Content uploaded by TeodorIon Petrut
Author content
All content in this area was uploaded by TeodorIon Petrut on Aug 25, 2016
Content may be subject to copyright.
Distributed Data Classiﬁcation in Underwater Acoustic
Sensors based on Local TimeFrequency Coherence
Analysis
Teodor Petrut, Cornel Ioana, D. Mauuary, Julien Mallet, Olivier Phillipe
To cite this version:
Teodor Petrut, Cornel Ioana, D. Mauuary, Julien Mallet, Olivier Phillipe. Distributed Data
Classiﬁcation in Underwater Acoustic Sensors based on Local TimeFrequency Coherence Anal
ysis. OCEANS ’14 MTS/IEEE Taipei  Ocean Regeneration, Apr 2014, Taipei, Taiwan. pp.978
1479936465/14, 2014. <hal00985751>
HAL Id: hal00985751
https://hal.archivesouvertes.fr/hal00985751
Submitted on 30 Apr 2014
HAL is a multidisciplinary open access
archive for the deposit and dissemination of sci
entiﬁc research documents, whether they are pub
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destin´ee au d´epˆot et `a la diﬀusion de documents
scientiﬁques de niveau recherche, publi´es ou non,
´emanant des ´etablissements d’enseignement et de
recherche fran¸cais ou ´etrangers, des laboratoires
publics ou priv´es.
Distributed Data Classification in Underwater
Acoustic Sensors based on Local TimeFrequency
Coherence Analysis
Teodor Petrut, Cornel Ioana
Grenoble Institute of Technology, GIPSAlab
Saint Martin d’Hères, France
[teodorion.petrut, cornel.ioana]@gipsalab.grenobleinp.fr
Didier Mauuary
CYBERIO
Meylan, France
didier.mauuary@cyberiodsi.com
Julien Mallet, Olivier Phillipe
OSEAN SAS
Toulon, France
[julien.mallet, olivier.phillipe]@osean.fr
Abstract – This paper introduces a stochastic approach that
considers the distributed classification problem for a network of
underwater acoustic sensors. The proposed classifier applies the
third order polynomial regression to the instantaneous frequency
extracted from timefrequency representation of different classes
of signals and represent the polynomial’s coefficients in a three
dimensional representation space. This automatic classifier is
then compared to a nonparametric classifier based on the
training of a standard neural network. The results of the
proposed method on real data illustrate the efficiency of this
algorithm, in terms of signal’s characterization and lower
communication bit rates between each sensor and the data
center.
Keywords — signal classification; neural network clustering;
pattern recognition; distributed signal processing; timefrequency
analysis.
I. INTRODUCTION
The design and the operation of networks with distributed
sensors are nowadays activities that increase the operational
performances of monitoring of large areas of observation. The
key point in the design of distributed networks of sensors is
the capacity of the distribution of signal processing algorithms
that are generally aimed to detect, localize and classify the
signals in the area of interest.
In our previous work [1], an architecture of network has
been proposed, with processing tools based on local time
frequency coherence, at sensor level, and detection and
localization, at central level. The stateoftheart in marine
intelligent sensing is the use of an energy detector (fixed or
adaptive) and the estimation of the timefrequency content
describing an event through either local Fourier analysis or
Wavelet multiresolution analysis. However, highly energetic
efficient methods were proven to be the timefrequency
coherence processing algorithms in [2]. One of the greatest
advantage of this detector is that it improves the
characterization of signals with fewer parameters than the
previous mentioned methods. These parameters are sent out to
the central processing unit that will reconstruct, for each
sensor, the corresponding Instantaneous Frequency Law (IFL).
The IFL is then modeled with a third order polynomial
regression whose coefficients are then used for classification
in a 3D space, similarly to the Kmeans clustering algorithm,
frequently used as a clustering technique in a 2D space.
The paper is organized into 6 Sections. The Section 2 puts
into light the timefrequencyphase coherencebased method
and describes the architecture of the distributed network used
for the detection, classification and localization of marine
acoustical sources. The Section 3 presents the context of the
classification problem and explicitly defines the classes of
frequency modulated marine acoustical signals. The Section 4
presents the results of classification using a nonparametric
method as it is the neural network. The Section 5 validates the
proposed parametric method for classification on realcontext
signals acquired during an offshore experimentation. Finally,
the Section 6 compares the two methods and concludes this
paper with analyzing the advantages of the proposed
classification method.
