Академический Документы
Профессиональный Документы
Культура Документы
2, FEBRUARY 2013
809
Manuscript received November 27, 2011; revised March 22, 2012, June 30,
2012, and September 6, 2012; accepted October 5, 2012. Date of publication
October 10, 2012; date of current version February 12, 2013. This work was
supported in part by the National Natural Science Foundation of China under
Grant No. 61201225 and in part by the Natural Science Foundation of Shanghai
under Grant No. 12ZR1450800. The review of this paper was coordinated by
Prof. B. Hamdaoui.
X.-L. Huang is with the Department of Information and Communication
Engineering, Tongji University, Shanghai 201804, China (e-mail: xlhuang@
tongji.edu.cn).
G. Wang is with the Communication Research Center, Harbin Institute of
Technology, Harbin 150001, China (e-mail: gwang51@hit.edu.cn).
F. Hu is with the Department of Electrical and Computer Engineering, The
University of Alabama, Tuscaloosa, AL 35487 USA (e-mail: fei@eng.ua.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2012.2223767
I. I NTRODUCTION
ODAY, the spectrum assignment policy in wireless communications is regulated by governmental agencies. The
huge band wireless spectrum is segmented and authorized to
licensed holders or services. With the dramatic increase of highdefinition audio/video applications through wireless access,
hundreds of megahertz to many gigahertz of wireless bandwidth
are required, which causes scarcity of the limited wireless spectrum resource. On the other hand, according to a report from the
Federal Communications Commission (FCC) [1], the temporal
and geographical variations in the utilization of the licensed
spectrum are from 15% to 85%. This means that much of the
spectrum is not efficiently utilized. The increasing high-quality
service requirement, limited available spectrum, and inefficient
spectrum utilization necessitate a new communication pattern
to exploit the existing wireless spectrum opportunistically [2],
[3]. Dynamic spectrum access (DSA) has been proposed to
solve the spectrum inefficiency problems and is implemented
in cognitive radio networks (CRNs) [4]. In a CRN, through
the opportunistic use of free spectrum (also called spectrum
holes), a device can gain access to more wireless bandwidth,
which is the main goal of the FCC regulations [5].
Cognitive radio (CR) techniques provide the capability of
detecting spectrum holes and sharing the spectrum in an opportunistic manner. DSA techniques can select the best available
channel from the spectrum pool for CR devices to operate [4],
[6]. More specifically, CR enables secondary users (SUs) to
perform a series of operations as follows: 1) spectrum sensing
to predict what spectrum is available and recognize the presence
of the primary user (PU) when a PU reoccupies the licensed
channel; 2) spectrum management to select the best available
channel from the spectrum pool for special services; 3) spectrum sharing to coordinate access to all available channels with
other SUs; and 4) spectrum mobility to vacate the channel as
soon as possible when a PU is detected [4]. Spectrum sensing
is one of the most important components in the cognition cycle
(see Fig. 1).
In Fig. 1, the spectrum-sensing module helps the SUs to
recognize the radio environment, i.e., identifying the spectrum
occupancy states of both PUs and other SUs. The spectrum
information can be further used by spectrum analysis and spectrum decision modules to analyze the available channel quality
and then make a channel assignment decision, respectively [4].
Recently, many signal processing techniques have been developed for spectrum sensing, and these can be classified as
either noncooperative detection or cooperative detection. In
noncooperative detection methods, different spectrum-sensing
methods can be chosen by individual SUs [8], such as match
810
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
811
first assumption imposes a stringent requirement on the synchronization among all SUs, which is difficult to implement in a
large-scale CRN. The second assumption imposes that the PUs
are very powerful or the whole CRN is deployed in a small
area, and no interference and deep fading exists in the radio
environment. Those conditions may not be realistic in many
CRN applications.
In this paper, we consider the cluster-based CRN shown
in Fig. 2. In Fig. 2, the SUs are deployed in a wide area,
and the clusters are formed based on some metrics, such as
location, mobility, etc. (please refer to our previous work [3]
on CRN clustering strategies). This paper assumes the same
clustering criterion as in [3]. Therefore, we can assume that
all cluster members share the same spectrum map due to their
close distances to each other. However, due to spectrum-sensing
noise and errors in each cluster member, there could be minor
discrepancies among their sensing results. Thus, an efficient
spectrum fusion algorithm is needed to reach a consensus in
terms of the entire clusters spectrum patterns. In the spectrumsensing stage, all SUs in the same cluster first individually
operate CS sampling in synchronization mode based on the
spectrum management commands sent from the CH. The CH
then collects the local result from each cluster member to make
a fusion result and then broadcasts the fused information to the
CHs in other clusters.
