TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6297 ~ 6312
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.593
8
6297
Re
cei
v
ed Ma
rch 9, 2
014;
Re
vised
Ma
y 19, 2014; Accepted June 5,
2014
Three Decades of Development in DOA Estimation
Technology
Zeesh
an Ah
mad*
1
, Iftikh
ar Ali
2
1
School of Co
mmunicati
on E
ngi
neer
in
g, Ch
ong
qin
g
Un
iver
sit
y
, P.R.Chi
n
a
2
Militar
y
Col
l
eg
e of Signa
ls, Nation
al Un
iver
s
i
t
y
of Scie
nce
&
T
e
chnol
og
y,
Pakistan
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: engr.zees
ha
n@h
o
tmail.co
m
1
, ia
y
ous
afzai
@
hotmai
l
.com
2
A
b
st
r
a
ct
T
h
is p
aper
pr
esents
a br
ief
overvi
ew
of
narro
w
b
a
n
d
di
rection
of arr
i
val (DOA)
estimatio
n
alg
o
rith
ms
an
d
techn
i
qu
es. A
compre
hens
ive
study
is carr
ie
d o
u
t in
this
p
a
per to
inv
e
stig
ate a
n
d
eva
l
u
a
t
e
the p
e
rfor
manc
e of var
i
ety of
alg
o
rith
ms for
DOA esti
matio
n
. T
w
o categor
ies of
DOA esti
mati
on
al
gorith
m
s
are c
onsi
der
ed
for d
i
scussi
on
w
h
ich
are
Cl
a
ssical
metho
d
s
an
d S
ubsp
a
c
e
b
a
sed
tech
ni
ques.
Class
ica
l
meth
ods
incl
u
de Su
m-
an
d-
Delay
metho
d
and
Ca
po
n
’
s
Mini
mum Va
rianc
e Distorti
onl
ess Res
p
o
n
se
(MVDR) w
h
ile
Subsp
a
ce b
a
s
ed tech
niq
ues
are
multi
p
l
e
si
gna
l class
i
ficat
i
on (MUSIC)
a
nd T
he Mi
ni
mum
Nor
m
Techni
q
ue. Also ESPIRIT techniqu
e is eval
uat
ed. Inefficie
n
cies ar
e poi
nted o
u
t and so
lutio
n
s
ar
e
sugg
ested to
overco
me thes
e shortfal
l
s. Si
mu
lati
on re
s
u
lt
s show
s that the MUSIC
alg
o
rith
m is a
b
l
e
to
better repr
ese
n
t the DOAs of
sign
als w
i
th
more pr
o
m
in
ent peaks.
T
h
e
Mi
n-Nor
m
alg
o
rit
h
m als
o
i
dentifi
e
s
the DOAs of si
gna
ls si
mi
lar to
t
he MUSIC
al
gorith
m
, b
u
t pr
oduc
es sp
urio
us pe
aks at
other l
o
cati
ons.
T
h
e
MVDR
meth
od
ide
n
tifies t
he
DOAs of sig
nal
s, but the
loc
a
ti
ons ar
e n
o
t re
p
r
esente
d
by
sh
arp p
eaks,
due
to
spectral l
eak
a
ge. T
he class
i
cal be
a
m
for
m
e
r
also pr
od
uce
s
several s
puri
ous p
eaks. MUSIC show
hi
ghe
r
accuracy a
nd
resol
u
tion th
an
the other al
g
o
rith
ms. It
should b
e
note
d
that MUSIC is mor
e
ap
plic
ab
l
e
beca
u
se it can
be use
d
for different array g
e
o
m
etri
es.
Ke
y
w
ords
:
na
rrow
band DOA
estimati
on, arr
a
y sign
al proc
e
ssing, music, E
SPRIT
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Accordi
ng to the definition
of
IEEE “Ante
nna is a transmitting or
receiving system
that is
desi
gne
d to radiate o
r
receive el
ectro
m
ag
neti
c
wave
s” It has b
een
long de
bat
ed in
electromagnetic system
s li
terature
whether antenna arrays play a
significant role in satellite,
RADA
R, G.P.S and lon
g
distan
ce
co
m
m
unication.
F
i
nding
s of
so
me re
ce
nt e
m
piri
cal litera
t
ure
sho
w
that prope
rly desi
g
ned anten
na
arra
y syste
m
, operating
autonomo
u
s
ly, along with
optimize
d
an
d robu
st algo
rithm plays an
instrume
nt
al role in upliftin
g
the perform
ance of satelli
te
navigation
an
d commu
nica
tion sy
stem
s. Whil
e e
ngin
eers have ge
nerally
re
ach
ed
a
con
s
en
sus
on the
ce
ntral rol
e
of a
n
t
enna a
r
rays in Satellit
e navigation sy
stem
s,
G.P.S, RADARS
a
n
d
comm
uni
cati
on system
s
gro
w
th, theoretical and e
m
pi
ri
cal wo
rk suppo
rting this co
ncept is still
very much in
prog
re
ss.
The ante
nna
array refe
rs t
o
a set of microph
one
s o
r
antenn
as
co
n
necte
d an
d a
rra
nge
d
in a re
gula
r
stru
cture to form a
singl
e
antenn
a
tha
t
is able to p
r
odu
ce
a re
q
u
ired
dire
ctio
nal
radiatio
n pattern, whi
c
h
we can
not achi
eve throug
h individual ant
enna
s. For some appli
c
ati
o
n
s
singl
e el
eme
n
t anten
na
s are u
nabl
e
to me
et th
e gai
n o
r
ra
diation
pattern requi
reme
nts.
Combi
n
ing
several
sin
g
le
antenn
a ele
m
ents in
an
array can
be
a
possibl
e solut
i
on [1]. In GP
S
and satellite navigation sy
stem we ofte
n requi
re
very high dire
ctivity and the
singl
e-el
eme
n
t
antenn
a fails to achieve
this requi
rem
ent be
ca
u
s
e
the ra
diatio
n pattern of
singl
e-elem
ent
antenn
a is compa
r
atively wide an
d ha
s low di
re
ct
ivity (gain). Th
ough hig
h
di
rectivity can
be
achi
eved by enlargi
ng th
e dimen
s
ion
s
of singl
e
element ante
nna but it may leads to the
appe
ara
n
ce o
f
multiple side
lobes a
nd te
chn
o
logi
cally inco
nvenient
sha
p
e
s
and d
i
mensi
o
n
s
[2].
