TELKOM
NIKA
, Vol.12, No
.3, Septembe
r 2014, pp. 7
17~724
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i3.101
717
Re
cei
v
ed Fe
brua
ry 26, 20
14; Re
vised Ju
ly 20, 20
14; Acce
pted Au
gust 5, 201
4
A Cognitive Radio Spectrum Sensing Algorithm to
Improve Energy Detection at Low SNR
Agus Sub
ekti
1,2
, Sugihartono
1
, Nana
Rachm
a
na S
1
, Andriy
an
B.Suksmono
1
1
School of Ele
c
trical Eng
i
ne
e
r
ing a
nd In
form
atics, Institut
T
e
kno
l
og
i Ban
d
ung
Jala
n Ganes
ha
No. 10, Band
u
ng 40
13
2,
Indo
nesi
a
, Ph./F
ax: +
6222-2
5
0
166
1/253
41
34
2
Research C
e
nter for Informatics, Indon
esi
an Institute of Scienc
es (LIPI)
Jl.Sangk
uria
ng
21/154
D, Ban
dun
g 40
135, In
don
esia, Ph/F
a
x
:+
62
22-2
5
0
4
7
11/25
04
712
e-mail: a
gus.su
bekti@l
ipi.
go.id
A
b
st
r
a
ct
Energy d
e
tecti
on is a
m
on
g th
e most po
pul
ar
s
pectru
m
sens
ing
met
hod for
spectru
m
se
nsi
ng d
u
e
its low
compl
e
xity. Unfortuna
tely, its perfor
m
a
n
ce
is
po
or
at low
SNR. In this pa
per
w
e
propos
ed
a
spectru
m
se
nsi
ng
meth
od for
cogn
itive ra
dio
netw
o
rk
that impr
oves th
e p
e
rfo
rmanc
e of ener
gy detecti
on
.
T
he
pro
pos
ed meth
od bas
ed on
d
i
strib
u
tion
ana
lysis
us
in
g ku
rtosis as
tes
t
statistic. This com
e
s from
t
h
e
fact that d
i
strib
u
tion
of r
e
ceiv
ed s
i
gn
al
w
h
e
n
a
ch
ann
el
is
occup
i
ed
w
ill
b
e
d
i
fferent fro
m
v
a
ca
nt ch
an
nel
.
Noise
ten
d
s to
hav
e
a Gauss
i
an
distri
buti
on.
Sig
nal
w
h
ich
faces
multip
ath
fadi
ng
dur
ing
the trans
missio
n
w
a
y w
ill hav
e
non
Gaussi
an
distrib
u
tion. S
e
nsin
g a
l
gor
it
h
m
w
a
s teste
d
usin
g ca
pture
d
DTV sig
nal.
R
e
sult
show
s that o
u
r
metho
d
p
e
rfor
ms w
e
l
l
at
low
SNR.
It ach
i
e
v
es pr
oba
bil
i
ty of d
e
tection
o
f
90
% for
10
%
Proba
bil
i
ty of false a
l
ar
m for l
o
w
SNR, belo
w
-20 dB.
Ke
y
w
ords
: co
gnitiv
e
radi
o, spectru
m
sens
in
g, DT
V signal,
kurtosis
1. Introduc
tion
Comp
are to a conventio
nal
radio, cogniti
ve r
adio intro
duces two m
a
in differe
nt feature
s
,
cognitive capability and
re
configurability [1],[2]. Cogni
tive radio allows same
f
r
equency bands
to
be u
s
ed
simu
laneo
usly by
prima
r
y user
and
se
con
dary use
r
[3]. Primary u
s
e
r
, a
s
the o
w
n
e
r
of
spe
c
tru
m
li
ce
nse,
ha
s the
first
usage
pri
o
rity. Seco
nd
ary u
s
e
r
m
a
y
only u
s
e th
e
vacant
sp
ect
r
um
that is not b
e
ing u
s
ed b
y
primary u
s
er. This
sp
e
c
trum
sha
r
in
g cog
n
itive radio ha
s be
en
prop
osed
to
be
u
s
e
d
in
TV’s white space
b
and
s
oppo
rtuni
stically [4]. A n
e
w
ra
dio
access
standard IEEE 802.22 has bee
n issued
lately [5]. A cognitive radi
o platform
may
transmit at t
he
spe
c
tru
m
hol
es in the TV band
s a
s
lon
g
as the
chan
nel is not bei
ng used by the prima
r
y use
r
.
As req
u
ire
d
by the standard, before tran
smitti
ng, se
conda
ry use
r
has to en
su
re that its
transmission will not cause harmful
int
e
rference to the primary us
er by its sens
ing capabil
i
ty.
The
spe
c
tru
m
sen
s
in
g h
a
s to a
c
cu
rat
e
ly detect th
e existing
sp
ectru
m
hol
es to avoid ha
rmful
interferen
ce.
In the othe
r h
and, mi
stake
n
dete
c
tion
will re
sult in lo
wer spe
c
trum
hole
s
utilization.
