Wireless Sensor Network, 2009, 3, 171-181
doi:10.4236/wsn.2009.13023 ctober 2009 (http://www.SciRP.org/journal/wsn/).
Copyright © 2009 SciRes. WSN
Published Online O
Research on DOA Estimation of Multi-Component LFM
Signals Based on the FRFT
Haitao QU1, Rihua WANG2, Wu QU3, Peng ZHAO4
1Beijing University of Posts and Telecommunications, Beijing, China
2Communication University of China, Beijing, China
3K-Touch Corporation, Beijing, China
4Beijing Research Institute of China Teleco m Co., Ltd, Beijing, China
E-mail: quhaitao2007@gmail.com
Received April 10, 2009; revised May 20, 2009; accepted May 25, 2009
Abstract
A novel algorithm for the direction of arrival (DOA) estimation based on the fractional Fourier transform
(FRFT) is proposed. Firstly, using the properties of FRFT and mask processing, Multi-component LFM sig-
nals are filtered and demodulated into a number of stationary single frequency signals. Then the
one-dimensional (1-D) direction estimation of LFM signals can be achieved by combining with the tradi-
tional spectrum search method in the fractional Fourier (FRF) domain. As for the multi-component LFM
signals, there is no cross-term interference, the mean square error (MSE) and Cramer-Rao bound (CRB) are
also analyzed which perfects the method theoretically, simulation results are provided to show the validity of
our method. The proposed algorithm is also extended to the uniform circular array (UCA), which realizes the
two-dimensional (2-D) estimation. Using the characteristics of time-frequency rotation and demodulation of
FRFT, the observed LFM signals are demodulated into a series of single frequency ones; secondly, operate
the beam-space mapping to the single frequency signals in FRF domain, which UCA in array space is
changed into the virtual uniform circular array (ULA) in mode space; finally, the DOA estimation can be
realized by the traditional spectral estimation method. Compared with other method, the complex
time-frequency cluster and the parameter matching computation are avoided; meanwhile enhances the esti-
mation precision by a certain extent. The proposed algorithm can also be used in the multi-path and Doppler
frequency shift complex channel, which expands its application scope. In a word, a demodulated DOA esti-
mation algorithm is proposed and is applied to 1-D and 2-D angle estimation by dint of ULA and UCA re-
spectively. The detailed theoretical analysis and adequate simulations are given to support our proposed al-
gorithm, which enriches the theory of the FRFT.
Keywords: DOA Estimation, The Fractional Fourier Transform, UCA, ULA, LFM
1. Introduction
In various applications of array signal processing such as
radar, sonar, communications, and seismology, there is a
growing interest in estimating the DOA of LFM signals
by dint of time-frequency analysis tools. G. Wang [1]
proposed an iterative algorithm based on time-compen-
sation, but the initial estimate is necessary. Using inter-
polation in the spatial time-frequency distribution matri-
ces (STFD’s) [2], Gershman [3] extended the signal
subspace technique and estimated effectively DOA of
LFM signals, however Gershman’s approach presences
model biases in addition to time consuming. The above
Wigner-Ville distribution (WVD) based methods conse-
quentially suffer from the disturbance of cross-terms in
the presence of multi-component signals.
Using a new time-frequency analysis tool-FRFT, di-
rection estimation of LFM signals has been proposed in
Reference [4]. However, only maximal energy concen-
tration point is selected as estimate data, easily interfered
by surroundings. In this paper, a new FRFT based algo-
rithm is proposed. Firstly, Observed signals are separated
into a number of single components by adding an adap-
H. T. QU ET AL.
172
tive filter in the FRF domain. Secondly, the separated
components are demodulated into stationary signals. Fi-
nally, the 1-D DOA of LFM signals can be estimated by
the traditional spectrum search method. This algorithm
digs two dimensional time and frequency information
without the initial estimate, frequency focusing and pa-
rameter partnership. With the increasing of the Sig-
nal-to-Noise ratio (SNR), the MSE is quite closed to the
CRB [5], for multi-component signals, cross-terms and
non-linear optimize operation are also avoided.
For the UCA widely used in the third generation mo-
bile communication system, the time-frequency charac-
teristics of the FRFT are combined with the beamform-
ing technology in FRF domain, an algorithm for the 2-D
DOA estimation of the multi-component LFM signals is
also proposed. Compared with other methods, the preci-
sion is enhanced by a certain extent. Simulation verifies
the method to be effective in the multipath and Doppler
frequency shift existed complex channels.
2. Background Knowledge of FRFT
2.1. Definition and Properties of FRFT
Recently the FRFT attracts more and more attention in
the signal processing society, in 1980, Namias [6] firstly
introduced the mathematical definition of the FRFT.
Then Almeida [7] analyzed the relationship between the
FRFT and the WVD, and interpreted it as a rotation op-
erator in the time-frequency plane. This characteristic
makes FRFT especially suitable for the processing of
LFM signals [8–9].
As a generalization of the standard Fourier transform,
the FRFT can be regarded as a counterclockwise rotation
of the signal coordinates around the origin in the time-
frequency plane. If the traditional Fourier transform of a
signal can be considered as a /2
counterclockwise
rotation from the time axis to the frequency axis, the
FRFT can be accordingly considered as a counterclock-
wise rotation from the time axis to the axis with an
angle
u
, as illustrated by Figure 1.
The FRFT of signal()
x
tis represented as
()[()]() (,)
P
X
uFxtxtKtudt