Another ap
proach is to increa
se the ele
c
tri
c
al si
ze of an antenn
a b
y
constructin
g
an assem
b
l
y
of
radiatin
g ele
m
ents in a p
r
ope
r ele
c
tri
c
al and geo
m
e
trical
config
uration – a
n
tenna a
r
ray. Not
necessa
rily but for sake of simplicity an
d conv
e
n
ien
c
e, the array e
l
ements a
r
e
mostly assu
med
to be identica
l
. The individual eleme
n
ts
may be of
any type like wire dipole
s
or l
oop
s, apertu
res,
etc
[2].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 629
7 –
6312
6298
Duri
ng the
twentieth
ce
n
t
ury the world ha
s be
co
me in
cre
a
si
n
g
ly depe
nde
nt on
electromagnetic syst
em
s.
Satellites orbiting the eart
h
pr
ovide communi
cations links vital to
comm
erce a
nd gove
r
nm
e
n
t. Rada
r
system
s help t
o
navigate
ai
rcraft and
shi
p
s a
s
well a
s
to
control the traffic of these
vehi
cle
s
in ve
ry crowd
ed
skie
s and
har
b
o
rs. In
wartim
e, the effective
coo
r
din
a
tion
and
control o
f
land, se
a, and ai
r force
s
re
qui
re reli
able
comm
u
n
icatio
ns.
Ra
dar
system
s a
r
e
use
d
to locate and tra
c
k e
nemy forces,
guide frie
ndl
y force
s
to their targ
ets, a
n
d
direc
t
s
hell and miss
ile fire [3].
Estimation of paramete
r
s is one of the majo
r appli
c
ations of array signal pro
c
e
ssi
ng
whe
n
si
gnal
s are im
ping
e
d
on the
arra
y. Numbe
r
of
sign
als, ma
g
n
itude
s, frequ
enci
e
s, di
re
ction
of arrival
(DO
A
), dista
n
ces and
spe
e
d
s
of sig
nal
s
are
so
me
co
mm
on p
a
ramete
rs that
are to
be
identified by
the anten
na
array syste
m
. Of all
these
para
m
eters,
the DOA
estimation is v
e
ry
importa
nt an
d attra
c
ts m
o
st attention,
esp
e
ci
ally
in
far-field
sig
n
a
l
appli
c
ation
s
, in whi
c
h
ca
se
the wave fro
n
t of the
sig
nal may
be
treated
pla
n
a
r
, indicating t
hat the
dista
n
ce
is irrelev
ant.
Thus, thi
s
p
a
per p
r
e
s
e
n
ts
the detaile
d i
n
vesti
gatio
n
of DOA e
s
ti
mation an
d a
d
vancement i
n
it
with time in the pas
t three dec
a
des
.
2. DO
A
Estim
a
t
i
on
DOA e
s
timati
on is th
e p
r
o
m
inent figu
re
in the field
of array sig
nal
pro
c
e
ssi
ng a
pplied i
n
rada
rs,
sona
r’s, sei
s
mi
c and co
mmun
i
cation
sy
st
e
m
s. Va
riou
s
types of
info
rmation
can
be
extracted
fro
m
an i
n
comin
g
wave impi
n
ged
on
anten
na a
r
ray which are the
cou
p
led
sig
nals
at
different poi
nts in spa
c
e [4]
.
There a
r
e t
w
o types
of d
a
ta involved, one is th
e tra
i
ning data fro
m
whi
c
h the
ad
aptive wei
ght
s a
r
e
cal
c
ulat
ed an
d the
ot
her i
s
p
r
ima
r
y data from
wh
ich va
riou
s type
of information
can be extra
c
ted like dete
c
tion an
d parameter e
s
tim
a
tion (an
g
le, rang
e, Dop
p
l
e
r
estimation
), inclu
d
ing thei
r directio
n of arrival (DOA) [
5
].
There are ma
ny applicatio
ns wh
ere accurate e
s
timation of a signa
ls dire
ction of
arrival
(DOA) is of particul
a
r inte
rest. Rada
r, sonar,
an
d mo
bile comm
uni
cation a
r
e but
a few examples
of the many possible ap
plications. DOA method
s
ca
n be use
d
to de
sign an
d ada
pt the directiv
ity
of array ante
nna
s; for example,
an ante
nna array ca
n be desi
gne
d to accept si
gnal
s from so
me
spe
c
ific direction, while
rej
e
cting
si
gnal
s fro
m
a
ll oth
e
r
di
re
ction
s
and de
clared
as
interfe
r
en
ce
[6].
The main rea
s
on for
cho
o
sing aspe
cts o
f
DOA es
timation for res
e
arc
h
is
that majority of
system
s
no
wadays solely
rely on
thi
s
u
n
ique
te
chnol
ogy for its
su
ccessful
ope
rations, li
ke
th
e
US Global P
o
sitioni
ng System (GPS),
Russi
an GLONASS etc and Euro
pe, China, Japan and
India are in process of developing
navigation satellite
system
s [7].
3.
Signal Mode
l for Narro
w
b
and An
ten
n
a
Array
In this se
ctio
n we b
r
iefly introdu
ce the
basi
c
sig
nal
model for n
a
rrowban
d a
n
tenna
arrays whi
c
h
will be used throu
gho
ut the pape
r. St
ructure of del
ay propa
gation,
forming sp
atial
covari
an
ce m
a
trix and
its
spe
c
tral
de
co
mpositio
n
a
r
e the m
a
in
contents of thi
s
sign
al mo
d
e
l.
For
simpli
city we will
use the uniform linear arra
y onl
y for discussi
on. Subspaces are formed by
con
s
id
erin
g a
s
soci
ation
s
of eigenvalue
s
and eige
nv
ectors with the
sign
al and no
ise compo
n
e
n
ts
of the signal.