The mai
n
cha
llenging
of co
gnitive ra
dio
spe
c
tru
m
sen
s
ing i
s
it
s re
q
u
irem
ent to b
e
able to
det
ect
a sig
nal
eve
n
at very lo
w level
with
a stri
nge
nt a
c
ceptabl
e pe
rforma
nce. T
he pe
rforma
nce
usually is
measured by probabilit
y of detection (Pd) and
probability of false al
arm
(Pf). IEEE
802.22
dema
nds
a stri
nge
nt sen
s
in
g re
quire
ment. F
o
r maximu
m
prob
ability of false al
arm
of 10
%, a sensing
algorithm
should ac
hi
eve probability of detection 90
%
for
signal
as
low
as -20 dB
s
i
gnal to
noise ratio (S
NR) [5],[6]. This
means
that
some lic
e
ns
ed s
i
gnals
have to be sens
ed at
a very l
o
w
SNR. In
ad
d
i
tion, wi
rele
ss
cha
nnel
fa
ding
and
noi
se flu
c
tuatio
n will
al
so
b
r
ing
difficulties.
Several
se
n
s
ing
meth
od
s h
a
ve
bee
n p
r
op
osed
to co
mbat t
he a
bove
m
entione
d
chall
enges such as m
a
tched filtering
[7
],[8], feature detecti
on approach
[9,[10]
an
d ene
rgy
detectio
n
[8],[11]-[14]. The
matche
d filtering can m
a
ximize th
e SNR. It can
dete
c
t sig
nal at v
e
ry
low S
NR.
Howeve
r, mat
c
he
d filterin
g
method
re
q
u
ire
s
info
rma
t
ion su
ch
a
s
pilot an
d frame
stru
cture of p
r
imary
sign
al. For the fe
ature d
e
tectio
n
method
whi
c
h relie
s o
n
cyclostatio
narit
y,
the suffici
ent
sign
al info
rmation mu
st
be given
as well. Th
e re
quire
d info
rm
ation intro
d
u
c
e
s
compl
e
xity in
implem
entat
ion, such
as if the
co
gnit
i
ve ra
dio
ha
s to
dete
c
t
several
differe
nt
prima
r
y sig
n
a
ls. On th
e
other
hand,
energy dete
c
tion doe
s n
o
t req
u
ire i
n
formation ab
out
the
sign
al to be
detecte
d. Thi
s
metho
d
ex
ploits
en
ergy differen
c
e b
e
twee
n occu
pied an
d vacant
cha
nnel cond
ition. It compare
s
the ene
rgy of rece
ive
d
signal
with pre
-
defin
ed thre
shol
d. Du
e to
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 71
7 – 724
718
this low com
p
lexity, energy detection
is the mo
st
prefe
r
abl
e method. Ho
we
ver, this method
woul
d be pro
ne to the false detection
si
nce it onl
y rel
i
es on the si
g
nal ene
rgy feature
s
[13],[14]
.
Whe
n
si
gnal
fluctuate
s
or
noise is la
rg
e
,
this me
thod
is likely to fail to distingui
sh b
e
twe
en t
h
e
absen
ce
and
the p
r
e
s
en
ce of the
si
gn
al [8],[
13]. Eventhou
gh it
g
i
ves a
go
od
perfo
rman
ce
at
high SNR, bu
t its perform
a
n
ce i
s
very poor at low S
N
R.
Many metho
d
s h
a
ve b
e
e
n
propo
se
d
to ov
ercome
the short
c
o
m
ings of the
ene
rgy
detectio
n
app
roa
c
h. Some
method
s ba
sed on the ei
g
envalue
s fro
m
the covaria
n
ce mat
r
ix of
the
received si
gn
al were pro
p
o
se
d in [15]-[17]. Same
as in energy de
tection, the method
s also
do
not req
u
ire
primary sig
nal i
n
formatio
n. Howeve
r,
the correspon
ding
comp
utation
a
l com
p
lexitie
s
are
quite l
a
rg
e. In this pa
p
e
r,
we p
r
o
p
o
s
ed
a
simpl
e
r meth
od
usi
ng 4th
orde
r
moment, i.e.
its
kurto
s
i
s
valu
e. Kurtosi
s
of re
ceived
sig
nal is
used
a
s
test
statisti
cs. Thei
r valu
e
s
a
r
e
com
pared
with a pred
e
f
ined thre
sho
l
d to distingu
ish
between
occupi
ed sp
e
c
trum an
d white spa
c
e. A
threshold
is
cal
c
ulate
d
from em
piri
cal
estima
tio
n
of system’
s
noise. With
o
u
t kn
owl
edg
e of
sign
al’s pa
ra
meters, the
target
of
kurto
s
is an
alysi
s
i
s
the
outp
u
t o
f
sign
al’s FFT
. Our expe
rim
ent
sho
w
s that it provide
s
o
p
timal pe
rform
a
nce, e
nablin
g
detectio
n
at
very low S
N
R pri
m
ary
sig
nal
as verifie
d
b
y
simulation
results. The
rest
of this pape
r will
descri
be
sub
s
eq
uently ab
out
formulatio
n of
sp
ect
r
um
se
nsi
ng problem
fo
ll
owed
with expl
anation
of p
r
opo
se
d
sen
s
in
g
method, re
sul
t
s of perform
ance evaluati
on by simu
lati
on usi
ng re
al DTV sig
nal a
nd co
ncl
u
si
on
.
2. Rese
arch
Metho
d
Whe
n
sp
ect
r
um se
nsi
ng
works to det
ect wh
ite
sp
ace b
a
sed o
n
the re
ceive
d
sign
al,
there will b
e
2 possible
co
ndition
s, cha
nnel is va
can
t
or occupie
d
. Deci
sio
n
has to be be mad
e
if the receive
d
si
gnal
is compri
se
of p
r
imary
sign
als with i
nhe
ren
t
noise o
r
it j
u
st
com
p
ri
se
of
sytem noi
se.
