Krzysztof Konopko
1 Introduction
J
a
,
(6)
t,J) = f x(t+-)x
2 2
W(
r (t--)e
r -J'2tif:r dr
*
(2) where:
2 2
-00
t
where W( ,J) is the Wigner-Ville distribution (WVD) of 2
XT (J) = f x (u)e-J2"fo du (7)
the signal x(t)
[2, 4, 6]. The WVD exhibits the highest T
signal energy concentration in the time-frequency plane
for linearly modulated signals, but has drawbacks in the is the Fourier Transform of the time domain signal x(u).
cases of nonlinear frequency modulated signals. The The additional variable a leads to a two-dimensional
WVD also contains interfering cross terms between each representation S; (J ) , in the bi-frequency plane or if,
pair of signal components. The main efforts of
a) plane [3]. Thus, a cyclostationary process is
researches is directed on working out new
completely characterized by the cyclic spectrum
transformations which could enable to remove the cross
terms [9]. For the signal processing a windowed version S; (J ) or equivalently by the cyclic autocorrelation
of the WVD, the pseudo WVD (PWVD), and smoothed function R; (r) .
pseudo Wigner-Ville distribution (SPWVD) are widely
used [10].
Pseudo Wigner-Ville Distribution is defmed as the 2.3 Examples of LPI signals analysis
following:
For analysis of the presented methods efficiency the
PWV ( ,J) =
t triangular modulation of a Frequency Modulated
(3) Continuous Wave (FMCW) and Binary Phase Shift
Keying (BPSK) have been used. The WVD, PWVD,
SPWVD and CSD of the following FMCW code signal:
where h(r) is a window function. carrier frequency - 1000 Hz,
Smoothed Pseudo Wigner-Ville Distribution can be sampling frequency - 4000 Hz,
written as: modulation bandwith - 250 Hz,
SPWV (t,J) = modulation period - 20 ms
+00 +00 (4) are shown on Fig. 2, and the WVD, PWVD, SPWVD
= fq(u-t) fh(r) x(t+)x*(t-)e-J27rfT dr du
-00 -00
and CSD of the following BPSK code signal:
carrier frequency - 1000 Hz,
where q(u-t) and h(r)
are window functions.
sampling frequency - 4000 Hz,
In addition to the WVD a number of bilinear
number of Barker bits - 7,
distributions have been proposed [2, 5, 6] for removing
number of carrier periods per Barker bit - I
the cross terms.
are shown on Fig 3.
2.2 Cyclostationary signal processing
'000
. ...
'000
'000
1400
'400
r=
'200
i'...
i
Ir
... ...,
000
GOO
"'"
400
200
200
0 00 00 . .1> .2
0 0 .. O. 010 02 IItnI(SI
ClIne lsi
' '''''
". . .
I
" f l lt I ' 111. ' . ' . .. ." . .... ' . .. 11 t 1 1 . 1 1 1 .,. ' '''' . . .
'"'"""11111"'1111111111111""'111"'"1111111'1'1""111IHIIIIIIII
f .. " , . , . .... U ... . . .... . .. . . . ... . ,., .... . . .. ... . . ....... .
.
. , .................... ................ " ...................... ,
................ - ... _ ....... ------... - .... ......
...... -
., 1,,,.,,,,,.,.,., .. . ,,1,, , ,1 , ,.1,,,",1111'., ' tll
:: t ::':. ::: I':: II I
...
o.
00 .1> .2
wne(11
wnelsl
2000 ri-------
''''''
' '''''
1400 .,.,......... II , ,. ""I" " " <" '.
, h .I. I .'.I.I.,. I . t t l. I .I.I.l t ' I I.III I I .'. 1 1 1 1 1 1 1 11 1 111 1 1 1 , 1 ' 1 , 1... ..
,.fHllllfl ... ... III IIII".I.11,".1
-I
... III.......II.IIIIIIII.'
_1200 'It h h h h h . ... hh,
:2
' f I , I I I I I t ' , , I t ' , , , I , I , , , I , I I I I I
1000 1...... . .... ' 11HHi1H ...... '. i l l t, IHH...... .. .. .. 11I H IIH ......... H H .. I HHH I
1'1 I I I If' I 1 '1 I" I I" I" 1,1' 1 '1 I I II
" , , ,', ,." 4
' '
"" 100 I't ','.', ,'.'.'.', ',','f'.',' ,','::,',',',','.','.',',',',',',':",',' ' :,',',',',',' ' ' .....,':,.,','. ,',',',
'
'
, , ,. ,',',':::.
'
tl,',I"II,I,I, ,II"',.,I,I'."""'I,I,I'I"I ,III,I'.II"'I'I"
GOO .. .. ....... "....... . .................. ..... , .... .... . . .
. . .
05 . .. .2
liIINIl'l
lime lsi
o
C)'dItr(Hz}
Fig.2. The WVD, PWVD, SPWVD and CSD of the Fig.3. The WVD, PWVD, SPWVD and CSD of the
FMCW signal. BPSK signal.
