Documentos de Académico
Documentos de Profesional
Documentos de Cultura
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
Faculty of Electronic Information and Electrical Engineering, Dalian University of Technology, Dalian 116024, China
School of Electronic & Information Engineering, Liaoning University of Technology, Jinzhou 112021, China
a r t i c l e in f o
abstract
Article history:
Received 25 August 2014
Received in revised form
14 December 2014
Accepted 15 December 2014
Available online 24 December 2014
In this paper, a novel algorithm for fractional time delay estimation is proposed. It is based
on the maximum correntropy criterion and the Lagrange fractional delay filter (FDF). The
instantaneous correntropy is introduced to measure the similarity between received
signals and estimated ones in the proposed algorithm. It leads to an effective performance
under both Gaussian and impulsive noises. The performances, including the convergence
of the algorithm and the variance of the time delay estimation, are theoretically analyzed.
Simulations demonstrate that the time delay estimation precision of the proposed
algorithm is higher than that utilizing the mixed modulated Lagrange explicit time delay
estimation (MMLETDE) especially under low generalized signal-to-noise ratio (GSNR).
& 2014 Elsevier B.V. All rights reserved.
Keywords:
Fractional time delay estimation
Correntropy
Maximum correntropy criterion
Lagrange FDF
1. Introduction
Fractional time delay estimation (referred to as FDE), also
known as subsample time delay estimation, is widely
utilized in high precision location (such as the radar [1,2],
the medical auxiliary diagnostic [3], the satellite navigation
[4] and the fault location [5]) and the time delay compensation (such as the digital pre-distortion [6,7], sensors calibration [8] and the beam forming [9]).
Well-known fractional time delay estimation algorithms
include the interpolation algorithm [10], the Hilbert transform
based algorithm [11], the explicit adaptation of time delay
(ETDE) algorithm [12], etc. The interpolation algorithm interpolates some approximate values based on a fitting function
between the estimated adjacent integral time delays to
estimate the fractional time delay. Although it has low
calculation complexity, the accuracy of the interpolation
algorithm is not sufficient for high precision applications.
The Hilbert transform based algorithm is fast since it consists
of only one scalar multiplication, but it is only applicable to
n
hDk
^ n
M2
^ i
D
;
ni
i M1
M1 r n r M2
ia n
yk sk D w2 k
M2
n M1
hDk
^ nxk n
6
where
^
g n Dk
gv ej0 v f v j0 sin cv
sin cv
cos v sin cv
^
v
Dk
1
V^ L; X; Y xk yk
Lk1
10
M2
n M1
0
hDk
^ nxk n
11
h
2
The instantaneous
correntropy ec k V yk; M
i
n M1
0
nxk
n
serves
as
the
error
function
in
the
MCCL
hDk
^
"
#
ec k V yk;
M2
n M1
0
1
B
p exp@
2
0
hDk
^ nxk n
M2
yk n
!
ek2
1
p exp
2 2
2
2 1
0
h
nxk
n
C
^
M 1 Dk
A
2 2
12
2
where ek yk M
^ nxk n.
n M1 hDk
The iterative formula of the MCCL is derived as follows:
en kec k
^ 1 Dk
^
Dk
c
13
^
Dk
where
!
enc kec k
ec k
2Re enc k
^
^
Dk
Dk
!
1
ek
2Re ec kec k 2 ek
^
Dk
!
2
2
ek
14
2 ec k Re ek
^
Dk
2
^
^
Since ek=Dk
M
n M 1 g n Dk xk n in
n
^
[14], the ec kec k =Dk
can be written as
!
2
n
M2
2ec k
ec kec k
^
Re
ek
g
n
Dk
xk
n
^
2
Dk
n M1
n
3. MCCL algorithm
3.1. Correntropy and maximum correntropy criterion
Correntropy [16] proposed by Weifeng Liu et al. in 2007
is a theory for non-Gaussian signal processing. The cross
correntropy is a generalized similarity measurement
between two arbitrary scalar random variables X and Y
defined by
V X; Y E X Y
L
data xk; yk k 1 are available, the sample estimator of
correntropy is shown as
223
9
p
2
2
where X Y 1= 2 exp jX Y j =2 , and 40
is the kernel size. In practice, since only finite numbers of
15
^
where g n Dk
is defined in (7). Using (13) and (15), the
iterative formula of the MCCL can be deduced as
!