Den
o
te the
discrete tim
e
received
sig
nal by r(n)
d
u
ring th
e sen
s
ing
stag
e. The
unde
rlying primary sign
al
is
de
noted
by
s(n)
whil
e w(n
)
i
s
ad
ditive white
Gau
ssi
an
noi
se
(AWG
N). Signal re
ceived
by spe
c
trum
sen
s
in
g is
.
There are t
w
o possibl
e
condtions. If there i
s
no pri
m
ary signal, detecto
r will
only receive noise.
If primary user occupied
cha
nnel by transmitting
si
gnal
s, detect
o
r will rece
ive sign
al and
noise. Th
ere are 2 diffe
rent
con
d
ition
s
on
each time i
n
stan
ce, H
0
: si
gnal is
ab
sen
t
and H
1
: sign
al is present. Suppo
se the
r
e
are
N sampl
e
s for
detectio
n
, the pro
b
le
m can
th
en b
e
model
ed a
s
equatio
n of b
i
nary hypoth
e
s
is
testing:
:
0,
1,
…
,
1
(1)
:
0,
1,
…
,
1
(2)
It is assume
d
that signal a
nd noi
se a
r
e
uncorre
late
d and ea
ch i
s
a
n
i.i.d. seque
nce. Th
e
spe
c
tru
m
sen
s
ing
p
r
oble
m
is the
r
efo
r
e t
o
dete
r
min
e
wheth
e
r the
signal
s
(
n
) exi
s
ts
or not,
ba
sed
on the re
ceiv
ed sig
nal sa
mples
r
(
n
) [11],[15].
Spectrum se
nsin
g ha
s to
be a
c
curately
makin
g
a d
e
c
isi
on
whethe
r a chan
nel i
s
vacant
or occupied.
Deciding occupied
as vacant will result
in prohi
bited transmi
ssion which causes
harmfull inte
rf
eren
ce to
pri
m
ary u
s
er.
While wron
g de
cisi
on of o
c
cu
pied for va
ca
nt make
s lo
wer
spe
c
tru
m
h
o
l
e
s
utilization.
A sp
ect
r
um
se
nsi
ng m
e
thod’s pe
rformance
i
s
ev
aluated
by its
probability of detection
and pro
bability
of false al
arm
. The
detectio
n
alg
o
rithm in spe
c
trum
sen
s
in
g wo
rks ba
sed on Neym
an-Pe
arson’
s theorem, wh
ich
states that the objective is to
maximize
probability of detection
in a fixed probability of false
alarm
=
. T
he dete
c
tor m
a
ke
s de
ci
sion
of H1 if it
fulfills likeli
hood
ratio test (L
RT
) of:
|
|
(3)
Whe
r
e thresh
old
is cal
c
ul
a
t
ed by:
|
:
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Cognitive Radio Spe
c
tru
m
Sensing Algorithm
to
Impro
v
e Ene
r
g
y
Detectio
n …. (Agus Sub
e
kti)
719
Suppo
se the
prima
r
y sig
n
a
l
is a wi
de
-se
n
se
station
a
ry
Gaussia
n
random proce
ss with
varian
ce of
and n
o
ise is
additive wi
de
Gau
ssi
an n
o
i
s
e
with vari
a
n
ce
of
,
li
kelih
ood ratio
of
of equatio
n 3
can be exp
r
e
s
sed a
s
:
∑
∑
(5
)
Takin
g
the na
tural logatitmi
c of
, we will g
e
t the log-like
lihood ratio:
ln
∑
(
6)
Detector
will make
dici
sion of H
1
if
∑
. This dete
c
to
r
cal
c
ulate
s
en
ergy of
received
dat
a an
d
comp
are
s
it
with
a p
r
edefin
ed
thre
sh
old. It’s
calle
d e
n
e
r
gy det
ection
o
r
radio
m
eter.
Energy of th
e re
ceived
si
gnal
w
ill be
comp
osed of
energy of si
gnal a
nd noi
se’
s
varian
ce if th
e chann
el i
s
being
ou
ccup
ied. But if the
prim
ary
sign
al is
ab
sent it
will b
e
e
qual
to
noise vari
an
ce. To g
e
t its
perfo
rman
ce
metrix (P
f
and
P
d
), we
ne
e
d
to defin
e th
e dist
ributio
n
fo
r
both co
nditio
n
s. For th
e la
rge
sampl
e
s,
we can ma
ke
assumption
on the statisti
cal di
stributio
n of
for both conditions to be Gaussan, then, t
he probability of false al
arm and the
probability
of detection
can be expressed a
s
:
;
(7)
Whe
r
e
Q(
) i
s
a q
-
fun
c
tio
n
. As a
cla
s
s of Nyema
n
-Person d
e
tector, we
set th
e thre
shol
d from
equatio
n 7 as:
1
(8)
whe
r
e
is se
t to certain
accepta
b
le required valu
e, mostly set
at 0,1. We can
see fro
m
equatio
n 8 that the thresh
old is a fun
c
tion of t
he noise’s varia
n
ce and the num
ber of sample
s. A
prima
r
y si
gna
l in the t
a
rg
et frequ
en
cy b
and
will b
e
e
a
sie
r
to
be
d
e
tected
if its l
e
vel is re
ceiv
ed
highe
r tha
n
n
o
ise’
s va
rian
ce. Sign
al to
noise radio
(SNR) is u
s
ed
to mea
s
u
r
e
a ratio
betwe
en
the power le
vel of receiv
ed sig
nal (
)
and noi
se’
s
varian
ce. Pro
bability of detection is
a
function of th
e SNR,
and
can be expresed as:
(9)
Energy d
e
tection is
quite p
o
we
rful
spe
c
t
r
um
sen
s
in
g method wh
en
the
si
gnal
is stron
g
e
r
than n
o
ise’s
varian
ce
(hig
h SNR). But,
its p
e
rf
o
r
ma
nce
will
be
p
oor
at lo
w S
NR. At lo
w S
NR,
noise un
certa
i
nty make
s it impossibl
e to detect t
he p
r
ese
n
se of sig
nal even with
high num
ber of
s
a
mples
[8],[14].