II. DISTRIBUTED SIGNAL PROCESSING
The acoustic network is composed by two main
elements: the set of sensors and the centrallevel processing
unit, as seen in Fig. 1a.
Each sensor contains an embedded processing algorithm
that allows the detection of a signal of interest and the
extraction of the key parameters describing the signal in an
analysis window W. This processing algorithm uses the time
frequency coherence analysis to locally estimate the frequency
modulations in the window W.
The timefrequencyphase algorithm assumes that the
signal is scanned locally with windows of N samples. For each
window W, we firstly look for the local Linear Frequency
Modulation (LFM) that approximate the best the local time
9781479936465/14/$31.00 ©2014 IEEE9781479936465/14/$31.00 ©2014 IEEE
This is a DRAFT. As such it may not be cited in other works.
The citable Proceedings of the Conference will be published in
IEEE Xplore shortly after the conclusion of the conference.
frequency behavior. For this purpose, the
a
(AF) is used in the LFM estimation f
o
approximates the best the signal x an
d
demodulation operation is applied in (2) to
signal [2], [3].
() ( )
FunctionAmbiguity
f
enxnxC
−
∑
−=
τ
τ
*
maxarg
2
1
()
2
2
)(
Cti
filt
etxtx
⋅−
⋅=
π
Applying this method in the analysis f
o
windows of the signal, for jth analyzed
received signal is modeled in (3) as a
P
Signal (PPS):
()
()
()
[
]
()
()
⎢
⎣
⎡+∈∝
+∈∝
++
+
+
+
2
)1(;
1,;
2
1,21,1
2
21
2
1
2
W
jtetx
WjjWtetx
tCtC
i
j
tCtCi
j
jj
jj
π
π
Mathematically, the detection based o
n
frequency coherence is defined in relation (4
()( )
[]
()
ontransmissino
o
transmissiCC
CCCCdIf
Else
jj
Then
jjjj
⎯⎯→⎯
⎯⎯→⎯
≤
++
21
1,21,1,2,1
,
,,,
where d is the Euclidian distance betwe
e
coefficients estimated in two adjacent ti
m
detection is called when the distance betwee
n
sets are inferior to a statistically chosen thre
s
Greater improvement in detection perf
o
the lowering of the false alarm rate,
implementing the processing algorithm
o
sensors, as illustrated in Fig. 1b. This fig
principle of distributed processing which
g
[1]. Besides, the improvement of IFL
e
achieved by increasing the number of F
F
used for the computation of the Ambigui
t
also,
b
y increasing the order of ap
p
polynomials.In order to do that, high comp
u
are required to do massively parallel comput
III. CLASSIFICATION OF MARINE
A
SOURCES
The key features of the proposed classificati
o
obey to the following three rules: the acous
b
e described by as small as possible num
b
maintaining its robustness; the separ
modulations classes must be as cle
a
distinguishable; the interferences phen
o
cancelled out after the post
p
rocessing o
polynomial regression.
a
mbiguity function
o
rmula (1) which
d
, eventually, the
obtai
n
the filtered
nfi⋅
−
π
2
(1)
(2)
or
half overlapped
windows Wj, the
P
olynomial Phase
]
()
⎥
⎦
⎤
+2
2,
;
W
j
W
(3)
n
the local time
) [1], [3].
o
n
⇒
≤
δ
(4)
e
n the set of two
m
e windows. The
n
both coefficients
s
hold δ.
o
rmance and, thus,
is achieved by
o
n a twochannel
ure illustrates the
g
ives better results
e
stimation is also
F
T frequency bins
t
y Functions and,
p
roximating local
u
tational resources
ations [7].
A
COUSTICAL
o
n algorithm must
tical sources must
b
er of parameters,
ation of signal
a
rly as possible
o
mena must be
f an IFL with a
Fig. 1 a: Overview of the model of d
i
channel sensor model with simplified
in the expe
r
The research study is con
description of the following
which their mathematical mo
d
(5)(8).
MonoChromatic Signals
(C
engines etc.
