TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 11, Novembe
r
2014, pp. 77
8
5
~ 779
7
DOI: 10.115
9
1
/telkomni
ka.
v
12i11.64
71
7785
Re
cei
v
ed
Jul
y
11, 201
4; Revi
sed Aug
u
st
23, 2014; Accepted Sept
em
ber 6, 201
4
Time Varying Autoregressive Model Parameters
Estimation using Discrete Energy Separation Algorithm
G.Rav
i
Shan
kar Re
dd
y
*
1
, Ram
esh
w
a
r
Rao
2
1
Dept.,of ECE, CVR College of
Engineer
ing,
Hy
derabad, India
2
JNT
Universit
y
, H
y
d
e
rab
ad, Indi
a
Corresp
on
din
g
author, emai
l:
ravigosula_ece39@y
a
hoo.co.in
1
, ramesh
w
a
r
_
rao@hotmail.
com
2
A
b
st
r
a
ct
T
i
me V
a
ryin
g Autoregr
essive
(T
VAR) mod
e
l
for the Ampli
t
ude a
nd F
r
eq
uency
mo
du
lat
ed (AM-
F
M) signa
l is
prese
n
ted
in
this p
a
p
e
r. T
VAR p
a
ra
me
ter
s
of AM-F
M si
gna
l ar
e esti
mated
usin
g D
i
s
c
ret
e
Energy S
e
paration (DESA) A
l
gorithm
. The
performanc
e
of
DESA method is shown to be com
p
arable
to
the existi
ng
ba
sis functio
n
me
thod for AM, F
M
, AM-F
M signal
mode
ls. T
he pro
pos
ed
me
thod is s
i
mpl
e
r
t
o
execute
in h
a
r
d
w
a
re an
d co
n
s
umes co
nsid
e
r
ably l
e
ss
co
mputatio
na
l reso
urces co
mpare
d
to the
meth
o
d
using Adaptive and the B
a
sis
function
m
e
t
h
ods. .It is
dem
onstrated that the prop
osed
technique based on
DESA has cert
ain distinct advantages
over
the conven
tional
m
e
t
hod employing basis f
unctions. Another
adva
n
tag
e
is that the prese
n
t meth
od w
o
rks w
e
ll w
i
th quickl
y varying si
gna
ls
Ke
y
w
ords
:
amplitu
de
an
d fre
que
ncy
mo
dul
a
t
ed si
gna
l, ti
me
varyin
g a
u
tore
gressiv
e
mo
del
, basis fu
nctio
n
,
discrete e
nerg
y
separati
on a
l
gorith
m
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
Most tem
poral sig
nal
s e
n
c
ou
ntere
d
in
real
appli
c
a
t
ions h
a
ve ti
me-varyin
g
statistics,
whi
c
h m
a
ke
them no
nstat
i
onary [1,
2]. The
probl
e
m
of time
de
pend
en
cy in
non
stationa
ry is
bypasse
d by assumin
g
local stationa
ry over a rel
a
tively short time
interval, in which
stationa
ry
system i
denti
f
ication a
nd
analysi
s
te
ch
nique
s a
r
e a
pplied. Ho
we
ver,
this assumption
i
s
n
o
t
alway
s
valid for real
life signals like
spee
ch, Electro
c
a
r
dio
g
ra
m (ECG
) and
Electro
e
n
c
ep
halog
ram
(EEG), be
cau
s
e su
ch
sign
a
l
s may have
time varyin
g amplitud
e
and
freque
ncy
[3,
4].
No
nstatio
nary sign
als whi
c
h are
a compou
nd of con
s
tituent
s with
time
-varying
amplitude
an
d fre
quen
cie
s
can
be
mo
deled
by a
m
plitude
and
freque
ncy m
o
dulated
(AM
-
FM)
s
i
gnals
[5].
The be
st fre
quen
cy re
sol
u
tion for stati
onar
y si
gnal
s is obtained
by using pa
rametri
c
method
s. Th
e sig
nal i
s
fitted in to a
n
Autoreg
r
e
ssiv
e
(AR) o
r
a
moving ave
r
age
(MA) o
r
an
Autoreg
r
e
ssiv
e
Moving
Averag
e
(ARM
A) mo
del. It i
s
sho
w
n
that
the p
a
ramet
r
ic metho
d
yi
elds
very high
fre
quen
cy resol
u
tion in th
e
spe
c
tral
e
s
timation fo
r e
v
en very
sm
all length
of
the
stationa
ry si
g
nal [6]. Spect
r
al a
nalysi
s
o
f
nonsta
tio
n
a
r
y signal
s, wit
h
high f
r
eq
ue
ncy-re
solutio
n
is
obtaine
d by using the time
varyi
ng auto
r
egre
s
sive (T
VAR) process.