(1)
where is called the order of the FRFT,
p/2p
,
denotes the FRFT operator and
[
p
F](, )
K
tu
is the
kernel function of the FRFT
22
1cot
exp( cot
22
(, )csc ),
(), 2
(), (21)
jtu
j
Ktu jtu n
tu n
tu n







(2)
(, )
s
Wtf
(,)
p
s
Wuv
u
v
Figure 1. FRFT and WVD.
This has the following properties,
*
(, )(, )
K
tuK tu

(3)
*' '
(, )(,)()
K
tuKtu dtuu

 
(4)
Hence, the inverse FRFT is
()[( )]( )(,)
P
x
tFXu XuKtudu


(5)
Equation (5) indicates that signal ()
x
t can be inter-
preted as decomposition to a basis formed by the or-
thonormal LFM functions in the domain, and the
domain is usually called the fractional Fourier domain, in
which the time and frequency domains are its special
cases. The FRFT is a one-dimension linear transform and
has the rotation-addition property. Essentially, the repre-
sentation of a signal in the fractional domains contains
the information in both time and frequency domains of
the signal; Thus the FRFT is considered as a time-fre-
quency analysis method and has close relationships with
other time-frequency analysis tools.
u u
In Reference [10], some important characteristics are
expressed as
22
/2 1tan tan
[] exp(
1tan 21tan
pjct juc
Fe cc
)


(6)
2
2
[()](sin )4
exp[(sincoscos )]
2
pjvt
p
F
xteX u vbac
v
juv

 

(7)
2
[()](cos )
exp[(sincos/ 2sin)]
pp
Fxt Xu
ju

 
 
(8)
2.2. Discrete FRFT Computation
In engineering applications, the discrete FRFT (DFRFT)
is usually required. According to the definition of the
FRFT, it is obvious that the numerical computation of
the DFRFT is much more complicated than that of DFT.
So far, there have been several DFRFT algorithms with
Copyright © 2009 SciRes. WSN
H. T. QU ET AL. 173
x
x
x
Figure 2. Normalized time-frequency support region.
different accuracies and different complexities. In this
paper, we select the decomposition algorithm proposed
in Reference [11]. This algorithm decomposes the com-
putation of DFRFT to a convolution which can be com-
puted by FFT, and the result is very close to the output of
continuous FRFT. In this algorithm, the signal represen-
tation in time domain and frequency domain should be
approximately constrained with an interval of [/2T,
and a bandwidth of
/2]T[/2,/2FF]
respectively,
viz. the time-bandwidth product of the signal is NTF
,
and according to the uncertainty principle, con-
stantly. If the sampling rate is selected as
1N
/
sTT N
,
the discrete representations of the signal in time domain
and frequency domain will have the same length, which
is called the dimensionless normalized process and the
principle can be shown in Figure 2.
Therefore, Equation (1) can be expressed as
22
cot2 csccot
() ()
jujut jt
X
uAeee xtd
 


t (9)
where
1cot
2
j
A
For 0.5 1.5p, signal has a bandwidth
which is at most
2cot ()
jt
ex

t
2
F
and can be represented using
Shannon formula
222
cotcot/ ( 2)
()sin 2
22
N
jtjn F
nN
nn
ext excFt
F
F
 

 

 
 
(10)
Substituting Equation (10) into Equation (9) and ex-
changing the sequence of the integral and the summation,
we have
22
cot2csc/ (2)cot/ ( 2)
() ()
22
P
N
juj unFjnF
nN
XuF xt
An
ee ex
By quantizing the variable in the fractional Fourier
domain, Equation (11) can be finally discredited as
u
222
(2)/(2)
() 22 2
N
Pjmmnn
nN
A
nn
XmF xex
F
F
FF
 



 

 

 

