See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/303773062

Neural Passive Sonar Signal Classification Using Independent Component

Analysis

Conference Paper · March 2016

DOI: 10.21528/CBIC2011-17.2

CITATION READS

1 172

3 authors, including:

Eduardo Furtado de Simas Filho

Universidade Federal da Bahia
154 PUBLICATIONS   522 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Simulador de Sinais Acústicos Submarinos View project

Intelligent decision support systems for non-destructive testing applications View project

All content following this page was uploaded by Eduardo Furtado de Simas Filho on 19 July 2019.

The user has requested enhancement of the downloaded file.

 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

NEURAL PASSIVE SONAR SIGNAL CLASSIFICATION USING

INDEPENDENT COMPONENT ANALYSIS

Eduardo F. de Simas Filho

Electrical Engineering Department, Poli/UFBA
[email protected]

Natanael Nunes de Moura, José Manoel de Seixas

Signal Processing Laboratory, COPPE/UFRJ
{natmoura,seixas}@lps.ufrj.br

Abstract – Sonar systems use the underwater sound propagation to detect, identify and locate targets (such as vessels or shoals
of fishes). One of the most important tasks in passive sonar signal processing is target identification, which relies on sonar op-
erators who listen to the acoustic signatures and assign to them a certain class of vessel. When there are acoustic signals from
multiple targets arriving at adjacent directions, target identification becomes a harder task due to cross-channel interference. The
purpose of this work is to develop a neural design support system for target identification, specially in multiple target applications.
In order to improve the discrimination performance when cross-interference is present, independent component analysis (ICA)
was used as a preprocessing step. It is shown that the proposed approach improved considerably the discrimination efficiency in
an experimental two-target problem.

Keywords – Passive Sonar System, Spectral Analysis, Neural Classification, Independent Component Analysis.

1. INTRODUCTION
Sonar systems [1] are used in several applications for military and civilian purposes such as underwater surveillance, target
detection, location and identification, determination of water depth, bottom contours and composition and shoals of fishes lo-
cation. Active sonar consists on emitting acoustic pulses and analyzing the received echoes. Otherwise, passive sonar systems
record and process the acoustic signals emitted by other vessels aiming at target detection and identification.
In order to detect and classify signals against background noise, passive sonar systems [2] listen to the noise radiated by
targets using an array of hydrophones. The background noise may be produced by the sea ambient noise or the self-noise of the
sonar platform. From the acquired signals, the direction of arrival (DOA) is estimated, in order to inform the eventual presence
of a target in a determined direction (bearing). After DOA estimation, relevant features of the target may be extracted from a
given direction. For passive sonar systems, target identification is usually performed by experienced human operators (sonar
operators).
LOFAR (Low Frequency Analysis and Recording) [3] which is a broadband analysis usually employed in passive sonar
systems, estimates the noise vibration of the target machinery. The LOFAR analysis is based on spectral estimation and supports
targets classification.
Depending on the bearing resolution and the geographical positioning of the targets, signal interference may occur for adjacent
directions, contaminating the acquired signals and making even more difficult the target classification task.
As for passive sonar systems target classification is a very important task, some works have been developed in order to
produce decision support systems to help sonar operators in this job. Neural networks have successfully been used in passive
sonar signal classification [4–7]. Additionally, other techniques have been applied to perform the classification task, in [8] a
hidden Markov model with Hausdorff similarity measurement was used to detect and classify targets. Another way to perform
the detection and classification of targets is achieved through the Prony’s method [9].
This work proposes a decision support system for target identification, specially in the multiple target case. For this, indepen-
dent component analysis (ICA) [10] is used to minimize the cross-channel interferences. ICA transformation has been applied
recently as a preprocessing step in classification problems [11–13]. By projecting the multidimensional signals into a new data
space where the statistical dependency among the variables is eliminated ICA may (in some cases) favor the signal classification
task. In our particular problem, ICA contributes to reduce the cross-interference between adjacent bearings, revealing the original
acoustic signatures of the vessels, as demonstrated in previous works [4, 14]. Here ICA is applied after LOFAR analysis as a
preprocessing step for classification, in order to recover the acoustic characteristics of the original sources and thus improve the
discrimination efficiency. For automatic classification, a neural network (multi-layer perceptron architecture) [15] was used.
The paper is organized as it follows. Section 2 describes how the signals were acquired, and provides a brief description of
LOFAR analysis. The proposed methodology is detailed in Section 3. Section 4 shows the results obtained before and after ICA
preprocessing. Finally, Section 5 brings the conclusions.
1
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

2. THE PROBLEM
A passive sonar system is typically used by submarines to perform the surveillance in a given operation area. The beamform-
ing purpose is to estimate the DOA from a given target as captured by the hydrophone array. The Figure 1 shows the beamforming
display for a given experimental setup. The main purpose of the DOA is to estimate the target energy for a particular direction
of interest. The horizontal axis represents the bearing position (−180o to 180o ) and the vertical axis represents time (waterfall
display). In this case, an acquisition window of one second was applied [16].

