Passive Sonar Signal Detection
and Classification Based on Independent
Component Analysis
N. N. de Moura, J. M. de Seixas and Ricardo Ramos
Federal University of Rio de Janeiro, COPPE/Poli, Signal Processing Laboratory,
Rio de Janeiro
Brazil
1. Introduction
Sonar systems use the sound propagation in underwater environments for detection,
communication and navigation. The main purpose of these systems is to analyse the
underwater acoustic waves received from different directions by a sensor system and
identify the type of target that has been detected in a given direction. Sonar systems may
either be passive or active. Both passive and active sonar systems are mainly employed
in military settings, although they are also used in commercial and scientific
applications, e.g. detecting shoal fishes, performing tomography on sea to exploit a
given area, to measure the depth of a region, and so on (Burdic, 1991).
In order to detect and classify signals against background noise, passive sonar systems
(Waite, 2003) listen to the noise radiated by targets (ships or submarines) 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 targetmay be
extracted froma given direction.
There are two types of analysis that can be implemented to obtain the signal relevant
features: DEMON (Detection Envelope Modulation On Noise) (Nielsen, 1991) and
LOFAR (Low Frequency Analysis and Recording) (Di Martino, 1993). The DEMON is
a narrowband analysis that furnishes the propeller characteristic: number of shafts, shaft
rotation frequency and blade rate of the target. On the other hand, LOFAR, which is a
broadband analysis, estimates the noise vibration of the target machinery. Both analysis
are based on spectral estimation and support detection and classification of targets.
Depending on the bearing resolution, signal interference may occur for neighbour
directions, which contaminates the acquired signals and makes even more difficult the
target detection and classification tasks. To minimize these interferences, algorithms
using ICA (Independent Component Analysis) (Hyvarinen, 2001), ( Jutten, 2004)may
be applied to recover the original sources of the resulting signal mixture and obtain
optimal target detection and classification for each direction.
The detection is implemented using the classical signal demodulation to obtain the
propeller characteristics. On the other hand, efficient classification is often obtained
through neural networks (Soares Filho, 2001). For both DEMON and LOFAR analysis,
a signal preprocessing using ICA may be implemented to reduce the signal interference
in neighbour directions.
On the other hand, underwater acoustic signals suffer fluctuations as a function of the
sea features (salinity, temperature, and so on). In addition, noise radiated from targets
may vary according to operational conditions. As a consequence, the stationarity of
passive sonar signals may be affected. Therefore, it is necessary to monitor changes in
the statistics of the passive sonar signals in real time. When changes are detected, the
independent component extraction phase may be reloaded for updating the feature
extraction procedure.
An unsupervised clustering method using a modified ART (Adaptive Resonance
Theory) neural network can perform such monitoring task.
The chapter is organised as it follows. In section 2, both DEMON and LOFAR analysis
are detailed. Section 3 addresses signal interference and its removal in frequency-
domain using ICA. Section 4 briefly presents the independent component analysis and
the algorithm that was used to implement the blind source separation scheme. Section 5
gives some signal detection and classification results from experimental data, which
were acquired from a passive sonar system that has been installed in a submarine.
Finally, Section 6 presents conclusions and the perspectives for passive sonar signal
processing.
2. Spectral analysis
In this section, both DEMON and LOFAR analysis are described. Their aim is to detect
and classify the targets from a given DOA.
2.1 DEMON analysis
For signal detection, DEMONanalysis is normally applied. DEMONis a narrowband
analysis that operates over the cavitation noise of the target propeller (Nielsen, 1991).
As resulting parameters provide a detailed knowledge of the target propellers (the
propeller noise is characteristic for each target), often efficient detection is achieved.
Figure 1 shows the block diagram of classical DEMON analysis.
