Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
9
, No
.
1
,
J
an
ua
ry
201
8
,
pp.
1
39
~
1
4
5
IS
S
N:
25
02
-
4752
,
DOI: 10
.11
591/
ijeecs
.
v9.i
1
.
pp
1
39
-
1
4
5
139
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Perform
ance of
Principa
l Comp
on
ent An
alys
i
s and
Or
th
ogon
al
Least Sq
uare on
Optimi
zed Fe
atu
re Set in
Clas
si
fyin
g
Asphyxiat
ed Inf
ant Cr
y Using Su
pp
or
t Ve
ctor Ma
chine
R.
Saha
k
1
, W.
Mans
or
*
2
, Khua
n
Y.
Lee
3
,
A.
Z
ab
idi
4
1
,2,3
,4
Facul
t
y
of
E
le
c
tri
c
al E
ngin
eering,
Univ
ersit
i
Te
knologi MA
RA
40450
Shah
A
la
m
,
Sel
angor
,
Malay
s
ia
2
,3
Com
puta
ti
onal Int
e
ll
ig
ence
De
t
ec
t
ion
RIG,
PLS
CORE,
Univ
ers
it
i Te
kno
logi
M
ARA
,
40450
Shah
Alam,
Sela
ng
or,
Malay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
2
6
, 201
7
Re
vised
N
ov
2
, 201
7
Accepte
d
Nov
20
, 201
7
An
inve
stiga
t
ion
int
o
opti
m
iz
ed
support
vec
tor
m
ac
hine
(SV
M)
int
egr
a
te
d
with
princ
ip
al
c
om
ponent
ana
l
ysis
(PCA
)
and
o
rthogona
l
l
ea
st
s
quar
e
(OLS)
in
c
la
ss
if
y
ing
a
sph
y
xiated
infant
cr
y
w
as
per
f
orm
ed
in
th
is
stud
y
.
Thr
e
e
appr
oac
h
es
wer
e
used
in
the
c
la
ss
ifi
c
at
ion
;
SV
M,
PC
A
-
SVM
,
and
OLS
-
SV
M.
Vari
ous
num
ber
s
of
fe
atures
ext
r
ac
t
ed
fr
om
Mel
-
fre
quen
c
y
Cepstr
a
l
coe
ffi
ci
en
t
(MF
CC)
were
te
sted
to
obta
in
the
o
pti
m
al
par
amet
e
rs
of
S
VM
ker
nel
s.
Onc
e
th
e
opti
m
al
f
ea
tu
r
e
set
is
obt
ai
ned
,
PC
A
and
OLS
sele
c
te
d
th
e
m
ost
signifi
ca
nt
fe
at
ur
es
and
the
opti
m
iz
ed
SV
M
the
n
c
la
ss
ified
t
he
select
ed
cr
y
p
atter
ns.
In
PC
A
-
SVM,
ei
genva
lu
e
-
one
-
criter
ion
(
EOC),
cumulat
iv
e
per
ce
n
ta
g
e
v
aria
nce
(CPV
)
and
t
he
Scre
e
te
st
(S
CREE
)
were
used
to
sel
ect
the
m
ost
signifi
ca
nt
fe
at
ur
es.
SV
M
with
ra
dia
l
basis
func
ti
on
(
RBF
)
ker
nel
was
chose
n
in
the
class
ifi
cati
on
stage
.
Th
e
cl
assifi
ca
t
ion
accura
c
y
and
computat
ion
tim
e
were
comp
ute
d
to
ev
al
ua
t
e
th
e
p
erf
orm
a
nce
of
each
m
et
hod.
The
b
est
m
et
hod
for
cl
as
sif
y
ing
asph
y
x
iated
infa
n
t
cr
y
is
PC
A
-
SVM
with
EOC
sinc
e
it
produc
es
t
he
high
est
c
la
ss
ifi
c
at
ion
accurac
y
which
i
s
94.
84%.
Us
ing
PC
A
-
SVM,
the
cl
assifi
ca
t
ion
proc
ess
was
per
form
ed
in
1.
98s
onl
y
.
The
re
sults
al
so
show
th
at
e
m
plo
y
ing
fe
a
tur
e
sel
ec
t
ion
te
chn
ique
s
cou
ld
enha
nc
e the
class
ifi
er
p
erf
orm
an
ce
.
Ke
yw
or
d
s
:
Asphyxia
Infan
t C
ry
Or
t
hogonal Le
ast
Squar
e
,
Pr
inci
pal Com
pone
nt Analy
sis
Suppor
t
V
ect
or Mac
hin
e
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
W.
