You are on page 1of 13

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Joint 2D-DOA and Carrier Frequency


Estimation Technique Using Nonlinear
Kalman Filters for Cognitive Radio
SAMAR ELARABY1 , HEBA Y. SOLIMAN 1 , (Member, IEEE), HEBA M. ABDEL-ATTY 1 ,
(Member, IEEE), AND MOHAMED A. MOHAMED 2
1
Electrical Engineering Department, Faculty of Engineering, Port Said University, Port Said, Egypt
2
Electronics and Communication Engineering Department, Faculty of Engineering, Mansoura University, Mansoura, Egypt
Corresponding author: Samar Elaraby (e-mail: samar.elaraby@eng.psu.edu.eg).

ABSTRACT The problem of jointly estimating carrier frequencies and their corresponding two-dimension
direction of arrivals (2D-DOA) of band-limited source signals is considered in this paper for cognitive
radio (CR). The main problem of estimating carrier frequencies spread over a wideband spectrum is the
requirement of high sampling rates. Thus, Kalman filters are applied in the spatial domain instead of
the temporal domain in the proposed algorithm to relax hardware complexity. The proposed algorithm
exploits both azimuth and elevation angles instead of a single DOA to increase the spatial capacity. Two
approaches are proposed using two different types of nonlinear Kalman filter: extended Kalman filter (EKF)
and unscented Kalman filter (UKF). Using simulations, the factors that affect the performance of both
two filters are discussed. Scaling the estimated parameters to the same range and the proper tuning and
initialization of the filters are crucial factors to prevent the filter divergence. Although UKF is supposed
to have a better performance than EKF, reducing the inter-element spacing of the employed arrays and the
proper filter initialization can make EKF approach the performance of UKF. On the other hand, UKF suffers
from high processing time. In overall, both two filters are able to converge to the true values of the unknown
parameters using a number of relaxed analog-to-digital converters (ADCs) equal to the number of the array
elements in the employed arrays. However, both two approaches cannot detect a number of source signals
higher than one-third of the total number of the array elements.

INDEX TERMS 2D-DOA Estimation, Cognitive Radio, Extended Kalman Filter, Spectrum Sensing,
Unscented Kalman Filter.

I. INTRODUCTION spectrum sensing. Besides, while CR is transmitting on a


particular frequency band, it continues to sense the spectrum
ECENTLY, the problem of radio frequencies shortage
R has grabbed the attention of researchers as radio de-
vices have increased tremendously. In the next few years,
to check for the presence of the licensed user and prevent the
interference with it. This way, the spectrum can be exploited
wisely and many CRs can share the limited resources. In the
radio applications and devices are expected to exponentially
last decade, many proposals for spectrum sensing have been
grow due to the growth of Internet of things (IoT) and
presented in the literature [2] and [3].
machine to machine (M2M) communication. In IoT, millions
of edges and radio devices are going to communicate on To increase the number of CRs that can share the un-
radio frequencies, however the existing spectrum is limited occupied frequency bands in the same area, multiantenna
and cannot be expanded to embrace the rising demand. As a techniques have been suggested to exploit the spatial domain.
result, the opportunistic spectrum access has been suggested Then, CR can detect the presence of the licensed users and
to solve the frequency band shortage. CR is one of the their transmitting directions. As a result, the spectral and spa-
technologies that implement opportunistic spectrum access tial domains can be exploited efficiently. The main challenge
[1]. CR is a radio device that detects the frequency bands facing the problem of estimating both spectral and spatial do-
left unoccupied by their licensed users. This process is called mains is the need to sample a wideband spectrum at Nyquist

VOLUME 5, 2017 1

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

rates. Nyquist rates require a high-speed ADC and generate rectangular array (URA). Moreover, the PARAFAC analysis
a large number of samples to be processed. Thus, many has been proposed for carrier frequency and 2D-DOA esti-
proposals have been presented at sub-Nyquist sampling rates. mation with a conformal array in [20] and [21]. In [22], the
In [4] and [5], the authors have proposed new architectures authors have proposed iterative least square methods for both
that can sample a wideband spectrum at sub-Nyquist rates frequency and 2D-DOA estimation problem.
with relaxed hardware requirements. Using ESPRIT [6] and In this paper, the problem of jointly estimating the carrier
MUSIC [7] algorithms, the authors have succeeded to esti- frequencies and their corresponding azimuth and elevation
mate carrier frequencies and the corresponding DOAs. In [8], angles of uncorrelated band-limited source signals is consid-
the authors have refined the estimated carrier frequencies and ered. To the best of our knowledge, Kalman filter (KF) has
the corresponding DOAs using 2D-iterative grid refinement. not been exploited in this problem before. KF is a recursive
In contrast, the authors in [9] have executed sub-Nyquist algorithm that can estimate unknown parameters from a state
sampling using a modulated wideband converter (MWC) space model with noisy measurements [23], and the state
channel [10]. The authors have then proposed two algorithms space model can be formulated in any domain. Since the
to detect carrier frequencies and the corresponding DOAs. considered problem has to tackle the sampling of a wideband
The first algorithm depends on compressive sensing, which spectrum, we propose investigating the spatial domain using
was exploited to reconstruct the spectrum before estimating KF instead of relying on sub-Nyquist methods in the tem-
the desirable parameters. The second one is based on parallel poral domain. Two L-shaped uniform arrays are employed
factor (PARAFAC) analysis method [11]. A third algorithm, in our proposal to form the spatial state space model that
which depends on ESPRIT algorithm, has been proposed in KF would follow to predict the desired unknown parameters.
[12]. Thus, one time sample at each array element is only required
The problem of these proposals is the limitation on the to identify the time delay encountered by any source signal
degrees of freedom; thus the number of detected sources between any two successive array elements. As a result, our
is limited to the number of array elements. Increasing the proposal does not have to sample a wideband spectrum at
degrees of freedom requires employing sparse antennas. So, Nyquist or sub-Nyquist rates leading to gaining an advantage
the authors in [13] have employed a 2D nested array to raise over the other methods in the literature in terms of hardware
the degrees of freedom. However, the proposal in [14] has in- complexity. In other words, there are not any restrictions
creased the degrees of freedom using several MWCs at each on the sampling rate, and hence the speed of the employed
element in the antenna. Since there is no need to reconstruct ADCs has no restrictions as well. Furthermore, the required
the whole spectrum to detect the existing carrier frequencies, number of the relaxed ADCs by sub-Nyquist methods is
the authors in [15] have decided to detect carrier frequencies reduced to the number of the employed array elements. On
and their corresponding DOAs from a reconstructed 2D the other hand, since our proposal relies on a uniform dense
power spectrum. The 2D angular power spectrum has been array, it has limited degrees of freedom. The degrees of free-
reconstructed from sub-Nyquist samples generated by multi- dom are found to be equal to one third of the employed array
coset samplers. Besides, the authors have compressed the elements. To implement our proposal, we have exploited two
spatial measurements as well using a minimum redundancy different types of nonlinear Kalman filters, extended Kalman
linear array to increase the degrees of freedom. filter (EKF) [23] and unscented Kalman filter (UKF) [24].
Furthermore, a more efficient scenario has addressed the Since nonlinear KFs are sub-optimal estimators, they may
idea of investigating 2D-DOA instead of a single DOA, tend to converge to wrong values or even completely diverge.
which represents a single direction around CR. Then, a CR The performance of the two algorithms is examined under
can share the same carrier frequency and azimuth angle with different conditions using simulations, and the factors that
a licensed user, but they differ in the elevation angles. This, can affect their performance are discussed.
in turn, increases the spatial capacity. The problem of jointly The paper is organized as follows. Section II describes the
estimating 2D-DOA and carrier frequencies of primary users two L-shaped array model. Kalman filter and its different
(PUs) has been discussed in the literature. ESPRIT algorithm types are then visited in Section III. Section IV declares
is widely applied on this problem. In [16], the authors have a detailed analysis of our proposal. Proposed spatial state
relied on ESPRIT algorithm to detect carrier frequencies space model is first derived. Then, a detailed description of
from sub-Nyquist samples of source signals that impinge on a both EKF-based and UKF-based approaches is presented.
uniform circular array. Then, 2D-DOA can be estimated. The Simulations with discussions are addressed in Section V.
authors in [17] have also implemented ESPRIT algorithm to Finally, Section VI derives final conclusions.
estimate 2D-DOA using an L-shaped array. However, carrier
frequencies have been estimated by singular value decompo- II. SYSTEM MODEL
sition technique (SVD). In [18], ESPRIT algorithm has been Consider two L-shaped uniform arrays located in the x-z
extended to estimate both carrier frequencies and 2D-DOA. plane and the y-z plane as shown in Figure 1. Each single
In [19], another ESPRIT-based algorithm has been proposed uniform linear array (ULA) in this structure has N elements
for estimating the carrier frequencies and their corresponding with an inter-element spacing of d. The element at the origin
2D-DOA of multi-band signals from the outputs of a uniform is considered as a reference point. Suppose that L different
2 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