M2
2
2
^
^ 1 Dk
^
xk n
Dk
2 ec k Re ek g n Dk
n M1
16
224
4. Performance analysis
ek yk
M2
n M1
hDk
^ nxk n
^
^
sk D w2 k s k Dk
w1 k Dk
17
M2
n M1
^
g n Dk
xk n
M2
n M1
n M1
^
g n Dk
xk n T 1 T 2 T 3 T 4
19
where
^
T 1 sk D sk Dk
^
T 2 w2 k w1 k Dk
M2
n M1
M2
n M1
^
T 3 sk D sk Dk
^
g n Dk
sk n
^
g n Dk
sk n
M2
n M1
20
^
g n Dk
w1 k n
22
^
T 4 w2 k w1 k Dk
M2
n M1
^
g n Dk
w1 k n
23
Thus,
h
i
h
i 2 h
i
2
^ 1 E Dk
^
E Dk
2 E ec k ReT 1 T 2 T 3 T 4
24
According to the properties of the correntropy [16],
"
#
M2
1
0
0 oec k V yk; hDk
25
^ nxk n r p
2
n M1
p
ec(k) reaches its maximum value 1= 2 if and only if
0
M2
yk n M1 hDk
nxk n.
^
Thus Eq. (24) can be rewritten as
h
i
h
i 22
^ 1 E Dk
^
E Dk
2 EReT 1 T 2 T 3 T 4
26
p i
where A 0; 1= 2 .
Since the signal and the noises are uncorrelated, we
have both ReET 2 0 and ReET 3 0. ReET 1 is
evaluated as follows:
M2
n M1
^
g n Dk
sk n
M2
n M1
sk n
h
i
^
^
^
j0 sin c n Dk
ej0 n Dk f n Dk
n M1
"
M2
n M1
^
^
j0 ej0 n Dk sin cn Dksk
n
^
^
ej0 n Dk sin cn Dksk
n
^
Dk
M2
n M1
M2
n M1
^
j0 n Dk
^
sin cn Dke
sk n
i
^
^
ej0 n Dk sin cn Dksk
n
3
^
^
ej0 n Dk sin c n Dk
sk n
5
4
^
Dk
n M1
2
27
Since
^
j0 n Dk
2
^
^
M
s k Dk
, Eq.
n M1 sk n sin c n Dk e
(27) is simplified as
^
s k Dk
M2
^
28
g n Dk sk n
^
Dk
n M1
21
^
sin cn Dksk
n
^
Dk
M2
We get
M2
M2
j0
^
g n Dk
sk n w1 k n
18
ek
n M1
j0
Since
ej0 n Dk
"
M2
^
s k Dk
^
Dk
^
A k Dk
^
^
^
j0 A k Dk
ej0 k Dk ej0 k Dk
^
Dk
29
For a narrowband signal, the envelope A(k) varies
^
insensitively to Dk,
and (30) is the approximate equation
of (29).
^
s k Dk
^
^
ej0 k Dk
j0 A k Dk
^
Dk
^
30
j0 s k Dk
So we obtain
M2
n M1
^
^
g n Dk
sk n j0 s k Dk
31
ReET 4 Re E
" "
Re E
M2
n M1
M2
n M1
^
g n Dk
w1 k n
##
##
0
^
^
hDk
n f n Dk
j0 sin c n Dk
^
Re f 0 j0 sin c0 f 0 0
34
!
36
h
i
22
^
^
E Dk
D D0
D 1 2 2s 20
37
=22 2s 20 .
Thus
^
^ k 2DEDk
D2
ED
38
In Appendix A, we have
k 1 Ck B
39
where
42 2 2 42 4 4 4
4 s 0
2 s 0
40
41
2
20
q 2 2s 20 2 20 2s 20 2
3
1 EG
42
M2
n M1
B
1C
45
47
^ is seriously
From (47), we could see that the varD
influenced by the dispersion parameter . However, for the
proposed correntropy based algorithm, if the kernel size
is chosen properly, the estimation accuracy could be
improved.
5. Simulation results
Suppose that the interesting signal s(k) is double sideband suppressed carrier (DSB-SC) signal. Its carrier frequency is 10 MHz and the sampling frequency is 60 MHz.
The actual delay D is set as 0.3 sample interval. Both w1(k)
and w2(k) are noises subjected to SS distribution with the
location parameter a 0. The accuracy of the delay estima^ is the delay estimation
tion (P) is defined in (48), where D
value and D is the true value.
1
0
^
D D
A 100%
P @1
48
D
The GSNR is defined in
22 4
q
4
^ k
varD
k-1
46
It is 0 o o 4 = 2s 20 .
^ is given by
The delay variance varD
!k
h
i
^
When 0 o1 22 = 2 2s 20 o1 and k-1, E Dk
-D.
h
i
^
It means that the E Dk
will converge to the actual delay
C 1
0 o o 4 = 2s 20 , 0 o o 2 4 = 2s 20 and a 4 = 2s 20 .