Ma
nsor
,
Faculty
of Elec
tric
al
Engineer
ing
,
Un
i
ver
sit
i Te
knol
og
i M
ARA
,
4045
0
S
ha
h A
lam
, S
el
angor,
Ma
la
ysi
a
.
E
m
a
il
:
wah
ida
h231@salam
.u
itm
.ed
u.
m
y
1.
INTROD
U
CTION
Asphyxia
is
re
ferred
to
res
pir
at
or
y
fail
ure,
a
conditi
on
ca
use
d
by
ina
de
quat
e
intake
of
ox
y
gen
[1
]
.
It
is
im
po
rtant
to
dia
gnos
e
as
ph
y
xia
in
in
fa
nts
at
early
bir
th
since
it
is
a
m
ajo
r
ca
us
e
of
infa
nt
m
or
bid
i
ty
.
If
m
isa
pp
r
opriat
e
treatm
ent
is
giv
en
to
the
in
fa
nt,
hypoxia
will
resu
l
t,
wh
ic
h
cou
l
d
le
ad
to
s
erio
us
c
om
plication,
su
c
h
as
dam
age to
in
fa
nt’s bra
in, or
ga
ns
, tis
su
es
or e
ven fa
ta
li
t
y.
Asphyxia o
ccu
rs
in
in
fan
ts wit
h
ne
urolo
gical
le
vel
disturba
nce,
wh
ic
h
is
f
ound
to
af
fect
the
sou
nd of
the
cry
pr
oduc
ed
by
t
he
inf
a
nts
[
1].
T
he
ef
fect
causes
t
he
cry
sig
nals
to
hav
e
disti
nct
patte
rn
s
com
par
ed
t
o
healt
hy
infa
nt
cry.
These
ha
ve
been
pro
ve
n
by
pr
e
vious
stud
ie
s
[
2].
The
researc
hers
su
c
cessf
ul
disti
nguishe
d
betwee
n
the
he
al
thy and as
phyxia
te
d
in
fan
t
cry us
i
ng Com
pu
te
r
-
Ba
se
d
a
na
ly
sis
.
B
asi
cal
ly
,
Com
pu
te
r
-
Ba
sed
analy
sis
co
ntains
th
ree
c
om
po
ne
nts;
pre
-
pr
ocessin
g,
featu
re
ext
racti
on
and
patte
r
n
cl
a
ssific
at
ion
.
Th
e
cry
sig
nals
wer
e
pre
-
proc
essed
first
befor
e
cry
featur
e
s
extracte
d
in
featu
r
e
extracti
on
sta
ge
.
Th
e
po
pu
la
r
m
e
tho
d
us
e
d
f
or
e
xtr
act
in
g
f
eat
ur
es
from
cry
sign
al
s
is
Me
l
-
fr
e
quency
Ce
ps
tral
Coef
fici
ent
(M
FCC
)
[3
-
5].
O
nce
the
featu
r
es
are
e
xtracted,
patte
rn
cl
as
sifie
r
will
cl
assify
the
cry
pa
tt
erns
accor
ding
to
ty
pes
of
cry.
H
ow
e
ve
r,
rece
nt
researc
her
s
a
pp
e
nded
a
sta
ge
after
the
fe
at
ur
e
extra
ct
io
n
sta
ge,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
47
52
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
1
,
Jan
ua
ry
201
8
:
1
39
–
1
4
5
140
cal
le
d
feature
sel
ect
ion
[6
-
8].
Norm
al
l
y,
in
featur
e
e
xtracti
on,
the
e
xtract
ed
feat
ur
es
ha
ve
a
la
rg
e
dim
ensio
n
that
so
m
eti
m
e
s
enco
m
passes
le
ss
sign
ific
a
nt
and
redu
ndant
featu
res.
Hen
ce
,
featu
re
sel
ect
ion
m
eth
od
is
require
d
t
o
sel
ect
the
m
os
t
sign
ific
a
nce
c
r
y
featu
re
wh
ic
h
the
n
en
ha
nc
es
the
cl
assi
ficat
ion
acc
ur
a
c
y
and
co
m
pu
ta
ti
on
ti
m
e as w
el
l
.
In
this
stu
dy,
or
t
hogonal
le
a
st
square
(
OL
S)
a
nd
pr
i
ncip
al
com
po
ne
nt
analy
sis
(P
CA
)
hav
e
bee
n
e
m
plo
ye
d
t
o
s
el
ect
the
m
os
t
sign
ific
a
nt
cry
featu
res
w
hich
e
xtracted
f
rom
the
analy
sis
of
MFC
C.
O
LS
has
sh
ow
n
t
o
be
a
ble
to
sel
ect
s
ign
ifi
ca
nt
feat
ur
es
f
ro
m
MFC
C
su
ccess
fu
ll
y
in
pre
vious
stud
ie
s
[
9
-
11
]
.