To realize our cooperative decentralized spatiotemporal data
mining, we assume the following: 1) Sampling synchronization
in the whole CRN is not required, and synchronized sampling
only occurs inside each cluster. When a cluster starts spectrum sensing, it may also receive the signal from neighboring
clusters. Such signal can be seen as the interference from the
viewpoint of spectrum sensing. Hence, we need to identify
which channels are occupied by the PUs or SUs from the neighboring clusters. 2) The transmission power of PUs may not be
high enough to cover the whole CRN area, and different clusters may have different spectrum occupancy status due to the
geographical-dependent PUs. Hence, the sparseness spectrum
is determined by the geographical-dependent PUs as well as the
812
n = 0, 1, . . . , M 1 (1)
where rjP (n) = i=1 hi,j (n)xi (n) and rjC (n) = i gi,j (n)
yi (n) correspond to the received signals from a total of I PUs
and the interference from neighboring clusters, respectively.
hi,j (n) is the CIR between PU i and SU j (j = 1, 2, . . . , N ),
and () denotes convolution. gi,j (n) is the CIR between a
neighboring cluster i and SU j. xi (n) and yi (n) correspond
to the original transmitted signal from PU i and neighboring
cluster i, respectively. wj (n) represents the additive white
Gaussian noise.
After M -point discrete Fourier transform (DFT) for the
observed signals, (1) can be further rewritten as [12]
Rj (k) =
I
i=1
M
N j,k |0, 01 k1
1 P
1 C
1
Rj + j FM FM
Rj + j FM FM
Wj
j = j FM FM
(3)
(7)
k=1
(5)
(4)
ba a1
exp(b0 ).
(a) 0
(8)
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
813
where
j = j Tj j
1
.
j = Tj j + A
(10)
N
log p(j |)
j=1
N
p(j |j , 0 )p(j |, 0 )p(0 |a, b)dj d0
log
j=1
1
2
N
(11)
1
sj,k = Tj,k Bj,k
j,k
(mj +2a) log jT Bj1 j +2b +log |Bj |
j=1
+ Const.
1
qj,k = Tj,k Bj,k
j
1
ej,k = jT Bj,k
j + 2b.
To update k in each iteration, we fix the other hyperparameter k as the latest value, differentiate the likelihood function
(k ) with k , and set the result to zero, i.e.,
(k )
k
2
2
N s
j,k sj,k qj,k /ej,k /k (mj +2a)qj,k /ej,k +sj,k
=
2 /e
2(k +sj,k ) k +sj,k qj,k
j=1
j,k
= 0.
(19)
Since k is the precision of the Gaussian distribution, we have
k > 0. We assume k sj,k (this is an empirical result stated
in [29]), and thus, k + sj,k sj,k in (19). Then, we can derive
the new k from (19) as
(12)
if
Here
Bj = E + j A1 Tj
Tj
Bj = E + j A
=E +
n1 j,n Tj,n + k1 j,k Tj,k
n=k
(14)
1
1
Bj,k
j,k Tj,k Bj,k
1
k + Tj,k Bj,k
j,k
(16)
Then, the contribution of the basis vector j,k in the likelihood function (12) can be separated from others, i.e.,
N
1
1
j + 2b
(mj + 2a) log jT Bj,k
2 j=1
+ log |Bj,k | + Const
N
1
log 1 + k1 sj,k
2 j=1
2
/ej,k
qj,k
+ (mj + 2a) log 1
k + sj,k
() =
= (k ) + (k )
(17)
N
2
(mj + 2a)qj,k
/ej,k sj,k
>0
2 /e
sj,k (sj,k qj,k
j,k )
j=1
(13)
(18)
N
N
2 /e
(mj +2a)qj,k
j,k sj,k
j=1
2 /e
sj,k (sj,k qj,k
j,k )
(20)
else
k = .