We
co
nsi
d
e
r
anothe
r fe
atu
r
e to
imp
r
ove
the
ene
rgy d
e
tection’
s
pe
rforman
c
e. In
wirel
e
ss
comm
uni
cati
on, re
ceived
signal
will
be faded
du
e to multipath cha
nnel. If we che
ck t
he
distribution of
received signal, mostly they w
ill have a certain non
Gaussi
an such as a Rayleigh
distribution.
The di
stri
buti
on of
the received signal (contamina
ted with noi
se) will
be
differen
t
from noi
se o
n
ly. Spectru
m
sen
s
in
g problem th
e
n
can be d
r
a
w
n
as hypoth
e
ses te
sting of
two
con
d
ition
s
, a vacant chan
n
e
l as a Gau
s
sian di
st
rib
u
tion re
ceived
sampl
e
s, an
d
occupie
d
as a
non-Gau
s
sia
n
distrib
u
tion:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 71
7 – 724
720
:
~
̅
,
;
:
̅
,
(10
)
Then we ca
n dedu
ce
con
d
i
t
ion based on
the signal’
s
d
i
stributio
n, a Gau
ssi
an or
not.
Kurtosi
s
is o
n
e
of quite easy measu
r
e of
gau
ssi
anity [12]. Kurtosis i
s
defined a
s
:
3
(11)
Suppo
se that r has zero mean, and r has been n
o
rmali
z
e
d
so
that its variance is equa
l to
one:
1
. Equati
on 16 will b
e
:
3
(12)
Since r i
s
gau
ssi
on, its fourth moment eq
ual to:
3
3
(13)
So, for the samples
with Gaus
sian distri
bution, they will
have kurt
osi
s
equal to 0.
The Kurto
s
i
s
value of Guassian
sampl
e
s w
ill n
o
t exactly equal
to 0 if the numbe
r of
sampl
e
s is
n
o
t long
enou
gh. For the li
mited sample
s of
N:
̅
…
…
, esti
mation of
the kurt
osi
s
is:
∑
̅
3
∑
̅
(14)
Since, th
e
sensi
ng
meth
ods mo
stly p
e
rform
d
e
tect
ion in
fre
que
ncy d
o
main,
received
sampl
e
s r(n)
is tra
n
sfo
r
me
d to freq
uen
cy domain
R(f) by an
FFT
block. Th
e re
sult is compl
e
x-
value
sampl
e
s
R(f). T
he
F
FT outp
u
t co
mpri
sed
of re
al an
d ima
g
in
ary pa
rts. F
o
r the p
u
rp
ose
o
f
out sen
s
in
g method
s, we
set the test statistics
(T) in
to 3 kind
s, and its re
spe
c
t
i
ve perform
a
n
ce
will be evaluated:
;
;
(15
)
We
perfo
rme
d
the
non
-G
aussia
n
ity test for th
e
re
al pa
rts
and
imagina
ry pa
rts. T
h
e
prop
osed spe
c
trum
sen
s
in
g algorith
m
is performed a
s
follows:
Step
1.
R(f) i
s
got by performin
g
N-poi
nt
FFT to the received sam
p
l
e
frame x(n). FFT is
curre
n
tly wid
e
ly use
d
for
comm
uni
cati
on sy
stem e
m
ploying m
u
l
t
icarrie
s
(OF
D
M), the
system
which
is also u
s
e
d
by digital bro
adcastin
g
system.
Step
2.
Kurtosi
s
i
s
ca
lculate
d
for e
a
ch frame
of
M sam
p
le
s of
real p
a
rt a
s
well a
s
ima
g
i
nary pa
rt
usin
g equ
atio
n 20.
Step 3.
Cal
c
ulation of
the test
statistics (T
) usi
n
g equatio
n 21
.
Step
4.
Reje
ct the
n
u
ll hypothe
si
s
H
0
in
favor of
the p
r
e
s
en
ce
of prim
ary
sig
nal tran
smissi
on if T
> thre
shol
d. Otherwise, accept H
0
an
d d
e
cla
r
e the ab
sen
c
e of the
prima
r
y use
r
’
s
sig
nal.
In the application of energy
detection for spectrum
sensing, initial
s
ensi
ng phase will be
started
by noi
se
calib
ration
to mea
s
u
r
e it
s vari
an
ce in
orde
r to b
e
a
b
le to
set the
threshold. T
h
i
s
is also appli
e
d in our p
r
o
p
o
se
d metho
d
. The thre
sh
ol
d is taken em
pirically from
noise sa
mple
s.
To get the threshold, the detector
will
perform
detection in the
condition of H
0
. Dete
ction
ev
ents
will be co
unt
ed to get the
detecto
n rate
which
sh
o
w
s the probabilit
y of
false alarm.. As required
by the
stand
a
r
d, its value
should
be
un
d
e
r
10 %.