(12)
where ()
X
m
denotes the DFRFT of signal ()
x
t,
cot
, csc
. This algorithm can be imple-
mented by FFT, and has a computation complexity of
2
og )N(lN
[11].
2.3. Two Special FRF Domain
WVD is an important non-stationary signal analysis tool,
which has a very simple relationship with FRFT; viz. the
WVD of FRFT is the coordinate rotation of the original
signal’ WVD, while the shape of WVD keeps unchanged
in the rotation. Therefore, a lot of the WVD-based signal
processing methods can be substituted by FRFT. The
relationship of the two time-frequency analysis tools can
draw a conclusion that “time width” and “fre-
quency width(
()u
)v
” will change with the difference of
the rotation angle. Considering two extreme cases,
0,uv
 or0,vu
, from the above
analysis, the former corresponds to the rotation an-
gle 1
cot


cot

, LFM signal becomes an impact func-
tion, which domain is called energy concentrated FRF
one. The latter corresponds to the rotation an-
gle , LFM becomes a single fre-
quency signal, which domain is called demodulated FRF
one and is the base of the proposed algorithm in this pa-
per. By dint of the time-frequency rotation property of
FRFT, the detection, extraction and parameter estimation
of LFM signals can be easily achieved.
1/2

2.4. The FRFT of Gaussian White Noise
Theorem 1: The FRFT of zero-mean Gaussian white
noise is still Gaussian white noise.
Proof: let subject to the ()nt2
(0, )N
distribution,
and is its FRFT, the mean is ()
p
Nu

()[()][()]0
pp
p
EN uEF ntFEnt
 (13)
Because the FRFT is the linear transform, does not
change the distribution characteristics of Gaussian noise.
Therefore, the noise is still a zero mean Gaussian noise.
2
F
F
 



As for the second-order statistical properties of noise,
the correlation of the white noise can be defined
as:
()nt
(11)
*2
()()()Entnt

 (14)
C
opyright © 2009 SciRes. WSN
H. T. QU ET AL.
174
The correlation of is defined as: ()
p
Nu



*
**
**
2*
2*
() ()
()(, )( )( ,)
()()(, )( ,)
()(,)(, )
(,) (,)
pp
pp
pp
pp
pp
EN uNuv
EntKtudtnKuvd
EntnK tuKu vdtd
tKtuKuvdtd
KuK uvd


 

 
 
 
 
 
 




 




(15)
Submit Equation (2) to Equation (15), and obtain:

2
*
2
20.5(2)cot
2
2
() ()
1cot ee
2
1cot
2sin()
2
pp
juvv jv
EN uNuv
jd
jv

csc
 



(16)
Due to Equation (16), we can see that the FRFT does not
change the time-domain white characteristics of noise,
while noise energy does not be changed.
Assume the array noise is the zero-mean airspace one,
viz. as for the array element k, the output noise is
unrelated:
(kl)



**
() ()()()0
klk l
EntntEnt Ent (17)
The cross-correlation of the noise in FRF domain is


*
**
()[ ()]
()()(, )(, )
0
pp
kl
kl pp
EN uN u
EntnK tuKudtd
 
 

(18)
The above equation shows, FRFT does not change the
airspace white characteristics of noise. Therefore, we can
draw a conclusion that FRFT does not change the statis-
tical properties of Gaussian white noise, the theorem
certification has completed.
Inference: as for the
M
antenna array element, if the
array output noise is zero mean and variance2
, the
noise covariance matrix in FRF domain is:

*
() ()
p
Npp
RENuNu I

2
M
(19)
3. 1-D DOA Estimation Algorithm
3.1. ULA Array Model
Let a ULA of
M
sensors receive LFM sources from
the unknown directions
D12
{,,,}
D

 , as illustrated by
d
()
k
s
t
k
Figure 3. ULA and array model.
Figure 3. The observed signal at the output of the th
sensor can be described as
i
1
()[] ()
D
ikik
ki
x
tst n

t
(20)
1, 2,,1iM
 1, 2,,kD
where,
2
()exp[ (/2)]
kkk
stj tt

 (21)
(1)cos /
ik k
id c
(22)
k
, k
are initial frequency and FM rate, ()
k
s
t is the
th source in reference sensor
k1
x
. is the additive
white Gauss noise with variance
()
i
nt
2
, which is assumed
to be statistically independent with signal sources. ik
is the th’s path delay, is light velocity and is
sensor spacing.
kcd
From (20) and (21), we get the direction matrix is
time-variant; however the traditional estimation method
is merely suitable for time-invariant signal model.
Therefore, the traditional method cannot be used to the
direction finding of LFM signals directly.
3.2. 1-D Estimation Algorithm Description
In this section, the main work is how to make the direc-
tion matrix time-invariant. The FRFT is actually a “Ro-
tation” of signal in time-frequency plane. An LFM signal
can be turned into an impulse in a proper fractional do-
main, for the ULA model, signal ()
k
s
t
'
k
will present an
impulse while the rotation anglecot k