3.1.
Propaga
tion Dela
y
s
in Uniform Linear
Array
s
Con
s
id
er a
system of
M
element
s uni
form line
a
r a
rray, nu
mbe
r
ed 0, 1, …,
M
- 1.
Con
s
id
erin
g
half-a
-
wavele
ngth spa
c
ing
betwe
en th
e adja
c
e
n
t array eleme
n
ts, it ca
n
be
assume
d tha
t
signal
s re
ceived by t
he array
ele
m
ents a
r
e
correlated. Half-a-wavel
en
gth
(
d/
λ
=1/
2
) is
often referre
d
to as the
desi
gn wave
l
ength of the
array
sin
c
e
it chara
c
te
ri
ze
comp
romi
se
betwe
en
a n
a
rro
w
beam
wi
dth an
d g
r
ati
ng lo
be
s. A b
a
se
ban
d
sign
al
s
(
t
) i
s
re
cei
v
ed
by ea
ch a
r
ra
y element
at a different t
i
me in
stant.
The p
h
a
s
e
o
f
the ba
seb
a
nd
signal,
s
(
t
),
received at el
ement 0 is ta
ken a
s
zero and the ph
ase of
s
(
t
)
recei
v
ed at other
element
s will
be
cal
c
ulate
d
wit
h
re
spe
c
t to this. To me
asure the
p
h
a
s
e differen
c
e, i
t
is necessa
ry to measu
r
e
the
differen
c
e in
the time the
sign
al
s
(
t
) a
r
rives at elem
e
n
t 0 and th
e
time it arrive
s at elem
ent
k
.
From Fig
u
re
1 the time delay betwee
n
the 0
th
eleme
n
t and k ele
m
ent usin
g b
a
si
c trigon
om
etry
can b
e
com
p
uted as [6]:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Thre
e De
ca
d
e
s of De
vel
o
p
m
ent in DOA Estim
a
tion Techn
o
log
y
(Z
e
e
sh
an Ahm
ad)
6299
∆
(1)
Whe
r
e
C
i
s
th
e spe
ed of lig
ht.
If w
e
Su
pp
os
e
s
(
t
) to b
e
a
narro
wband
digitall
y modulate
d
sign
al
with
lowp
ass
equivalent
s
l
(
t
), carrier freq
uen
cy
f
c, and
symbol pe
rio
d
T
. It can be written a
s
:
(2)
The sig
nal re
ceived by the
k
th element i
s
given by:
∆
∆
(3)
Figure 1. Unif
orm Lin
ear Array
No
w su
ppo
se that the re
ceived
sign
al
at the
k
th element i
s
d
o
wn
co
nverte
d to the
baseba
nd. In that case, the base
ban
d re
ceived
sign
al is:
∆
∆
(4)
No
w sam
p
le the re
ceived b
a
se
ban
d sig
n
a
l with symbo
l
period T
se
cond
s i.e.,
∆
∆
(5)
In prac
tic
e
,
≫
∆
,
0
,
1
,
2
,
3
,
……..,
1
(6)
So Equation (5) ca
n be rewritten as:
∆
(7)
Whe
r
e
, wh
ere
is the
wavelength
of thepro
pag
atin
g wave. T
he
element
spa
c
ing can
be com
puted
in wavele
ngth
s
as
d
=
D
/
λ
.
Usi
ng the
s
e
Equation, (7
) can b
e
writte
n as:
(8)
To avoid alia
sing in
spa
c
e,
D
≤
λ
/2. Equation (8
) is
simplified to:
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 629
7 –
6312
6300
In dis
c
rete time notation
with time index
n
Equation (9)
can b
e
wri
tten as:
(10)
Let the
n
th sample of the baseb
and si
gnal at the
k
th element be denoted a
s
.
Whe
n
there a
r
e
r
si
gnal
s p
r
esent, the
n
th symbol of the
i
th signal will be denoted
for
i
= 0,
1,
…
,
r
– 1. The ba
seb
and,
sample
d sig
nal at the
k
th element can
be expre
s
sed
as:
∑
(
1
1
)
If the propag
ating sign
al is not digitally
modulat
e
d
an
d is narro
wba
nd, the appro
x
imation sho
w
n
in (8) i
s
still valid.
Equation (11) can be
writte
n in
matrix form as follows:
x
x
.
.
x
.
.
.
.
.
.
..
.
.
.
..
.
.
.
.
.
.
.
.
.
.
(12)
Whe
r
e additi
ve
noise,
, is con
s
id
ered
at each ele
m
ent. Equation
(12
)
can be
written
in
comp
act mat
r
ix notation, as follows:
x
..
.
(13)
Whe
r
e;
=
1
vec
t
or
A =
matrix
=
sign
al vect
ors, an
d
= noi
se v
e
ct
o
r
.
The matrix
A
c
o
mp
os
ed
of c
o
lu
mns
θ
, are called th
e steeri
ng ve
ctors (di
r
e
c
tion
vectors) of the sign
als
.
The set of all possi
ble st
eerin
g vectors is kn
own as the
arra
y
m
anifold
. Th
e array m
a
nifold
can
b
e
compute
d
in two wa
ys that i
s
analytically
and
experim
entall
y
. Mostly fo
r linea
r, pla
nar, or
ci
rcu
l
ar array co
nfiguratio
ns,
it is comp
uted
analytically, while
it ca
n
be comp
ut
ed
expe
ri
me
ntally for m
o
re
com
p
lex
antenn
a a
r
ray
geomet
rie
s
. In the ab
sen
c
e of noi
se, the sig
nal
recei
v
ed by each
element of th
e array can b
e
written a
s
:
x
(14)
From a
bove
equatio
n it is clear th
at
linear
combi
nati
on of the col
u
mns of
A
forms
the
data vector
.
These eleme
n
ts sp
an the
sign
al sub
s
p
a
ce
. In the abse
n
ce of noise, one ca
n
obtain o
b
serv
ations
of several ve
ctors
x
n
and o
n
ce
we e
s
timate
r
li
nearly i
nde
pe
ndent ve
ctors, a
basi
s
for the
sign
al su
bspa
ce can be
cal
c
ulate
d
.