z
ml (t) space model and measurement space model as follows
N b

xn+1 = Axn + wn

. . . . . .
(2)
b
Yn = Hxn + un
1 b θl where xn+1 and xn denote posterior and prior estimate of
b b b b b y the state variable respectively. Matrices A and H represent
1 b 0 1 . . . . . . N transition matrix and observation matrix respectively. How-
b
.

φl ever, wn and un respectively denote process noise and mea-


.

b
.
.
.

surements noise signals, which are assumed to be Gaussian


.

N b

x distributed with zero mean and covariance matrix of Q and


FIGURE 1: Two L-shaped uniform array model R respectively. The process noise covariance matrix Q repre-
sents uncertainty in process space model. However, the mea-
surement noise covariance matrix R represents uncertainty
in observed measurements. While KF estimates the state
band-limited signals from L different uncorrelated sources
variable, it can estimate unknown parameters as well. As a
are impinging on the two L-shaped arrays. Each source signal
result, it is used intensively in estimation problems where the
is transmitted from a different direction on a different carrier
unknown parameters are concatenated to the state variable.
frequency. Then, the output of each element can be evaluated
Then, they can be estimated with it through iterations.
as
In many estimation problems, the state space model may
L cos φl sin θl
have a nonlinear process model and/or a nonlinear measure-
−j2π(n−1)d
X
rxn (t) = ml (t)e λl
+ ηxn (t) ment model. The nonlinear state space model has the form
l=1 of
L
X −j2π(n−1)d
sin φl sin θl xn+1 = f (xn ) + wn
ryn (t) = ml (t)e λl
+ ηyn (t) (1) (3)
l=1
Yn = h(xn ) + un
L cos θl where f (.) and h(.) represent nonlinearity in the process and
−j2π(n−1)d
X
rzn (t) = ml (t)e λl
+ ηzn (t) measurement models respectively. For nonlinear state space
l=1 models, traditional KF fails to converge to the true values
and modifications to KF are required resulting in sub-optimal
where rxn (t), ryn (t) and rzn (t) are the outputs of the nth element performance. Two of the renowned nonlinear KFs are EKF
in the ULA located on x-axis, y-axis and z-axis respectively. and UKF.
The signal ml (t), with l = 1, 2 . . . L, denotes the received In EKF, nonlinear models are linearized about the esti-
source signal at the reference point from the lth source. mated trajectory. It relies on Taylor series to execute lin-
The signal ml (t) has a wavelength of λl and arrives from a earization and the resultant is approximated to the first order
direction with an azimuth angle of φl and an elevation angle derivatives. The detailed EKF algorithm can be found in [23].
of θl . Noise signals ηxn (t), ηyn (t) and ηzn (t) are assumed to be The linearization process results in significant errors in the
complex Gaussian white noise with zero mean and variance estimated state variable and these errors are accumulated
of σn2 . through iterations. On the other hand, the filter updates
Since CRs are not allowed to have any prior information the estimated state variable each iteration based on noisy
about licensed users being detected, the problem considered measurements. This, in turn, leads the estimated trajectory
in this paper is a blind estimation problem. The CR has to to follow the noisy measurements. As a result, the filter may
blindly estimate the carrier frequency and the corresponding converge to erroneous values or even completely diverge.
2D-DOA of the surrounding licensed users. For this problem, To enhance the performance of EKF, the filter should be
we propose Kalman filter for estimating λl , φl and θl of each tuned and initialized properly. Tuning the filter means to
licensed user. choose a proper estimate for Q and R matrices to wisely
build confidence in both process and measurement models
III. KALMAN FILTERS: AN OVERVIEW respectively. Furthermore, we propose wisely choosing the
KF is an algorithm that optimally estimates an unknown state parameters to be concatenated to the state variable instead
variable from a state space model based on noisy measure- of directly choosing carrier frequencies, azimuth angles and
ments. First, KF algorithm predicts the posterior estimate of elevation angles themselves. In other words, the filter may
the state variable from the previous estimate. Then, the pre- converge to true estimates and its performance approaches
dictions are adjusted according to observed measurements. the performance of traditional KF, if the unknown parameters
The two steps are repeated till the filter converges to the true are scaled to the same range with a small variance relative to
estimate of the state variable. KF is an optimal algorithm as a region where the model is relatively linear. In this range,
it depends on linear state space models. A linear state space linearization errors can be reduced and the filter performance
model can be generally described as a combination of process can be enhanced. Simulations are presented to prove the
VOLUME 5, 2017 3