Adj
u
stment
of thresh
old
will
be
mad
e
u
n
til th
e
prob
ability of false ala
r
m a
c
hieve 1
0
%. This th
re
shol
d then will
be
taken a
s
d
e
tection th
re
sh
old
to be
co
mpa
r
ed
with te
st statistic in
eq
u
a
tion 1
0
. In t
h
is
pro
p
o
s
ed
metho
d
, the
detecto
r
wo
rks
blindly as it d
oesn’t need i
n
formatio
n of sign
al param
eters.
3. Results a
nd Discu
ssi
on
The pe
rform
ance wa
s m
easure
d
from
its detec
tion
rate in low sign
al-to
-
noi
se ratios
(SNR). Thi
s
sho
w
s the se
nsitivity of propo
s
ed m
e
th
od in dete
c
ti
ng a lo
w level prima
r
y si
gnal
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Cognitive Radio Spe
c
tru
m
Sensing Algorithm
to
Impro
v
e Ene
r
g
y
Detectio
n …. (Agus Sub
e
kti)
721
existen
c
e. Th
e prima
r
y sig
nal so
urce fo
r the ex
pe
rim
ent wa
s take
n from real capture
d
digit
a
l
television
(DTV) sig
nal [20]. The re
aso
n
for
the
choi
se i
s
twofold
s
, for
the type and
its
cha
r
a
c
teri
stics. For the
DT
V sign
al type, it re
p
r
esents one the
spe
c
trum
sh
ari
n
g
cog
n
itive ra
d
i
o
stand
ard, IEE
E
802.22,
wh
ich o
p
e
r
ate
s
i
n
the TV
whit
esp
a
ce.
And the
sig
nal wh
ich wa
s
ta
ke
n
from real broadcasting
signals carrier also w
ill represent the real env
ironment of the wave
prop
agatio
n esp
e
ci
ally
its multipath
fadi
ng chan
nel
’
s
effect.
The capture
d
sign
al’s
frequ
en
cy
is
545 M
H
z
with 6 MHz ba
n
d
width. Give
n the sam
p
in
g rate 50 M
H
z and th
e oversampli
ng fa
ctor
8/7 x bandwi
d
th, we ca
n cal
c
ulate the
sample
rate
or numb
e
r
of sample
s p
e
r seco
nd. T
h
e
output of A
D
C is sample
s with
spe
ed
equal to
68
5
7
sample
s p
e
r 1
ms.
Nu
mber of sam
p
les
taken for
se
nsin
g is pa
rticula
r
ty important
since o
u
r metho
d
works ba
se
d on the re
cei
v
ed
sampl
e
s di
stribution. Even for the abse
n
t of primar
y
sign
als, the d
i
stributio
n of received si
gn
al
will not be Ga
ussian if the numbe
r of sa
mp
les
we too
k
is not suffici
enty long.
Prior to
test t
he p
r
op
osed
detectio
n
met
hod,
we n
eed
to set up th
e
thre
shol
d. As in th
e
energy dete
c
tion, to defin
e the thresh
old for
ou
r
detecto
r, we
perfo
rme
d
a kin
d
of n
o
ise
calib
ration. T
he pu
rp
ose o
f
noise calib
ration in th
e e
nergy
dete
c
tion i
s
to m
e
a
s
ure the
noi
se’s
varian
ce. In orde
r to get it,
the energy detecto
r
wo
rks in the conditi
on of no prim
ary sign
al (H
0
).
Once we g
e
t its value, the
varian
ce’
s
th
en u
s
ed
to set up the th
re
shol
d for th
e
detecto
r. In t
h
e
prop
osed m
e
thod, inste
a
d
of mea
s
ure
the varian
ce,
we too
k
the
noise’s
ku
rt
osi
s
. Since the
measured signal
is
taken
at the output
of fast
Fouri
e
r transform
(FFT
), each
sampl
e
will have
real a
nd imag
inary pa
rts. Kurtosi
s
valu
e
s
we
re
ta
ken
from both p
a
rts. Acco
rdin
g
to equation 2
2
,
t
e
st
st
at
i
s
t
i
c
s
of
T
1
, T
2
, and T
3
fo
r n
o
ise
only a
r
e cal
c
ul
ated.
Cal
c
ulatio
n
wa
s pe
rform
e
d
seq
uently to each frame
of M numbe
r of sample
s.
The test st
atistic for e
a
c
h fram
e’s t
hen
comp
ared wit
h
a pred
efine
d
threshold.
We
set the thre
shol
d to 0.04. False de
cisi
on (H
1
) wi
ll
occur
wh
en t
he test
statist
i
c is a
bove t
he thre
sh
old.
The num
be
r of false de
ci
sion
devided
b
y
total num
ber
of test
statisti
cs shows the
false
alarm
rate or the probability of fal
s
e alarm
(P
f
). W
e
set the numb
e
r of FFT poi
nts to 2048 a
nd 819
2, t
he same n
u
mb
er used by the
DTV stan
dard.
Figure 1. False alarm rate as a fun
c
tion
of M
Figure 1 sho
w
s the result, the false ala
r
m rate
a
s
a functio
n
of the numbe
r of sample
s.