()
()
[
()
[
,,0
,,
,
1
2
1
⎪
⎩
⎪
⎨
⎧
+∈
+∈
=
⎢
⎣
⎡
−
∈=
j
TFj
T
jjs
j
j
tCi
j
j
Ttt
T
tttf
C
ttets
j
π
where Tr is the pulse repetitio
n
L
inear frequency modulat
e
case, two waveforms sub
c
chirp and a double ‘V’ chir
p
()
()
,
2
2
21
⎢
⎣
⎡
∈=
+
j
tCtCi
j
ttets
jj
π
Second order frequenc
y
representing parts of mam
m
()
()
=
++
,
3
3
2
21
2tCtCtCi
j
ets
jjj
π
i
stributed network of sensors; b: Two
phasebased embedded algorithm, used
r
imental tests
cerned with the detection and
marine acoustical sources, for
d
el are defined, in the relations
(C
lass C1), used by sonars, boat
() ()
()
]
]
0,0,
2
,
2
32
==
⎥
⎦
⎤
+
−
jj
r
j
TF
j
TF
j
j
TF
CC
jT
T
T
t
T
(5)
n
,
()
j
TF
T
the duration of the IFL.
e
d signals (Class C2). In this
c
lasses are considered: a simple
p
s.
() ()
0,
2
,
2
3
=
⎥
⎦
⎤
+−
j
j
TF
j
j
TF
C
T
t
T
(6)
y
modulations (Class C3)
m
als’ vocalizations.
() ()
⎥
⎦
⎤
⎢
⎣
⎡+−∈ 2
,
2
j
TF
j
j
TF
j
T
t
T
tt
(7)
Third order frequency modulations (clas
s
also, for example, parts of und
e
vocalizations;
()
()
⎢
⎣
⎡−∈=
+++
,
4
4
3
3
2
21
2
j
tCtCtCtCi
j
ttets
jjjj
π
During research, it was revealed that
regression
p
olynomial order, which succes
s
three rules, is the third order one, as a
l
examples from the Fig. 2ad.
The automatic signal classifier was desi
g
the classes of frequency modulations (Clas
Euclidian distances between the esti
m
coefficients giving the order of clustering,
2e. In brief, the principle consists of p
r
amount of data and representing them in a
3
using the first three coefficients of the regr
defined in the relation (9).
kkKClass
tctcfPIFL
2
2
3
3
)(
+
+==
Fig. 2: ad. The third order polynomial regressi
studied signal classes; e: The clustering of subsets (cl
a
representation space of the regression polynom
i
Fig. 3: Flow chart showing the proposed
p
rincipl
classification is achieved
s
C4) representing
e
rwater mammal’
() ()
⎥
⎦
⎤
+2
,
2
j
TF
k
j
TF
T
t
T
(8)
the most robust
s
fully obeys to the
l
so shown in the
g
ned to separate all
ses 24) using the
m
ated polynomial
shown in the Fig.
r
ocessing a large
3
D vector space by
ession polynomial
kk
ctc
01
+
+
(9)
on of the IFL of the
a
sses) of signals in the
i
al’s coefficients
e after which the
where t simply denotes the ti
m
kkk
ccc
123
,,
are the coeffici
e
classification algorithm.
Thus, each IFL is loca
t
analyzing its Euclidian dist
a
coordinates, as it is shown in
t
now clear that the classificat
i
only three parameters, i.e.
coefficients, instead of using
done in the matched filterin
g
with supervised learning.
IV. CLASSIFICATION
W
Neural Networks (NN) t
e
linear and nonlinear classi
f
robustness to outliers and nois
e
in conjunction with energy
transforms [4], however we w
i
frequencyphase estimator.
Classification rules are
d
feedforward network model
neurons. As schematically s
h
network used in this paper i
s
neuron layer, a database wit
h
neuron layer. The IFLs are u
n
small as possible number of n
We tested this architecture on
a
varying the number of neurons
clustering of the IFL classes.
In Fig. 5a
b
it is presente
d
MeanSquare Errors (MSE), w
h
the inputs and the outputs of
t
mentioned in the Fig. 4. As t
h
p
ossible number of neurons f
o
it is achieved that for at leas
t
good results. The figure 6a
b
c
the performances in terms of
c
equally distributed in the 4 cl
a
Fig.6a, as the number of neur
error to identify the second cl
a
The reason is the network’
singularity posed by “V” c
u
doubling the number of neuro
n
4 classes is obtained.
