the craft are shown in Figs. 11a

2
1.5
1
0.5
Y

722

0
-0.5
-1
-1.5
-2
-2

-1.5

-1

-0.5

0.5

1.5

25

30

35

40

(a)
2

-1

1.5
1

-1.5

0.5
-2

-0.5

z [m]

-1

-2.5
-1.5
-2
-3
-2.5
-3
-2

-1

-3.5
0

(a)

(b)
-1

10

15

20
t [s]

Fig. 11. Path followed by vehicle with unknown change in damping and
online learning (ppm solid line, cpm dotted line): (a) horizontal projection;
(b) depth prole versus time.

-1.5

and b. Although the neural network feedforward controller

does not follow the reference path perfectly, a considerable
increase in performance is observed. Again the corresponding changes in the neural network estimate of the damping
will show local improvements while the approximation
might be worse globally. However, while operating the
vehicle further the state space visited will gradually become
a representative set. If online learning is applied the neural
networks will thus learn to represent the new damping
eventually.

z [m]

-2

-2.5

-3

-3.5
0
(b)

10

15

20

25

30

35

40

t [s]

Fig. 10. Path followed by vehicle with damping approximated from noisy
measurements and unknown change in damping (ppm solid line, cpm
dotted line): (a) horizontal projection; (b) depth prole versus time.

5. Comparison with parameters obtained with a classical

model with linear and quadratic terms
To investigate the identication performance obtained
with the proposed identication method further, a classical

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

723

Table 1
Comparison of damping matrices identied with classical model and neural network identication scheme
Model
Dl

116
0
0
0 15
B 0
116 0 15
0
B
B
B 0
0 116 0
0
B
B 0
15
0 26
0
B
B
@ 15 0
0
0
25
0

Dq

0
0
0
0
0

C
C
C
C
C
C
C
C
A

NN

1
112
0
0 8 27 1
B 1
128 0
31
1
1 C
B
C
B
C
B 1
1 115 4 1
0 C
B
C
B 0
12 1 22 2
0 C
C
B
B
C
@ 12 1 3 3
19
0 A

26
1

0
B
B
B
B
B
B
B
B
@

100

100

0
0

0
0

100 0
0 30

0
0

30

0 C
C
C
0 C
C
0 C
C
C
0 A

30

B
B
B
B
B
B
B
B
@

LS

5
0

0
0

0
0

101 0
0 49

0
0

39

0 C
C
C
0 C
C
0 C
C
C
0 A

32

19

20

21

0
0
B
B
B
B
B
B
B
B
@

29
1

102

98

0
0

0
0

101 0
0 36

0
0

25

0 C
C
C
0 C
C
0 C
C
C
0 A

26

Table 2
Errors in estimated damping matrices for NN algorithm and LS algorithm

(18)

and the maximum error of those elements that are zero in

the model matrix, relative to the largest element in the
matrix:
 

Aest i; j
E Zero-elements max abs
,
i;j
maxAModel
AModel i; j 0,

22
1

98

the maximum relative error of those off-diagonal elements

that are non-zero in the model matrix:
 

Aest i; j  AModel i; j
E Off -diag max abs
,
i;j
AModel i; j
iaj; AModel i; ja0,

where diagDq is a diagonal matrix with the six

quadratic damping terms on the main diagonal. With the
data obtained from Eq. (14), a least squares (LS)
approximation of the linear and quadratic damping
matrices is calculated. The same matrices are then extracted
from the neural network by presenting all values for m,
obtained from the simulation, to the neural network. From
m and the neural network output, Dm, a linear and
quadratic damping matrix can be calculated, again using
an LS algorithm. Table 1 presents the thus obtained
matrices.
To compare the accuracy of the two identication
methods, three error functions are dened. The criteria
used are the maximum relative error of the elements on the
main diagonal of the matrices:
 

Aest i; i  AModel i; i
E Diag max abs
,
i
AModel i; i
AModel i; ia0,

103

model with a linear and quadratic damping term is used to

identify the damping parameters.
D Dl diagDq diagjmj,

1
113 1
0 7 17 1
B 1
128 0
18
1
1 C
B
C
B
C
B 1
1 115 4 1
0 C
C
B
B 0
12 1 25 2
0 C
C
B
C
B
@ 12 1 3 2
22
0 A

Dq

Dl

E Diag (%)
E Off -diag (%)
E Zero-elements (%)