False
ala
r
m
rate
sho
w
s t
he n
u
mbe
r
o
f
times th
e
detecto
r m
a
d
e
wron
g d
e
cision
of va
ca
nt
cha
nnel
as
o
c
cupie
d
. Thi
s
will mi
ss th
e
oppo
rtunity
for
cog
n
itive radio to m
a
ximize
sp
ectru
m
holes ulitization. Thi
s
will
be mai
n
ly a f
unction of
the threshold. I
f
we
set the threshol
d hi
gher,
this will
result
in lower fal
s
e alarm, but
will al
so
make lower detection rate
when later it has
to
work to d
e
te
ct p
r
ima
r
y si
gnal
s. As a
gene
ral
sp
ectrum
sen
s
in
g
re
quirement,
false
ala
r
m
rate
sho
u
ld not ex
cee
d
P
f
maximum of 10 %, as sho
w
n in
the figure a
s
a dash
ed lin
e. By setting the
fixed thre
shol
d, we inte
nd t
o
mea
s
u
r
e th
e othe
r facto
r
s, incl
udin
g
n
u
mbe
r
of
sa
mples in a f
r
a
m
e
use
d
in the test statisti
cs
cal
c
ulatio
n and also t
he effect of FFT size. The six li
nes in the fig
u
re
sho
w
th
e result of expe
ri
ment for ea
ch num
be
r
of
FFT p
o
ints a
nd fo
r 3
test
statisti
cs.
F
a
lse
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 71
7 – 724
722
alarm rate wi
ll
decrea
s
e a
s
nu
mbe
r
of sampl
e
s
i
n
crease. In mea
n
s that fo
r a
fixed threshol
d,
increasing number
of
sam
p
les will
make detector perform
be
tter due to the
closer to it
s
Gausian
distrib
u
tion fo
r large
sam
p
l
e
s. As
we
m
entione
d
bef
ore, for the
real Ga
usssia
n noi
se
sam
p
les,
output of the FFT
will
compri
se
of the real
pa
rt and imagi
nary
part, ea
ch wi
ll have G
a
ussian
distrib
u
tion a
nd its kurt
osi
s
sho
u
ld be
0. Sinc
e we
have 2 kurt
osi
s
values
of the real and
imagina
ry p
a
r
t, we
can
u
s
e
both
or o
m
it one
of
t
hem. F
r
om
the figu
re, if
we
co
mpa
r
e
the
performance
of the te
st
st
atistics, we will get t
hat the test stati
s
tic
of T
1
perfo
rm
s bette
r than
T
2
and T
3
. Usi
ng the both
part by avera
g
ing give the
highe
st perf
o
rma
n
ce, i.e. the lowe
st false
alarm rate for the sam
e
number of
samples
(M). Since the vari
an
ce of kurtosis estimates
will
decrea
s
e
wit
h
the in
cea
s
i
ng num
be
r o
f
sample
s
(M
), we n
eed t
o
get the M
minimum
whi
c
h
fulfills the requirem
ent. And if we used t
he T
1
, an ave
r
age
ku
rtosi
s
of the real a
n
d
the imagin
a
r
y
sampl
e
s as the test
statist
i
c, the result also
su
ge
st that the nu
mb
er of
sampl
e
sho
u
ld at le
a
s
t
3000
0 to achi
eve false ala
r
m rate belo
w
10 %.
After we
set t
he thresh
old
and it ha
d be
en c
onfirm
e
d
its re
sp
ectiv
e
pro
bability
of false
alarm
i
s
und
e
r
10
%, we p
e
rform
ed sim
u
lation
to
get
the main
pe
rforma
nce met
r
ic, the
dete
c
t
i
on
rate. A captu
r
ed DTV Sign
al is as
prim
a
r
y sign
al.
The
main purpo
se of simulatio
n
is to mea
s
u
r
e
the dete
c
tor’
s sen
s
itivity. The sen
s
itivity
here
me
a
n
s
how l
o
w th
e level of prim
a
r
y sign
al can
be
detecte
d by the pro
p
o
s
ed
method. The
prima
r
y
sign
a
l
was a
dded
by generated
sequ
en
ce
s of
white Ga
ussi
an noi
se. Sig
nal’s level
wa
s adju
s
ted
to rep
r
e
s
ent
the
intended sig
nal-to
-
noi
se ratio
(SNR) value
s
particul
a
ry to represent the sam
p
les value at low SNR. For each samples in a
frame, fast F
ourie
r tra
n
sfo
r
m (FF
T
) wa
s pe
rfor
m
ed.
Estimation o
f
kurto
s
is
wa
s cal
c
ul
ated
for
each re
al sa
mples
and i
m
agina
ry sa
mp
les,
an
d then
use
d
for te
st
statistic
(T
1
).
The te
st stati
s
tic
of frame
s
(M
) of 60,0
00
sample
s ea
ch
wa
s calculat
ed. The te
st
statistics
(T
1
) for ea
ch f
r
a
m
e
then’s
com
p
a
r
ed to the thresh
old ab
ove
(0.04)
. Th
e prop
osed det
ector
will ma
ke de
ci
sion
of
cha
nnel i
s
be
ing occu
pied
if the
test statistic is a
bov
e the thre
sho
l
d. The total numbe
r of te
st
statistics whi
c
h are above
the thre
sh
old
devide by the total numbe
r of
frames o
r
the ratio of test
statistic
abov
e thre
shol
d to the all ava
ilable test
statistic value
s
is a dete
c
ti
on rate fo
r the
respective SNR (probability of detection).