Next we
will
compute the
spatial
covari
ance ma
t
r
ix
of the antenna
array. A
s
sume that
the si
gnal
an
d noi
se
vecto
r
s
are u
n
correlat
ed
and
ze
ro m
ean. Al
so the
noi
se v
e
ctor
is
a vector
of Gaussia
n
, white noi
se
sample
s wi
t
h
ze
ro mea
n
and co
rrelat
ion matrix
σ
.
Let
.
Then we ca
n write the
sp
atial covari
an
ce matrix a
s
:
σ
(15)
Since the ma
trix
can be
unitarily deco
m
posed and
has re
al eige
nvalue
s beca
u
se it
is
Hermitian
(compl
ex conjug
ate
tra
n
sp
ose).
Usi
ng the
data
matrix
X,
we
ca
n find
th
e
eigenve
c
tors
of the autoco
rrel
a
tion matrix by an
alternative metho
d
. The ro
ws
of the matrix
X
are
co
mplex
co
njug
ate transpo
se
of the d
a
ta
ve
ctors obtai
ned
from th
e a
r
ray of
sen
s
o
r
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Thre
e De
ca
d
e
s of De
vel
o
p
m
ent in DOA Estim
a
tion Techn
o
log
y
(Z
e
e
sh
an Ahm
ad)
6301
Suppo
se that
the data ma
trix
X
contain
s
K
sn
ap
shot
s of data o
b
tained from
N
s
e
ns
or
s
in
a
linear a
r
ray. The matrix
X
is
and can b
e
written a
s
:
(
1
6
)
Whe
r
e;
U
is a
matrix
who
s
e
colum
n
s are orth
on
ormal,
D
is a dia
gon
al
matrix, and
V
is an
matri
x
whose col
u
mns a
r
e al
so
orthon
orm
a
l.
This d
e
comp
osition i
s
kn
own a
s
the
sing
ular valu
e decompo
si
tion (SVD).
The SVD of
X
is
related to th
e sp
ectral de
comp
ositio
n (eigen
de
co
m
positio
n) of t
he spatial
co
varian
ce m
a
trix
.
The colu
m
n
s of the mat
r
ix
V
will be e
i
genve
c
tors o
f
and the dia
gonal el
emen
ts of the
matrix
D
will
be squa
re
roots of the eigen value
s
of
. In practic
e
, the
smallest
eigenvalu
e
s
will not be preci
s
ely
σ
; rather, they will all have sm
all values compared to the
sign
al Eigen
values. Thi
s
is be
cau
s
e the matrix
is not kno
w
n pe
rfectly, but must be
estimated fro
m
the data. A common e
s
timator for
t
he sp
atial co
varian
ce mat
r
ix is the sa
mple
spatial
covariance matrix, which is o
b
tained
by a
v
eragin
g
ra
n
k
-o
ne data
matrices of the
form
, i.e.
∑
x
x
(
1
7
)
Whe
r
e
K
is t
he total num
ber of
sna
p
shots of data
available fro
m
the
se
nsors. Although t
h
e
discu
ssi
on so far has fo
cuse
d on the uniform line
a
r
array, the princi
ple
s
of si
gnal and n
o
i
s
e
sub
s
p
a
ces a
l
so
apply to
othe
r a
r
ray geo
metrie
s su
ch
a
s
t
he u
n
iform
plana
r a
n
d
the
semi
sph
e
ri
cal
array
s
.
4.
Classific
a
tio
n
of DO
A
There are m
any ways to
cla
ssify the DOA
estimatio
n
method
s. Here we have
broa
dly
categ
o
ri
zed
Dire
ction of A
rrival (DOA) e
s
timation into
four gro
u
p
s
that are [8]:
a) Conve
n
tional
Tech
niqu
es
b)
Subsp
a
ce Ba
sed Te
ch
niqu
es
c)
Maximum Likelihoo
d Tech
nique
s
d)
Integrated T
e
ch
niqu
es (Combi
ne
Property
Resto
r
al Te
ch
niqu
es a
nd Sub
s
pa
ce
Based T
e
chni
que
s)
A large nu
m
ber of ele
m
ents a
r
e req
u
ired to a
c
h
i
eve high re
solutio
n
in case of
Conve
n
tional Tech
niqu
es since
they
are based
on
cl
a
ssi
cal be
amfo
rming techniq
ues. Sub
s
pa
ce
based metho
d
s a
r
e high
resol
u
tion sub
-
optimal te
ch
nique
s which
exploit the Eigen st
ruct
ure
of
the input dat
a matrix. Maximum likelih
ood techni
q
u
e
s are the o
p
timal techni
que
s whi
c
h
show
tremen
dou
s perfo
rman
ce unde
r
lo
w
S
NR co
ndition
s even
but a
r
e
comp
utationally inten
s
i
v
e.
The inte
grate
d
app
roa
c
h
u
s
e p
r
o
perty restoral
b
a
se
d tech
niqu
es to se
parate
multiple
sign
als
and e
s
timate
their
spatial
sign
ature
s
from wh
ich their di
re
ction
of arrival
(DOA)
ca
n b
e
estimated u
s
i
ng su
bspa
ce
techni
que
s [6
-9].
DOA e
s
timati
on is on
e of t
he mai
n
focu
sing
co
ntent
and a
r
e
a
of rese
arch i
n
a
r
ray sign
al
pro
c
e
ssi
ng,
a
nd exp
a
n
s
ive
l
y applie
d in
the fiel
d
of
radar,
sona
r,
GPS and
wa
s exten
ded
to
comm
uni
cati
on in the la
st decad
e. Th
ere a
r
e
two types of tech
nique
s availa
ble to do DOA
estimation, which a
r
e curre
n
tly attracting
focus
of the rese
arche
r
s towards thi
s
technolo
g
y.
