Transactions on Information and Communications Technologies vol 19 1998 WIT Press, www.witpress.

com, ISSN 1743-3517

Ship's classification by its magnetic signature:

a neuro-genetic approach
J. A. A. Arantes do Amaral & P. L. Botelho
Instituto de Pesquisas da Marinha, RJ, Brazil
N.F.F. Ebecken & A. E. Xavier
COPPE-UFRJ, Ilha do Fundao, RJ, Brazil
L. P. Caloba
Escola de Engenharia-UFRJ, Ilha do Funddo -RJ, Brazil

Abstract
The ship's classification by its magnetic signatures is of great importance in the
development of magnetic sea mines. This work concerns the use of neural
network classification system combined with the relevant features method to
solve this problem. Alternatively we use genetic algorithm techniques to the
train neural network. We compare both approaches in order to find the best
characteristics of each one.

Introduction

Nowadays modern magnetic mines need to have the ability to destroy specific
kinds of targets. From a military point of view it is interesting to have the option
of setting up mines to destroy an specific class of ship when a fleet is passing
by. This paper explores the use of neural network classification system in spite
of the use of traditional classification algorithms. We will study two different
ways of training neural networks, the backpropagation learning rule and genetic
training .

Transactions
on Information
and Communications Signatures
Technologies vol 19 1998 WIT Press, www.witpress.com, ISSN 1743-3517
2 Ship's
Magnetic

Ship's movement generates acoustic signals and alterations in local magnetic

and pressurefields.These alterations can be measured by appropriated sensors.
This paper will evaluate only the alterations of the surrounding magnetic field
measured by magnetometers. Ship's magnetism is classified as either permanent
or induced. Permanent magnetization is a function of constitutive material of the
ship, related to the crystalline structure of the metal and the place and the
process that ship was built ^. Induced magnetization is function of local
geomagnetic field that actuate over the ship. The magnetic signature of a ship is
the vetorial sum of permanent magnetism inherent to the ship and the induced
magnetism as related its current environment. To classify ship by its magnetic
signatures it is necessary to collect signatures from various courses of the same
ship .
3

Data Base Characteristics

We used the data base formed by thirty-two signatures collected from eight
ships that belong to four distinct class. Each class is formed by two ship's that
have the same characteristic dimensions. The signatures were collected through
the passage of the ship over an arrays of magnetometers. For each ship we
collected four signatures referents to the course of approximation north-south
,east-west, south-north and west-east. Ship's magnetic signatures don't have
abrupt variations and besides its period is related to the ship's velocity, the
depth and the distance from the magnetometers.
4

Some Considerations

The major problem of using neural network in military real time systems is the
time consumed by performing the necessary calculations. If the final
architecture has a large number of neurons its utilization will be unfeasible. The
ideal is to have few units of neurons, hence one will get small matrices, which
yield to low time reduced processing. To diminish the length of neural network
,we must reduce the signal size presented to it. The question is how to reduce
the signal size. It can be made by use of Relevant Features Method.
5

The Relevant Features Method

This method evaluates the relevance of each component of the signal in the
discrimination made by neural network. In other words, we tryfigureout which

324

component
of signal
be used
for training
neural
networks
and which
Transactions
on Information
and should
Communications
Technologies
vol 19 the
1998
WIT Press,
www.witpress.com,
ISSN 1743-3517
ones must be discarded. The method can be described in the following steps:
step 1 -Train the neural network with the original signals and then figure out the
best architecture. Save the weights matrix for this best architecture.
step 2 - Calculate the relevance of each component applying by the following
expression^.

|| OUt(xi) - OUt(xi | Xi, j =< Xj '.

>
N
The first parameter of this expression is the output of neural network for each
pattern presented to it. The second term is the same output vector ,when the jcomponent of input signal is replaced by the mean value ( N is the number of
pattern presented)
step 3- Apply this method to all set of signals to find the relevance of each
signal component. After doing so, we can select only the components of higher
relevance and re-train the network. On this way we can reduce the dimension of
the signal without significant losses of discrimination capacity of the network.

The Experiment

The neural network model used is two-layered feedforward network and the
learning rule is backpropagation with momentum term. It is not necessary to use
the full signal length to train the network. Based on Discrete Fourier Transform
(DFT) we select an initial step to collect points [1]. The numbers of points to
give the necessary information of the signal is 20. At the beginning of
experiment it was analyzed several architectures to find the best one, that reach
the highest score of correct classifications. The figure 1 shows that the best
architecture of neural network that gives the 100% of correct classification was
obtained by the use of 12 neurons in the intermediate layer, 4 neurons in the
final layer, learning rate of 0,1 and momentum term equal to 0.