. There
will be the energy concentration, consequently a distinct
peak will appear in that FRF domain, whereas the noise
energy is distributed much more symmetrically in the
entire time-frequency plane and will not be concentrated
in any FRF domain [12].
Using (20) and (21), we get that path delay can not
change the FM rates, so the impulse corresponding rota-
tion angles of signal ()
k
s
t are same in every sensor.
Then Equation (20) is rotated with angle '
k
by the
FRFT from two sides:
'' ''
'' '
()()()()
kk kk
D
iik ili
lk
Wu YuYuVu
 
 
'
(23)
Copyright © 2009 SciRes. WSN
H. T. QU ET AL. 175
where, presents an impulse, and
are approximately considered as LFM signal
and the white Gauss noise respectively.
''
()
k
ik
Yu
)
''
()
k
D
il
lk
Yu
''
(
k
i
Vu
Therefore, a mask operation is applied to (23) accord-
ing to the peak position , which is a narrowband
filter with central frequency , and with a properly
selected bandwidth, most energy of the signal
will be removed. This procedure can be re-
garded as an open loop adaptive time-varying filter
whose central frequency varies linearly following the
peak position .
ik
m
ik
m
2L
''
()
k
ik
Yu
ik
m
Signal is performed the FFT (viz. FRFT
of ). According to the rotation-addition property
[10], the two procedures above are equivalence to one
time rotation with angle
''
()
k
ik
Yu
1p
k
viz.
3/2cot
kk

;
tan
kk
(24)
Using (6), (7) and (8), signal ()
k
s
t is rotated with angle
k
by the FRFT can be expressed as
2
2
()
1tan(sin )tan
exp[ ]
1tan2 1tan
exp[(sincos/ 2cos)]
k
k
kkkk
kk kk
kkkkk
Su
ju
ju
k
 

 



2
1tan exp[(sincos/2)]
1tan
exp( cos)exp( cos)
kkkk
kk
kk kk
jj
juB ju





(25)
where,
2
1tan exp[(sincos/ 2)]
1tan
kkkk
kk
j
Bj



(26)
From (25) and (26), it can be seen that LFM signal ()
k
s
t
has been transformed into the single frequency signal
in the FRF domain.
()
k
k
Su
Similarly, the FRFT of path delayed signal ()
k
st
with rotation angle k
can be expressed as
2
2
[()]exp(sincos/ 2)
exp(cos) exp[(cossin)]
kkkk
kkkk k
FstB j
jju

 


(27)
In practice,
is too small viz.
sin cos
kk
2
exp(sincos/ 2)0
kk
j

(28)
Substituting (28) into (27), we get
2
2
[( )]
exp(cos) exp(cos)
exp(cos)( )
k
k
k
kk k
kkk
Fst
Bj ju
jSu
k
 
 

(29)
From the above analysis, Using (25) and (29), the ob-
served signals described by (20) are performed the FRFT
with rotation angle k
from two sides
() ()()
kk k
ik ikik
X
uSuNu
 
 (30)
1, 2,,1iM

Equation (30) can be compactly represented by matrix
form as follows
()() ()
kkkk
kkkk
X
uASuNu

 (31)
1
[,,,, ]
kT
kkikMk
Aaaa
 (32)
where, denotes the transpose of matrix.
T
2
2
exp(cos )
2
exp(cos(1) cos)
ikik kk
kk
aj
jid


(33)
12
() [(),(),,()]
kkk k
kkkMk
X
uXuXuXu
 

12
() [(),(),,()]
kkk k
kkkMk
NuNuNuNu

 (34)
From (32) and (33), the direction matrix k
k
A
is only
relative to the direction information k
, so the observed
signal model has been time-variant in the FRF domain.
In the FRF domain, the covariance matrix of the ob-
served signal can be defined as
2
[() ()]
kk kkkk
HH
XXkkkSS k
REXuXuARA
 
I
 (35)
where,
H
denotes the conjugate transpose of matrix.
k
SS
R
is the auto-correlation matrix of signal sources. The
composite covariance matrix (35) has the same structure
as the covariance matrix arising in the case of stationary
signals. Therefore, the DOA can be estimated by per-
forming eigendecomposition to k
X
X
R
. Using the signal
subspace k
N
S
and the noise subspace k
N
E
, the space
spectrum function of the th source in the FRF domain
can be given by [13]
k
()1/( )
kkkk
HH
kkNN
PAEE

k
A
(36)
()
k
P
is performed an 1-D search and k
can be obtain
by the maximal peak rotation angle. Similarly, all the
Direction of LFM signals can be estimated in turn. This
k
 