4.1. Non
-
Subsp
a
ce
Tech
niqu
es
These metho
d
s de
pen
d on spatial
spe
c
trum, an
d lo
cation
s of pe
aks in the sp
ectru
m
determi
ne th
e DOA
s
of
si
gnal
s. The
s
e
method
s a
r
e co
nceptuall
y
simple
but
offer mod
e
st
or
poor p
e
rfo
r
m
ance in terms of resolutio
n
. One of
the main advantag
es of these te
chni
que
s is th
at
can b
e
used i
n
situation
s
where
we la
ck
of
information
about pro
p
e
r
ties of sign
al [10].
4.2.
Subspace T
echnique
s
There a
r
e
ce
rtain limitation
s
in
re
sol
u
tio
n
wh
ich i
s
hin
derin
g the
g
r
owth
of no
n-subspa
ce
or cl
assi
cal
method
s of DOA estimatio
n
. They
do n
o
t exploit the stru
ct
ure of
narro
wba
nd i
nput
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 629
7 –
6312
6302
data mo
del
of the m
e
a
s
urem
ents wh
ich
give ri
se
to certai
n l
i
mitations. S
ubspa
ce-ba
s
ed
method
s dep
end on ob
se
rvations con
c
ernin
g
the Ei
gen de
com
p
o
s
ition of the covarian
ce m
a
trix
into a
s
i
gnal
s
u
bs
pace and a noise
s
u
bs
pac
e
. Two of thes
e
methods
MUSIC
and ESPRIT were
applie
d here to determi
ne
DOA [10-11].
5.
DO
A Estimation Base
d on Classical
Metho
d
Cla
ssi
cal
dire
ction
of arriva
l (DOA) m
e
th
ods
are e
s
se
ntially based
on be
amformi
ng. The
two
cla
ssi
cal
tech
niqu
es for
DOA
are t
he d
e
lay-a
n
d
-
su
m m
e
thod
and
the
min
i
mum va
rian
ce
distortio
n
le
ss
respon
se
(M
VDR) metho
d
.
The basi
c
id
ea behi
nd the
classi
cal met
hod
s is to sca
n
a beam th
ro
ugh spa
c
e a
nd mea
s
u
r
e t
he po
we
r re
ceived from e
a
ch di
re
ction.
Dire
ction
s
from
whi
c
h the largest amo
unt of powe
r
is re
ceived a
r
e ta
ken to be the
DOA
s
[9-12].
5.1.
Dela
y
and Sum Method
Delay-and
-Su
m
metho
d
i
s
the
simple
st
cla
s
sical me
thod b
a
sed o
n
be
am fo
rm
ing for
estimation of
DOA. Figure 2 sho
w
s classica
l narrowb
and be
a
m
forme
r
stru
cture
whe
r
e
the
output
sign
al
y(k) is given
by a lin
early
weig
ht
ed
su
m of the
sen
s
or ele
m
ent
s output [1
3].
That
is:
(18)
Figure 2. Del
a
y-and
-Sum
Method
The total outp
u
t powe
r
of the above conv
entional b
e
a
m
forme
r
ca
n be expre
s
sed
as:
|
|
|
|
(19)
Whe
r
e
is the auto co
rrelation matrix
of the array inpu
t data and co
ntains u
s
eful
informatio
n
about both th
e array re
spo
n
se ve
ctors a
nd the si
gnal
themselve
s
,
and by caref
u
l interp
retati
on
of
we can estimate
sign
al para
m
eters. This
e
quat
io
n
plays a
n
imp
o
rtant role in
the in all the
conve
n
tional
DOA e
s
timation algo
rithm
s
.
Con
s
id
er a
si
gnal
impingi
ng on
the a
r
ray at an a
n
g
l
e
θ
0
. Us
ing
th
e
n
a
r
r
o
w
b
and
input data mo
del, the powe
r
at the beam
former
output
can be exp
r
e
s
sed a
s
:
|
|
|
|
|
|
(20)
Whe
r
e:
= Steering vector a
s
so
cia
t
ed with the DOA angle
θ
0
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Thre
e De
ca
d
e
s of De
vel
o
p
m
ent in DOA Estim
a
tion Techn
o
log
y
(Z
e
e
sh
an Ahm
ad)
6303
= Noi
s
e vect
or at the arra
y input and
and
are the
sign
al po
wer
and noi
se p
o
w
er
re
spe
c
tively.
The above e
q
uation cl
early
demon
strate
s
that output power is m
a
ximized
whe
n
.
Therefore
all
of the p
o
ssibl
e
weight ve
ctors,
th
e
re
cei
v
er ante
nna
has the hi
gh
est g
a
in
in the directio
n
θ
0
, when
. This i
s
be
ca
use
align
s
the
pha
se
s of the
sign
al
comp
one
nts arrivin
g
from
θ
0
at the sen
s
ors,
cau
s
ing t
hem to add constructively.
In classi
cal b
eamformi
ng
approa
ch to
DOA e
s
timation, the beam
is scan
ned
over the
angul
ar
regio
n
of intere
st i
n
discrete ste
p
s by fo
rming
weig
hts
for different
θ
, an
d th
e
output po
we
r is mea
s
u
r
e
d
.
Usin
g Equa
tion (19
)
,
the
output po
we
r at the cla
s
si
cal b
eamfo
rm
er
as a fun
c
tion
of the angle o
f
arrival is giv
en by:
(21)
Therefore if we have an e
s
timate of autoco
r
relation
matrix and kn
ow the stee
ri
ng vector
for all
θ
’s of
intere
st (eithe
r throug
h cali
bration o
r
an
alytical com
p
utation), it is possibl
e to
estimate the
output po
we
r
as a fu
nctio
n
of the angl
e o
f
arrival
θ
. T
h
e output p
o
wer a
s
a fu
ncti
on
of the angle o
f
arrival is oft
en terme
d
a
s
the spat
ial
spectrum. Cle
a
rly the direct
ion of arrival can
be estimate
d by locating p
eaks in the s
patial sp
ectru
m
defined in
Equation (21).