In the modeling of nonstati
onary si
gnal
s by a TVAR p
r
ocess, two
method
s may be used
for estimatin
g
the TVAR para
m
eters: the ada
ptiv
e algorith
m
ap
proa
ch a
nd t
he ba
sis fun
c
tion
approa
ch. Ad
aptive Algorit
hms, such as the l
east me
an sq
ua
re (L
MS) and the
recursive lea
s
t
squ
a
re
(RLS), use a dyn
a
m
ic mo
del fo
r ada
pting th
e TVAR pa
ra
meters and
are
cap
able
of
tracking time
-varying frequ
ency, provid
e
d
that t
he va
riation is sl
o
w
. These me
thods work well
with slo
w
ly varying si
gnal
s but fail to track rapi
d va
riation [5]. If
the coeffici
en
ts cha
nge fa
st
enou
gh; co
m
pare
d
to the
algorith
m
’s
Converg
e
n
c
e ti
me, the adap
tive algorithm
will not be a
b
le
track the time
varying para
m
eters.
The
ba
sis fu
nction
meth
o
d
, in
whi
c
h
the time
-vari
a
nt pa
ramete
rs a
r
e
expa
nd
ed a
s
a
summ
ation of
the weig
hte
d
time-fun
ctio
ns, are ca
pa
ble of trackin
g
both the fa
st (o
r) the
sl
ow
time-varying f
r
equ
en
cie
s
. In the ba
si
s fu
nction
ex
pan
sion, two issu
es n
eed to
be
re
solved. Fi
rst,
a gene
ral
cla
ss
of basi
s
fu
nction
s i
s
ch
o
s
en
and the
n
, the signifi
ca
nt basi
s
fun
c
tions
need to
be
sele
cted.
Sev
e
ral cla
s
se
s of
functio
n
s have
b
een
propo
sed
in
clu
d
ing
polynom
ial, wavel
e
t a
nd
prolate
sphe
roidal fu
nctio
n
s
[15,
16].
Howeve
r, no
uniform
rule exists to
indi
cate which
cla
ss
sho
u
ld b
e
ad
opted. Mo
reo
v
er, the app
roach of c
hoo
sing th
e si
gni
ficant ba
si
s functio
n
s i
s
b
a
se
d
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
85 – 779
7
7786
on trial and e
rro
r [16]. The
sele
ction of the ex
pan
sio
n
dimensi
on is questio
nabl
e
since there is
no fun
dame
n
t
al theo
rem
o
n
ho
w to
cho
o
se
them.
It i
s
id
eally exp
e
c
ted th
at whe
n
the
expan
si
on
dimen
s
ion i
s
infinite, the result of the
freque
nci
e
s
estimation f
r
om any ba
si
s fun
c
tion is the
same,
which will
exactly
equal to
the true frequency. But this
i
s
im
pract
i
cal,
since t
he
comp
utation
may requi
re i
n
finite memory, and infinite computatio
n
a
l time con
s
u
m
ption.
Modelin
g by a TVAR pro
c
ess is a ge
ne
ral
app
ro
ach,
and all other
non
stationa
ry models
(AM, FM, AM-FM and mo
re) ca
n be sh
own to be sp
ecial cases o
f
the general
approa
ch [5]. A
real
AM-FM
signal
ca
n b
e
modele
d
u
s
in
g a
2-ord
e
r T
VAR p
r
o
c
e
s
s, and
a
sig
nal
co
mpo
s
e
d
of
p
real
co
mpon
ents
will
req
u
ire
a 2
p
-
order
TVAR p
r
oce
s
s [14]. I
n
the m
odeli
ng by a
TV
AR
pro
c
e
ss,
the
estimation
of
the TVA
R
p
a
ram
e
ters
re
quire
the i
n
ve
rsio
n of
an
covarian
ce
ma
trix
of size [2p (q
+1) x 2
p
(q
+1)], whe
r
e q
is the
requi
re
d numb
e
r of
basi
s
fun
c
tio
n
s to represe
n
t
each TVAR p
a
ram
e
ter.