190 205 076

Figura 1: Bearing time display of a passive sonar system.

When there are multiple targets and signal interference at neighbor bins occurs, as it may be the case for bearings 190◦ and
205◦ (see Figure 1), the original target features may be masked and thus the target classification efficiency decreases. Beyond
that, the self-noise, bearing 076◦ , may interfere on both target signals. Thus, a preprocessing scheme aiming at reducing signal
interferences would be attractive, as it may, facilitate target classification.
2.1 LOFAR Analysis

The LOFAR [2] is a broadband analysis that provides the machinery noise spectra to the sonar operator and goes from DC to
15,625 Hz. The block diagram of the LOFAR analysis is shown in Figure 2.

HANNING SPECTRUM DC
BEARING FFT
WINDOW MODULE REMOVAL

Figura 2: LOFAR analysis block diagram.

After bearing, the signal is convolved to a Hanning window to emphasize the frequency range of interest [17]. Then, the sig-
nal is separated in blocks of 1,024 samples, which are transformed into frequency-domain using a short time Fourier transform.
A spectrum module is implemented and the spectra are normalized using the TPSW algorithm [18]. The normalization was im-
plemented by estimating the background noise that is present at each spectrum and computing a normalized frequency bin using
this estimation as normalization factor. This procedure removes the spectrum bias and equalizes the spectrum amplitude [19].

3 PROPOSED CLASSIFICATION SYSTEM

In passive sonar systems, target identification is usually performed by trained operators that listen to the noise radiated in a
given direction and assign it to a certain class of ship. This procedure is susceptible to human errors and becomes more difficult
when additive noise or multiple targets are present. In this work it is proposed a neural network based classifier to assist sonar
operators by indicating the most probable class of a acoustic signal. As illustrated in Figure 3, considering that two acoustic
signals are arriving at adjacent bearings (190o and 205o ), the received signatures are firstly processed by LOFAR analysis. The
obtained average spectra are used as inputs to an ICA algorithm. The estimated independent components are then used to feed a
neural (multi-layer perceptron) classifier.
Frequency Independent
Bearings contents components
D190 LOFAR D190 D190 Neural Decision
ICA
D205 LOFAR D205 D205 Classifier

Figura 3: Proposed classifier system block diagram.

In the following subsections, the main concepts involved in both ICA and neural classification are described.
2
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

3.1 Independent Component Analysis

The basic ICA model considers that a set of N observed signals x(t) = [x1 (t), ..., xN (t)]T is generated by a linear combina-
tion of unknown sources s(t) = [s1 (t), ..., sN (t)]T :
x(t) = Ws(t) (1)
where W is the N×N mixing matrix [10].
Formulated this way ICA is also referred to as Blind Source Separation (BSS) and its purpose is to estimate the source signals
s(t) using only the observed data x(t). A solution is obtained if one can find the inverse of the mixing matrix B = W−1 and so:

s(t) = Bx(t) (2)

A general principle for estimating the matrix B can be found by considering that the source signals are statistically indepen-
dent. There are many mathematical methods for estimating the coefficients bij . The nonlinear decorrelation and the maximally
nongaussianity are the most applied ones [20].
In this work, JADE algorithm [21] was used for independent components estimation. In a previous work [14], it was demon-
strated that, among popular ICA algorithms (such as FastICA and Akuzawa’s Newton-like method [10]) JADE presents a com-
promise between good separation performance and fast convergence.