Eve
n
though
PC
A
ha
s
been
us
e
d
in
the
pre
vious
infan
t
cry
a
na
ly
sis,
app
li
cat
ion
of
PC
A
on
both
norm
al
and
asp
hyxiate
d
c
r
y
on
ly
has
not
been
i
nvest
igate
d.
F
urt
her
m
ore,
dif
fer
e
nt
ap
proac
hes
of
P
CA
sel
ect
ion
s
kn
ow
n
as
ei
genvalue
-
one
-
crit
eri
on
(
EOC)
,
cum
ulativ
e
per
ce
ntage
var
ia
nce
(CP
V)
a
nd
sc
ree
te
st
(S
CR
EE)
ha
ve
not
been
e
xam
ine
d
ye
t
in
the
pr
evio
us
stu
dy.
To
cl
assify
inf
ant
cry
sign
al
s
,
a
capab
le
suppo
rt
vector
m
achin
e
(S
VM
)
is use
d.
SV
M pe
rfor
m
s
cl
assifi
cat
ion
ta
sk
s b
ase
d
on
the p
ri
nciple of b
inary
cl
as
sif
ic
at
ion
.
S
VM o
f
fer
s
m
or
e
adv
anta
ge
s
su
ch
as
gl
obal
ly
op
tim
al
,
sm
a
ll
sa
m
p
le
-
siz
e,
good
gen
e
rali
zat
ion
abili
ty
and
resist
ant
to
the
ov
e
r
-
fitt
ing
pro
blem
, th
an oth
er classi
fiers
[1
2
-
15]
.
This
pa
pe
r
des
cribes
the
asp
hy
xiate
d
infa
nt
cry
cl
assifi
cat
i
on
us
in
g
SV
M
with
rad
ia
l
ba
sis
fu
nctio
n
(RBF)
ke
rn
el
.
Thr
ee
ap
proac
hes
know
n
as
SV
M,
PCA
-
S
VM
a
nd
OL
S
-
SV
M
wer
e
use
d
in
the
cl
assif
ic
at
ion
.
The
opti
m
iz
at
i
on
process
is
carried
ou
t
to
ob
ta
in
op
ti
m
al
par
am
et
er
s.
The
optim
iz
at
io
n
of
in
put
featur
e
set
extracte
d
from
MFC
C
analy
si
s
is
per
f
or
m
ed
first
in
the
SVM
app
r
oac
h.
T
he
optim
iz
ed
featur
e
set
ob
ta
i
ned
is
then
us
ed
as
a
ref
ere
nce
in
pu
t
to
OLS
-
S
VM
an
d
PCA
-
SV
M
a
ppr
oa
ches.
Cl
assifi
c
at
ion
accu
racy
and
com
pu
ta
ti
on
ti
m
e w
ere c
om
pu
te
d
a
nd c
om
par
ed
to
ac
hieve
the
best cla
ssif
ie
r
pe
rfor
m
anc
e.
2.
RESEA
R
CH MET
HO
D
The
data
base
of
norm
al
and
asph
y
xiate
d
inf
ant
cry
in
this
stud
y
is
ob
ta
in
ed
from
the
Un
ive
rsity
of
Mi
la
no
–
Bi
co
c
ca
[16].
T
he
w
ho
le
process
of
cl
assify
ing
as
ph
y
xia
te
d
in
fa
nt
cry
is
s
how
n
in
Fig
ure
1.
Ini
ti
al
l
y,
pre
-
pr
ocessin
g
was
car
ried
ou
t
w
her
e
the
sig
nals
w
ere
norm
al
iz
ed
and
sam
pled
at
ei
gh
t
kHz
a
nd
pr
e
-
e
m
ph
asi
zed
be
fore
div
idi
ng
them
into
seg
m
ents
of
one
seco
nd
each
.
Fr
om
this
pro
cess,
316
se
gm
ents
of
norm
al
cr
ie
s an
d 2
84
of asph
yxia
te
d
cries
were
gen
e
rated
.
Af
te
r
pr
e
-
pr
oc
essing,
MFC
C
analy
sis
was
carried
ou
t
wh
e
re
the
se
gm
ented
sign
al
s
wer
e
fi
rst
m
ul
ti
plied
with
a
H
am
m
ing
window
with
a
width
of
25
m
s
and
an
ov
e
rlap
ped
of
50%
betwee
n
suc
cessi
ve
fr
am
es.
The
outp
u
t
sig
nal
w
as
then
pr
oces
sed
by
Fast
F
ourier
Tra
nsfo
r
m
(F
FT).