In this pa
per
we h
a
ve est
a
blish
ed the
relation bet
we
en the p
a
ra
m
e
ters
of the
AM-FM
sign
al and th
e para
m
eters of the TVAR pro
c
e
ss.
We
have used th
e Discrete En
ergy Sepa
rati
o
n
Algorithm (DESA-1) to est
i
mate
the TVAR coeffici
ents and the m
odulating signals of the T
VAR
pro
c
e
ss [14].
The estimati
on tech
nique
presented
h
e
re is
con
c
e
p
tually simpl
e
r and e
a
si
e
r
to
impleme
n
t than the metho
d
based on b
a
si
s functio
n
s.
The p
ape
r i
s
orga
nized
as
follows: In se
ction 2, th
e T
VAR re
pre
s
e
n
tation of the
AM-FM
signal is
presented. In section 3, the comple
te estim
a
tion procedure based
DESA is provi
ded.
The
revie
w
of estim
a
tion
usi
n
g
ba
sis functio
n
s i
s
presented
i
n
Se
ction
4. In
se
ction
5
experim
ental
pro
c
e
dure i
s
discu
s
sed.
The ex
p
e
rim
ental results
for the AM,
FM, and A
M
-FM
signals with the
DESA ba
sed
techni
que and basis
fu
nction technique are pr
esented in section6.
Finally, in Section 7, co
ncl
u
sio
n
is provided.
2. TVAR Rep
r
esen
ta
tion of AM
-FM Signal
The AM-FM
signal is given
by:
cos
(1)
Whe
r
e
and
are the
pha
se and a
m
plitude of the
si
gnal respe
c
tively. The AM-FM
sign
al is give
n by a 2-orde
r TVAR proce
ss [14].
1
2
(2)
Whe
r
e
is the
predictio
n error. The sequ
ence
is assu
med to be of
zero mean a
nd varian
ce
. Let
be the time instant
at which the tran
sient
re
sp
onse has be
come insig
n
ificant. For
≫
(
→0
), the role
of
in
is in
sig
n
ificant
comp
ared to th
e role of
1
and
2
.
Then, the Eq
uation (2
) can
be written a
s
,
1
2
(3)
Or,
1
cos
1
2
cos
2
(4)
As
s
u
ming d
1
,
and d
1
1
2
,
We obtai
n:
sin
1
(
5
)
Then,
can be
written a
s
:
cos
1
cos d
s
in
1
sin d
(6)
From Equ
a
tio
n
(5) a
nd (6
), we w
r
ite:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
Varying
Autoregressi
ve
Model Parameters
Es
timation us
ing
…
(G.Ravi S
han
kar
Red
d
y
)
7787
cos
1
cos
2
(7)
Comp
ari
ng Equation (7)
wi
th (4), we g
e
t:
(
8
)
(
9
)
Once
are esti
mated, the si
gnal
can be reco
nstructe
d by Equation (3).
3. TVAR Par
a
meter Es
timation Usin
g Discre
t
e E
n
erg
y
Separation Algori
t
hm
For both cont
inuou
s
a
nd d
i
screte
tim
e
signal
s, Kaiser ha
s defin
ed
a nonli
nea
r e
nergy
tracking o
p
e
r
ator
[7].For the discrete tim
e
ca
se, the e
nergy op
erator for
is defin
ed as,
≜
1
1
(10)
For the si
gnal
,
cos
,
(11)
We have:
(12)
And,
|
|
(13)
Whe
r
e,
.
(14)
Whe
n
one
of the variable
s
(or)
is co
nstant, we can get the ot
her vari
able
with a
scaling of
.
So, the energy operato
r
can e
s
timate
the modulat
ing sig
nal, or more
pre
c
isely its
scaled
versio
n, wh
en
either AM
(o
r)FM
is
pre
s
e
n
t [7
]. When
both
AM an
d FM
are
pre
s
ent sim
u
ltaneou
sly, three alg
o
rith
ms are de
scribed in [7]
to estimate
and
sep
a
rately. T
he be
st amo
ng the thre
e
algorithm
s
according to
performan
ce
is the discrete
energy separation algorithm1(DESA-1).
T
he DESA-1 i
s
defined as f
o
llows:
S (n) =
(n)
−
(n
−
1).