3.1.1 JADE algorithm

In JADE (Joint Approximate Diagonalization of Eigenmatrices) algorithm [21], second and fourth order statistics are applied
for independent component estimation through a tensorial approach. Cumulant tensors are matrices containing the cross cumu-
lants [22]. Considering this, the second-order cumulant tensor is the covariance matrix (C) and the fourth-order tensor (T4 ) is
formed by the fourth-order cross cumulants cum(xi , xj , xk , xl ), that for zero mean random variables are defined as:

cum(xi , xj , xk , xl ) = E{xi , xj , xk , xl } − E{xi , xj }E{xk , xl } − E{xi , xk }E{xj , xl } − E{xk , xj }E{xi , xl } (3)

The tensor T4 is a four dimensional array, in which for each element Qijkl = cum(xi , xj , xk , xl ), the indexes i, j, k, l vary
from 1 to N (where N is the number of signals).
The ICA tensorial methods are derived through a procedure analogous to the diagonalization of a covariance matrix C,
which produces signal decorrelation. Considering that T4 is a fourth-order counterpart of C, independence can be achieved
by diagonalizing T4 , as for independent signals the only non-zero fourth-order cross cumulant appears when i = j = k = l.
Analogous to the second order case, diagonalization of the fourth-order tensor can be achieved through eigenvalue decomposition
(EVD) [23].
Although the idea of the ICA tensorial methods is quite simple, the diagonalization of T4 through an analytical procedure de-
mands very large computational resources and is prohibitive in most applications. Considering this, JADE algorithm proposes an
approximate method for diagonalization of T4 . If the input data (x) is pre-whitened (using for example PCA transformation [24])
it can be demonstrated that the cumulant tensor of x has a special structure and its eigenmatrices are described through [10]:
T
M = wm wm (4)
where m = 1, ..., N and wm are the rows of the ICA mixing matrix W (x = Ws).
JADE algorithm uses the linear transformation Fij applied to M:
X
Fij (M) = mkl cum(xi , xj , xk , xl ) (5)
kl

where mkl is an element of M, and searches for a matrix W that makes Q = WF (Mi )W as diagonal as possible.

3.2 Neural Classifier

In this work, multi-layer perceptron networks with a single hidden layer were used. The number of hidden neurons was
chosen after testing exhaustively the discrimination performance of each network.
The particular experimental setup comprises two distinct acoustic signatures (two ship classes) arriving at adjacent directions.
Considering this, the designed classifier has a single output neuron (with training target output 1 for class 1 and -1 for class 2).
For training, the resilient back-propagation algorithm (RPROP) [25] was used. In order to account for statistical variations
in the available signals, a cross-validation procedure was adopted by restarting the training procedure 10 different times using
different sample arrangements into training, validating and testing sets.

3
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

4. RESULTS
An experimental run comprising the recordings of approximately 220 seconds was used. Two ships (targets) were detected
at adjacent directions (190o and 205o ). LOFAR analysis was performed using time-windows of ∼6 ms in length (resulting
in ∼3600 windows). The average LOFAR spectra, considering all time windows, was computed for both directions and is
illustrated in Figure 4-(left). The JADE algorithm was applied to the time-windowed LOFAR spectra and two frequency-domain
independent components were estimated for each time-window (Figure 4-(right) illustrates the average LOFAR+ICA spectra). It
can be observed that the LOFAR spectra for both directions present similar characteristics (probably due to the existing cross-
interference). After ICA, the frequency contents in different bearings are quite different, favoring the target classification task.
LOFAR LOFAR + ICA

6 8
D190 D190
6
4
4
Normalized Energy

Normalized Energy
2 2
0
0 5000 10000 15000 0 5000 10000 15000

6
D205 4 D205
4
2
2
0
0 5000 10000 15000 0 5000 10000 15000
Frequency (Hz) Frequency (Hz)

Figura 4: LOFAR analysis (average spectrum) applied to raw data (left) and after ICA (right).

For designing the classifiers, the available LOFAR time windows (∼3600) were split into training, validation and testing
sets. In order to evaluate the effects of ICA pre-processing, the neural classifier was trained using both LOFAR spectra and the
frequency-domain independent components (LOFAR+ICA).
In order to compare both classifier design approaches, discrimination performance was evaluated through both ROC (Receiver
Operating Characteristic) curve [26] and SP product. The ROC illustrates how both the detection (PD ) and false alarm (PF )
probabilities vary with respect to the decision threshold. The SP product index is defined as [27]:

Ef1 + Ef2 p
SP = × (Ef1 × Ef2 ) (6)
2
where Ef1 = PD is the detection efficiency for class 1 and Ef2 = (1 − PF ) is the corresponding efficiency for class 2. The
threshold value that maximizes the SP tend to produce both high PD and low PF .
In order to chose the optimum number of hidden neurons for the MLP classifiers an iterative procedure was employed.
Initially a neural network with a single hidden layer was trained and the efficiency estimated (through SP). New neurons are
added to the hidden layer and the training procedure restarted until the SP index stabilizes near its maximum. As illustrated in
Figure 5, for the classifiers based on LOFAR the maximum SP was achieved for five hidden neurons and in the LOFAR+ICA
approach three neurons were enough to reach the maximum efficiency. It can also be observed that the use of ICA results on
higher discrimination efficiency.