(21)
2
Hence, the SU calculates ((mj + 2a)qj,k
/ej,k sj,k )/
2
(sj,k (sj,k qj,k /ej,k )) in each iteration and broadcasts such
value to its CH. From (20) and (21), we can update the hyperparameter k (k = 1, 2, . . . , M ) after each iteration. After reaching the upper bound of iteration times, or if the increment value
of the likelihood in (17) is less than a threshold (which means
that we almost reach the maximum value of the likelihood),
the CH obtains the spectrum decision for its cluster. In (21),
k = means j,k = 0 (j = 1, 2, . . . , N ), and the subcarrier
k is available to SUs.
From the foregoing analysis, one can see that: 1) the member
nodes in one cluster seek a consensus spectrum map based on
the multitask BCS model, and 2) the information exchanged
among member nodes can be used to derive the shared hyperparameter = {1 , 2 , . . . , M }. An advantage of our proposed
hierarchical prior [see (7) and (8)] is to collect the spatial
contribution from all member nodes to derive the common
sparseness spectrum and thus remove ISI channel fading.
After several iterations, the result = {1 , 2 , . . . , M }
will converge, and we can then make a binary spectrum decision
of dPU and dSU [to be discussed in (57) and (58)], which
represents the spectrum occupancy states of PUs and SUs.
814
observations from different clusters. In Section III, one cluster is assumed to share one common sparseness spectrum.
However, the CRN may be deployed over a large-scale area,
and the sparseness spectrum decisions may vary in different
positions due to the geographical-dependent PUs and signal
attenuation along a path. Hence, different clusters may not be
statistically interrelated to each other, and the CS observations
from different clusters may not be appropriate for sharing. For
example, one cluster may be located near a high TV tower (base
station), which makes less IEEE 802.22 channels available for
SUs. In the multicluster case, we should design an efficient
algorithm that first groups the CS observations from different
clusters (multiple clusters CS observations may belong to the
same group as long as they obey the same spectrum statistics)
and then uses the multitask BCS model (see Section III) in
each group to discover the common sparseness spectrum within
each group. For this purpose, we introduce a DP prior to the
hierarchical BCS model that has been discussed in Section III.
The DP prior [30] has shown a powerful capability of automatically classifying different samples into groups based on their
statistical patterns. In our application, the DP prior will be used
to realize both spectrum grouping and CS inversion.
k k
(27)
k=1
where
k = k
k1
(1 n )
(28)
n=1
k | Beta(1, )
k |G0 G0 .
(29)
(30)
A. DP
M
Ga(c, d).
(31)
k=1
DP is a distribution over probability measure and has two parameters: 1) precision and 2) base distribution G0 [30]. In the
multicluster spectrum-sensing case, different clusters may have
different hyperparameters, that is, i = {i1 , i2 , . . . , iM },
the cluster ID i = 1, 2, . . . , C (C is the total number of clusters
in the CRN). We assume {i , i = 1, . . . , C} is drawn identically from distribution G, which is a random draw from the DP,
i.e.,
i |G G,
i = 1, . . . , C
(22)
G DP (, G0 )
(23)
E(G) = G0 .
(24)
Equation (22) is the likelihood function for G, and the hyperparameter i has been derived in the multitask BCS model (see
Section III). Equation (23) is the prior knowledge of G.
When we integral out G according to (22) and (23), i
obeys the base distribution G0 . In our cluster-based CRN, when
one cluster collects the hyperparameter information i =
{1 , 2 , . . . , i1 , i+1 , . . . , C } from other clusters, the base
distribution G0 is updated, and we have [36]
1
G0 +
p(i | , , G0 ) =
+C 1
+C 1
i
C
k (25)
k=1,k=i
k in {
k }K
are ni
k=1 . Then,
k
(25) can be further written as
p(i |i , , G0 ) =
1
G0 +
+C 1
+C 1
K
k=1
ni
k . (26)
k
J
lk k
(32)
k=1
J
k=1
J
p(i |G) = p i |{lk }k=1,J , {k }k=1,J =
lk k (33)
k=1
(34)
j |j , 0 N j j , 01 E , j = 1, 2, . . . , C
, k = 1, 2, . . . , M
j,k |zj ,k N 0, 01 z1
j ,k
0 Ga(a, b)
j |{lk }k=1,J , {k }k=1,J
J
lk k
k=1
zj Categorical(l1 , l2 , . . . , lJ )
{l1 , l2 , . . . , lJ } Dir(1 , 2 , . . . , J )
(35)
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
q(l)
C
q(zj )
j=1
log p(l|) +
C
log p(zj |l) + log p j |zj
j=1
log q(l)
C
log q(zj )
j=1
dzdl.