(15)
≃c
o
s
1
(16)
|
|
≃
(
1
7
)
Thus,
can b
e
estimated usin
g the Equation (1
6) and (17), and the TVAR
coeffici
ents a
r
e estim
a
ted
with the follo
wing e
quatio
n.
(
1
8
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
85 – 779
7
7788
(
1
9
)
The
coefficient
s
th
us
estimate
d
exhibit
ri
pple
s
a
nd
th
erefo
r
e
req
u
ire
sm
oothing
u
s
ing
a
filter [8]. Rip
p
les
ca
n be
redu
ce
d by
us
in
g the
smoothing. S
m
oothing
ca
n be d
one
u
s
ing
binomial filter with filter co
e
fficients (1,6
,15,20,15,6,1
)
.The
si
gnal
can b
e
re
co
n
s
tru
c
ted by
equatio
n(3
)
u
s
ing the estim
a
ted value
s
o
f
.
4. TVAR Par
a
meter
s
Esti
mation Usin
g Basis Fun
ctions
The coefficie
n
ts
(n) of TV
AR mod
e
l in
Equation
(2)
are a
s
sum
e
d
to be sm
oot
h in the
sen
s
e that if the first deri
v
ative of each coeffi
ci
ent may be arbit
r
arily
large, the highe
r order
derivatives n
e
ce
ssarily v
anish. So, the
coeffici
ents
(n) can be ap
proximate
d
by a
set of basis
function
s.
The no
n stat
ionary di
scre
te-time sto
c
h
a
stic
pro
c
e
ss
is
r
e
pr
es
en
te
d
b
y
p
th
orde
r
TVAR model
as:
∑
,
(
2
0
)
Her
e
,
are tim
e
-varyin
g
coe
fficients a
nd
is a
stationa
ry white n
o
ise pro
c
e
s
s an
d
who
s
e mea
n
is zero an
d varian
ce is
.
Acco
rdi
ng to the time-varyi
ng coeffici
ent
s evolution,
TVAR is likel
y to be cate
gori
z
ed i
n
to two
gro
up’
s i.e. adaptive method an
d
basi
s
fun
c
tion
approa
ch.
TVAR mod
e
l
based
on t
he ba
si
s fun
c
tion te
chni
q
ue is able
to
trace a
stro
ng no
n-
stationa
ry si
g
nal. In this te
chni
que, e
a
ch of
its time
-varying
coeffi
cient
s a
r
e m
odele
d
a
s
lin
ear
combi
nation
of a set of basis fun
c
tion
s [15].
The pu
rpo
s
e
of the basi
s
i
s
to pe
rmit fast
and
smoot
h time variati
on of the coe
fficients.
If we denote
,
as the ba
si
s functio
n
and
consi
der a
set
of (q + 1) fu
nction for a g
i
ven model,
we can state
the TVAR co
efficients in g
eneral as:
,
∑
,
(
2
1
)
From (2
1) we examine that, we have to calculat
e the set of paramete
r
s
for
{k=
1
,2,.......
.,
p; m=
0,1,2,.......
....
.,q;
=1} i
n
order to co
mpute the
T
VAR co
effici
ents
,
,and
the TVAR model is
abs
o
lutely s
p
ec
ified
by this
s
e
t.
The TVAR
coefficient
s are de
sign
ed a
s
follo
ws
,
we
con
s
id
er
sin
g
le re
alizatio
n of the
p
r
oc
es
s
.For a given reali
z
ation of
we can analy
z
e (2
0) as a time
-varying linea
r predi
ction
error filter an
d con
s
id
er
to
be the predi
ction error.
=
-
(22)
Whe
r
e,
≜
∑
,
(
2
3
)
The total squ
a
red
predi
ction error,
whi
c
h
is a
s
well as the error i
n
modelin
g
, is now
specified
by:
∑|
|
Substitute (2
1) in (2
3) an
d
the predi
ctio
n error
can be written a
s
:
∑∑
,
(
2
4
)
The total squ
a
red p
r
e
d
ictio
n
error
can b
e
formulate
d
as:
∑
∑∑
,
(
2
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
Varying
Autoregressi
ve
Model Parameters
Es
timation us
ing
…
(G.Ravi S
han
kar
Red
d
y
)
7789
For m
odelin
g
the non
stati
onary
sto
c
ha
stic p
r
o
c
e
s
s
,
covari
an
ce te
chni
que,
we
make
no a
s
sumption
s on t
he data out
si
de [0, N-1]. In equatio
n (2
5)
is the interval over whi
c
h
the summatio
n
is perform
e
d
and set
,
1
.