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

vital role of the proper initialization in enhancing the filter where xn+1 n+1
2l−1 and x2l represent real and imaginary parts of
n+1
performance. Xl . Similarly,
In UKF, Gaussian-distributed state variable is captured by   sin φl sin θl   sin φl sin θl  

 n+1  cos 2πd sin 2πd n

a minimal set of sample points, called sigma points. These y2l−1  λ l λ l  y2l−1
points are actually capturing both the mean and the variance  =


 
n+1 sin φl sin θl  sin φl sin θl  y2ln
y2l
  
of the state variable. After sigma points are propagated − sin 2πd cos 2πd
λl λl
through the nonlinear filter, their outputs can be used to (6)
perfectly recover the estimated Gaussian state variable. In and
each iteration, the posterior estimates and covariance matrix   cos θl   cos θl  

are evaluated as weighted mean and covariance of the sigma
 n+1 
z2l−1 cos 2πd sin 2πd n

 λ l λl  z2l−1
point outputs. Through iterations, the filter tends to converge  = 

 
to the true values of the state variable. UKF, like EKF, needs n+1
z2l
  cos θ l
  cos θ l
 n
z2l
to be properly tuned and initialized to enhance the filter − sin 2πd cos 2πd
λl λl
convergence. The complete algorithm of UKF is proposed (7)
in [24]. In comparison to EKF, UKF is accurate to the third The state variable xs ∈ R6L×1 is then formed at any
order derivative. Besides, UKF, unlike EKF, does not rely element as a concatenation of the real and imaginary values
on evaluating partial derivatives. However, the complexity of in the three ULAs as follows
UKF is higher than the complexity of EKF as UKF needs to xs = [x1 , x2 . . . x2L−1 , x2L , y1 , y2 . . .
calculate sigma points in each iteration. (8)
y2L−1 , y2L , z1 , z2 , . . . z2L−1 , z2L ]T
Since Kalman filters are considered to predict unknown
IV. PROPOSED APPROACHES
parameters, these parameters should be concatenated to the
In this section, we first derive the spatial state space from state variable. Then in each iteration of the kalman filtering,
Eq. (1). Then, the proposed EKF-based and UKF-based the parameters are estimated as well as the posterior estimate
approaches are presented in detail. of the state variable. The parameters to be estimated in our
problem are carrier frequency, azimuth and elevation angles
A. PROPOSED SPATIAL STATE SPACE of the L sources. These parameters are not going to be
The main challenge of wideband spectrum sensing is to sam- concatenated to the state variable directly as they vary over
ple the signals at high Nyquist sampling rates, which require different wide ranges and may force EKF to diverge. So,
sophisticated high-speed ADCs. Since the array geometry we propose selecting related parameters that share the same
provides an opportunity to exploit the spatial domain, we range to speed up the filter convergence and boost the filter
propose forming a spatial state space instead of a temporal performance. The parameters are selected to be
state space to reduce the complexity associated with the lat- cos φl sin θl sin φl sin θl
ter. The spatial state space is formed from the time delay that , bl =
al =
λl λl
each source signal encounters between any two successive (9)
cos θl
array elements. The time delay is expressed as a phase shift cl =
λl
between the different versions of the source signals arrive the
where l = 1, 2 . . . L. By estimating these parameters using
elements. Therefore, the lth source signal that reaches the
Kalman filter, φl , θl and λl can be evaluated for each source.
(n + 1)th element in the three different ULAs Xln+1 , Yln+1
Moreover, the inter-element spacing d is set to a fraction
and Zln+1 can be determined as
of the minimum wavelength. Reducing d as possible can
cos φl sin θl
result in expanding the sinusoidal model in Eq. (5) - (7)
−j2πd
Xln+1 = e λl
Xln regarding to the unknown parameters. This, in turn, leads
n+1 −j2πd
sin φl sin θl
n the filter to search through a region where the model is
Yl =e λl
Yl (4)
relatively linear. Under this scenario, both EKF and UKF
cos θ
−j2πd λ l n can overcome their sub-optimal performance and converge
Zln+1 = e l Z
l
to more precise values. Using simulation in Section V, an
n n n th enhanced performance is proven while decreasing d.
where Xl , Yl and Zl are the versions of the l source
th Now, the state variable becomes xs ∈ R9L×1 and the
signal reaches the n element. In matrix notation, Eq. (4)
process model of the spatial state space model can be defined
can be reformulated as follows  
αx 0 0 0 0 0
cos φl sin θl cos φl sin θl  0 αy 0 0 0 0
     
 n+1  cos 2πd sin 2πd
 xn2l−1
 
x2l−1 λ λ

0

 l l
n+1 0 αz 0 0 0  xns
 =   x s = 
0 (10)
0 0 I 0 0
 
n+1 cos φ sin θ cos φ sin θ n

x2l x2l
    
l l l l
− sin 2πd cos 2πd
 
λl λl 0 0 0 0 I 0
(5) 0 0 0 0 0 I
4 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

Algorithm 1 EKF-based proposed algorithm


L×L
where I ∈ R is the identity matrix. The submatrices αx , 0: Intialization:
αy and αz ∈ R2L×2L are defined as x̂0 = E[xs ]
   
αx1 0 0 αy1 0 0 P0 = E[(xs − x̂0 )(xs − x̂0 )T ]
αx =  0 .. ..
0  ,αy =  0 0 ,
   
. . 0: loop:
0 0 αxL 0 0 αyL Prediction Step:
 
αz1 0 0 ∂f (xn−1 )
s
.. Fn =
αz =  0 ∂xn−1
  n−1
. 0  s xs =x̂n−1
0 0 αzL x̂−
n = f (x̂n−1 )
(11) P− T
n = Fn Pn−1 Fn

where Updating Step:


cos φl sin θl  cos φl sin θl 
   
cos 2πd sin 2πd Kn = P− T − T −1
n H (R + H Pn H )
λl λl
x̂n = x̂− −
 
αxl = n + Kn (Yn − H x̂n )
 

cos φl sin θl cos φl sin θl  Pn = P− −
n − Kn H Pn
    
− sin 2πd cos 2πd
λl λl if convergence:
for each source:
sin φl sin θl  sin φl sin θl 
    b 
cos 2πd sin 2πd φl = tan−1
l
 λl λl 
al
αyl =
 
 b 
l
sin φl sin θl sin φl sin θl  θl = tan−1
    
− sin 2πd cos 2πd cl
λl λl
cos θl
λl =
cl
cos θl  cos θl 
  
cos 2πd sin 2πd break loop
 λl λl 
αzl = 
 

  cos θl   cos θl 

− sin 2πd cos 2πd
λl λl The spatial state model is now formulated and it obviously
(12) contains a nonlinear process model as in Eq. (3) and a linear
measurement model as in Eq. (2). Thus, sub-optimal EKF
with l = 1, 2, . . . L. and UKF are proposed for the nonlinearity of this problem.
While applying these filters to the proposed state model, the
However, the measurement model of the spatial state space process noise covariance matrix Q is set to zero giving full
represents the measured output at the nth element in all trust in the proposed process model.
arrays. The measured output at the nth element is expected
to be the sum of all incident source signals on this element. B. PROPOSED EXTENDED KALMAN FILTER-BASED
It is, however, exposed to measurement errors and hence the APPROACH
measurement model can be declared in matrix notation as
n n n n n n
In the first approach, we propose EKF algorithm, described in
Yn = [Yx,re Yx,im Yy,re Yy,im Yz,re Yz,im ]T [23], for estimating the state variable xs from the formulated
state space model. First, EKF is initialized with an initial
estimate x̂0 and initial covariance matrix P0 . The initial
 