3 Noise reduction where NTFD1, NTFD2 are normalized time-frequency
distributions.
A two dimensional signal representation may be
D(NTFD1,NTFD2)=
The Kullback-Leibler distance is defmed as:
NTFD ( t , f)1
subject to noise and interference from several sources,
I N TFD 1 [
NTFD2(t,f)1Jdtdf INTFD21
(14)
including sensor noise and channel transmission errors.
Generally the noise has a higher-spatial-frequency
= II [ ( t , f) In (/ > 0
spectrum. Hence, simple low-pass filtering can be
effective for noise cleaning. In general, the
representation TF(m, n) is convolved by a filter
-eJO-eJO
o INTFD21=O
function FF(i, j) giving the smooth representation [7]:
k 1
S (m,n)= L L TF(m-i,n-j)FF(i,j)
i=-k )=-1
(8)
5 Statistical detector
f f tP rTFD(t,J)dtdf (9)
where Fis n-dimensional feature vector, and f(F)
Mpq = specifies the probability density function for the feature
vector of noise. The threshold value K can be computed
wherep, q=O,I, ... -00.
on the basis on following equation:
P( F
The central moments of a time-frequency distribution
TFD(t,f) is defmed as: j( ) S K )= ao (16)
+00 +00
P
is probability function, and ao is significance
f.1pq= f f (t-TY( f- l)qTFD(t,J) dtdf (10)
where
level and it represents probability of false alarm (a false
alarm occurs where a non-signal event exceeds the
where t and f are defmed as: detection criterion and is identified as a signal).
t-= MIO , --
Moo
-
f= MOl
Moo
(11)
In the proposed method the Radial Basis Function (RBF)
or the Hyper Radial Basis Function (HRBF) neural
networks are applied as probability density function
The second order moments (1-4)2, f.!lh f.!20) may be used to (pdt) estimators [10]. The RBF neural network [11]
determine an important image feature orientation. In consists of three layers: an input, a layer of hidden
general the orientation describes the directions of the neurons and an output layer. The input layer corresponds
principal axes. In terms of moments the orientations of to the input vector space. Each input neuron is fully
(
the principal axes o are given by: connected to the neurons of hidden layer. The layer of
J
1 I 2f.lt1 hidden neurons consists of m neurons. Each neuron of
O=-tan - (12)
2 hidden layer computes an activation function which is
J.izo- f.1oz
x
[ j
usually the Gaussian:
where 0 is the angle of the principal axis nearest to the
axis which takes on values in the range -1t/4 s Os 1t/4. G(x)= (2;r )"I1 . t exp Ilx2-CzillZ (17)
The Kolmogorov distance is defmed as: Za i
a
D(NTFDp NTFDz)=
+00 +00
where Ci
- center of i-th activation function, a; - width
2
= f f IINTFDI(t,J)-
I I NTFDz(t,J) l dtdf
(13) of i-th activation function, n - dimension of the
estimated pdf.
The output layer consists of n neurons which correspond 6 An Example of detection of LPI signal
to the classification problem. The output layer is fully
connected to the neurons of hidden layer. The output For the purpose of analysis of the algorithm detection
following signal was used:
( ( t }}
layer neuron computes a linear weighted sum of the
following form:
m
X (t )=sin 2" fo+ (20)
fj(x)= LWij Gi(x) j=1,2,... ,n (18)
i=1 where:
carrier frequency - 1 kHz,
where wij is the weight value between the neuron of i-th sampling frequency - 1 kHz,
the hidden layer and the j-th
neuron of the output layer. k - 20 kHZ/s,
For multidimensional pdf, the activation function can be modulation period - 25,5 ms
given by: The detection algorithm was based on the Wigner-Ville
G(x) =
( (x-cY )
Distribution and noise reduction procedure with using the
l Gaussian two-dimensional filter (8) where:
1 I
Q - (X -C) I
(19)
= 2 exp - FF
( 2" t det(Q) 2 g (21)
where Q is the covariance matrix.
The examples of two-dimensional pdf estimates are where:
shown in the Fig 4 and 5.
(22)
and n], n2 are the number of rows and columns in FF
0.04 respectively. The degree of smoothing is determined by
standard deviation 0" of the filter.
003
The signal was detection based on a feature vector which
( )
0.02 consists of two components:
M
0.01 F; = Ol , F; = ! tan -l 2Pll
. (23)
M 2
oo P20-P02
o
10 The density function of feature vector was estimated by
10
using the proposed HRBF neural network estimator. 20
neurons in the middle layer and 1000 learning samples
-10 -10
were used. The detection performance of the algorithm
was tested on the basis of 1000 Monte Carlo
Fig. 4. The estimate of two-dimensional normal pdf simulations. The probability of right detection as the
N(0,2) (40 neurons in the middle layer and 10000 function of signal noise ratio (SNR) is shown in the Fig.
learning samples). 6.
Probability of right detection
0.8
0_6
0.4
0.2
-4 -4
References