Figure 2. Det
e
ction
Rate for N-FFT
=8,1
92, M=60,0
0
0
Figure 2
sho
w
s the
re
sult
of the
experimen
t. The
d
e
tection
rate
of pro
p
o
s
ed
method
wa
s com
pare
d
with the energy dete
c
tio
n
, both usin
g
simulation fo
r several sig
n
a
l to noise ra
tio
(SNR). T
he
prop
ose met
hod’
s pe
rformance’s bet
t
e
r tha
n
en
ergy detectio
n
and It al
so g
i
ves
accepta
b
le
p
e
rform
a
n
c
e.
They p
e
rfo
r
m well for l
o
w S
N
R of
unde
r
-20
d
B
. As the
e
nergy
detectio
n
me
asu
r
e
s
the
e
nergy
differe
nce, it
s
pe
rfo
r
man
c
e dete
r
iore
s
rapidly once
the pri
m
ary
sign
al’s po
wer i
s
clo
s
e
o
r
bel
ow the
noise’s varia
n
ce.
Our me
thod
whi
c
h b
a
se
d o
n
kurt
osi
s
measures th
e differen
c
e
of its
st
atisti
cal di
stributio
n
.
The
distri
bu
tion of n
o
ise
is G
a
u
ssi
an.
We
can
say that i
n
the
con
d
itio
n of H
0
, the
d
i
stributio
n of t
he FFT
outp
u
t sh
ould
be
Gau
ssi
an. Th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Cognitive Radio Spe
c
tru
m
Sensing Algorithm
to
Impro
v
e Ene
r
g
y
Detectio
n …. (Agus Sub
e
kti)
723
approa
ch
will
result in rob
u
st result a
s
long a
s
the
n
u
mbe
r
of sa
mples i
s
sufficient, so that
its
distrib
u
tion wi
ll remain a
s
i
n
the origin
al distrib
u
tion.
Figure 3. Det
e
ction rate co
mpari
s
o
n
on
numbe
r of sa
mples
Figure 4. Det
e
ction
Rate for Several
N-FFTs
Number of samples to calcul
ate the test
statistic
also has im
pact on probability of
detectio
n
. Since the va
ri
ance of ku
rtosi
s
esti
m
a
tion is a fun
c
tion of 1/M, we cond
uct
an
experim
ent t
o
compa
r
e
th
e pe
rform
a
n
c
e for
differe
nt numb
e
r of
sample
s in
a f
r
ame
(M
). If
we
modify the n
u
mbe
r
of
sa
mples to be
twice
we
get
a comp
ari
s
o
n
re
sult. In F
i
gure
3, we
plot
simulatio
n
re
sult for M
=
3
0
,
000 an
d M
=
60,000
re
spe
c
tively. This f
i
gure
sh
ows t
hat the samp
le
size of 60,00
0 is accepta
b
l
e to detect primary
si
gnal
at low SNR.
The 60,00
0 sample
s is eq
ual
to sen
s
ing ti
me of 10 m
s
. We also co
ndu
cted an
e
x
perime
n
t to measure the
effect of FF
T
.
Figure 4 sh
o
w
s the effe
ct
of N-FFT to
the detecto
r
perfo
rman
ce.
When n
u
mb
er of FFT set as
variable,
we
can see that
probability of
detecti
on increases
with t
he in
creasing of FFT
point
s.
This resul come
from
the fact
that the
higher the FF
T point the
m
o
re accurate it
will
represent
the sig
nal in
time domai
n, includi
ng
its dist
ri
butio
n. The N-FF
T of 8192
g
i
ves a
c
cepta
b
le
perfo
rman
ce.
It performs
well for low SNR of unde
r -2
0 dB.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 71
7 – 724
724
4. Conclusio
n
In this pape
r we pro
p
o
s
e
d
a new spe
c
trum
sen
s
in
g method ba
sed o
n
ku
rto
s
is. Ou
r
method i
s
abl
e to wo
rk
blin
dly without p
r
imary
si
gnal
kno
w
le
dge. Its pe
rform
a
n
c
e improves t
he
energy dete
c
tion a
nd it is co
mply
with t
he sta
ndard re
qui
rement with
o
u
t adding
m
u
ch
compl
e
xity. Two mai
n
fact
ors
have be
e
n
investigate
d
, numbe
r of
FFT point
s
and nu
mbe
r
o
f
sampl
e
per
testing fram
e. Numbe
r
of FFT poin
t
s gives mo
re sig
n
ifica
n
t impact on the
perfo
rman
ce
with mini
mu
m value i
s
8,
192
point
s.
Since
thre
sh
old
is
got from
real me
asure
m
ent
of noise
sam
p
les, ou
r met
hod can ad
ap
t to various n
o
ise
cha
r
a
c
te
ristics.
Referen
ces
[1] FCC.
Spectrum
Policy Task Force Report
. in Procee
di
ngs
of F
CC ’02, F
CC Doc.ET
Docket No. 02-
135 N
o
v. 200
2.
[2]
S. Ha
y
k
in. C
o
g
n
itive R
a
d
i
o: B
r
ain-Emp
o
w
e
re
d W
i
reless
Co
mmunicati
ons.
IEEE Journal on Selected
Areas In Co
mmu
n
ic
ations
. 2
005; 23(
2).
[3]
J. Mitola, G.
Maguire. Cognitive radio:
Makin
g
soft
w
a
re radi
os mor
e
pers
ona
l.
IEEE Personal
Co
mmun
icati
o
ns
. 1999; 6(
3).
[4] LE.Doy
le.