C
opyright © 2009 SciRes. WSN
H. T. QU ET AL.
176
algorithm is considered as FRFT based demodulation
method.
To summarize, the proposed algorithm can be formu-
lated as follows:
1) The observed signals at all sensors are rotated with
a continuously variable angle
by the FRFT; per-
form a 2-D peak search in the(, )m
plan to obtain
the maximal peak position and corresponding
rotation angle
ik
m
'
k
respectively.
2) Mask operations are applied according to at
every sensor, then the filtered
ik
m
2
L
points are per-
formed the FFT to obtain stationary signals conse-
quently.
3) Get the covariance matrix of the stationary signals
and perform eigendecomposition in the FRF do-
main, construct the spectrum function()
k
P
ac-
cording to (36).
4) Perform 1-D peak search to()
k
P
and obtain the
DOA of the th LFM signal.
k
5) For multi-component LFM signals, all the direction
can be estimated by repeating the above proce-
dures.
4. 2-D DOA Estimation Algorithm Using UCA
4.1. Introduction
UCA has many advantages which the linear array cannot
match. E.g. UCA can be implemented with all-direction-
funding; its precision measurement does not change with
the azimuth significantly and is fit for the system cor-
recting. UCA is the main receiving antenna of base sta-
tion system in the third generation mobile communica-
tion system. Thus, the UCA based DOA estimation has
been a research hotspot in array signal processing.
Mathwes [14] proposed an UCA-RB-MUSIC method,
which can be only suitable for the stationary signals;
however, the actually existed signals are non-stationary
ones which are represented by LFM. Tao ran [4] pro-
posed an algorithm of LFM signal DOA estimation.
However, the method does not apply to the UCA.
Due to the above analysis, we propose a novel DOA
estimation algorithm based on FRFT using UCA, as for
the multi-component LFM signals, using the characteris-
tics of time-frequency rotation and demodulation of
FRFT. Firstly, the observed signals are demodulated into
a series of single frequency ones; secondly, operate the
beam-space mapping to the single frequency signals in
FRF domain, which UCA in array space is changed into
the virtual ULA in mode space; finally, the DOA estima-
tion can be realized by the traditional spectral estimation
method. The proposed algorithm mines the time, fre-
quency and spatial information maximally; compared
with other method, the complex time-frequency cluster
and the parameter matching computation are avoided;
meanwhile enhance the precision [15]. As for the
multi-component LFM signals, there is no cross-term
interference, the proposed algorithm is also applicable
for the multi-path and Doppler frequency shift channels.
4.2. UCA Array Model
Assuming independent LFM signals and the pitch
and azimuth angle is
D
1122
{( ,),(,),,(,)}
DD


N
re-
spectively, the array element number of UCA is and
radius is , the center is the reference point of receiving
antenna, as shown in Figure 4. Then the output of the
th sensor is:
r
i
1
()[] ()
D
ikik
ki
x
tst n

t (37)
1, 2,,iN
 1, 2,,kD
where,
2
()exp[(/ 2)]
kkk
stj tt


(38)
sincos()/
ikkk i
rc

(39)
2( 1)/
iiN
(40)
()
k
s
t is the kth LFM source, and k
and k
are the
initial frequency and FM rate respectively, ik
is the
path delay and is the light velocity. is the ad-
ditive white Gaussian noise with zero mean and variance
c()
i
nt
2
, which is independent with signals.
From the Equations (37) and (38), the direction matrix
of observed signals is time-varying in UCA, while the
traditional DOA estimation algorithm is only suitable for
the time-invariant model, which cannot be used to deal
with LFM signal directly.
4.3. 2-D Estimation Algorithm Description
From Equations (26) and (28), operate the FRFT to Equa-
tion (37) with the rotation angle k
from two sides:
() ()()
kk k
ik ikik
X
uSuNu
 
 (41)
DOA
X
Y
Z
r
k
θ
k
i
Figure 4. Uniform circular array.
Copyright © 2009 SciRes. WSN
H. T. QU ET AL. 177
The matrix form of Equation (41) is:
()() ()
kkkk
k
X
uASuNu


(42)
1
[,,,, ]
kT
kikNk
Aaaa
  (43)
where, donates the transpose of matrix.
T
2
2
exp(cos )
exp(2 sin cos()cos/)
ikik kk
kki k
aj
jr

 

(44)
From Equations (43) and (44), in appropriate FRF
domain, direction matrix k
A
is only related to angle
information
,
, viz. the observed signals have been
transformed into unvaried smooth signal model. There-
fore, the mode excitation method can be used to estimate
the DOA of LFM signals.
The spatial beam former
H
r
F
in FRF domain is de-
fined as
H
H
rH
F
QCeR (45)
where,
H
denotes the conjugated transpose of matrix.
101
,,, ,,,
M
Cediag jjjjj