( ) =
k (k ) =
J
k=1
C
k (k )
C
j,k log
Bj,k = E +
M
1
1
k,t
j,t Tj,t + k,n
j,n Tj,n
t=1,t=n
1
= Bj,k,n + k,n
j,n Tj,n
(41)
, k,2
, . . . , k,n1
, k,n+1
, . . . , k,M
}. In
lated effects of {k,1
1
1
Bj,k,n
j,n Tj,n Bj,k,n
B 1
k,n
j,k,n j,n
(43)
C
1
1
j,k (mj + 2a) log jT Bj,k,n
j + b
2 j=1
+ log |Bj,k,n | + const
C
1
1
(37)
j,k log p (j |k )
j=1
obtain the optimal values = {1 , 2 , . . . , J }, we decompose the matrix Bj,k in the same way as what we have done in
(14), i.e.,
(36)
815
+ k k,n
= k k,n
(44)
where
1
sj,k,n = Tj,n Bj,k,n
j,n
p(j |j , 0 )p (j |k , 0 )
.
Ak = diag k,j
j=1,M
1
qj,k,n = Tj,n Bj,k,n
j
j=1
(39)
(40)
1
ej,k,n = jT Bj,k,n
j + 2b.
(45)
1
1
j,k
sj,l,k,n
log 1 + k,n
k k,n =
2 j=1
l=1
2
/ej,l,k,n
qj,l,k,n
+(mj,l + 2a) log 1
. (46)
k,n + sj,l,k,n
816
erj,l,k
l=1
Nj
J
j,k =
(47)
erj,l,m
m=1 l=1
1
+
j,k
J j=1
C
k =
(48)
where
rj,l,k = (k )
J
m
m=1
1
T
1
(mj,l + 2a) log j,l
Bj,l,k
j,l + b
2
(49)
+ log |Bj,l,k |
(x) =
log (x)
.
x
(50)
Maximizing k (k,n
) in (46) (i.e., (k (k,n
))/k,n
= 0)
cannot be solved directly in a close-loop format because the
denominator of each factor is a second-order polynomial
of
) in the M-step,
result stated in [29]). By maximizing k (k,n
we get
if
C
j,k
j=1
Nj
2
(mj,l + 2a)qj,l,k,n
/ej,l,k,n sj,l,k,n
>0
2
l=1 sj,l,k,n sj,l,k,n qj,l,k,n /ej,l,k,n
C
k,n
Nj j,k
j=1
C
j,k
j=1
Nj
2
/ej,l,k,n sj,l,k,n
(mj,l +2a)qj,l,k,n
l=1
2
sj,l,k,n (sj,l,k,n qj,l,k,n
/ej,l,k,n )
(51)
else
= .
k,n
(52)
N j
C. HMM
In Fig. 3, the relationship between hidden subcarrier states
and CS observations is plotted. Since the subcarrier states
should be time relevant, only a small number of subcarriers change their binary states between two consequent CS
observations.
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
Fig. 3.
817
(t)
(t) 2 < threshold
k,n
+ z,n
if
else
dn (t) = 1
(57)
dn (t) = 0
(58)
where n {1, 2, . . . , M }.
Fig. 4.
1))
k
z
z (t1)
Vz (t1)
(t 1)
Vz (t) (t) .
(t) = arg max
z (t)
(54)
(55)
(56)
818
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
Fig. 7.
Comparisons of spectrum detection performance. Number of measurements: (a) 40. (b) 50. (c) 60. (d) 70. (e) 80. (f) 90. (g) 100.
819
820
Fig. 8. Tradeoff between the number of member nodes and the number of
measurements.
In Fig. 7, one can also see the following: 1) When the number
of measurements increases, a higher probability of correct
detection can be achieved since a richer spectrum information
can be exploited in each SU. Meanwhile, the spectrum-sensing
performance curves start to deviate from the X-axis more obviously (see Fig. 7) since more spectrum information exchange
among SUs also helps to reduce the probability of false alarm.
Under such a case, Fig. 7 clearly shows that our scheme
has the lowest probability of false alarm among all schemes.
