Professional Documents
Culture Documents
Spectrum Sensing in Cognitive Radio
Spectrum Sensing in Cognitive Radio
I. I NTRODUCTION
After initially being proposed in [1], cognitive radio is
being tauted as the next Big Bang in wireless communications.
As [2] asserts, conventional fixed spectrum allocation results
in a large part of frequency band remaining underutilized.
Channels dedicated to licensed(primary) users are out of
reach of unlicensed users, while the licensed users hardly
occupy the channel completely, at all times. Cognitive radio
revolution hopes to tap into this inconsistency and attempts
to utilize the channel in its full capacity. In this paradigm,
either a network or a wireless node changes its transmission
or reception parameters to communicate efficiently avoiding
interference with licensed or unlicensed users. This alteration
of parameters is based on the active monitoring of several
factors in the external and internal radio environment, such
as radio frequency spectrum, user behaviour and network
state.[3]. While the above description is essentially that of
full cognitive radio, the present work will be primarily focused
on spectrum sensing cognitive radio. In this paradigm, the
secondary unlicensed users keep sensing the spectrum to
determine if a primary user is transmitting or not; and then
they occupy the idle band and leave it as soon as the primary
user kicks in. So it is extremely important that we employ
efficient and robust spectrum sensing techniques to determine
the presense of the signal from the primary user, and a review
of those techniques is the basis of this term paper.
In the description that follows, the following system model
has been used. Assuming there are M > 1 antennas at the
p(x|H1 )
,
p(x|H0 )
N
1
Y
k=0
p(x(k)|Hi ).
N 1
1 X
2
|x(k)| ,
N
k=0
Fig. 1.
N
1
X
xT (k)Rs (Rs + n2 I)
x(k).
k=0
1
N
1
X
x (k)x(k).
k=0
N
1
X
sT (k)x(k).
k=0
So, we notice that the LRT reduces to different known detectors under given constraints.
III. S IGNAL P ROCESSING T ECHNIQUES FOR
T RANSMITTER D ETECTION
Mainly three techniques are in vogue for transmitter detection, which are described below.
A. Matched-Filter
Matched filter optimizes detection by maximizing received
SNR. Given that it has a priori knowledge of the primary
signal, coherency makes sure that only O(1/SNR) samples
are needed for effective detection, thereby making detection
faster so that an idle channel can be quickly occupied without
delay. On the bleak side, knowledge of the primary signal
might not be available. Furthermore, in case there are a lot
of primary users, the radio receiver would have to have a
dedicated matched filter for each different one of them.
B. Energy Detector
To counter the above problem, non-coherent detection can
be done through energy-detectors ([5]). Consider Figure 1.
The basic idea is to compare the enrgy of the received signal
to a threshold to determine the presence of the primary signal.
It can also be implemented by averaging the FFT of the
signal over frequency bins and comparing with a threshold.
=Q
1
N .
n2
Similarly, for H1 , T (x) has a Gaussian distribution with 1 =
2
4
4
s2 + n2 and 12 = N1 [E|s(k)| + E|n(k)| (s2 n2 ) ].
Assuming s(k) to be a complex PSK modulated signal with
4
E|s(k)| = s4 , we have 12 = N1 (2 + 1)n4 , where =
s2 /n2 . Then, for threshold , probability of detection is
Pd () = P r(T (x) > |H1 )
!
s
N
=Q
1
.
n2
2 + 1
Hence, for a target false alarm of Pf a , probability of detection
1
1
Pd = Q
.
Q (Pf a ) N
2 + 1
Non-coherent detection requires more number of samples
(O(1/SNR2 )) to give the same performance as the coherent detector. [6] presents many more disadvantages of this
technique. The threshold is ambiguous as it is susceptible to
varying noise levels. In-band interference can give erroneous
results even if adaptive thresholding is followed. Also, the
detectors failure to recognize interference rules out the use of
any adaptive signal processing for interference cancellation.
C. Cyclostationary Detection
Digital modulated signals usually exhibit cyclostationarity, i.e. their statistical parameters vary periodically in time.
Since different types of signals have diffrent non-zero cyclic
frequencies, they can be identified from their signature. [7]
elaborates a few statistical tests to detect cyclostationarity.
One of the useful techniques is determination on the basis of
spectral-correlation density(SCD). For a source signal x(t),
the received signal after wireless transmission can be written
as
y(t) = x(t) h(t)
Sy (f ) = H(f + )H (f )Sx (f )
2
2
with being the cyclic frequency of x(t), and other symbols
have their usual meaning. While a signal takes nonzero values at a few nonzero cyclic frquencies, noise is devoid of
cyclostationarity, so that noise SCD is zero at all nonzero
cyclic frequencies, there by making analysis of SCD a valid
and effective technique. The advantages of the process include
robustness against noise and channel fluctuations. On the flip
side, it requires a high sampling rate, many samples for complex SCD computation and the possibility that sampling time
error and frequency offset could affect the cyclic frequencies.
IV. C OOPERATIVE S ENSING
Given a master node, many cognitive nodes can interact
amongst each other to detect the presence of primary signal,
leading to increased sensitivity due to their dispersed nature.
They can either send their collected data to the master, letting
the master do data fusion to decide on the presence of
the primary signal; or they could each send their individual
decisions and the master conducts decision fusion to take a
decision.
A. Data Fusion
Let each user employ energy detection and compute received source signal as
Ti (x) =
1
N
N
1
X
|xi (k)| ,
k=0
PM
with gi 0 and
i=1 gi = 1 (As one might notice, the
notations of the previous section have been modified a bit
for simplicity the M antennas are now being treated as M
cognitive nodes, without any implication on the validity of the
discussion). gi s are decided using a priori knowledge, if it is
available. [8] asserts that for low-SNR, optimal gi s are given
by
2
gi = PM i
, i = 1, . . . , M,
2
k=1 k
with i2 being the received source signal power of user i. The
decision statistic is then compared to a threshold to decide on
the presence of the primary user, just as in a general energy
detector case.