NN

LS

NN

LS

12
20
6

24
107
6

20
NA
NA

63
NA
NA

where maxA denotes the value of the largest element in

matrix A. Table 2 lists the errors thus obtained for the
newly presented algorithm and the LS algorithm.
Table 2 shows that the classical model yields less accurate
estimates than the newly proposed identication method
using neural networks even though the same LS algorithm
has been used to distill the damping parameters from the
information contained in the neural networks. Two reasons
can be identied. Firstly, an LS algorithm yields the best
linear unbiased estimate, only if the noise is uncorrelated. In
this case the velocity measurements were impeded with a
noise component proportional to the magnitude of the
velocity. A certain correlation thus exists. Due to generalisation, the neural network may have ltered this noise
component. Secondly, during data collection, a certain
amount of outliers is collected due to the craft being exposed
to boundary conditions, being the walls, oor and surface of
the pool. The sudden and stepwise changes in velocity,
occurring at these boundary surfaces, will affect the classical
model directly, whereas the neural network again suppresses
the inuence of such outliers due to generalisation.
6. Conclusions
In this work, the use of neural networks for
the identication of underwater vehicle dynamics was

ARTICLE IN PRESS
724

P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725

discussed. Neural networks are appealing candidates for

innovative identication methods because of their aptitude
for learning nonlinear mappings from input and output
data and because of the possibility to perform online
updates. Most literature sources describe the use of neural
networks in parallel with a mathematical representation of
the model, or even on its own. Learning thus becomes a
difcult task as, in general, the neural network will have to
represent a complicated nonlinear function. In this work it
is proposed to use several neural networks to identify
different parts of the model separately. The advantages are
that the learning process is signicantly simplied and
knowledge regarding the parameters and the model can
easily be incorporated. This procedure is illustrated in a
computer experiment in which the hydrodynamic damping
is identied. During these simulations, initially data are
gathered while a human operator controls the craft. The
data are then used to train the neural networks to represent
the damping. The accuracy of the identied damping is
tested by using the neural networks in a feedforward
control loop. To test the robustness of the neural networks,
noise and changing dynamics are introduced. Online
learning is used to improve the neural network performance if necessary. The resulting trajectories are compared
to reference trajectories, showing that the damping has
been identied accurately with the neural networks.
Additionally, the newly proposed identication method is
compared to a standard ofine least squares identication
algorithm and found to yield better estimates of the
damping parameters. As for the simulations a model as
proposed by Fossen (1994) was used, which is generally
regarded to be a truthful model, and as noise was
introduced in the simulations, the authors expect the
proposed method to be applicable to real world problems
with noisy measurements. Such experiments are part of
planned future work.
Acknowledgements
This work was supported by the EC Research Directorates through a project at the Marie Curie Training Site,
CyberMar at NTNU (HPMT-CT-2001-00382), and by the
Irish Research Council for Science, Engineering and
Technology: funded by the National Development Plan.
The rst author would also like to express his gratitude
towards the Irish Marine Institute for their nancial
support. The reviewers are thanked for their valuable
comments.
References
Akkizidis, I., & Roberts, G. (1998). Fuzzy modelling and fuzzy-neuro
motion control of an autonomous underwater robot. In 5th international workshop on advanced motion control (pp. 641646).
Campa, G., Sharma, M., Calise, A., & Innocenti, M. (2000). Neural
network augmentation of linear controllers with application to
underwater vehicles. In Proceedings of the 2000 American control
conference, Vol. 1 (pp. 7579).