The delay an
d sum metho
d
has many
disa
dvantag
e
s
. The width
of the beam and the
height of the sidelob
es li
mit t
he effectiveness whe
n
signal a
rri
ving from multiple dire
cti
o
n
s
and/or sources a
r
e
pre
s
e
n
t becau
se t
he si
gnal
ov
er a
wid
e
an
gular re
gion
contri
bute to
th
e
measured av
erag
e po
we
r at each lo
ok
dire
ction.
He
nce thi
s
tech
nique h
a
s p
o
o
r re
sol
u
tion [14].
Although it is possibl
e to i
n
crea
se the
resol
u
ti
on by
addin
g
mo
re
sen
s
o
r
ele
m
e
n
ts, increa
sin
g
the numb
e
r o
f
sen
s
ors in
crease the num
ber of re
ceive
r
s a
nd the a
m
ount of sto
r
age requi
red
for
the calib
ratio
n
data i.e.
.
5.2.
Capo
n’s Minimum Variance Method
This metho
d
has a
simil
a
ri
ty with the
p
r
ev
iously
de
scribed
d
e
lay-a
nd-sum
techn
i
que i
n
whi
c
h th
e
po
wer of th
e
re
ceived
sign
al i
s
m
e
a
s
ured
i
n
all
po
ssible
dire
ction
s
. In
simple
words in
forming the
beam in the
desi
r
ed lo
ok
dire
ction,
all the degree
s of freedom a
c
cessibl
e
to the
array were utilized. Thi
s
wo
rk go
es very
well wh
en the
single si
gnal
is available el
se co
ntrib
u
tio
n
from b
o
th de
sire
d a
nd u
n
desi
r
ed
si
gn
als i
s
contai
ned by th
e
array output
power.
Cap
o
n’s
Method contributes in solving the poor
resolution
p
r
oblem by usi
ng the idea to utilize som
e
of
the deg
ree
s
of freedo
m to
form a b
eam
in the de
sire
d loo
k
directi
on an
d at the
same tim
e
u
s
ing
the remai
n
ing
degre
e
s of freedom to form nulls in the
directio
n of interferi
ng si
g
nal [15-1
6
].
To mea
s
u
r
e t
he po
we
r fro
m
DOA,
θ
, th
e gain
of bea
mforme
r is
constraine
d to
be 1 i
n
that dire
ction
and
contrib
u
t
ion to the
ou
tput po
wer from the
sig
nal
s a
pproa
chin
g from
all
oth
e
r
dire
ction
s
i
s
minimized by
usin
g the
re
maining
deg
rees
of free
do
m. Mathemati
c
ally this pro
b
lem
is
kno
w
n
a
s
a con
s
train
e
d
minimi
zation
pro
c
e
s
s [1
5]. For eve
r
y probabl
e a
ngle,
θ
, the
po
we
r of
the signal is
minimized pe
rtaining to
subje
c
t to the
const
r
aint that
1
, this
is
the
basi
c
ide
a
be
hind the con
s
trained mi
nim
i
zation p
r
o
c
e
ss.
|
|
(22
)
After solving
this eq
uation
the weig
ht vect
or
whi
c
h i
s
obtai
ned, is termed
as M
i
nimu
m
Varian
ce
Di
stortionle
s
s
Resp
on
se
(M
VDR) b
eamf
o
rme
r
wei
g
h
t, since fo
r
a spe
c
ific lo
ok
dire
ction, it minimize th
e varian
ce
(ave
rage po
we
r)
o
f
the output si
gnal while p
a
ssi
ng the
sig
nal
comin
g
in the look di
re
ction witho
u
t distortion (u
nity gain an
d ze
ro
phase shift).
The ab
ove e
quation
(22
)
has
pea
ks fo
r the
certai
n
angle
s
, re
pre
s
ent
s the e
s
ti
mates of
the angle
s
of arrival of the
sign
als.
Usi
ng La
gra
n
ge multiplie
r, its
weig
hts are given by [17]:
(23)
No
w u
s
in
g th
e Capo
n’s be
am forming
method
as th
e outp
u
t po
wer of
the
arra
y as
an
angle of arriv
a
l’s fun
c
tion, given by the
Cap
on’
s sp
atial spe
c
tru
m
is as follo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 629
7 –
6312
6304
(24)
The DOA’
s can be asse
ssed by comput
ing and
plotti
ng the Cap
o
n
’
s spe
c
trum o
v
er the
whol
e ran
ge
of
θ
and dete
c
ting the pe
a
ks in the
spe
c
trum.
The dra
w
b
a
ck of this method inclu
d
e
s
the re
q
u
ire
m
e
n
t of an inverse matrix co
mputatio
n
whi
c
h may b
e
com
e
ill-con
d
itioned if highly correlat
ed sig
nals a
r
e pre
s
ent an
d expen
sive for
large a
r
rays. As com
pared
to the delay-and-su
m bea
m former thi
s
method provides hi
ghe
r a
n
d
better res
o
lution.
Suppo
se if o
t
her si
gnal
s
that are correlated
with t
he sig
nal of intere
st are
pre
s
ent
becau
se it i
n
advertently u
s
e
s
that
co
rrelation to
re
duce the
pro
c
e
s
sor outpu
t
power with
out
spatially n
u
lli
ng it. In othe
r word
s,
we
can
sa
y th
at the co
rrelate
d
co
mpo
nent
s may b
e
u
n
i
t
ed
detrime
ntally in the pro
c
e
s
s of
minimizi
n
g
the output p
o
we
r.
6.
Subspac
e
M
e
thod
s For
DO
A Estimation
Low
re
solutio
n
is the maj
o
r limiting fact
or, in spite o
f
broad
er u
s
e of cla
ssi
cal
beam-
forming
ba
se
d method
s
d
ue to the le
ss com
putat
ion
a
l com
p
lexity, affecting the
developm
ent
of
non-su
bspa
ce ba
sed
tech
nique
s fo
r the
DOA
estimat
i
on. The
efforts of the
re
se
arche
r
s
be
co
me
more o
n
the
sub
s
pa
ce b
a
se
d DOA e
s
timators to
achi
eve and
attain the hig
h
re
solution,
by
makin
g
use of the signal subspa
ce. The
s
e metho
d
s t
e
rme
d
as the
sign
al sub
s
p
a
c
e metho
d
s a
r
e
origin
ated du
ring the
re
se
arch on
spe
c
t
r
al e
s
timati
on
whe
r
e the e
s
timation of a
u
toco
rrelation
of
a sign
al and t
he noi
se mod
e
l is made a
n
d
then us
ed to achi
eve a matrix who
s
e
Eigen stru
ct
ure
prod
uces the
sig
nal
and
th
e noi
se
sub
s
paces. By
fu
nctioni
ng th
e
spatial
covari
ance m
a
trix, this
simila
r tech
ni
que can al
so
be used in array antenna
DOA estimatio
n
[18].