^
varD
225
2
hDk
^ n
43
44
GSNR 10log10 2s = dB
49
100
100
90
90
80
80
70
70
Accuracy (%)
Accuracy (%)
226
60
50
40
30
60
50
40
30
20
10
MCCL
MMLETDE
20
MCCL
MMLETDE
10
0
-10
-5
10
15
0
0.5
20
GSNR(dB)
1.5
alpha
100
100
90
90
80
80
70
Accuracy (%)
Accuracy (%)
70
60
50
40
60
50
40
30
30
MCCL
MMLETDE
20
MCCL
MMLETDE
20
10
10
0
0.5
0
-10
-5
10
15
20
GSNR(dB)
Fig. 1. Delay estimation accuracy versus GSNR: (a) relative bandwidth of
s(k) is 0.1% and (b) relative bandwidth of s(k) is 5%.
M2
n M1
#
0
hDk
^ nxk n
1
r p
2
50
p
ec(k) reaches its maximum value 1= 2 if and only if
0
M2
yk n M1 hDk
nxk n. It means that correntropy
^
could diminish the effect of outliers. So under low GSNR or
low values, the MCCL has better performance.
For the convergence analysis for both algorithms, the
MMLETDE is just suitable for a narrowband condition [14],
while the MCCL works well under both narrowband and
1.5
alpha
Fig. 2. Accuracy of the delay estimation versus : (a) relative bandwidth
of s(k) is 0.1% and (b) relative bandwidth of s(k) is 5%.
100
227
100
90
90
80
80
Accuracy (%)
Accuracy (%)
70
60
50
MCCL FM
MMLETDE FM
MCCL QPSK
MMLETDE QPSK
MCCL BPSK
MMLETDE BPSK
MCCL DSB-SC
MMLETDE DSB-SC
40
30
20
10
0
-10
-5
10
15
70
60
50
N=1
N=2
N=3
N=4
40
30
20
-10
20
-5
10
15
20
100
90
80
70
Accuracy (%)
GSNR (dB)
GSNR (dB)
60
50
6. Conclusion
40
30
20
10
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
1.8
kernel size
Fig. 4. Accuracy versus kernel size.
Acknowledgments
This work is partly supported by National Natural Science
Foundation of China (61172108, 61139001, and 81241059) and
the Science and Technology Support Program (2012BAJ18B06)
Appendix A
Taking the square and expectation on both sides of (16),
then using (19) we have
h 2
i
h 2 i 4 h
i h
i
2
^ k 1 E D
^ k E Dk
^
E D
E ec k ReT 1 T 2 T 3 T 4
2
i
4
42 h
E ec k Re2 T 1 T 2 T 3 T 4
4
A:1
h
i
2
Since ReT 12 Re TT n T 2 , where T T1 T2 T3
T4, we get
i
h 2
i
h 2 i 42 h
i
22 4 h
^ k 1 E D
^ k
^
E D
E Dk
EReT 4 E Re TT n T 2
2
p i
where A 0; 1= 2 .
A:2
228
E T 4 T n4 2 20 2
h
i
in the third term of
Then evaluating Re E TT n T 2
(A.2)
Re E TT n Re E T 1 T 2 T 3 T 4 T 1 T 2 T 3 T 4 n
Re ET 1 T n1 T 2 T n2 T 3 T n3 T 4 T n4
2Re ET 1 T n2 T 1 T n3 T 1 T n4 T 2 T n3 T 2 T n4 T 3 T n4
A:4
h i
h
i
Re E T 2 Re ET 1 T 2 T 3 T 4 2
h
i
Re ET 21 T 22 T 23 T 24
2ReET 1 T 2 T 1 T 3 T 1 T 4 T 2 T 3 T 2 T 4 T 3 T 4
n
A:5
Obviously, ET 1 T 2 , ET 1 T 2 , ET 1 T 3 , ET 1 T 3 , ET 1 T 4 ,
ET 1 T n4 , ET 2 T 3 , ET 2 T n3 , ET 2 T 4 , ET 2 T n4 , ET 3 T 4 , ET 3 T n4 ,
ET 22 , ET 23 are all equal to zero, for the signals and the
noises are uncorrelated.
The other terms are evaluated one by one as follows:
2
2 3
M2
^
^
ET 1 T n1 E4 sk D sk Dk
n 5
gn Dksk
n M1
2
^
E 2s j0 1 ej0 Dk D
A:6
(A.6) is simplified to
h
i
^
^
ET 1 T n1 E 4s 20 2 sin 2 0 Dk
D 4 sin 2 0 Dk
D=2
h
^
E 4s 40 Dk
D2
.
20 1 EG
3
2
A:11
A:12
42 2 2 h ^ i 22 4 4 4
DE Dk 4 2 s 0 k q
2 s 0
A:13
where
q 2 2s 20 4n 20 2s 20 2 2 =3 20 1 EG.
Then, we get
h 2
i
h
i
^ k 1 2DE Dk
^ 1 D2
k 1 E D
!