By minimizing the mean sq
u
a
red p
r
edi
ctio
n
error in (2
5)
we ca
n estim
a
te the time-varying para
m
eters
[16].
We can mini
mize the mea
n
squ
a
re
d predi
ction erro
r in (25
)
by
mean
s of setting th
e gradi
ent of
with respec
t
to
∗
ze
ro.
∗
∑
∗
∗
∑
∗
∗
0
(26
)
1,2
,
⋯
,
;
0,1,
⋯
,
Whe
r
e,
∗
∗
∑∑
∗
,
∗
∗
And the deriv
ative of
∗
with res
p
ec
t to
∗
∗
∗
,
∗
∗
Con
s
e
quently
(26) b
e
come
s,
∑
,
∗
∗
0
(
2
7
)
The a
bove
mentione
d
condition i
s
si
milar to th
e orthog
onality
law en
cou
n
tered
in statio
nary
sign
al modeli
ng. Substitute
24
in (27) we have:
∑
∑∑
,
,
∗
∗
0
(
2
8
)
No
w we d
e
fin
e
a function
,
as sh
own belo
w
,
,
≜
∑
,
,
∗
∗
(29
)
Usi
ng the ab
ove definition
in (28)
we ha
ve,
∑∑
,
,
0
(
3
0
)
The above e
quation rep
r
e
s
ent
s a syste
m
of p(
q+1) li
near e
quatio
ns. The a
bov
e system
of linear eq
ua
tions can be
efficiently re
p
r
esented in m
a
trix form as f
o
llows.
Define a
colu
mn vector
as
follows
:
⋯
,
(
3
1
)
0,1,
⋯
,
We can us
e the func
tion (10) to find the following matrix for
0
,
1,
1
1,2
⋯
1,
2,1
2,
2
…
2,
⋮⋱
⋮
,
1
,
2
⋯
,
(
3
2
)
The ab
ove matrix is of size pxp a
nd
all the differe
nt values for m and g re
sulting in
(q+1)x(q+1) such
matri
c
e
s
,
by mea
n
s of
these
ma
tri
c
e
s
,
we can no
w
d
e
scri
be a block
m
a
trix as
s
h
ow
n
be
lo
w
,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
85 – 779
7
7790
⋯
⋱
::
:
:
..
.
.
⋯
(
3
3
)
The ab
ove Block matrix
C has
(q
+1
)x(q+1
) el
e
m
ent
s an
d ea
ch e
l
ement is
a
matrix of
size pxp, whi
c
h implie
s the
Block mat
r
ix C of size p(q
+
1)x p
(
q
+
1).
N
o
w
w
e
descr
ibe a column vec
t
or
as sh
own b
e
lo
w:
1,0
2,0
⋯
,
0
(
3
4
)
0,1,
⋯
,
By using the
definition
s
fro
m
(31
)
-(34
) we
ca
n re
pre
s
ent the sy
ste
m
of linear
e
quation
s
in (30) in a c
o
mpac
t mat
r
ix form as
follows
:
⋯
⋮⋱
⋮
⋯
⋮
⋮
(35)
By solving th
e above matrix equation,
we
can obtai
n the set of
TVAR param
eters
(elem
ents of
), the predi
cto
r
co
efficient
s
,
can n
o
w b
e
cal
c
ulate
d
usi
ng (2
1).Th
e
matrix C is
of size p(q+1)xp(q+1),to
solve the a
b
o
ve
system
of linear e
q
u
a
tions
we re
quire
s O
(
13
computat
ions.
In the
ba
sis functio
n
exp
a
n
sio
n
, two
issues ne
ed to
b
e
resolved. Fi
rst
a g
ene
ral
cla
s
s of
basi
s
fun
c
tio
n
s is to b
e
chosen, whi
c
h
can su
itably
captu
r
e the
time variation
,
and then, the
signifi
cant n
u
m
ber
of ba
si
s fun
c
tion
s n
eed to b
e
se
lected. Seve
ral cla
s
se
s of
function
s h
a
ve
been p
r
o
p
o
s
ed in the literature
su
ch a
s
time bas
i
s
fu
nction
s, Leg
e
ndre
polynom
ial, Cheby
she
v
polynomial, Dis
c
rete prolat
e
sphe
roidal
(DPSS) s
e
quence, Fourier ba
s
i
s
,
disc
rete c
o
s
i
ne basis
,
Wal
s
h
basi
s
,
Multi wavele
t basi
s
fun
c
ti
ons.