6.1. Music
Algori
t
hm
In lots of
DO
A estimatio
n
algo
rithm
s
with ex
celle
n
t
perfo
rma
n
ce, one
of th
e
ea
rliest
prop
osed al
gorithm i
s
t
he multiple
sign
al
cl
assi
fication
(MUSIC) ba
se
d
on eig
env
alue
decompo
sitio
n
of the si
gnal covaria
n
ce m
a
trix [19]. MUSIC stand
s for Multiple Signal
Clas
s
i
fic
a
tion. MUSIC giv
e
s
the es
timation of nu
m
ber
of so
urce
s an
d he
nce
their di
re
ction
of
arrival. MUSI
C is a techni
q
ue based on
exploiting
the
Eigen stru
ctu
r
e of input co
varian
ce matrix.
Usi
ng Sin
gul
ar Val
ue
De
comp
ositio
n
(SVD) of the
data
matrix
or Eigen
d
e
com
p
o
s
ition
of
sampl
e
cova
riance matrix, we can obtai
n Eigen vecto
r
s ea
sily.
Due to
orth
o
gonality bet
ween the
sig
n
a
l su
bspa
ce
and n
o
ise
sub
s
p
a
ce a
s
sho
w
n
previou
s
ly, the MUSIC try
to find all the
possibl
e
ste
e
ring ve
cto
r
s of the inco
m
i
ng si
gnal
s li
e in
the signal
su
bsp
a
ce that are orth
ogo
n
a
l to the noise sub
s
pa
ce [
19-2
0
]. If
is
the steeri
ng
vector
co
rrespondi
ng to o
n
e
of
the in
co
ming
signal
s,
then
0
. The function fo
r M
U
SIC
s
p
ec
tr
um c
an b
e
w
r
itte
n
as:
(25)
The
above
fu
nction
will
a
s
sume
a
very l
a
rge
value
when
θ
i
s
equa
l to the
DOA
of one
of
the sig
nals. T
he MUSI
C al
gorithm
wa
s
prop
osed
in
1
979 by Schmi
d
t [21]. In first phase MUSI
C
estimate
s a
basi
s
for the
noise su
bspa
ce,
, afterwards, dete
r
min
e
s the
r pea
ks in (2
5); the
asso
ciated a
ngle
s
provid
e
the DOA esti
mates.
The MUSI
C a
l
gorithm may
be sum
m
ari
z
ed as [21]:
Step 1:
Estimate the input covari
a
n
ce matrix
in accor
d
a
n
ce t
o
{
,
1
,
2
,
3……
}.
∑
Step 2: Perform eige
n de
compo
s
ition o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Thre
e De
ca
d
e
s of De
vel
o
p
m
ent in DOA Estim
a
tion Techn
o
log
y
(Z
e
e
sh
an Ahm
ad)
6305
Whe
r
e
,
,
…
…
,
,
,
……,
are the eige
n values a
nd
,
,
……..,
are the corre
s
po
ndin
g
eig
en vectors of
.
Step 3: E
s
timate the
num
b
e
r
of si
gnal
s
D,
from the
multiplicity
K,
o
f
th
e sma
lles
t
e
i
ge
n
value
as:
Step 4: The MUSIC spe
c
trum can be o
b
tained a
s
fol
l
ow:
Whe
r
e
,
……,
Step 5:
D
is
the largest pea
k of
which co
rrespo
nd to estimates of the
Dire
ction
-
Of-
A
rriv
a
l.
6.1.1.
Disadv
a
ntag
es of M
U
SIC Algorithm
The MUSI
C algorith
m
ha
s good pe
rformance and
i
s
widely used
till now be
ca
use of its
sup
e
r re
sol
u
tion cap
ability. Although there are ma
n
y
positives in
MUSIC algo
rithm, there
are
nume
r
ou
s ba
rrie
r
s a
nd co
mpeting exist
i
ng sol
u
tion
s that are hin
derin
g the growth of MUS
I
C
algorith
m
. Inefficienci
e
s fro
m
which MUSI
C algorithm
is sufferi
ng a
r
e given bel
o
w
:
Its performan
ce de
graded
whe
n
the sig
nals a
r
e
correlated an
d so
is not able to
identify
DOA
s
of correlated si
gnal
s.
MUSIC al
gori
t
hm is al
so
computation
a
ll
y co
mplex a
n
d expen
sive
becau
se it in
volves a
sea
r
ch over t
he functio
n
for the pea
ks.
If the numbers of sou
r
ces
are ove
r
e
s
timated,
it is possible that MUSIC algorithm
gives
spu
r
iou
s
pea
ks an
d thi
s
happ
ened
u
s
ually when t
he
steeri
ng
vector i
s
not
in the
si
gn
al
sub
s
p
a
ce an
d is perpen
di
cula
r to som
e
of the noise
eigenve
c
tors
[20].
6.1.2. Proposed
S
o
lutions
The
above
m
entione
d in
efficien
cie
s
po
se challe
nge
for
su
stainin
g
the g
r
o
w
th of
MUSIC
algorith
m
. Ma
ny innovative techni
que
s
were p
r
opo
se
d in the pa
st to make the M
U
SIC alg
o
rith
m
more robu
st and efficie
n
t. These innova
t
ions
in
clud
e Prime MUSIC, Root MUSIC etc
.
[22].