2) In CRN, since interference to PUs is not allowed, one may be
only interested in the PCD = 1 case. For such a case (i.e., when
PCD = 1), Fig. 7 shows that the probability of false alarm in
our proposed algorithm is reduced by 70% compared with the
MT-CS algorithm. This is because more spectrum holes can be
accurately detected via our proposed algorithm.
C. Tradeoff Between the Number of Measurements and the
Number of Tasks
Since our proposed spatiotemporal data mining is a multitask
spectrum-sensing algorithm, we can strike a good balance
between the number of member nodes in each cluster and
the number of spectrum measurements. In Fig. 8, we plot the
normalized MSE performance for three different cases (i.e., 1,
3, and 5 member nodes in each cluster, respectively). From
Fig. 8, it can be seen that if the number of tasks is reduced
(i.e., a smaller number of member nodes in each cluster),
we can increase the number of measurements to collect more
information for the CS inversion process.
When the required normalized MSE values are 0.12, 0.13,
0.14, and 0.15, respectively, the number of measurements for
cluster size = 5 can be reduced by 7.4%, 7.7%, 9.4%, and
3.9%, respectively, compared with the case of cluster size = 3.
Hence, under the same performance requirements (here we use
normalized MSE), the CRN with more nodes in each cluster
can reduce its sampling rate and thus lower the sampling cost.
VI. C ONCLUSION
In this paper, we proposed spatiotemporal data mining
schemes for low-cost spectrum sensing in CRNs. First, we
employed the DP-based MT-CS method to group the observations from different clusters that may not share one common
sparseness spectrum. Meanwhile, the BCS inversion was used
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
N
j=1
=
=
N
j=1
N
&
' m2j
' 12
M &
2
2
=
0
0 k
$ k=1
1 T T
0
0
j Tj j +A
exp jT j
j j
2
2
1 T
Tj j +A j Tj j +A
j j
%
0
1 T
j j dj
+ jT j Tj j +A
2
mj
$ & ' 2
' 12
M &
2
2
=
0
0 k
$ k=1
1 T T
0
j Tj j +A
exp
j j
2 T
j j +A
%
1 T (
j Tj j +A
j j dj
)
1 T (
0
j j . (63)
exp jT E j Tj j +A
2
From (63), one can see that p(j |, 0 ) follows a zeromean Gaussian distribution, with covariance matrix {0 [E
j (Tj j + A)1 Tj ]}1 = (1/0 )(E + j A1 Tj ). Hence,
(62) can be further rewritten as
log
log
N
$
%
log p(0 |a, b) p(j |j , 0 )p(j |, 0 )dj d0
j=1
p(0 |a, b)
%
p(j |j , 0 )p(j |, 0 )dj
d0
where
p(j |,
0 )
p(j |j , 0 )p(j |, 0 )dj
&
' m2j
2
0
=
exp j j j 22
0
2
' 12
M &
2
0 k
j,k 22 dj
exp
0 k
2
k=1
& ' m2j
' 12
M &
2
2
=
0
0 k
k=1
0
exp (j j j )T (j j j )
2
0
exp jT Aj dj
2
=
' m2j
' 12
M &
2
2
=
0
0 k
k=1
0 T
j j jT j j jT Tj j
exp
2
T T
T
+j j j j +j Aj dj
j=1
&
()
log p(j |)
(62)
821
N
&
2
0
' m2j
1
1
E +j A1 T 2
j=1
j
1
0
exp jT E +j A1 Tj
j d0
2
$
N
mj
mj
1
2
=
log (2) 2
|a,
b)
p(
0
1
0
T
2
1
E +j A j
j=1
%
0 T
1 T 1
exp j E +j A j
j d0
2
$
N
mj
1
ba a1+ m2j
=
log (2) 2
1
(a) 0
E +j A1 T 2
j=1
j
$
1
exp 0 b+ jT
2
%
%
1 T 1
E +j A j
j d0
mj
1
ba
log (2) 2
=
1
E +j A1 T 2 (a)
j=1
j
1
mj
1 a+ 2
j
b+ 12 jT E +j A1 Tj
=
log
p(0 |a, b)
822
N
1
mj
log(2) logE +j A1 Tj
2
2
j=1
& a '
b
mj
+log
a+
2
(a)
1
1 T
j
log b+ j E +j A1 Tj
2
N
1
(mj +2a) log jT Bj1 j +2b
=
2 j=1
+logE +j A1 Tj
& a '
%
N $
b
mj
mj
log(2)+log
log2
+
+ a+
2
(a)
2
j=1
1
=
(mj +2a) log jT Bj1 j +2b
2 j=1
N
(64)
R EFERENCES
[1] Spectrum policy task force report, FCC, Washington, DC, Rep. Et
Docket 02-135, Nov. 2002.