325

Transactions on Information andPercentage

Communications
vol 19 1998 WIT Press, www.witpress.com, ISSN 1743-3517
ofTechnologies
correct classifications

1
0.9

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

8
10
12
14
neurons at second layer

16

18

20

Figure 1
To get this result was not easy. It was necessary to evaluate more than 500
neural network topologies .
The backpropagation learning rule is a gradient descendent method, and has a
problem of being trapped in a local minimum. One way to avoid this problem to
use a training based on genetic algorithms. We will study that at item 10 .
7

Extraction of Relevant Signal Components

Once founding the best architecture the next step is to analyze the relevance of
each signal component. The next four plots allow to verify: i) the curves of
mean square errors obtained by training and test of the best architecture (figure
2, the top plot) ,ii) the percentage of correct classifications (figure 2, the bottom
plot), iii) the relevance of each signal component and the relevance frequency.
Should be observed (figure 3 ) that half of signals components have a low
relevance frequency ( 8 components with relevance frequency of 0.1 and three
components with 0.2 ). The next step is to verify how the network behaves itself
when this low relevant signal components are extracted. First ,the eight
components of relevance 0.1 are extracted and the performance of network is
analyzed. After that, we will do the same way with the two components of
relevance frequency of 0.2.

326

Transactions on Information and Communications

Technologies
vol 19
1998 WIT Press, www.witpress.com, ISSN 1743-3517
Training
and Test
Errors

4
6
8
10
12
Percentage of correct classifications

14

16

x
0.9
0.8
0.7

14
16
12
8
10
epochs
Figure 2. The top plot shows the errors of training and test sets by epoch. The
bottom plot shows the increase of percentage of correct classifications by epoch.

Relevance of each signal component

0.5

0)
C

Hnnn
R n,nn
10
15
Relevance Frequency

20

25

i
0.8

8
6
4
2,
()

I
|
\
0.2

|
at

I
\
0.4

at

relevance

\
0.6

Figure 3. The top plot shows the relevance of each signal element. The bottom
plot shows the number of elements by its relevance ( relevance frequency).
8

Retraining

After reducing the dimension of signal from 20 to 12 it is necessary to re-train

the neural network and to find the best configuration. The followings three plot
show that ,the new best architecture has 6 neurons in intermediate layer, 4
neurons in output layer, learning rate of 0.1 and momentum term of 0.1. The
percentage of correct classification remains 100 % ( figure 4 and 5).

327

Percentage
ofTechnologies
correct classifications
Transactions on Information and
Communications
vol 19 1998 WIT Press, www.witpress.com, ISSN 1743-3517

neurons at intermediate layer

Figure 4

Choosing the stop training point

Figure 5. The top plot shows the errors of training and test sets by epoch. The
bottom plot shows the increase of percentage of correct classifications by epoch.
9

New Signal Reduction

Let's cut off the signal components that have relevance frequency of 0.2. By
doing this ,the signal dimension becomes equal to 9. Redoing the re-training of
neural network and after analyzing several architectures we found the best one,
which has five neurons in the intermediate layer ,four neurons in output layer,
learning rate of 0.1 and momentum term of 0.8.
Now the percentage of correct classifications decreased to 87,5%, ( figure 6 ).

328

correct vol
classifications
Transactions on Information andPercentage
Communicationsof
Technologies
19 1998 WIT Press, www.witpress.com, ISSN 1743-3517

neurons at intermediate layer

Figure 6
The reduction of number of signal components presented to the network (from
20 to 9) brings about a reduction of the network neuron number (from 12 to 5 in
the intermediate layer) as desired.
The price paid for getting the reduction of processing time is the reduction of
correct classification (from 100% to 87,5%). A trade-off must then be found
between the desired processing time and the correct classification expected *.
10. A Neuro-Genetic Approach
Our aim now is to verify the possibility of training the topology studied at item
6 by genetic algorithm techniques. That topology has 12 neurons at first layer
and 4 neurons at second. First of all, we create a population of 30 chromosomes
that represents neural networks and used genetic algorithms techniques to find
the best one. The best chromosome represents the neural network that gets the
best ship's classification results. It's necessary to find the best combination of
genetic parameters such as mutation rate, probability of mutation ( pm ) and
crossover ( pc ) and the number of neurons of each layer that will be crossover .
11 Finding the Best Combination of Parameters
11.1

Crossover Analysis

First of all, we had to choose which parameters not to change initially. We

decided to maintain mutation rate ( 10 %), pc and pm (both equal 1) and find
the number of crossover of each neural network layer.

329

After doing
a study
of the influence
ofvol
crossover
number
at each neural
Transactions
on Information
and Communications
Technologies
19 1998 WIT
Press, www.witpress.com,
ISSN 1743-3517
network layer we found out that the crossover of 2 neurons at second
layer brings better results. So we studied the influence of crossover
number atfirstlayer.
Table 1
Neurons to crossover at
1st. Layer
% Correct Classification
(training set)
% Correct Classifications
(test set)
Ideal generation to stop
training
Mean square error
(test set )
% Correct Classifications
(test set at ideal generation)
Fitness (maximum 256)