Ho
weve
r, no
uniform
rule
exist
s
to indi
cate
wh
ich
cla
ss
sho
u
ld
be ado
pted. The ap
pro
a
ch of choo
si
n
g
the signifi
cant numbe
r o
f
basis fun
c
ti
ons
(order
sele
cti
on) is
based
on trial an
d erro
r [15]:
Moreove
r
, the expan
sio
n
of the TVAR
para
m
eters i
n
to the basi
s
sequ
en
ce
s substa
ntia
lly increa
se
s the
numbe
r of model pa
ram
e
ters
that is to be estimated. T
o
compare the perfo
rmance of DESA with Basis function method
we
use di
screte
co
sine b
a
si
s functio
n
.
4.1. Discre
t
e
Cosine Basi
s Functio
n
,
=
α
(m)
co
s
π
Whe
r
e,
α
(m
)=
0
0
,1,2
…
…
.
(36)
n=
1,2.....N
For all the a
bove mentio
ned p
r
oble
m
s, the
TVAR model pa
ra
meter e
s
tima
tion usin
g
basi
s
fun
c
tio
n
app
roa
c
h
requi
re
s hig
h
com
putatio
nal complexi
ty. In this paper
we
ha
ve
establi
s
h
ed t
he rel
a
tion b
e
t
ween th
e pa
rameters of th
e AM-FM
sig
nal an
d the p
a
ram
e
ters of
the
TVAR pro
c
e
s
s. We h
a
ve u
s
ed the
Di
screte Ener
gy S
eparation Alg
o
rithm1
(DES
A-1) to e
s
tim
a
te
the TVAR c
o
effic
i
ents and the modulating
s
i
gnal
s of
the TVAR proc
es
s [11]. The DESA-1 bas
ed
TVAR pa
ram
e
ter e
s
timatio
n
is con
c
eptu
a
lly simpl
e
r a
nd e
a
si
er to
i
m
pleme
n
t th
an the
metho
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
Varying
Autoregressi
ve
Model Parameters
Es
timation us
ing
…
(G.Ravi S
han
kar
Red
d
y
)
7791
based on basis functions.
T
VAR parameter estimation usi
ng DESA requires O
(p (q+1)
2
)
comp
utation
s
wherea
s Basis functio
n
method re
qui
re
s O(
1
comput
ations.
5. Experimental Proced
u
r
e
Step 1: Cal
c
ulate the TV
AR pa
ramete
rs
usin
g Equ
a
tions
(18
)
, (19) a
nd fo
rm
the
coeffici
ents
,
usin
g
(21
)
.
Step 2: Solve the roots of t
he time-va
r
ying aut
o regre
ssive p
o
lyno
mial forme
d
b
y
TVAR
linear p
r
edi
ct
ion filter A(z; n)
1
∑
,
1
at each
instant n to find the time-varyin
g
pole
s
:
,
, i=
1, 2.....p.
Step 3: The instantan
eo
us freq
uen
cy
of t
he non stationa
ry sig
nal for ea
ch
sample
instant n can
be estimate
d from the inst
antane
ou
s an
gles of the p
o
les u
s
in
g the formula
,
=
arg
,
2
for
,
1.
Step 4: From time varying paramete
r
s
,
we
can p
r
e
d
ict non
stationary si
gnal
usin
g
(23) with initi
a
l
Conditions
; n
=
0,1,...p where p is the TV
AR model order
Step 5
:
The ti
me varying p
o
we
r sp
ect
r
al
density ca
n be estimate
d from time varying
para
m
eter
,
as
follows
:
P (f; n) =
2
1
∑
,
2
1
2
(36)
Whe
r
e
,
are T
VAR coeffici
e
n
ts and
is:
∑
∑
,
(
3
7
)
6. Simulation Resul
t
s
For
simulatio
n
, we have
con
s
id
ere
d
three
modul
a
t
ed sig
nal m
odel
s, the Discrete
Amplitude M
odulate
d
(A
M) sig
nal, Di
screte F
r
eq
u
ency Mo
dula
t
ed (FM
)
sig
nal and
Discrete
Amplitude an
d Freq
uen
cy modulate
d
(A
M-FM
) sig
nal
s.