[2] H. Khalife, S. Ahuja, N. Malouch, and M. Krunz, Probabilistic path selection in opportunistic cognitive radio networks, in Proc. IEEE GLOBECOM, Dec. 2008, pp. 15.
[3] X.-L. Huang, G. Wang, F. Hu, and S. Kumar, Stability-capacity-adaptive
routing for high-mobility multihop cognitive radio networks, IEEE
Trans. Veh. Technol., vol. 60, no. 6, pp. 27142729, Jul. 2011.
[4] I. F. Akyildiz, W. Lee, M. C. Vuran, and S. Mohanty, Next generation/
dynamic spectrum access/cognitive radio wireless networks: A survey,
Comput. Netw., vol. 50, no. 13, pp. 21272159, Sep. 2006.
[5] K. R. Chowdhury and M. D. Felice, SEARCH: A routing protocol for
mobile cognitive radio ad-hoc networks, Comput. Commun., vol. 32,
no. 18, pp. 19831997, Dec. 2009.
[6] G. Zhu, I. F. Akyildiz, and G. Kuo, STOD-RP: A spectrum-tree based
on-demand routing protocol for multi-hop cognitive radio networks, in
Proc. IEEE GLOBECOM, 2008, pp. 15.
[7] H. Kushwaha, Y. Xing, R. Chandramouli, and H. Heffes, Reliable multimedia transmission over cognitive radio networks using fountain codes,
Proc. IEEE, vol. 96, no. 1, pp. 155165, Jan. 2008.
[8] T. Ycek and H. Arslan, A survey of spectrum sensing algorithms for
cognitive radio applications, IEEE Commun. Surveys Tuts., vol. 11, no. 1,
pp. 116130, 1st Quart, 2009.
[9] R. Tandra and A. Sahai, Fundamental limits on detection in low
SNR under noise uncertainty, in Proc. IEEE WNCMC, 2005, vol. 1,
pp. 464469.
[10] S. Shankar, C. Cordeiro, and K. Challapali, Spectrum agile radios:
Utilization and sensing architectures, in Proc. IEEE DySPAN, 2005,
pp. 160169.
[11] M. Ghozzi, F. Marx, M. Dohler, and J. Palicot, Cyclostationarility-based
test for detection of vacant frequency bands, in Proc. IEEE CROWNCOM, 2006, pp. 15.
[12] F. Zeng, C. Li, and Z. Tian, Distributed compressive spectrum sensing
in cooperative multihop cognitive networks, IEEE J. Sel. Topics Signal
Process., vol. 5, no. 1, pp. 3748, Feb. 2011.
[13] M. R. Duarte, M. B. Wakin, D. Baron, and R. G. Baraniuk, Universal
distributed sensing via random projections, in Proc. IEEE IPSN, 2006,
pp. 177185.
[14] Y. Wang, A. Pandharipande, Y. Polo, and G. Leus, Distributed compressive wide-band spectrum sensing, in Proc. IEEE ITA, 2009, pp. 178183.
[15] M. E. Yildiz, T. C. Aysal, and K. E. Barner, In-network cooperative
spectrum sensing, in Proc. EUSIPCO, 2009, pp. 15.
HUANG et al.: MULTITASK SPECTRUM SENSING IN CRNs VIA SPATIOTEMPORAL DATA MINING
823
Fei Hu (M12) received the Ph.D. degree in signal processing from Tongji University, Shanghai,
China, in 1999, and the Ph.D. degree in electrical
and computer engineering from Clarkson University,
Potsdam, NY, in 2002.
He is currently an Associate Professor with the
Department of Electrical and Computer Engineering,
University of Alabama, Tuscaloosa, AL. He has published over 170 journal/conference papers and book
chapters. His research has been supported by U.S.
NSF, Cisco, Sprint, and other sources. His research
expertise is in cognitive radio networks and security.