42
42 4
22 4
k 1 2 2s 20 4 4s 40 4 q A:14
Defining
C 1
42 2 2 42 4 4 4
4 s 0
2 s 0
A:15
and
22 4
q
4
A:16
Then
k 1 Ck B
A:17
A:7
Similarly,
2
M2
n M1
!2 3
^
gn Dksk n 5
2
^
E 2s j0 1 ej0 Dk D
h
^
E 4s 40 Dk
D2
A:8
Then ET 2 T n2 is evaluated
2
2 3
2 M2
^
^
ET 2 T n2 E4 w2 k w1 k Dk
n 5
gn Dksk
n M 1
2
^
j0 k Dk
^
^
E w2 k w1 k Dk
j0 Ak Dke
2
^
2 2s 20 E jej0 k Dk
2 2s 20
A:9
n
ET 24
ET 3 T n3 E 1 hDk
n
^
s 0
are evalu-
n M1
2s 20 1 EG
ET 24 2 2 20
2
A:3
^
ET 1 2 E4 sk D sk Dk
n
^
M 1 hDk
A:10
References
[1] X. Zhang, Y. Bai, X. Chen, An improved cross-correlation method
based on fractional delay estimation for velocity measurement of
high speed targets, Proceedings of the World Congress on Engineering and Computer Science (WCECS) (2013).
[2] B. Feng, T. Wang, C. Liu, C. Chen, W. Chen, An effective CLEAN
algorithm for interference cancellation and weak target detection in
passive radar, 2013 Asia-Pacific Conference on Synthetic Aperture
Radar (APSAR) (2013) 160163.
[3] C.C. Took, S. Sanei, S. Rickard, J. Chambers, S. Dunne, Fractional delay
estimation for blind source separation and localization of temporomandibular joint sounds, IEEE Trans. Biomed. Eng. 55 (March 3)
(2008) 949956.
[4] M. Sirbu, E. Delfino, V. Koivunen, A novel method for time delay
acquisition in satellite navigation systems, in: Proceedings of the
2004 IEEE Eighth International Symposium on Spread Spectrum
Techniques and Applications (2004) 726730.
[5] P. Liao, M. Cai, Y. Shi, Z. Fan, Compressed air leak detection based on
time delay estimation using a portable multi-sensor ultrasonic
detector, Meas. Sci. Technol. 24 (March 3) (2013).
[6] Y. Ma, Y. Akaiwa, Y. Yamao, S. He, Test bed for characterization and
predistortion of power amplifiers, Int. J. RF Microw. Comput.Aided
Eng. 23 (2013) 7482. (May (5).
[7] H. Qian, H. Huang, S. Yao, A general adaptive digital predistortion
architecture for stand-alone RF power amplifiers, IEEE Trans. Broadcast. 59 (March 3) (2013) 528538.
[8] I. Bucher, O. Halevi, Optimal phase calibration of nonlinear, delayed
sensors, Mech. Syst. Signal Process. 45 (January 1) (2014) 424432.
[9] S.A. Fares, T.A. Denidni, S. Affes, C. Despins, Fractional-delay sequential blind beamforming for wireless multipath communications in
confined areas, IEEE Trans. Wirel. Commun. 7 (February 2) (2008)
629638.
[10] F. Viola, W.F. Walker, A spline-based algorithm for continuous timedelay estimation using sampled data, IEEE Trans. Ultrason. Ferroelectr. Freq. Control 52 (January 1) (2005) 8093.
[11] A. Grennberg, M. Sandell, Estimation of subsample time delay
differences in narrowband ultrasonic echoes using the Hilbert
transform correlation, IEEE Trans. Ultrason. Ferroelectr. Freq. Control
41 (May 5) (1994) 588595.
[12] H.C. So, P.C. Ching, Y.T. Chan, A new algorithm for explicit adaptation
of time delay, IEEE Trans. Signal Process. 42 (July 7) (1994)
18161820.
[13] T.L. Laakso, V. Valimaki, M. Karjalainern, U.K. Laine, Splitting the unit
delay [FIR/all pass filters design], IEEE Signal Process. Mag. 13
(January 1) (1996) 3060.
229
[14] Z. Cheng, T.T. Tjhung, A new time delay estimator based on ETDE,
IEEE Trans. Signal Process. 51 (July 7) (2003) 18591869.
[15] E. Hermanowicz, Explicity formulas for weighting coefficients of
maximally flat tunable FIR delays, Electron. Lett. 28 (September 20)
(1992) 19361937.
[16] W. Liu, P.P. Pokharel, J.C. Prncipe, Correntropy: properties and
applications in non-Gaussian signal processing, IEEE Trans. Signal
Process. 55 (November 11) (2007) 52865298.