100

100

100

100

100

100

50

75

75

56.2

62.5

62.5

22

231

148

88

261

94

1.07

1.19

0.92

1.29

2.2

2.2

93.7

81.2

917

81.2

68.7

56.2

255.9

256.0

252.6

256.0

2554

255.8

Finding the ideal generation to stop training is done by two steps. First we train
the neural network until the 500th generation. After that, we find out the
generation where the mean square error of test set is minimum. So, new training
is done, this time stopping at ideal generation. Compare the 3rd line of the table
with the 6th. It's easy to see that stopping the training at ideal generation brings
better percentage of correct classifications. This table shows us that the best
generation to stop the evolution is the 148th that give us 93.7 % of correct
ship's classification. The parameters that brought this result are pm= 1, pc =1,
mutation rate of 10%.
11.2

Choosing Pm and PC

Once chosen the crossover number of 4 neurons atfirstlayer e three at second

we will analyze the influence of pm and pc. We are dealing with pc and pm
values high, medium and small as following:
1-pm high (1.0) and pc (0.1), medium (0.4) and high (0.8) .
2-pm medium (0.4) and pc low (0.1), medium (0.4) and high (0.8) .
3-pm low (0.1) and pc low (0.1), medium (0.4) and high (0.8) .

330

Table 2vol 19 1998 WIT Press, www.witpress.com, ISSN 1743-3517

Transactions on Information and Communications Technologies
Pm and PC

Pm=1.0,pc=0.1
Pm=1.0,pc=0.4
Pm=1.0,pc=0.8
Pm=0.4,pc=0.1
Pm=0.4,pc=0.4
Pm=0.4,pc=0.8
Pm=0.1,pc=0.1
Pm=0.1,pc=0.4
Pm=0.1,pc-0.8

Training Gen.
and error
170(1.0002)
170(0.2717)
187(0.0096)
181 (0.2500)
190(0.0373)
172 (0.0414)
193(1.0070)
198(0.1649)
200 (0.0005)

Test Gen.
and error.
169(2.56)
194(3.31)
54(1.98)
167(3.08)
121(1.46)
31(2.51)
2(3.06)
178(1.34)
80(1.68)

Fitness %Correct %Correct

Test Set Training Set
56.25
31.25
239.9
87.50
43.75
251.7
100
75.00
255.8
93.75
56.25
251.9
93.75
50.00
255.4
100
50.00
255.3
75
25.00
239.8
100
253.4
68.70
100
56.25
255.9

Table 2 allows us to verify that high values of pm and pc gives us better result.
Doing simulations around these values we finally find 100 % of correct
classifications obtained with pc= pm = 0.9. The mutation rate study done after
that didn't bring us as good results.
The figure 7 allows us to have a clear idea that what is happening. See that
while the fitness is improving the mean square errors due to training set and test
set are decreasing. Notice that the error due to training set is lessen each
generation. It doesn't happen with the error due to test set. It oscillates around
an average value,so it's necessary to stop the training when this oscillation
brings to the smallest value of the mean square error.
12. Conclusions
The both training methods studied got very good results at ship's classification.
The combined utilization of backpropagation learning rule with relevant
features method seems to be quite adequate to this kind of problem. It's mainly
advantage is the signal reduction that gives the possibility of finding small sized
neural network able to classify -correctly the ships. Small sized neural network
will reduce the processing time required by a embedded real time system of a
magnetic sea mine. Finding the ideal topology by this way isn't an easy task.
We examined more than 500 topologies to find the best one. It's was
accomplished automatically by an software that we developed. For details how
we face this problem see the reference *. The well known disadvantage of
backpropagation is the possibility of finding local minimum in spite of the
global one. In order to reduce this problem we studied the genetic training. The
disadvantage of this method is the computational cost involved. The neurogenetic training is very slow by its nature. It's necessary to find a good

331

combination
of mutation
rate, crossover
probabilities
of crossover ISSN
and 1743-3517
Transactions
on Information
and Communications
Technologies rate,
vol 19
1998 WIT Press, www.witpress.com,
mutation and the number of neurons to crossover. It takes a lot of time and
involves a try and error process.
In short, we can say that both approaches are useful to solve this kind of
problem, but combination of both seems to be the ideal. We think that
backpropagation can be used to find the more interesting topologies and genetic
algorithm to avoid to be trapped in local minimum.
Fitness Analysis
260
240
Fitness
220
200
0

50
Train and test

100

150

50

100

150

8
6
Errors 4

2
0

Generations
Figure 7. The top plot shows the improvement of fitness along the training
process (the maximumfitnessallowed is 256 ). The bottom plot shows the mean
square errors at output layer of the neural network due to training set ( the
bottom curve ) and due to test set (the top curve ).
References
[1] J. A. A. Amaral, "Classification of Vessels through their Magnetic
Signatures: A Neuro-Genetic Approach ", Doctor Thesis, COPPE-UFRJ,
1998, Brazil
[2] L. P. Caloba, C. D. P. Poppe, "Identifying Relevant Features for a RealTime Discriminating Process Under High-Event Rate Conditions ",
COPPE/UFRJ/EE 1994, Brazil
[3] H. Souza, Magnetism Reports, IPqM, 1993, Brazil

332