100

95 LOFAR + ICA
LOFAR
SPx100

90

85

80
2 4 6 8 10
Number of hidden neurons

Figura 5: Discrimination efficiency (SP×100) variation as the number of hidden neurons is increased.

4
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

Figure 6 illustrates the neural network outputs (considering the number of hidden neurons that produced higher discrimination
efficiencies) for classifiers based on LOFAR and LOFAR+ICA. One can see that ICA favors the separation of distinct classes,
which implies on a higher area under the ROC curve as shown in Figure 7. Table 1 summarizes the obtained results.
LOFAR LOFAR + ICA

3
10 D205
3
10 D205
D190 D190

2 2
10 10

Counts
Counts

1 1
10 10

0 0
10 10

−1 −0.5 0 0.5 −1 −0.5 0 0.5 1

Classifier output Classifier output

Figura 6: Neural classifier outputs using LOFAR (left) and LOFAR+ICA (right).

0.8

0.6
SPx100 = 87,87
PD

PD= 87,03%
0.4 PF=11,29%

LOFAR
LOFAR + ICA SPx100 =99.97
0.2 PD= 99,94%
PF=0,00%
0
0 0.2 0.4 0.6 0.8 1
PF

Figura 7: ROC curves for both classifiers.

Tabela 1: Discrimination efficiency obtained for both classifiers

Preprocessing SP×100 Eff. Clas 1 (%) Eff Clas 2 (%)
LOFAR 87.87 87.03 88.71
LOFAR+ICA 99.97 99.94 100.00

5. CONCLUSIONS
Sonar systems are widely used in military and civilian applications. In passive sonar, target identification is usually per-
formed by trained human operators which listen to the acoustic signatures and also analyze frequency information obtained from
LOFAR analysis. The identification tasks becomes more difficult when acoustic signals from different vessels arrive at adjacent
direction of the hydrophone array. In this case, either the acoustic signatures and the LOFAR spectra are distorted by the cross-
interference. In this work was proposed the application of independent component analysis (ICA) in order to reduce the effects
of cross-interference when targets are at adjacent directions. ICA was applied in frequency-domain after LOFAR analysis and
a multi-layer perceptron classifier performed the hypothesis testing. The obtained results indicate that using a combination of
LOFAR and ICA the discrimination efficiency was considerably improved, from ∼87% for class 1 and ∼89% for class 2 to re-
spectively ∼99.9% and 100%. Different signal processing algorithms based on statistical paradigms such as the convolutive blind
5
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

separation [28] and the nonnegative matrix factorization [22] may also be applied to improve passive sonar target identification.
These aspects will be addressed in a future work.

ACKNOWLEDGMENTS
The authors would like to express their gratitude to CNPQ, CAPES and FAPESP for the financial support and to the Brazilian
Navy Research Institute (IPqM) for providing the used data and for fruitful discussions concerning this work.