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 estimate x̂0 is preferable to be selected close to the mean
 
0 0 ... 1 0 ... 0 0 0 ... 0
=  xns + un value of the state variable to speed up the convergence to
0 0 0 1 0 0 0 0

0
 the true values. Moreover, the matrix P0 should be carefully
0 0 0 1 0 0 0
0 0 0 0 0 1 0 0 selected as it may be set to a low value that is enough to force
(13) the filter to fully trust the posterior estimates and neglect the
effect of measurements.
n n n
where Yx,re , Yy,re and Yz,re are the real parts of the nth Then, EKF goes through several iterations and each itera-
element outputs in the x-array, y-array and z-array respec- tion consists of two steps: prediction step that follows the pro-
n n n
tively. The remaining terms Yx,im , Yy,im and Yz,im are the cess model and updating step that follows the measurement
imaginary parts of these outputs respectively. The vector un model as shown in Algorithm 1. In each prediction step, the
represents the measurement noise signals which has zero- nonlinear process model must be linearized around the prior
mean and covariance matrix R. estimates. Then, the approximated resultant to the first order
VOLUME 5, 2017 5

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

derivatives is calculated as a Jacobian matrix Fn predicted from their prior values x̂n−1 and Pn−1 respec-

f1 0 0 f4 0 0
 tively. The predicted estimates are then updated depending
 0 f2 0 0 0 f5 on the filter gain Kn and observation matrix H as shown in
∂f (xn−1 )

 0 0 f3 0

f6  0 Algorithm 1. The filter then goes through several iterations
s
Fn = n−1 = 
0 0 0 I
 of prediction and updating, till it finally tends to converge.
0 0

∂xs n−1
xs =x̂n−1 
0 0 0 0 0

I The filter converges to the true values of al , bl and cl since
0 0 0 0 I 0 the estimated parameters have a small variance compared to
(14) a region where the model is relatively linear. This, in turn,
where I ∈ RL×L is the identity matrix. The matrices forces EKF to approach the performance of traditional linear
f1 , f2 , f3 , f4 , f5 and f6 are defined as follows KF. The unknown carrier frequencies and 2D-DOA angles
  are finally calculated from the estimated al , bl and cl .
..
 . 0 0
f1 =  0 0 C. PROPOSED UNSCENTED KALMAN FILTER-BASED
β 1l , l = 1, 2, . . . L (15)
APPROACH

..
0 0 .
The second approach proposes UKF algorithm, described in
where [24], for the considered estimation problem. The complete al-
 ∂xn ∂x2l−1n  gorithm is described in Algorithm 2. First, the initial estimate
2l−1
 ∂x n−1 n−1 x̂a0 and covariance matrix Pa0 are initialized properly. The
 2l−1 ∂x2l 

β 1l =   initial estimate x̂a0 is a concatenation of the initial estimate of

 ∂x2l n
∂x2l n
 the state variable x̂0 and the initial estimate of measurement
noise. This, in turn, leads the initial covariance matrix Pa0
∂xn−1
2l−1 ∂xn−1 2l n−1
xs =x̂n−1
(16) to gather both the covariance matrix of the state variable
P0 = E[(xs − x̂0 )(xs − x̂0 )T ] and the measurement noise
covariance matrix Pυ (equivalent to R in EKF).
    
cos 2πdân−1
l sin 2πdân−1 l
  Then, the filter rotates in several iterations till the filter
=   

 converges to the true values. In each iteration, sigma points
− sin 2πdân−1l cos 2πdân−1 l χai,n−1 , with i = 0, 1 . . . 18L, are selected to perfectly cap-
ture the mean and variance of the state variable. Since a state
Similarly, the remaining matrices f2 , f3 , f4 , f5 and f6 are
variable of dimension M needs 2M+1 sigma points, the state
diagonal and their diagonals have the elements β 2l , β 3l ,
variable xs needs 18L+1 sigma points. These sigma points
β 4l , β 5l and β 6l respectively. In the same manner, they are
are evaluated
p as shown in Algorithm 2 by determining the ith
evaluated as 
     row of (6L + Λ)Pn−1 using Cholesky decomposition.
cos 2πdb̂n−1
l sin 2πd b̂n−1
l The parameter
i
Λ is a scaling parameter which equals to
 
β 2l = 
    
 α2 (6L + κ) − 6L where α denotes the spread of sigma points
− sin 2πdb̂n−1 cos 2πdb̂n−1 around x̂n and κ is a secondary scaling parameter (usually
l l
     equals to 0).
cos 2πdĉn−1
l sin 2πdĉ n−1
l After selecting the sigma points, the filter executes predic-
tion step in which the sigma points are propagated through
 
β 3l = 
 
the nonlinear process model, and the resultant is used to
   
n−1 n−1
− sin 2πdĉl cos 2πdĉl
predict the posterior estimate x̂− −
n . The posterior estimate x̂n

     
− sin 2πdân−1 cos 2πdân−1
 n−1
 and covariance matrix Pn are evaluated as weighted mean
l l  x̂2l−1
 and covariance matrix of the resultant. The employed weights