Given a direction of interest (bearing), a bandpass filter is implemented to limit the
cavitation frequency range (Burdic, 1991). The cavitation frequency goes fromhundreds
until thousands of Hz. Therefore, it is important to select the cavitation band and obtain
the maximum information for ship identification. Following, the signal is squared as in
traditional demodulation (Van Trees, 2001), (Yang, 2007). In most cases, the signal
sampling rate is relatively high, so that the band of interest is sampled with coarse
resolution with respect to observation needs. Thus, it is necessary to decimate the signal
for better observation in the range of interest (Rabner, 1983), as shown in Figure 2.
The signal is sampled at a frequency of 31, 250 Hz. When the first decimate by a factor
of 25 is applied the down sampling goes to 1, 250 Hz. In the sequence, a further
decimate by a factor of 25 is realized and the range from DC to 50 Hz is searched for.
This range contains the rotations of interest that go from 0 to 1, 500 (rpm). Finally, a
short-time Fast Fourier Transform algorithm (Diniz, 2010) is applied for performing
signal analysis in frequency-domain and the TPSW (Two Pass Split Window) algorithm
is used to reduce the background noise (Nielsen, 1991).
2.2 LOFAR analysis
The LOFAR (Waite, 2003) is a broadband analysis that provides the machinery noise to
the sonar operator and goes from DC to 15,625 Hz. The block diagram of the LOFAR
analysis is shown in Figure 3.
After bearing, the signal is multiplied by a Hanning window to emphasize the frequency
range of interest (Diniz, 2010). Then, the signal 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. The normalization may be implemented estimating the background noise that
is present at each spectrum and computing a normalized frequency bin using this
estimation as normalization factor. This estimation removes the spectrum bias and
equalizes the spectrum amplitude (Soares Filho, 2011).
3. Signal interference
A passive sonar system is typically used by submarines to realize the surveillance in a
determined operation area. The beamforming aims at estimating the direction of arrival
(DOA) from a given target. The Figure 4 shows the beamforming display. 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 considered (Krim,1996).
Fig. 4. Bearing time display of a passive sonar system.
When signal interference from neighbor bins occurs, as it may be the case for bearings
190◦ and 205◦ (see Figure 4), the original target features may be masked and both target
detection and classification efficiencies may be affected. Beyond that, the self-noise,
bearing 076◦, may interfere on the detection of both targets by the sonar operator. Thus,
a preprocessing scheme may be developed aiming at reducing signal interferences,
facilitating target identification and classification. Here, signal preprocessing is
addressed by independent component analysis (Hyvarinen, 2001), which is performed
in frequency-domain (see Figure 5). After beamforming a spectral analysis is
implemented at each direction of interest and an algorithm to extract the independent
components is implemented. Then both DEMON and LOFAR analysis are performed
over the independent sources (ICA space).
The ICA provides a linear representation of nongaussian data, so that components are
statistically independent, or as much independent as possible (Hyvarinen, 2001), (Yan,
2000).
The basic ICA model assumes the existence of n independent signals s1(t), ..., sn(t) and
the observation of asmany linear and instantaneousmixtures x1(t), ..., xn(t), as shown in
equation 1:
x(t) = As(t)
(1)
where s(t) = [s1(t), ..., sn(t)]T is a nx1 column vector corresponding to the source
signals, and x(t) similarly collects the n observed signals. A is the mixing matrix that
contains the mixture coefficients. The ICA problem consists in recovering the source
vector s(t) using only the observed data x(t), with the assumption of independence
between the entries of the source vector s(t). The ICA problem may also formulated as
the estimation of a nxn demixing matrix B, which allows original sources to be
recovered:
y(t) = Bx(t)
(2)
where y(t) becomes the estimated source vector.
Due to the well known central limit theorem(Barany, 2007), the ICA estimates can be
obtained from the determination of the directions for which the nongaussianity is
maximum. Efficient algorithms are available today (Hyvarinen, 2001), among which
Newton-like algorithm has been claimed to be very efficient (Akuzawa, 2001).