There are
nume
r
ou
s te
chni
que
s av
ailable
in th
e literatu
r
e
to overcome
these
deficie
nci
e
s. One su
ch
te
chni
que
s
i
s
kno
w
n
as
S
patial smoot
hing, which is
a
n
e
s
sent
ial
techni
que i
n
multipath p
r
o
pagatio
n envi
r
onm
ent an
d
can
be
appli
e
d to ove
r
com
e
this
proble
m
.
To
p
e
rfo
r
m spatial smooth
i
ng,
the array
mu
st
be
divi
ded up
i
n
to smaller, po
ssi
bly
overla
ppi
ng
sub
a
rrays an
d the
sp
atial
cova
rian
ce
matrix of
each suba
rray is averaged
to
form
a
sin
g
l
e
,
spatially sm
o
o
thed covari
ance matrix. The MUSI
C
algorith
m
is then appli
ed
on the spatia
lly
smooth
ed ma
trix [23].
6.2.
The Minimum Norm Method
Kumare
sa
n a
nd
T
u
fts,
p
r
o
posed a
m
e
thod call
ed
th
e
Minim
u
m Norm Metho
d
, whi
c
h
i
s
applie
d in a
manne
r sim
ilar to MUSI
C algo
rithm
over the DO
A estimation
probl
em an
d is
defined a
s
“t
he vector lyin
g in the noise sub
s
pa
ce
who
s
e first element is on
e
having minimum
norm
”
[24-25]
.
The vecto
r
is
given as:
(26)
As soo
n
as t
he minimum
norm ve
ctor i
s
kn
ow
n, the DOA
s
are
sp
ecified by the
large
s
t
pea
ks of the functio
n
as foll
ows [25]:
(27)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 629
7 –
6312
6306
No
w
the obje
c
tive
is
to de
termine and establi
s
h
th
e minimum norm
vector
“
g
”.
So for
that purp
o
se let
be the mat
r
ix who
s
e
col
u
mns
develo
p
the ba
sis fo
r the sig
nal subspa
ce.
,
can b
e
divide
d as [25]:
∗
(28)
In the meantime, the vector
g
lies in the
sub
s
p
a
ce of
noise and will
be orthogon
al to th
e
(sign
a
l su
bspace), so
we
come u
p
with
the equation
as follo
ws [25
]
:
1
0
(29)
The sy
stem o
f
equation
s
a
bove will b
e
unde
r
dete
r
mi
ned; therefore we
are
goin
g
to use
the minimum
Frob
eniu
s
no
rm sol
u
tion e
n
lighten
ed be
low:
(30)
F
r
o
m
Eq
ua
tion
(
2
9)
, w
e
can
w
r
ite
as
:
∗
(31)
From thi
s
equ
ation, we can
have:
∗
1
‖
‖
⁄
(32)
By using th
e
above Equ
a
tion (32),
we
can elimin
ate the calculation
of the inverse matrix
in
Equ
a
tion (30). No
w we
can
comp
ute
g
ba
se
d o
n
ly
on the
signal
sub
s
p
a
ce o
r
t
hono
rmal
ba
sis,
given belo
w
:
1
‖
‖
⁄
(33)
As soon
as
g
has
be
en co
mputed,
the evaluation of
Min-Norm fun
c
tion give
n a
bove i
s
done
an
d the
angl
es of a
r
rival are
al
so
spe
c
ified
by t
he
r
p
e
a
ks.
T
h
is te
ch
niqu
e
call
ed th
e Mi
n-
Norm te
chniq
ue is commo
nly refle
c
ted
as
a hig
h
-re
s
olution m
e
tho
d
althou
gh it i
s
infe
rior to b
o
th
MUSIC and ESPRIT algorit
hms
.
6.3.
ESPRIT Algorithm
A novel a
nd
a vital app
roa
c
h fo
r the
sig
nal pa
ram
e
te
r e
s
timation
p
r
oble
m
was p
r
opo
se
d
and then termed as “ESP
RIT”. ESPRI
T
is ali
k
e th
e MUSIC algorithm that works by exploit
i
ng
the unde
rlying data mo
del, then generates e
s
t
i
ma
tes that are effective
,
effective a
n
d
asymptoticall
y
unbiased. In addition to this,
it has nu
mero
us a
d
va
ntage
s over
MUSIC.
Roy and Kail
ath proposed this method for
the DOA estimation call
ed the ESPRIT which
stand
s for “E
stimation of Signal Parame
ter via Rotational Invaria
n
c
e Te
chni
que
” [26].
It is obse
r
ve
d that in term
s of array im
perfe
ct
ion
s
th
is algo
rithm i
n
more vigo
rous a
n
d
robu
st a
s
co
mpared
with
the MUSI
C
algorith
m
.
Other th
an that
its sto
r
ag
e
con
s
trai
nts
a
nd
comp
utation
compl
e
xity are lesse
r
than
MUSIC. Thi
s
is be
cau
s
e th
is algo
rithm d
oes n
o
t take i
n
extensive
se
arch th
rou
g
h
out all th
e p
r
obabl
e
steer
i
ng ve
ctors. Nonethel
ess
it investigate
s
t
h
e
rotational
inv
a
rian
ce
prop
erty gen
erate
d
by the
t
w
o
sub
-
a
rray
s
in
the
signal
su
bsp
a
ce, re
sul
t
ed
from the origi
nal array with a translation inva
riance st
ructure. ESPRIT
does not
need the exact
kno
w
le
dge of
the array manifold vecto
r
s unli
k
e the
MUSIC alg
o
ri
thm so the a
rray adj
ustm
ent
requi
rem
ents are n
o
t strict, so
with
the tw
o su
b array’s
co
rre
sp
ondi
ng
element
s, it is
decompo
se
d into equal si
zed two su
b-a
r
rays, expatri
a
t
e from each
other by a sta
t
ic transl
a
tion
al
distan
ce [26
-
27].
The TLS ESPRIT algorithm is
s
u
m up below [26-27]:
Step 1: Obtain an estimate
of
of
from
m
easurem
ent.
Step 2: Perform Eigen de
compo
s
ition o
n
, i.e.
Evaluation Warning : The document was created with Spire.PDF for Python.