 
M
(46)
(,,,,)
H
H
MoM
RNQQQ
  (47)
Select the central Hilbert matrix,
0
'
1[(),,(),, ()]
MM
Qv vv
M

  (48)
0
() [, ,,,,,]
jMjj jjM
ve eeee


 
(49)
'
2/
t
M

(50)
[,tMM ]
where, the largest model number ,
M
kr'21
M
M
.
Wave number 2/,k
is the initial frequency
corresponding center wavelength of LFM signals.
H
r
F
can change the UCA in the array space into the virtual
ULA in the mode space, and finally, the DOA estimation
can be achieved by the eigendecomposition based search
method.
Summarize the above and the main steps are as fol-
lows:
1) The observed signals are continuously operated by
FRFT; perform a 2-D peak search in the (, )m
plan to obtain the maximal peak position ik
m and
corresponding rotation angle '
k
of the kth
LFM signal respectively.
2) Select 2L points whose center is ik
m and cal-
culate the FFT (FRFT with1p), obtain the kth
single frequency signal()
k
ik
X
u
.
3) Let ()
k
ik
X
u
pass the beam switch
H
r
F
, viz.
() ()
() ()
kk
kk k
H
ikr ik
HH
rk r
Yu FXu
F
AS uFN u

 

,
And calculate its covariance matrix
[()()
kk
]
H
Yikik
REYuYu

.
4) Define Re( )
Y
RR
, perform eigendecomposition
to R and obtain the signal subspace S and
noise subspace G. Construct:
1
(,)(,) (,)
kk TT
ik k kik k k
PaGGa


,
where, (,)(,)
H
ikk krk k
aFa

k
, perform 2-D
spectrum search and obtain
and k
.
5) As for the multi-component LFM signal, repeat the
above process and obtain all the DOA of signals
respectively.
5. Performance Analysis and Simulation
5.1. FRFT Property Simulation
5.1.1. Simulation of FRFT and WVD
As we all know, WVD is also one of the most important
and most widely used time-frequency analysis tool,
which is bound to FRFT with the existence of close ties.
The derivation process is relatively complex; however,
there is a very simple relationship between FRFT and
WVD, that is, FRFT of WVD is the coordinate’s rotation
form of WVD of original signal [10].
In order to validate the relationship between the FRFT
and WVD, experiments of compute simulations are
given. We assume a wideband LFM signal ()
s
t with a
length of 1024, which is modeled as: initial frequency
and FM rates are9
M
Hz
,0.7/
M
Hz s
 , sample
frequency is 50
s
f
MHz
. The WVD of ()
s
t is shown
in Figure 5 (a), ()
s
t
5
is performed the FRFT by the rota-
tion angle 0.1
and get the transformed signal
. The WVD of the transformed signal
is shown in Figure 5(b). Compared the two figures, it can
be found that the WVD of is just the rotation
of the WVD of
0.15 ()Su
0.15 ()Su
0.15 ()Su
()
s
t by angle 0.15
, meanwhile the
figure shape is invariable. So the FRFT is testified a kind
of rotation arithmetic operators in the time-frequency
plane.
5.1.2. Two Special FRF Domain Simulation
2
( )exp[(2)]stjtt

, signal model is: 19
M
Hz
,
11400000 /
M
Hz s
. Sampling rate 50
s
f
MHz, the
number of snapshots is 1024. Perform continuous FRFT
to signal and operate spectrum peak search, in the appro-
priate FRF domain, ()
s
t shows the ergy property of en
C
opyright © 2009 SciRes. WSN
H. T. QU ET AL.
178
(a)
(b)
Figure 5. (a) The WVD of The WVD of ()t, (b) s0.15π()
S
u.
(b)
Figure 6. (a) Energy concentration property of FRFT, (b)
Demodulated property of FRFT.
concentration, as shown in Figure 6(a). The signal con-
tinues to be rotated 2
ain,
in FRF domain, viz. in the
demodulated FRF dom()
s
t shows the demodulated
property, as shown in Figure 6(b).
5.1.3. Gaussian White Noise Simulation
Assume the complex Gaussian white noise is:
( )(1,1024)(1,1024)w nrandnjrandn
continuous FRFT to it, the energy distrib
and perform
ution of
in different FRF domain is shown in Figure 7. We can
see that the Gaussian white noise does not show energy
concentration property in any FRF domain and can still
be regarded as white noise.s theorem 1 is verified.
()wn
Thu
Figure 7. Energy distribution in different FRF domain.
(a)
Copyright © 2009 SciRes. WSN
H. T. QU ET AL.
179
5.2. 1-D DOA Estimation Simulation
5.2.1. MSE and CRB Analysis
The FRFT is a 1-D linear transform [10]. In the FRF
domain is approximately considered as the
additive Gauss white noise. Therefore, the probability
density of signal
()
k
k
Nu
function()
k
k
X
u
represents normal
school and the corresponding likelihood function can be
expressed as
2
2
1
[]
(2)(/ 2)
1
exp{[][]}
k
kkk kkk
kMM
H
kkk kkk
LX
XASXAS