Essentia
ls of Cog
n
i
tive Ra
dio
. Ca
mbridg
e Un
iver
sit
y
Press. 20
0
9
.
[5]
IEEE std. 802.22-2011
IEEE
Stand
ard for
W
i
reless R
e
g
i
ona
l Are
a
Net
w
orks Part 22
: Cogn
itive
W
i
reless RA
N
MAC & PHY s
pecific
ations:
Polici
e
s a
nd
pr
oced
ures for
o
perati
on i
n
th
e
T
V
Bands.
201
1.
[6] SJ.
Shellhammer.
Spectru
m
Sensi
ng i
n
IE
EE 802.2
2
.
Pr
ocee
din
g
of IA
PR W
o
rksho
p
on C
ogn
itive
Information Pr
ocessi
ng. Sant
orini-Gre
e
ce. 2
008.
[7]
D. Cabr
ic, A.
T
k
achenko, R
W
. Broderse
n.
Spectru
m
sen
s
ing
meas
ure
m
e
n
ts of
pil
o
t, en
ergy, a
n
d
colla
bor
ative d
e
tection
. in Pr
o
c
.IEEE Militar
y
Commun
i
cati
o
n
s Confer
enc
e. 2006: 1
–7.
[8]
A. Saha
i, D.
Cabric.
Spectr
um se
nsin
g: f
und
a
m
ent
al
li
mits
an
d pr
ac
tical c
hal
len
g
e
s
. in IEEE
Internatio
na
l Symp
osi
u
m on
Ne
w
F
r
onti
e
rs D
y
namic S
pect
r
um Access Ne
t
w
orks. 2
005.
[9]
S. Enserink, D.
Cochran.
A
cy
clostatio
nary fe
ature
detector
.
in Pr
oce
edi
ng
s of Asil
omar
Confer
ence
on Sig
nals, S
ystems and Co
m
puters. 199
4;
2: 806–
810.
[10]
YP. Lin, C. H
e
.
Subsecti
on-
avera
ge cycl
o
s
tationary fe
ature d
e
tection
i
n
cog
n
itive r
a
dio
. in Proc.
Internatio
na
l C
onfere
n
ce o
n
Neur
al Net
w
o
r
ks and Sig
n
a
l
Processi
ng. 20
08: 604
–6
08.
[11] SM.
Kay
.
F
und
amenta
l
s of Statistical Si
gna
l
Processi
ng: De
tection T
h
e
o
ry
. Upper Sa
dd
le
River, Ne
w
Jerse
y
: Pr
entic
e-Hal
l
. 199
8.
[12]
SJ. Shell
hamm
e
r, S. S
hankar,
R.
T
andra, J. T
o
mcik.
Performa
nce of p
o
w
e
r detector se
n
s
ors of DT
V
sign
als in IEE
E
802.22 W
R
A
N
s
. Proceed
in
gs of the F
i
st Internat
i
o
n
a
l W
o
rksho
p
on T
e
chno
log
y
a
nd
Polic
y
for Acce
ssing Sp
ectru
m
(T
APAS). 2006.
[13]
A. Sonne
nsch
ein, P. M. F
i
sh
man. R
adi
om
etric detecti
on
of spre
a
d
spe
c
trum signa
ls
in no
ise
of
uncerta
int
y
po
w
e
r.
IEEE Trans. Aerospace E
l
ectron. Syst
, 1992; 28(
3): 654
–66
0.
[14]
R.
T
andra, A. Saha
i.
F
unda
me
ntal li
mits on detectio
n
in
l
o
w
SNR under
noise u
n
certai
nty
. in Proc.
Internatio
na
l Confer
ence o
n
W
i
reless Netw
o
r
ks, Co
mmu
nicati
ons an
d Mobil
e
Comp
u
t
ing. 200
5; 1:
464
–4
69.
[15]
YH. Z
eng,
YC
. Lia
ng. E
i
g
e
n
v
alu
e
b
a
se
d s
pec
trum s
ensi
ng
alg
o
rithms
for cog
n
itiv
e r
adi
o.
IEEE
T
r
ans. Commu
n.
2009; 5
7
(6): 178
4-17
93.
[16]
YC. Lia
ng, Y
H
. Z
eng, EC
Y. Peh, AT
.
Hoa
ng. Se
nsi
ng thro
ug
hput
tradeoff for
cogn
itive ra
dio
net
w
o
rks.
IEEE Trans. Wireless Comm
un.
200
8; 7(4): 132
6–1
33
7.
[17] H.
Urko
w
i
tz.
Energy d
e
tectio
n of unkn
o
w
n
deterministic si
g
nals
. Proc. IEEE. 1967; 55(
4): 523–
53
1.
[18] Apurva
M
o
d
y
.
Spectrum
Se
n
s
ing of th
e DT
V in the vic
i
nity
of
the Pilot Us
ing H
i
g
her Ord
e
r Statistics
.
IEEE 802.22-
0
7
/037
0r0. Jul
y
200
7.
[19] A.H
y
var
i
n
en,
e
t.al.
Indepe
nde
nt Compo
n
e
n
t Analys
is
. Wiley-Interscience. 2001
[20] I.Garrison,
et.al.
DT
V C
han
ne
l C
haracter
i
z
a
ti
on.
C
onfer
enc
e o
n
Inform
ati
on Sc
ienc
es
a
nd S
y
stems
(CISS) 2001, J
ohns H
opki
n
s Univers
i
t
y
. Mar
c
h 200
1.
Evaluation Warning : The document was created with Spire.PDF for Python.