Comoglio, R., & Pandya, A. (1992). Using a cerebellar model arithmetic

computer (cmac) neural network to control an autonomous underwater vehicle. In , International joint conference on neural networks,
Vol. 2 (pp. 781786).
Faltinsen, O. (1990). Sea loads on ships and offshore structures. Cambridge:
Cambridge University Press.
Farrell, J., Goldenthal, B., & Govindarajan, K. (1990). Connectionist
learning control systems: Submarine depth control. In Proceedings
of the 29th IEEE conference on decision and control, Vol. 4
(pp. 23622367).
Fossen, T. (1994). Guidance and control of underwater vehicles. New York:
Wiley.
Fossen, T. (2002). Marine control systems. Guidance, navigation and
control of ships, rigs and underwater vehicles (1st ed.). AS: Marine
Cybernetics.
Fujii, T., & Ura, T. (1990). Development of motion control system for auv
using neural nets. In Proceedings of the (1990) symposium on
autonomous underwater vehicle technology (pp. 8186).
Guo, J., Chiu, F., & Wang, C.-C. (1995). Adaptive control of an
autonomous underwater vehicle testbed using neural networks. In
Proceedings of IEEE OCEANS 95, Vol. 2 (pp. 10331039).
Ishii, K., Fujii, T., & Ura, T. (1995). An on-line adaptation method in a
neural network based control system for auvs. IEEE Journal of
Oceanic Engineering, 2(3), 221228.
Ishii, K., & Ura, T. (2000). An adaptive neural-net controller system for
an underwater vehicle. Control Engineering Practice, 8, 177184.
Ishii, K., Ura, T., & Fujii, T. (1994). A feedforward neural network for
identication and adaptive control of autonomous underwater
vehicles. In , 1994 IEEE international conference on neural networks,
Vol. 5 (pp. 32163221).
Kim, T., & Yuh, J. (2001). A novel neuro-fuzzy controller for autonomous
underwater vehicles. In Proceedings of the 2001 international conference
on robotics and automation (pp. 23502355).
Kodogiannis, P., Lisboa, G., & Lucas, J. (1996). Neural network
modelling and control for underwater vehicles. Artificial Intelligence
in Engineering, 1.
Labonte, G. (2002). Fast adaptive control of a non-linear system by an
adaline: Motion in a uid. In Proceedings of the 2002 international joint
conference on neural networks, Vol. 2 (pp. 18371841).
Li, J. -H., Lee, P. -M., & Lee, S. -J. (2002). Neural net based nonlinear
adaptive control for autonomous underwater vehicles. In Proceedings
of the 2002 IEEE international conference on robotics and automation,
Vol. 2 (pp. 10751080).
Molnar, L., Toal, D., Flanagan, C., & Hayes, M. (2004). The tethra
autonomous underwater vehicle platform: System description and
controller development. In Proceedings of 2004 ATUVadvances in
technology for underwater vehicles international conference (pp. 8287),
1617 March 2004. London, UK: ExCel.
Narendra, K., & Parthasarathy, K. (1990). Identication and control of
dynamical systems using neural networks. IEEE Transactions on
Neural Networks, 1, 427.
Pollini, L., Innocenti, M., & Nasuti, F. (1997). Robust feedback
linearization with neural network for underwater vehicle control. In
Proceedings of oceans 97. MTS/IEEE, Vol. 11 (pp. 1216).
Psichogios, D. C., & Ungar, L. H. (1992). A hybrid neural networkrst
principles approach to process modeling. AIChE Journal, 38,
14991511.
Ross, A., Fossen, T., & Johansen, T. (2004). Identication of underwater
vehicle hydrodynamic coefcients using free decay tests. In IFAC
conference on control applications in marine systems (pp. 363368) July
2004.
Seube, N. (1991). Neural network learning rules for control: Application
to auv tracking control. In Proceedings of the 1991 IEEE conference on
neural networks for ocean engineering (pp. 185196).
Srensen, A. J., & Ronss, M. (2000). Mathematical modeling of
dynamically positioned and thruster assisted anchored marine vessels.
The ocean engineering handbook. Boca Raton, FL: CRC Press
( pp. 176189).

ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725
Thompson, M. L., & Kramer, M. A. (1994). Modeling chemical processes
using prior knowledge and neural networks. AIChE Journal, 40,
13281340.
Ura, T., Fujii, T., Nose, Y., & Kuroda, Y. (1990). Self-organizing control
system for underwater vehicles. In Proceedings of OCEANS 90 (pp.
7681).
Van de Ven, P., Flanagan, C., & Toal, D. (2003). Articial intelligence for
the control of underwater vehicles. In Smart engineering system design,
Proceedings of the artificial neural networks in engineering conference
(pp. 559564).
Venugopal, K., Sudhakar, R., & Pandya, A. (1992). On-line learning
control of autonomous underwater vehicles using feedforward neural
networks. IEEE Journal of Oceanic Engineering, 17, 308319.
Wang, J., & Lee, C. (2002). Self-adaptive recurrent neuro-fuzzy control
for an autonomous underwater vehicles. In Proceedings of the
2002 IEEE international conference on robotics and automation
(pp. 10951100).
Wang, J., Lee, C., & Yuh, J. (1999a). An on-line self-organizing neurofuzzy control for autonomous underwater vehicles. In Proceedings

725

of the 1999 international conference on robotics and automation

(pp. 24162421).
Wang, J., Lee, C., & Yuh, J. (1999b). Self-adaptive neuro-fuzzy control
with fuzzy basis function network for autonomous underwater
vehicles. In Proceedings of the 1999 IEEE/RSJ international conference
on intelligent robots and systems (pp. 130135).
Wang, J., Lee, C., & Yuh, J. (2000). Self-adaptive neuro-fuzzy systems
with fast parameter learning for autonomous underwater vehicles. In
Proceedings of the 2000 IEEE international conference on robotics and
automation (pp. 38613866).
Yamamoto, I. (1995). Application of neural network to marine vehicle. In
Proceedings of the 1995 IEEE international conference on neural
networks, Vol. 1 (pp. 220224).
Yuh, J. (1990). A neural net controller for underwater robotic vehicles.
IEEE Journal of Oceanic Engineering, 15(3), 161166.
Yuh, J. (1994). Learning control for underwater robotic vehicles. IEEE
Control Systems Magazine, 14(2), 3946.
Yuh, J., & Lakshmi, R. (1993). An intelligent control system for remotely
operated vehicles. IEEE Journal of Oceanic Engineering, 18(1), 5562.