In conclusion, given its r
o
and outliers, this approach de
p
contained in an IFL and, th
u
Thus, the classification need
s
suitable number of neurons to
c
m
e and P(f) is the frequency and
e
nts used in our proposed
t
ed in such vector space by
a
nce to the theoretical centroid
t
he flow chart of the Fig. 3. It is
i
on procedure is based on using
the polynomial regression’s
all the signal’s samples, as it is
g
techniques or neural networks
W
ITH NEURAL NETWORKS
e
chniques are widely used in
f
ication process due to their
e
[6]. These are intensively used
estimators based on Wavelet
i
ll adapt this model to our time
d
eveloped using a Perceptron
with a single hidden layer of
h
own in the Fig. 4, the neural
s
composed of a single hidden
h
estimated IFLs and an output
n
dersampled in order to use as
eurons in the network’s design.
a
database of estimated IFLs by
in order to achieve a very good
d
the performances in terms of
h
ich is the squared difference of
t
he network, in the three stages
h
e interest is to use as small as
o
r the sake of computation time,
t
10 neurons the network gives
c
ompletes this section by giving
c
lustering precision of 40 IFLs,
a
sses mentioned in Section 3. In
ons is too small, there is 20 %
a
ss of IFLs (double “V” chirp).
s failure to approximate the
u
rve. However, in Fig. 6b, by
n
s, the perfect distribution of the
o
bustness with respect to noise
p
ends on the number of samples
u
s, on the number of neurons.
s
an extra step of finding the
c
orrectly classify the signals.
Fig. 4: Flow chart of the classification
p
rocedu
r
typical steps during the use of a neural network: the t
r
with the creation of the dictionary, the validation
a
Fig. 5: a. Performance with respect to the meansqua
r
network with 5 neurons (or perceptrons in this case):
t
give the expected results in the testing stage, i.e. wh
e
signals than those used in the training stage; b. Perfor
m
neural network with 10 neurons: the network gives mu
testing stage.
Fig. 6: a. Confusion matrix with the clustering results
u
with 5 neurons in the hidden layer: the green boxes sh
o
correctly identified signals and the pink boxes shows
r
e showing the three
r
aining, synonymous
a
nd the test steps
r
e errors for a neural
t
he network does not
e
n tested with other
m
ance in the case of a
ch better results in the
u
sing a neural networks
o
ws the number of the
the signals wrongly
identified; in this case, the results give
the “V” double chirps; the overall er
r
matrix of the clustering results obtaine
is observed that each classes contain
s
during e
x
V. R
E
Experiments were condu
c
Oursinières, in the SouthEa
s
described the trajectory sho
w
p
erimeter there placed three p
a
takes multiple repeated traject
o
with the emission of the follo
w
linear frequency modulations
frequency modulations and thi
r
The classes of signals emitte
d
in the Fig. 8ad.
During experiments, it is
u
consisting of 2 hydrophones
b
locks whose gain is 40 dB
acquisition (sampling frequen
c
s, length of analyzing window
a bandpass filtering between
1
the local signal processing, it
w
for intense computations
Functions and local polynomia
One such trajectory take
s
during which period the sens
data at 100 kHz sampling f
r
signals at each 10 seconds, 40
signals.
Once the detection algorit
h
p
resence of an acoustical sou
r
and the radio emitter transmit
t
the central unit processing. He
r
the classification algorithm
classification for the real
experiments are shown in F
i
figures a good 3D grouping of
frequency modulated signal.
I
confusion matrix it is proved
method to perfectly classify th
e
We finally assert that the p
constitutes a viable solution
f
approach to a threedimension
a
increasing the separability
p
classes of frequency modulati
o
an error of 20% when trying to identify
r
or in this case is of 5%; b. Confusion
d with 10 neurons in the hidden layer; it
s
the correct number of IFLs processed
x
periments
E
SULTS
c
ted offshore, in the Bay of
s
tern of France, where a boat
w
n in Fig. 7 and in which
a
ssive acoustic sensors. The boat
o
ries, each repetitio
n
coinciding
w
ing classes of signals: class 2
(chirps), class 3 second order
r
d order frequency modulations.