REFERENCES
[1] R. J. Urick. Principles of Underwater Sound for Engineers.. McGraw-Hill., 1983.
[2] A. D. Waite. Sonar for practicing Engineers.. John Wiley and Sons, Inc., third edition, 2003.
[3] J. D. Martino, J. Haton and A. Laporte. “Lofargram line tracking by multistage decision process.” IEEE International
Conference on Acoustics, Speech, and Signal Processing., vol. 1., pp. 27–30., April. 1993.
[4] N. N. de Moura, W. S. Filho and J. M. Seixas. “Passive Sonar Classification based on Independent components.” Proceed-
ings of the Brazilian congress of neural networks, pp. 1-5. (portuguese)., 2007.
[5] R. C. Torres, J. M. de Seixas and W. S. Filho. “Classificação de Sinais de Sonar Utililizando Componentes Principais
não-lineares.” Learning and nonlinear models., vol. 1., no. 4., pp. 208–222., 2004.
[6] J. M. Seixas, D. O. Damazio and W. S. Filho. “Wavelet transform as a preprocessing method for neural classification of
passive sonar signals.” International Conference on Electronic, Circuits and Systems., pp. 83–86., September. 2001.
[7] W. S. Filho, J. M. de Seixas and L. P. Caloba. “Principal Component Analysis for Classifyting Passive Sonar Signals.”
IEEE International Symposium on Circuits and Systems., 2001.
[8] H. Peyvandi, B. Fazaeefar and H. Amindavar. “Determining class of underwater vehicles in passive sonar using hidden
Markov with Hausdorff similarity measure.” Proceedings of 1998 International Symposium on Underwater Technology.,
pp. 258–262., April. 1998.
[9] L. Marple and T. Brotherton. “Detection and classification of short duration underwater acoustic signals by Prony’s method.”
International Conference on Acoustics, Speech, and Signal Processing., vol. 2., pp. 1309–1312., May. 1991.
[10] A. Hyvarinen, J. Karhunen and E. Oja. Independent Component Analysis. Wiley, New York, United States, 2001.
[11] V. S. Poblador, E. M. Moreno and J. S. Casals. “ICA as a Preprocessing Technique for Classification”. In Ind. Comp.
Analysis And Signal Separation, 5th Int. Conf., LNCS, pp. 1165–1172, Charleston, United States, Setembro 2004. Springer.
[12] F. Oveisi. “EEG Signal Classification Using Nonlinear Independent Component Analysis”. In Proceedings of the Interna-
tional Conference on Acoustics, Speech and Signal Processing, volume 1, pp. 361–364, Taipei, Taiwan, 2009.
[13] E. F. Simas Filho, J. M. Seixas and L. P. Caloba. “Modified Post-Nonlinear ICA Model for Online Neural Discrimination”.
Neurocomputing, vol. 73, no. 16-18, pp. 2820–2828, 2010.
[14] N. N. Moura, E. F. Simas Filho and J. M. Seixas. Advances in Sonar Signal Processing, chapter Independent Component
Analysis for Passive Sonar Signal Processing, pp. 91–110. In-Tech, Vienna, Austria, 2009.
[15] S. Haykin. Neural Networks and Learning Machines. Prentice Hall, New Jersey, United States, third edition, 2008.
[16] H. Krim and M. Viberg. “Two decades of array signal processing research: The parametric approach”. IEEE ASSP
Magazine, vol. 13., pp. 67–94, 1996.
[17] P. S. R. Diniz, E. Silva and S. L. Netto. Digital Signal Processing: Applications and Design.. Second edition, 2010.
[18] R. O. Nielsen. Sonar Signal Procesing.. Artech House Inc, Nortwood., 1991.
[19] W. S. Filho, J. M. de Seixas and N. N. de Moura. “Preprocessing passive sonar signals for neural classification.” IET Radar,
Sonar and Navigation., vol. 6, 2011.
[20] A. Hyvarinen and E. Oja. “Independent Component Analysis: Algorithms and Applications”. Neural Networks, vol. 13,
no. 4-5, pp. 411–430, 2000.
[21] J.-F. Cardoso and A. Souloumiac. “Blind Beamforming for Non-gaussian Signals”. IEE Proceedings-F, vol. 140, no. 6,
pp. 362–370, 1993.

6
 X Congresso Brasileiro de Inteligência Computacional (CBIC’2011), 8 a 11 de Novembro de 2011, Fortaleza, Ceará
°c Sociedade Brasileira de Inteligência Computacional (SBIC)

[22] A. Cichocki, R. Zdunek,

View publication stats A. H. Phan and S. ichi Amari. Nonnegative Matrix and Tensor Factorizations Applications to
Exploratory Multi-way Data Analysis and Blind Source Separation. John Wiley & Sons, United States, 2009.

[23] G. H. Golub and C. F. V. Loan. Matrix Computations. Johns Hopkins Studies In The Mathematical Sciences, United States,
1996.

[24] I. Jolliffe. Principal Component Analysis. New York: Springer-Verlag, second edition, 2002.
[25] M. Riedmiller and H. Braun. “A direct adaptive method for faster backpropagation learning, the RPROP algorithm”. Proc.
of Int. Conf. on Neural Networks, pp. 586–591, 1993.

[26] H. L. V. Trees. Detection, Estimation, and Modulation Theory, Part I. Wiley, New York, United States, 2001.
[27] A. dos Anjos, R. Torres and J. Seixas. “Neural triggering system operating on high resolution calorimetry information”.
Nuclear Instruments and Methods in Physics Research A, vol. 559, no. 1, pp. 134–138, 2006.
[28] M. Syskind, D. L. Wang, J. Larsen and U. Kjem. “Separating Underdetermined Convolutive Speech Mixtures”. In Inde-
pendent Component Analysis And Signal Separation, 8th International Conference, edited by J. Rosca, D. Erdogmus, J. C.
Principe and S. Haykin, Lecture Notes In Conputer Science, pp. 674–681. Springer, 2006.