6.1. Discre
t
e
AM Signal
Con
s
id
er the
discrete Ampl
itude modul
ated (AM)
sign
al,
1
cos
cos
(38)
For n=1,2….
.
N, whe
r
e k=0.8,
=
128
,
=
6
a
nd N = 51
2.
The IF law of the above si
g
nal is given b
y
:
2
(39)
The
coeffici
e
n
ts of the
TVAR process
1
and
2
of the
discrete
AM
Signal a
r
e
estimated
u
s
ing (18),
(19
)
an
d Equati
on (
21)
and
sho
w
n i
n
Figure 1.
When the
TV
AR
coeffici
ents a
r
e estim
a
ted usin
g basi
s
functio
n
s,
p
=
2
and q=8 di
screte
co
sine
basi
s
functio
n
s
are fo
und
to
give be
st results. Figu
re
1
sh
ows
the
T
VAR pa
ram
e
ters
estim
a
te
d by the
DES
A
-1
and u
s
in
g ba
sis
metho
d
. T
VAR pa
ramet
e
r e
s
timation
usin
g ba
si
s f
unctio
n
ap
pro
a
ch
re
quires O
(p
3
(q
+1
)
3
) co
mputation
s
, whe
r
e a
s
DE
SA based a
p
p
roa
c
h
req
u
ires O (p (q
+1
)
2
) comp
utatio
ns.
Usi
ng
step
2
in expe
rime
ntal procedu
re we
can
co
mpute time v
a
rying
pole
s
,
the time
varying pole
s
are plotted in Figure
(2).
From Fig
u
re
(2)
we ob
se
rve that the p
o
les a
r
e clo
s
e to
the unit ci
rcl
e
. For every
sample in
stant
n, we no
w
come a
c
ross t
he an
gles
of the pole
s
an
d
divide by 2
to find the IF e
s
timate of the AM compon
ent. The true
IF & estimated IF of the AM
comp
one
nt a
r
e
sho
w
n
in
Figu
re
(3
). Fro
m
Fi
g
u
re (3
)
we
ob
serve
that t
he TVA
R
b
a
s
ed
techni
que
ha
s resulted in
really ni
ce IF
estimatio
n
. The me
an
sq
uare
erro
r (M
SE) amon
g t
h
e
true IF
,
and e
s
timated IF
,
fo
r n=2, 3
…
51
2 is cal
c
ul
ate
d
to be -86.4
847dB.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
85 – 779
7
7792
The TVAR
co
efficients
,
can
also b
e
used
to predi
ct the
non statio
nary process
by
mean
s of Eq
uation (23
)
. The di
screte
AM Signal
in additio
n
to the TVAR
predi
ction
are
sho
w
n in Figure (4
), and we observe that
the TVAR model has
effectively predi
cted
.The
averag
e sq
ua
red p
r
edi
ction
erro
r is
calcu
l
ated to be 0.1753.
The time-va
r
ying power
spectral de
nsit
y of di
screte
AM is co
mpu
t
ed usin
g (36
)
, A plot
of
the
time
-freque
ncy distribution (TF
D
) of
discret
e A
M
for th
e TV
AR mo
del i
s
obtaine
d in
Fi
gure
(5). At
every
sampl
e
in
sta
n
t, the TF
D i
s
p
r
oj
ecte
d
to
co
mpri
se
pe
aks at
the IF
estimate
s
at that
instant. To d
e
mon
s
trate t
h
is, we al
so
illustrate
the
analo
gou
s flat time-freque
ncy view of the
TFD in Fig
u
re
(6).
Figure 1. The
estimate of the TVAR
coeffici
ents
,
,
,
for the AM sig
nal
Figure 2. Traj
ectory of Tim
e
-Varyin
g
pol
es
use
d
for discrete AM Signal
Figure 3. Tru
e
and Estimat
ed IF of Discrete
AM Signal
Figure 4. Orig
inal AM Signal and Pre
d
ict
ed
AM Signal
Figure 5. Time Varying Po
wer Sp
ectru
m
of the
Discrete AM Signal
Figure 6. Time-F
requ
en
cy View of the TFD of
Discrete AM Signal
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
Varying
Autoregressi
ve
Model Parameters
Es
timation us
ing
…
(G.Ravi S
han
kar
Red
d
y
)
7793
6.2. Discre
t
e
FM Signal
Con
s
id
er the
discrete F
M
signal:
cos
∑
1
cos
(
4
0
)
For n
=
1, 2….