d
during experiments are shown
u
sed 3 receivers, each of them
equipped with preamplifying
, a Blackfin DSP is used for
c
y 100 kHz, acquisition period 1
256 samples). Also, it was used
1
.3 kHz and 48.8 kHz, and, for
w
as used a MPPA Board © [7]
of HigherOrder Ambiguity
l approximations.
s
10 minutes to be completed,
ors are continuously recording
r
equency. The boat emits one
such emissions per one class of
h
ms at sensor level validate the
r
ce, the DSP processes the data
t
he key parameters of the IFL to
r
e, after the data fusion process,
is applied. The results of
signals acquired during the
i
g. 9ad. We observe in these
data with a specifically class of
I
n the Fig. 10, by showing the
the ability of this classification
e
studied classes.
olynomial regression of the IFL
f
or extending the classification
a
l space representation, and thus
p
erformance between different
o
ns.
Fig. 7: One complete trajectory during the emissi
o
frequencymodulated signals
Fig. 8: ab. The real signals analyzed during expe
r
frequency modulations (Simple Chirp and double ‘
V
second order frequency modulations; d. Class 4 th
i
modulations
o
n of a single class of
r
iments: Class 2 linear
V
’ Chirp); c. Class 3
i
rd order frequency
Fig.9: a. The results of clustering usi
n
on the reconstructed IFL at central

classification on t
h
Fig. 10: The confusion matrix in
polynomial classification method; all
and grouped i
VI. CON
In this paper we presen
t
method for acoustical under
w
feasibility for marine source c
order of frequency modulatio
n
method is the fact that it wor
k
frequencyphase algorithm f
o
p
recision is proved in our ex
p
neural network in this cont
e
cumbersome. However, impr
o
take into account the trains o
n
g the polynomial regression of order 3

level processing; bd. The results of
h
e projection planes
the case of the 3
rd
order regression
the IFLs were successfully recognized
nto 3D clusters
CLUSIONS
t
ed a very simple and robust
w
ater signals and we proved its
lassification with respect to the
n
. The greatest advantage of this
k
s in conjunction with the time
o
r the IFL estimation and its
p
erimental results. The use of a
e
xt would be computationally
o
vements are required when we
f clicks and transients used by
mammals for echolocation. These improvements will consist
in matching the results with a dictionary of key parameters
from the timefrequencyphase algorithm.
ACKNOWLEDGEMENT
This work is supported by the French DGADGCIS under
a RAPID project “GREENAR” fund.
REFERENCES
[1] F. Dadouchi, C. Ioana, C. Gervaise, J. I. Mars, D. Mauuary, J. Mallet, O.
Philippe, S. Jespers “Distributed network of underwater acoustic sensors
based on local timefrequency coherence analysis”, Proceedings of
OCEANS 2013, San Diego, CA, USA, 2013;
[2] Hanling Zhang, Qingyun Liu, Zhishun Li, “Parameter Estimation of
Multicomponent Polynomial Phase Signals”, Intelligent Computing in
Signal Processing and Pattern Recognition, Lectures Notes in Control and
Information Sciences Volume 345, 2006, pp 10081012;
[3] C. Ioana, J.I. Mars, A. Serbanescu and S. Stankovic, “Time – Frequency –
Phase Tracking Approach: Application tot Underwater Signals in a
Passive Context” – Proceedings of ICASSP 2010, Dallas, TX, USA,
2010;
[4] Paul L. D. Roberts, Jules S. Jaffe, and Mohan M. Trivedi, “Multiview,
Broadband Acoustic Classiﬁcation of Marine Fish: A Machine Learning
Framework and Comparative Analysis”, IEEE Journal of Oceanic
Engineering, vol. 36, no. 1, January 2011;
[5] ChinHsing, Chen, JiannDer Lee and MingChi Lin “Classification of
Underwater Signals Using Neural Networks”, Tamkang Journal of
Science and Engineering, Vol. 3, No. 1, pp. 3148 (2000);
[6] Fausett, Laurene, “Fundamentals of Neural Networks, Architectures,
algorithms And Applications”, Prentice Hall Inc.: New York, 1994;
[7] http://www.kalray.eu/products/mppaboard/mppaboardemb01/