β 4l = 2πd  
     n−1 are defines as
− cos 2πdân−1 − sin 2πdân−1 x̂2l
l l (m) Λ
w0 =
6L +Λ
     
n−1 n−1
− sin 2πdb̂l cos 2πdb̂l
 n−1 
  ŷ2l−1
β 5l = 2πd   
     n−1 (c) Λ
− cos 2πdb̂n−1 − sin 2πdb̂n−1 ŷ2l w0 = + (1 − α2 + β) (18)
 
l
 
l
  6L + Λ
− sin 2πdĉn−1 cos 2πdĉn−1
 n−1 
 l l  ẑ2l−1 (m) (c) 1
β 6l = 2πd    wi = wi = i = 1, 2, . . . 18L
     n−1
ẑ2l 2(6L + Λ)
− cos 2πdĉn−1 l − sin 2πdĉ n−1
l
where β represents the state variable distribution (β = 2 for
(17) Gaussian distribution).
After performing linearization by evaluating Jacobian matrix Then, the filter executes updating step where the posterior
Fn , the posterior estimate x̂− n and covariance matrix P −
n are estimate and covariance matrix are corrected by the measure-
6 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

ments. They are updated based on the difference between the


predicted observations yn− and the actual measurements Yn .
The predicted observations are evaluated as a weighted mean
Algorithm 2 UKF-based proposed algorithm of the resultant of the measurement model after the sigma
0: Intialization: points are propagated through it. The complete algorithm
x̂a0 = [x̂T0 0]T used in prediction and updating steps are shown in Algorithm
  2. The filter keeps going in iterations of selecting the sigma
P0 0
Pa0 = points, prediction and updating, till it converges to the true
0 Pυ values. Finally, the carrier frequencies, azimuth angles and
0: loop: elevation angles are evaluated.
Selecting Sigma Points: UKF tends to have a better performance than EKF, as
UKF is accurate to the third order derivatives. Furthermore,
χa0,n−1 = x̂an−1
p  UKF, unlike EKF, is a derivative-free. However, it requires
χai,n−1 = x̂an−1 + (6L + Λ)Pn−1 calculating sigma points using Cholesky decomposition in
i
every iteration. This leads UKF to have higher processing
i = 1, 2 . . . 6L
p  time.
χai,n−1 = x̂an−1 − (6L + Λ)Pn−1
i
V. RESULTS AND DISCUSSIONS
i = 6L + 1, . . . 12L Numerical simulations of the proposed approaches are pre-
Prediction Step: sented in this section. We first present the simulation model,
and then we discuss the results. Finally, a comparison among
χxi,n|n−1 = f (χxi,n−1 )
our proposal and the proposals in [13] and [19] is presented.
12L
(m)
X
x̂−
n = wi χxi,n|n−1 A. SIMULATION MODEL
i=0
12L
For simulations, each ULA on each axis has 150 elements
T
(N = 150) with an inter-element spacing of one-fourth of
 
(c)
X
P−
n = wi χxi,n|n−1 − x̂−
n χxi,n|n−1 − x̂−
n
the minimum wavelength. Twelve different source signals
i=0
are impinging on the arrays. The source signals are traveling
Updating Step: from different angles and are carried on different carrier
Ψi,n|n−1 = H χxi,n|n−1 + χni,n|n−1 frequencies. The normalized carrier frequencies of the source
12L signals are {0.463, 0.217, 0.294, 0.716, 0.607, 0.861, 354,
(m) 0.119, 0.67, 0.294, 0.52, 0.4} respectively. The azimuth
X
ŷn− = wi Ψi,n|n−1
i=0 angles of the sources in degrees are {30, 47, 93, 52,
12L   T 127, 173, 60, 205, 146, 26, 237, 100} and the elevation angles
(c)
X
Pỹn ỹn = wi Ψi,n|n−1 − ŷn− Ψi,n|n−1 − ŷn− in degrees are {5, 84, 23, 37, 53, 127, 94, 151, 89, 163, 77, 41}
i=0 respectively. The signal-to-noise ratio (SNR) is selected to be
12L   T 5dB
(c)
X
Px̃n ỹn = wi χxi,n|n−1 − x̂−
n Ψi,n|n−1 − ŷn− The initial estimates in both EKF and UKF are set to values
i=0 that are close to the mean value of the state variable. Thus,
Kn = Px̃n ỹn P−1
ỹn ỹn all elements in the initial estimate vector x̂o are set to 0.03.
x̂n = x̂− − For simplicity, we assign all the elements to the same values,
n + Kn (Yn − ŷn )
however they can randomly and differently be chosen around
Pn = P− T
n − Kn Pỹn ỹn Kn the mean value of xs . The initial covariance matrix P0 is
if convergence: also set to a proper value which prevents the filter from fully
for each source: trust the estimates and ignoring the observations. Thus, we
b 
l initialize it with its true value. Then, the filters are tuned
φl = tan−1 carefully to guarantee the filter convergence. So, we properly
 al  set the matrices R and Pυ in EKF and UKF respectively.
−1 al
θl = tan
cl
cos θl B. SIMULATION RESULTS
λl = Under this scenario, Monte-Carlo simulations have been
cl
executed with both EKF and UKF for 500 snapshots. Both
break loop
filters succeeded in detecting carrier frequency and 2D-DOA
of all source signals. The root mean square error (RMSE)
in the estimated parameters is shown in Figure 2a. Although
EKF and UKF are sub-optimal estimators, they successfully
VOLUME 5, 2017 7

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

EKF, Azimuth EKF, Azimuth

RMSE in Estimated Angles

RMSE in Estimated Angles


UKF, Azimuth UKF, Azimuth
2 EKF, Elevation
0 EKF, Elevation 10
10 UKF, Elevation UKF, Elevation

10 -1

10 -2 10 1

0 50 100 150 0 50 100 150


No. of Iterations No. of Iterations
10 0 10 1
RMSE in Estimated Wavelength

RMSE in Estimated Wavelength


EKF EKF
UKF UKF

10 -1

10 0
-2
10

10 -3 10 -1
0 50 100 150 0 50 100 150
No. of Iterations No. of Iterations

(a) initial estimate = 0.03 (b) initial estimate = 0.7

FIGURE 2: RMSE in the estimated azimuth, elevation and normalized wavelength at different initial estimates