 
(51)
Using Reference [5], the CRB of the proposed method
in the FRF domain can be represented as
1()RB
1
2kkkk
k
HH H
2Re ()( )()( )
k
H
kkkkk k kk
C
SdwIAAAAdwS
k
 

 


(52)
where, 2
is the noise variance and
I
is unit matrix,
() /
k
kk
d wdAdw
.
Similarly, the MSE of the ed algorithm in the
FRF domain can be represented as
propos
1
()VAR
 1
2
121 11
11
2[ ()[()]
()]/{[ ][ ()]}
kk kk
kkkkk
HHH
MU kkkkkk
H
kXXXXkkX
dw IAAAA
dwRRA AR
 



where,
11X
(53)
k
X
X
R
is covariance matrix of the observed signals,
e MSE of
the proposed method will be more and more closed to the
CRB with the increasing of the sensor number and the
SNR.
5.2.2. MSE and CRB Simulation
tod
u
impinging from
11 denotes the first row and first line element of matrix.
From (52) and (53), it can be obtain that th
[]
In order validate the proposed metho, experiments of
compute simulations are given. We assme the ULA of
6M 2D
far field wideband
ar
LFM signals with a length of 1024, which is modeled as:
initial frequency and FM rates e1200
H
z
1900 /
H
zs
 ;2200
H
z
2300 /
H
zs
0
270
respec-
,
and angles
tively.
of arri
Sample fre
val are 1
quency is
0
30
s900
f
Hz
e FRF dom
, the mask snap-
in. The input Sshots are 23L00 in thaNR
varies fr15dB to 29dB with an interval 2dB, at each
level of the SNR, we run 100 Monte-Carlo experiments,
MSE of the proposed method and original method are
Figure 8. MSE of proposed and original method.
Figure 9. MSE and CRB of proposed method.
shown in Figure 8. Obviously, the accuracy of our
method has certain improvement comparing with the
method proposed in the Reference [4].
In same assumption, the input SNR various from
–15dB to 6dB with an interval 3dB, 100 times
Monte-Carlo simulations are performed at each level of
the SNR, MSE of the first signal and CRB are shown in
Figure 9. It can be seen, the MSE of proposed method is
closed to the CRB even at the lower SNR.
5.3. 2-D DOA Estimation Simulation
5.3.1. 2-D Estimation RMSE Simulation
2D
two far-field LFM sources shoot the 20N
00
50 ),
1
UCA
(3
om
the
with the angle information
}. The signa
11
{(60 ,

l model is:
00
11
0, 70)


C
opyright © 2009 SciRes. WSN
H. T. QU ET AL.
180
200
H
z
, 1300 /
H
zs
; 2200
H
z
, 2
900 /
H
zs
. Sampling rate is s
f
900
H
z, number
of snapshots is 1024, and the cover filter length is
2300L. The Figure 10(a) gives the 2-D DOA
tio
RMSE (root
ean square error, RMSE) comparison curves of the
be seen in
igure 10(b). The accuracy of our method has certain
thm.
estima-
n of signal one in the 0dB SNR.
Change the input SNR range from 0dB to 20dB with
the interval 5dB, firstly perform big step search to obtain
the rough DOA estimation. Then run the high differen-
tiation search with the 0.001rad step. Run 300 time
Monter-Carlo experiment respectively, the
m
proposed algorithm and literature one can
F
improvement compared to the original algori
(a)
(b)
Figure 10. (a) 2-D DOA estimation using UCA, (b) RMSE
comparison curves using UCA.
5.3.2. 2-D Estimation Performances in Complex
Channel and Simulation
In mobile communication system, the proposed algo-
rithm is applied to the complex channel which the multi-
path and Doppler shift is existed simultaneously. In the
same simulation conditions, viz. the random signal
source model is:
(54)
where,
2
1
()exp() exp[(()()/2)]
E
keekeke
e
stMjftj tt