B. Decision Fusion
While [9] asserts that the Chair-Varshney rule based on a
log likelihood ratio test is the optimal decision fusion rule,
it also describes two other simplistic rules. OR fusion rule
predicts the presence of the primary signal if at least one of
M
Y
(1 Pd,i ),
i=1
Pf a = 1
M
Y
(1 Pf a,i ),
i=1
Pf a =
M
Y
Pf a,i .
i=1
The majority rule asserts that if more than half of the received
decisions are 1s, then the master rules in favor of the presence
of the primary signal. In general, if the master decides 1 when
K out of the M decisions are 1, then the probabilities become
M
K
X
M
M Ki
K+i
Pd =
(1 Pd,i )
(1 Pd,i )
,
K
+
i
i=0
Pf a =
M
K
X
i=0
M
M Ki
K+i
(1 Pf a,i )
(1 Pf a,i )
.
K +i
DT =
X1 . . . XM
X12 C11 . . . X1 XM C1M
Z=
X2 X1 C21 . . . X2 XM C2M
...
2
XM X1 CM 1 . . . XM CM M
So, Z is formed by appending X with raster scanned form of
XXT C. Thus the first M elements of Z are the elements
of X, the next M are the first row of XXT C etc. p is the
vector formed by appending h with A in raster scanned form.
So, we essentially have to solve for optimal p that maximizes
deflection.
Now Z has zero mean under H0 . So, the deflection of S(Z)
is given by
2
pT
DS = T
p Kp
with = E1 (Z) and K = E0 (XXT ). The problem now
essentially is to optimize over ps that are not in the singular
space of K.
C. 2-Node Cooperation
In the previous subsection, the basic paradigm was communication of data or decision to a master node which has the
authority to take the final decision. [11] describes a 2-node
cognitive user model, where decision is taken by cooperation
between the nodes without the involvement of the master. For
notational convenience, lets define the system model for this
section as follows:
y = fx + w
where x is the sent signal, y is the received signal, f and w are
fading coefficient and noise respectively, modelled as complex
Fig. 2.
Fig. 3.
Gaussian random variables. Now consider, for simplicity, a 2user netwrok. Let the radios be U1 and U2 . The two need to
vacate the band dedicated to a primary user as soon as the
primary signal is transmitted. However, if a user is situated at
the decodability boundary (U1 in Figure 2), the time to detect
the primary signal is too large (more samples needed as SNR
is too low) for comfort. But intuitively, we notice that U2 can
get a better detection time and can relay that to U1 , thereby
reducing the total time of detection by U1 .
Assuming that both the users communicate with a single
base station, we can assign U2 the role of relaying U1 s signal.
By the use of slotted transmission, U1 and U2 can transmit
in successive slots. From Figure 3, U1 transmits in time
slot T1 , while U2 listens. In T2 , U2 relays the previous slot
information. Incorporating the possibility that the primary user
is transmitting, we have the following expression for the signal
received by U2 from U1 in slot T1 ,
y = hp2 + ah12 + w,
E[|ah12 | ] = P G12 ,
2
assures that
2
E[|
y | ] = P2 + P G12 + 1
yh12 + hp1 + w
P
2
E[|
y| ]
2 P
M
[
C m.
m=1
C m C n = , C =
P
.
2 + P G12 + 1
Y = H + W
which of the following holds
H1 : = 1
m
H1m : xm
j (t) = s(t) + vj (t)
with j = 1, 2, . . . , dm , m = 1, 2, . . . , M, t =
m
1, 2, . . . , Ti . s(t), xm
j (t), vj (t) are the primary signal and
received signal and the noise at jth node in ith cluster,
respectively. Ti is determined from time-bandwidth product.
It is assumed that s(t) and {vjm (t)} are independent of each
other for all m, j, t. Again, noise is assumed to be zero-mean
white, may be non-Gaussian. The spectrum sensing SNR is
defined as
Z T0
1 1
2
|s(t)|
s = 2
v T0 0
where T0 , v2 are signal duration and power of sensing noise.
Using the local decision function : CT0 +11 {0, 1}, every
node makes a decision. The decision function can be any thing,
ranging from energy detector, coherent detector, wavelet-based
detector or cyclostationary detector. The local decision of each
radio is denoted by bm
j such that
T
m
0
({xm
j (t)}t=0 ) = Hi : bj = i
H0 : = 0.
Now, a simple energy detector can be used to detect the
presence of the primary signal. The model can be further
extended for more than 2 nodes a detailed discussion of
which can be found in [11].
D. Wideband Detection
The methods discussed above usually pertain to detection
of primary user in a narrow fixed band. Wideband spectrum
sensing would require much more resourses, and as such
using only one cognitive node for sensing the whole band is
counter-intuitive. [12] propose a cooperative shared spectrum
sensing model for wideband detection in a noisy channel. The
idea is essentially to form a number of groups of cognitive
radios, each group dedicated towards detecting a particular
band in the whole of frequency spectrum. The nodes in a
single group cooperate amongst themselves, taking decisions
together, thereby effectively allocating system resourses.
First, the wide bandwidth of f = [f1 , fM +1 ) to be scanned
is divided into M subbands, viz.
f m = [fm , fm+1 ), f =
M
[
m=1
f m.
f (ym |H0 )
f (ym |H1 )
N
1
X
log g(xn ) + 2U
n=0