detect the different sources with RMSE in all the estimated spacing. A new experiment has been executed with an initial
parameters close to 10−1 . The reason is that the estimated estimate of 0.7 to show the effect of the inter-element spacing
values are located in the same range and the initial estimates in a worse case. For different inter-element spacing values,
are set to the center of that range. These factors helped the simulations have been carried out and the results are shown in
filters to converge to the true values. However, both EKF Figure 4. When the inter-element spacing is reduced to one-
and UKF showed a degraded performance when the initial tenth of the minimum wavelength, the performance of both
estimates were set to 0.7. In this case, RMSE deteriorated EKF and UKF is enhanced and the RMSE is reduced. More-
markedly as shown in Figure 2b. To explicitly declare those over, EKF performance starts to approach UKF performance
findings, the estimated carrier frequency and 2D-DOA in the and the gab between their performances is tremendously
two cases are shown in Figure 3. In Figure 3a, the estimated diminished. This indicates that the linearization error has
parameters are obviously close to the actual values when the been reduced and the process model tends to be relatively lin-
initial estimate is 0.03. However, when the initial estimate ear. Reducing the inter-element spacing, however, produces
was set to 0.7, EKF converged to erroneous values which mutual coupling in the array elements.
are far apart from the true values and UKF converged to the Another experiment has been carried out to find out the
true values with massive errors as shown in Figure 3b. This effect of the number of source signals. The simulations have
proves that initial estimates play a crucial role in the filter been repeated for 5, 12 and 24 source signals with inter-
convergence and may lead the filter to completely diverge element spacing of λmin /10 and initial estimate of 0.03. The
if they are not set properly. Another thing can be found in filters were still able to detect all the carrier frequencies with
Figure 2. When the initial estimate is 0.7, UKF outperforms their corresponding elevation and azimuth angles as shown
EKF. That is expected, as UKF is accurate to the third deriva- in Figure 5. The filters can continue to detect a larger number
tive while EKF is accurate to the first derivative. However, of sources up to the degrees of freedom of two L-shaped
EKF outperforms UKF when the initial estimate is 0.03. This uniform arrays. The degrees of freedom are limited to the
indicates that the filter operates over a relatively linear model number of array elements in one axis. Thus, the maximum
where the linearization error is massively reduced. This also number of source signals that can be detected with this
reflects the vital role of initial estimates. system is 149. Since the detection of PUs is a blind estimation
The inter-element spacing of the array can enhance the problem in CR, the CRs have no prior information about the
performance of EKF and UKF. As reducing the inter-element number of existing sources. In this case, the filters are set
spacing expands the process model relating to the estimated to detect 149 different source signals and finally detect the
parameters al , bl and cl . Expanding a nonlinear model can actual number of the sources, as the remaining signals would
produce relatively linear characteristics over the region where be zero. As depicted in Figure 5, our approaches can detect
those parameters are defined. This is proven by the improved azimuth angles up to 300◦ . As two L-shaped uniform arrays
performance accomplished by reducing the inter-element provide the ability to detect a range of azimuth angles of 360◦
8 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

180 180

160 160

140 140

120 120

Est. Elevation

Est. Elevation
100 100

80 80

60 60

40 Actual 40 Actual
EKF EKF
UKF UKF
20 20

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Est. Azimuth Est. Azimuth

0.9 0.9
Actual Actual
0.8 EKF 0.8 EKF
UKF UKF
Est. Normalized Frquency

Est. Normalized Frquency


0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
0 50 100 150 200 250 0 50 100 150 200 250
Est. Azimuth Est. Azimuth

(a) initial estimate = 0.03 (b) initial estimate = 0.7

FIGURE 3: Estimated carrier frequencies and their corresponding 2D-DOA

10 3
RMSE in Estimated Elevation

EKF, d = λ min /4
RMSE in Estimated Azimuth

EKF, d = λ min /4
UKF, d = λ min /4 UKF, d = λ min /4
EKF, d = λ min /10 2 EKF, d = λ min /10
UKF, d = λ min /10 10
10 2 UKF, d = λ min /10

10 1
10 1

0 0
10 10
0 50 100 150 0 50 100 150
No. of Iterations No. of Iterations
(a) RMSE in Azimuth (b) RMSE in Elevation
RMSE in Estimated N. Wavelength

EKF, d = λ min /4
10 1 UKF, d = λ min /4
EKF, d = λ min /10
UKF, d = λ min /10

10 0

10 -1

0 50 100 150
No. of Iterations
(c) RMSE in N. Wavelength

FIGURE 4: RMSE in the estimated azimuth, elevation and normalized wavelength with initial estimates of 0.7 and different
inter-element spacing

and a range of elevation angles of 180◦ . In Figure 5c, There them since they differ in the other parameters.
are many pairs that share the same frequency or the same From simulations, we found that EKF consumes time 20
elevation, but the two filters were able to distinguish between times lower than the time consumed by UKF to detect 5 dif-
VOLUME 5, 2017 9

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

90 80 90
Actual Actual Actual
EKF EKF 80 EKF
80 UKF 70 UKF
UKF
70
70 60
60
60 50
Est. Elevation

Est. Elevation

Est. Elevation
50
50 40
40
40 30
30
30 20
20

20 10 10

10 0 0
0 10 20 30 40 50 60 70 80 90 0 20 40 60 80 100 120 140 160 180 200 -50 0 50 100 150 200 250 300
Est. Azimuth Est. Azimuth Est. Azimuth

1 1 1
Actual Actual Actual
0.9 EKF 0.9 EKF 0.9 EKF
UKF
UKF UKF
0.8 0.8 0.8
Est. Normalized Frquency

Est. Normalized Frquency

Est. Normalized Frquency


0.7 0.7 0.7

0.6 0.6 0.6

0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
0 10 20 30 40 50 60 70 80 90 0 20 40 60 80 100 120 140 160 180 200 -50 0 50 100 150 200 250 300
Est. Azimuth Est. Azimuth Est. Azimuth

(a) L = 5 (b) L = 12 (c) L = 24

FIGURE 5: Estimated carrier frequencies and their corresponding 2D-DOA for different number of sources

10 1
EKF, Azimuth EKF, Azimuth
RMSE in Estimated Angles

RMSE in Estimated Angles

UKF, Azimuth UKF, Azimuth


EKF, Elevation 0 EKF, Elevation
10
UKF, Elevation UKF, Elevation
10 0

10 -1

10 -2

10 -1 10 -3
0 50 100 150 0 50 100 150
No. of Iterations No. of Iterations

10 0 10 1
RMSE in Estimated Wavelength

RMSE in Estimated Wavelength

EKF EKF
UKF UKF
0
10
10 -1

10 -1

10 -2
10 -2

10 -3
10 -3
0 50 100 150 0 50 100 150
No. of Iterations No. of Iterations

(a) SNR = 1dB (b) SNR = 13dB

FIGURE 6: RMSE in the estimated carrier frequencies and their corresponding 2D-DOA at different SNRs

ferent sources. When the number of source signals is raised element spacing of one-tenth of the minimum wavelength.
to 12, the ratio between the time consumed by UKF and EKF The SNR level was set to 1dB and 13dB respectively, and
roughly becomes 40. Again, this ratio becomes around 70, RMSE in the estimated parameters was evaluated at each
when the number of sources becomes 24. Although UKF is level. The results are gathered in Figure 6. At each SNR
a derivative-free, it consumes much time in selecting sigma level, the filters should be tuned again to handle the new
points. Since the number of sigma points depends on the level of uncertainty in the measurements. The filter tuning
number of source signals, the time consumed by UKF to is a trial-and-error process, where we keep on changing the
calculate and process these sigma points increases rapidly matrix R till the filter converges. Therefore, the resultant
with the increase in the number of sources. RMSE follows both the SNR level and the quality of tuning
as well. In overall, Figure 6 shows an improvement in the
A further experiment has been carried out to examine our
filter performance with the increase in SNR. EKF still follows
proposal at different levels of SNRs. Thus, we have repeated
the same behavior as in Figure 3a and outperforms UKF
the simulations with an initial estimate of 0.03 and inter-
10 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