12
1, 0.9MM
,
12
2ff
Doppler frequency shift is
0,
, multi-path delay is 12
0,1/ 900
 .
When the SNR is 0dB, the simulation result of signal one
in most powerful path can be shown in Figure 11(a).
(a)
(b)
Figure 11. (a) 2-D DOA estimation in complex channel, (b
RMSE curves in complex chl.
)
anne
Copyright © 2009 SciRes. WSN
H. T. QU ET AL.
Copyright © 2009 SciRes. WSN
181
Change the input SNR range from –21dB to 0dB with
the interval 3dB. Run 300 time Monter-Carlo experiment
respectively, the RMSE comparison curves of signal one
can be seen in Figure 11(b), which can show that the
proposed algorithm is also effective in complex channel.
6. Conclusions
Analyzing the definition and characteristics of the FRFT,
a novel DOA estimation algorithm has been presented
the implementation of the method, mask operation is
introduced to simply the filtering procedure with no ac-
curacy degradation. Demodulation operation is us
extend the application range of the traditional estimate
method without performance loss. Compared with other
methods, the veracity has certain improvement while th
cross-terms and interpoided. The prop
is also expanded to thetimation using UCA,
addition, the pro-
po
aking this method more reliable in theory and in prac-
rich the principle and applicatio
he optimization, the 2-D Cramer-R
is-
2000.
. S. Zhou, “A novel method for the DOA
” IEEE Transactions on ASSP, Vol.
p. 2395–
. In
37, No. 5, May 1989.
[6] V. Namias, “The fractional Fourier transform and its
ed to
application in quantum mechanics [J],” IMA Journal of
Applied Mathematics, No. 25, pp. 241–265, 1980.
[7] L. B. Almeida, “Fractional Fourier transform and
time-frequency representations [J],” IEEE Transactions
e
olation are av
2-D DOA es
osed
on Signal Processing, Vol. 42, No. 11, pp. 3084–3091,
1994.
[8] Y. Q. Dong, R. Tao, S. Y. Zhou, et al., “SAR moving
target detection and imaging based on fractional Fourier
which is suitable for the multi-path and Doppler fre-
quency shift complex environment. In
sed the method can be also applied to DOA estimation
of LFM signals in colored noise or near-field environ-
ment, which is not described in this paper.
The theoretical analyses about the error and CRB are
also provided and verified by simulation results thus
fil
m
tice, meanwhile en
e FRFT. As for t
n of
ao
[10] R. Tao, L. Qi, and Y. Wang, “Principle and application of
the fractional Fourier transform,” Tsinghua Publishing
Company, Beijing, 2004.
[11] H. M. Ozaktas, O. Arikan, M. A. Kutay, et al., “Digital
th
Bound of the proposed algorithm is the further research
direction.
7. Acknowledgements
The authors would like to thank the reviewers for their
detailed comments on earlier versions of this paper.
8. References
[1] G. Wang and X. G. Xia, “Iterative algorithm for direction
of arrival estimation with wideband chirp signals,” IEE
Proceedings of Radar, Sonar, Navig, Vol. 147, No. 5, pp.
233–238, 2000.
[2] A. Belouchirani and M. G. Amin, “Time-frequency MU-
SIC [J],” IEEE Signal Processing Letters, Vol. 6, No. 5,
pp. 109–110, 1999.
[3] A. B. Gershman and M. G. Amin, “Wideband direction
of multiple chirp signals using spatial time-frequency d
tributions,” IEEE Signal Processing Letters, Vol. 7, pp.
152–155, June
4] R. Tao and Y[
estimation of wideband LFM sources based on FRFT,”
Transactions of Beijing Institute of Technology, Vol. 25,
No. 10, pp. 895–899, 2005.
[5] P. Stoica and A. Nehorai, “Music, maximum likelihood,
and cramer-rao bound,
transform,” Acta Armamentarii (in Chinese), Vol. 20, No.
2, pp. 132–136, 1999.
[9] L. Qi, R. Tao, S. Y. Zhou, et al., “Adaptive time-varying
ter for linear FM signal in fractional Fourier domain,”
Proceedings of the 6th ICSP, Posts and Telecommunica-
tions Press, Beijing, pp. 1425–1428, 2002.
computation of the fractional Fourier transform,” IEEE
Transactions on Signal Processing, Vol. 44, No. 9, pp.
2141–2150, 1996.
[12] L. Qi, R. Tao, S. Y. Zhou, et al., “Detection and parame-
ter estimation of multicomponent LFM signal based on
the fraction Fourier transform [J],” Science in China (Se-
ries E), Vol. 47, No. 2, pp. 184–198, 2004.
[13] X. D. Zhang, et al., Modern Signal Processing (Second
editor), Tsinghua Publishing Company, Beijing, 2002.
[14] C. P. Mathews, “Eigenstructure techniques for 2-D angle
estimation with uniform circular arrays,” IEEE Transac-
tions on Signal Processing, Vol. 42, No. 9, p
2407, September 1994.
[15] L. M. Yang “DOA estimation for wideband sources
based on UCA,” Journal of Electronics, China, Vol. 23,
No. 1, January 2006.