.
N, whe
r
e
=0.
2
,
=
128
,
=
6
and N=51
2.
The IF law of the above si
g
nal is given b
y
:
∗
cos
2
(41)
The coeffici
e
n
ts of the TVAR pro
c
e
ss
1
and
2
of the
discrete FM sign
al are
estimated
u
s
ing (18),
(19
)
an
d Equati
on (
21)
and
sho
w
n i
n
Figure 7.
When the
TV
AR
coeffici
ents a
r
e e
s
timated
usin
g ba
sis f
unctio
n
s,
p
=
2
and q=14 di
screte
co
sine
basis fun
c
ti
ons
are
foun
d to
give be
st
re
sults. Fig
u
re
7
sh
ow
s the
T
VAR
param
eters e
s
timat
ed by
the
DE
SA-
1and u
s
ing b
a
si
s method.
TVAR param
eter estim
a
ti
o
n
usin
g basi
s
function app
roach req
u
ire
s
O
(p
3
(q
+1
)
3
) co
mputation
s
, whe
r
e a
s
DE
SA based a
p
p
roa
c
h
req
u
ires O (p (q
+1
)
2
) comp
utatio
ns.
Usi
ng
step 2
in experim
ent
al pro
c
e
dure
we
can
com
p
ute time varyi
ng pole
s
, T
r
a
j
ectory
of Time-varyi
ng Poles u
s
e
d
for discrete
FM Signal
are plotted in Figure
(8). Fro
m
Figure
(8)
we
observe th
at the p
o
le
s a
r
e
clo
s
e to
the
u
n
it circle
a
s
a
n
ticipate
d
. Fo
r eve
r
y sampl
e
in
stant n,
we
now
come
a
c
ro
ss the an
gles of the p
o
les a
nd divi
de by 2
to find the IF est
i
mate of the F
M
comp
one
nt.
The true
IF
& estimate
d I
F
of the
FM
comp
one
nt
a
r
e sh
own
in
Figure (9).
From
Figure (9
)
we
observe that
the TVAR b
a
se
d techni
q
ue ha
s resulted in really ni
ce IF e
s
timati
on.
The mean sq
uare erro
r (M
SE) among the true IF
,
and estimated IF
,
for n=2, 3… 512 is
cal
c
ulate
d
to be -96. 87
dB.
The TVAR
co
efficients
,
can
also b
e
used
to predi
ct the
non statio
nary process
by
mean
s of equation (23
)
. The discrete
FM Signal
in addition to the TVAR
predi
ction are
sho
w
n in Fi
gure
(10
)
, an
d we o
b
serve
that
the TVAR mod
e
l ha
s effectively predicte
d
. The
averag
e sq
ua
red p
r
edi
ction
erro
r is
calcu
l
ated to be 0.1065.
The time
-varying po
wer
spectral de
nsit
y of
discrete
FM Signal i
s
comp
uted u
s
i
ng (36),
A plot of the
time-fre
que
n
c
y di
stributio
n (T
FD)
of d
i
screte FM si
gnal
fo
r
the
TVAR
mo
del
is
obtaine
d in Figure (11
)
. At
every sampl
e
instant,
the TFD is p
r
oje
c
ted to compri
se pea
ks at the
IF estimates
at that instant. To demonstrate th
is, we also illu
strate the anal
ogou
s flat ti
me-
freque
ncy vie
w
of the TFD
in Figure (12
)
.
Figure 7. The
Estimate of the TVAR Co
efficients
1
,
,
2
,
for the FM Signal
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
85 – 779
7
7794
Figure 8. Traj
ectory of Tim
e
-varyin
g
Poles
use
d
for discrete FM Signa
l
Figure 9. Tru
e
and Estimat
ed IF of discrete
FM Signal
Figure 10. Ori
g
inal FM si
gn
al, and pre
d
icted
FM sign
al
Figure 11. Time Varying P
o
we
r Spe
c
tru
m
of
the discrete F
M
Signal
Figure 12. Time-F
req
uen
cy View of the TFD of Di
screte FM Signa
l
6.3. Discre
t
e
AM-FM Signal
Con
s
id
er the
discrete AM
-FM sign
al,
1
cos
cos
∑
1
cos
(
4
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.