because of selecting the initial estimate around zero and the rate. Since both Kumar and See [13] and Kumar and Chandra
small inter-element spacing. [19] operate at sub-Nyquist rates to sample a wideband
spectrum, their implementations require a large number of
C. COMPARATIVE STUDY relaxed ADCs. The number of ADCs required for Kumar and
References [13] and [19] are considered for the sake of com- See [13] is around 1.5 times the number of array elements.
parison. In [13], the authors have proposed a new architecture Kumar and Chandra [19] required a number of ADCs equal
with a 2D nested array to implement sub-Nyquist sampling. to the total number of the employed array elements and
The 2D nested array consists of two rectangular arrays: a delay channels together. As a result, the number of ADCs
sparse rectangular array and a dense one. To achieve sub- may be 3 or 4 times the number of array elements, which
Nyquist rate, each element in the dense array is followed by can be considered as the price that has been paid for the
a single ADC that samples the signals at sub-Nyquist rate. high degrees of freedoms. However, our proposal exploits the
However, each element in the sparse array is followed by two spatial domain instead of the temporal domain to get rid of
paths: a direct path with an ADC and a delayed path with an- the need to sample the signals at Nyquist rates, and hence
other ADC. Then, a proposed algorithm is executed to obtain it only requires a number of ADCs equal to the number of
the carrier frequency and a single DOA for each source. In the employed array elements without any restrictions on the
[19], the authors have exploited a standard URA for detecting speed of the ADCs. So the number of ADCs required for
both the carrier frequency and the corresponding 2D-DOA our proposal is always lower than the number required by
of each source. Then, the authors have proposed adding a Kumar and See [13] and Kumar and Chandra [19] for the
number of delay channels after a single array element of the same number of array elements.
employed URA. Each delay channel has been provided with Moreover, both our proposal and Kumar and Chandra [19]
an ADC. Then, the authors have exploited the spatial and detect two angles for each source signal instead of a single
time delays to detect the unknown parameters using ESPRIT angle as in Kumar and See [13]. This gives them another
algorithm. In contrast, our proposal does not rely on this large advantage as it increases the spatial capacity. As many CRs
number of ADCs as the processing is executed spatially. can share two parameters with PUs at the same time without
Table 1 shows the differences among our proposal, Kumar interfering them as they differ in the third parameter.
and Chandra [19] and Kumar and See [13]. Kumar and Chan-
dra [19] can detect a number of source signals higher than VI. CONCLUSIONS
the number of array elements if the number of the employed In this paper, EKF and UKF are proposed for the problem of
delay channels exceeds 4. However, the simulations in [19] estimating carrier frequencies and 2D-DOA of uncorrelated
have proven that the higher the number of delay channels, band-limited source signals. Exploiting two different angles
the higher the performance would be accomplished. Thus, in the space increases the spatial capacity of CR. The main
the number of delay channels would be twice or triple the challenge of wideband sensing is the need to high Nyquist
number of array elements. In this context, the degrees of free- rates and high hardware requirements. In order to decrease
dom of Kumar and Chandra [19] increase exponentially with hardware complexity, KF is applied in the spatial domain
the increase in the number of the employed array elements. where the incident signal on one array element is predicted
Under these circumstances, it outperforms our proposal and from its version impinges on the previous array element. As
Kumar and See [13]. Although the latter has degrees of a result, the proposed algorithm does not require complex
freedom that also increase exponentially, they rise at a slower hardware to perform Nyquist or sub-Nyquist rates. Since
rate than the degrees of freedom of Kumar and Chandra EKF and UKF are sub-optimal filters, they may converge
[19]. That is predictable since Kumar and Chandra [19], to sub-optimal values or even diverge. To overcome this
unlike the others, has no constraints on detecting a number of problem, the parameters to be estimated are scaled to the
source signals higher than the number of the employed array same range with small variance, and the filters should be
elements. However, Kumar and See [13] needed at least 5 properly initialized. Simulations have proven the vital effect
array elements to start detecting source signals. Besides, it of the initial estimates and the inter-element spacing on the
started to detect a number of source signals higher than the performance of the filters. The smaller the inter-element
number of the employed array elements when the number of spacing, the higher the performance is. Since the proposal
array elements exceeded 13. Our proposal, however, started depends on two L-shaped uniform arrays, it has limited de-
to detect signals with 4 array elements, and its degrees of grees of freedom restricted to the number of array elements.
freedom increase linearly with the increase of the array However, it has the availability to detect source signals from
elements. Besides, its degrees of freedom are restricted to all directions in the space.
one-third of the total number of the employed array elements.
Overall, our proposal has the lowest degrees of freedom
REFERENCES
among the three methods as it does not rely on sparsity on
[1] S. Haykin, “Cognitive radio: brain-empowered wireless communications,”
any domain.
IEEE Journal on Selected Areas in Communications, vol. 23, no. 2,
On the other hand, our proposal has employed a lower pp. 201–220, 2005.
number of ADCs as there are no restrictions on the sampling [2] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for

VOLUME 5, 2017 11

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

TABLE 1: Comparison among our proposed approaches, Kumar and Chandra [19] and Kumar and See [13]

KF-Based
Kumar and Chandra [19] Kumar and See [13]
Proposed Approaches
A URA with N elements A 2D nested array Two L-shaped uniform arrays
Architecture
and M delay channels with N elements with 3N − 2 elements
NM l N m2
Degrees of
−1 N−1
Freedom 4 4
Minimum (3N + 1)B∗
(N + M)B∗ No restrictions
Sampling Rate 2
No. of ADCs (3N + 1)
N+M 3N − 2
employed 2
Estimated 2D-DOA DOA 2D-DOA
Parameters Carrier Frequency Carrier Frequency Carrier Frequency
* B denotes the bandwidth of band-limited source signals.

