Professional Documents
Culture Documents
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.
ARTICLE IN PRESS
716
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)
(2)
717
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,
M MRB MA .
(3)
(4)
(5)
(6)
(8)
(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 ()
+
NN
.
M
1
^
^ fs CM mm Cmm
M
^
gM g g^ gg.
DM mm Dmm
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)
12
(13)
(14)
ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725
Hand
Control
Unit
Control
Allocator
719
N
X
yd k yk2 ,
(15)
k1
Propulsion
System
ROV model
Virtual
World
Interface
ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725
720
D ()
M-1
.
+
+
gM ()
1
0.5
-0.5
-1
-1.5
-2
-2
-1.5
-1
-0.5
0.5
1.5
(a)
(17)
^
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
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]
-1
-1.5
-2
z [m]
-2.5
-3
-3.5
-4
0
(b)
10
15
20
25
30
35
40
t [s]
ARTICLE IN PRESS
P.W.J. van de Ven et al. / Control Engineering Practice 15 (2007) 715725
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
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.
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)
22
1
98
103
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
ARTICLE IN PRESS
724
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