cognitive radio applications,” IEEE Communications Surveys & Tutorials, tion based on uniform circular array,” in IEEE International Symposium
vol. 11, no. 1, 2009. on Phased Array Systems and Technology, pp. 547–552, 2003.
[3] F. Salahdine, N. Kaabouch, and H. El Ghazi, “A survey on compressive [17] X. Qi, L. Gan, P. Wei, and C. Ren, “Joint frequency and 2-D DOA
sensing techniques for cognitive radio networks,” Physical Communica- estimation using pseudocovariance matrices,” in International Conference
tion, vol. 20, pp. 61–73, 2016. on Communications, Circuits and Systems, pp. 398–401, IEEE, Jul 2009.
[4] A. A. Kumar, S. G. Razul, and C. M. S. See, “An efficient sub-Nyquist [18] R. J. Weber and Y. Huang, “A wideband circular array for frequency and
receiver architecture for spectrum blind reconstruction and direction of ar- 2d angle estimation,” in IEEE Aerospace Conference, pp. 1–8, Mar 2010.
rival estimation,” in IEEE International Conference on Acoustics, Speech [19] A. A. Kumar, M. G. Chandra, and P. Balamuralidhar, “Joint frequency and
and Signal Processing (ICASSP), pp. 6781–6785, 2014. 2-D DOA recovery with sub-Nyquist difference space-time array,” in Pro-
[5] A. A. Kumar, S. G. Razul, and C.-M. S. See, “Spectrum blind reconstruc- ceedings of the 25th European Signal Processing Conference (EUSIPCO),
tion and direction of arrival estimation at sub-Nyquist sampling rates with pp. 420–424, Sep 2017.
uniform linear array,” in IEEE International Conference on Digital Signal [20] L.-T. Wan, L.-T. Liu, W.-J. Si, and Z.-X. Tian, “Joint estimation of 2D-
Processing (DSP), pp. 670–674, 2015. DOA and frequency based on space-time matrix and conformal array,”
The Scientific World Journal, vol. 2013, pp. 1–10, 2013.
[6] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via
[21] Y. Zou, H. Xie, L. Wan, and G. Han, “High accuracy frequency and 2D-
rotational invariance techniques,” IEEE Transactions on Acoustic, Speech
DOAs estimation of conformal array based on PARAFAC,” Journal of
and Signal Processing, vol. 37, no. 7, pp. 984–995, 1989.
Internet Technology, vol. 16, no. 1, pp. 107–119, 2015.
[7] R. Schmidt, “Multiple emitter location and signal parameter estimation,” [22] L.-y. Xu, X.-f. Zhang, Z.-z. Xu, and M. Yu, “Joint 2D-DOA and frequency
IEEE Transactions on Antennas Propagation, vol. 34, no. 3, pp. 276–280, estimation for L-shaped array using iterative least squares method,” Inter-
1986. national Journal of Antennas and Propagation, vol. 2012, pp. 1–8, 2012.
[8] A. A. Kumar, S. G. Razul, and C.-M. S. See, “Spectrum blind recon- [23] G. Welch and G. Bishop, “An introduction to the Kalman filter,” pp. 1–16,
struction and direction of arrival estimation of multi-band signals at sub- 2006.
Nyquist sampling rates,” Multidimensional Systems and Signal Process- [24] S. J. Julier and J. K. Uhlmann, “New extension of the Kalman filter to
ing, pp. 1–27, Sep 2016. nonlinear systems,” in Proc. SPIE, vol. 3068, pp. 182–193, 1997.
[9] S. Stein, O. Yair, D. Cohen, and Y. C. Eldar, “Joint spectrum sensing
and direction of arrival recovery from sub-Nyquist samples,” in IEEE
16th International Workshop on Signal Processing Advances in Wireless
Communications (SPAWC), pp. 331–335, 2015.
[10] M. Mishali and Y. Eldar, “From theory to practice: sub-Nyquist sampling
of sparse wideband analog signals,” IEEE Journal of Selected Topics in
Signal Processing, vol. 4, pp. 375–391, Apr 2010.
[11] Z. Xiaofei, L. Jianfeng, and X. Lingyun, “Novel two-dimensional DOA
estimation with L-shaped array,” EURASIP Journal on Advances in Signal
Processing, vol. 2011, Dec 2011.
[12] S. Stein Ioushua, O. Yair, D. Cohen, and Y. C. Eldar, “CaSCADE:
compressed carrier and DOA estimation,” IEEE Transactions on Signal Samar Elaraby received her B.Sc. in Electronics
Processing, vol. 65, pp. 2645–2658, May 2017. and Communication Engineering from Port Said
[13] A. A. Kumar, S. G. Razul, and C. M. S. See, “Carrier frequency and direc- University, Egypt, in 2011. She is currently pursu-
tion of arrival estimation with nested sub-nyquist sensor array receiver,” ing a master’s degree in Electronics and Commu-
in Proceedings of the 23rd European Signal Processing Conference (EU- nication Engineering with Faculty of Engineering,
SIPCO), pp. 1167–1171, 2015. Port Said University, Egypt. Her current research
[14] A. Lavrenko, F. Romer, S. Stein, D. Cohen, G. Del Galdo, R. S. Thoma, interests focus on cognitive radio, wireless com-
and Y. C. Eldar, “Spatially resolved sub-Nyquist sensing of multiband sig- munication and signal processing.
nals with arbitrary antenna arrays,” in IEEE 17th International Workshop
on Signal Processing Advances in Wireless Communications (SPAWC),
pp. 1–5, 2016.
[15] D. D. Ariananda and G. Leus, “Compressive joint angular-frequency
power spectrum estimation for correlated sources,” in Proceedings of the
21st European Signal Processing Conference (EUSIPCO), pp. 1âĂŞ–5,
2013.
[16] C. Hui, W. Yongliang, and W. Zhiwen, “Frequency and 2-d angle estima-

12 VOLUME 5, 2017

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2768221, IEEE Access

Heba Y. Soliman is an assistant professor in


Electrical Engineering Department at Faculty of
Engineering, Port Said University. She received
the BSc. degree in Electrical Communications in
1999 from Suez Canal University. She received
the MSc. Degree in Wireless Communications in
2004 from Suez Canal University. She received the
PhD degree in Smart Antenna Systems in 2010
from Port Said University. Her research interests
include adaptive antenna systems, MIMO systems
and cognitive radio.

Heba M. Abdel-Atty (received the Ph.D. degree


in Electronics and Communications Engineering
from the Faculty of Engineering-PortSaid Univer-
sity, Egypt 2012. She is working as an assistant
professor at the Electronics and Communications
Engineering Department from 2012 until now. She
is founder and counselor of IEEE in Port-Said
University student branch. Her current research
interests are in wireless communication systems,
mobile communication, multimedia processing,
cognitive radio, security of wireless networks and ?eld programmable gate
array (FPGA) applications.

Mohamed Abdel-Azim received the Ph.D. de-


gree in Electronics and Communications Engi-
neering from the Faculty of Engineering, Man-
soura University, Egypt by 2006. After that he
worked as an assistant professor at the electronics
& communications engineering department, then
awarded the associate professor degree in 2012
until now. He has 130 publications in various
international journals and conferences. His cur-
rent research interests are in multimedia process-
ing, wireless communication systems, and field programmable gate array
(FPGA) applications. He had awarded the best Ph.D. thesis at Mansoura
University at 2007. He is the executive director of scientific computing
center and the consultant for IT in Mansoura University.

VOLUME 5, 2017 13

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like