T
he
resu
lt
ed
spe
ct
r
um
was
processe
d
t
hro
ugh
t
rian
gu
la
r
filt
er
ba
nks
to
trans
form
it
in
to
a
Me
l
-
s
pect
ru
m
.
Finall
y,
DCT
was
a
pp
l
ie
d
to
pro
du
ce
t
he
c
orres
pondin
g
coeffic
ie
nts.
I
n
this
st
ud
y,
t
en
to
20
c
oeff
ic
ie
nts
wer
e
pro
duced
from
var
i
ou
s
nu
m
ber
s
of f
il
te
r ban
ks
ra
ng
e
d betwee
n 20 a
nd 30 w
hich f
orm
ed1
21 f
eat
ure set
s.
Ther
e
we
re
t
hree
m
et
ho
ds
use
d
i
n
the
cl
assifi
cat
ion
of
asp
hyxiate
d
i
nfant
c
ry.
T
he
first
m
et
ho
d
cal
le
d
SV
M
was
pe
rfo
rm
e
d
to
ob
ta
in
t
he
op
ti
m
al
feat
ur
e
set
.
SV
M
perform
s
cl
a
ssific
at
ion
ta
s
k
us
i
ng
hype
rp
la
ne
to
diff
e
ren
ti
at
e
betwee
n
tw
o
cases.
Th
us,
it
is
i
m
po
rtant
to
co
ns
tr
uct
an
opti
m
a
l
separ
at
ing
hype
rp
la
ne.
T
he
hype
r
plane
is
sai
d
t
o
be
op
ti
m
al
wh
en
the
r
e
is
a
la
r
ge
distance
betwee
n
hyp
e
rp
la
ne
a
nd
data
po
i
nt.
By
choosin
g
ap
pro
pr
i
at
e
reg
ularizat
ion
pa
ram
et
er
(
C
)
and
kern
el
fu
nctio
n,
a
n
optim
al
separ
at
ing
hype
rp
la
ne
co
uld
be
obta
ine
d
[17,
18
]
.
T
he
kernel
fun
ct
ion
is
us
ed
to
transfo
rm
data
po
ints
into
hig
h
dim
ension
al
s
pace
s
o
t
hat
t
he
data
ca
n
be
li
near
ly
se
pa
rated.
T
he
hy
perplane
is
then
buil
t
into
these
trans
form
ed
data
based
on
the
value
of
C
tha
t
has
to
be
sel
ect
ed
pro
per
ly
by
a
us
er.
I
f
C
is
too
la
rg
e,
it
m
ay
ov
e
rf
it
the
dat
a an
d
if
it
too s
m
al
l, it
m
a
y under
fit t
he data
.
The deci
sio
n
f
un
ct
io
n
ca
n be
expresse
d
as:
N
i
j
i
i
i
q
K
y
s
i
g
n
y
1
,
)
(
(1)
wh
e
re
N
is t
he
to
ta
l of
su
pp
or
t v
ect
or
num
ber
,
i
is a Lagr
an
gia
n
m
ult
ipli
er,
i
is
a com
po
nen
t v
ect
or
of
α
and
i
y
is
the
la
bel
associat
ed
with
i
;
+1
or
-
1,
j
is
su
pp
or
t
ve
ct
or
,
q
is
the
hy
perplane
a
nd
j
i
K
,
is t
he ker
nel fu
nction.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
Princip
al Co
mpo
nen
t
An
aly
sis a
nd O
rt
hogonal Le
as
t
Square
on… (
R.
Sahak
)
141
S
T
A
R
T
I
n
f
a
n
t
c
r
y
s
i
g
n
a
l
r
e
t
r
i
e
v
a
l
P
r
e
-
p
r
o
c
e
s
s
i
n
g
M
F
C
C
a
n
a
l
y
s
i
s
S
V
M
c
l
a
s
s
i
f
i
c
a
t
i
o
n
(
I
n
p
u
t
f
e
a
t
u
r
e
o
p
t
i
m
i
z
e
d
?
)
P
C
A
s
e
l
e
c
t
i
o
n
C
l
a
s
s
i
f
i
e
d
c
r
y
p
a
t
t
e
r
n
S
V
M
c
l
a
s
s
i
f
i
c
a
t
i
o
n
(
F
e
a
t
u
r
e
s
e
t
i
s
o
p
t
i
m
i
z
e
d
?
)
E
N
D
O
L
S
s
e
l
e
c
t
i
o
n
S
V
M
P
C
A
-
S
V
M
O
L
S
-
S
V
M
S
V
M
c
l
a
s
s
i
f
i
c
a
t
i
o
n
(
F
e
a
t
u
r
e
s
e
t
i
s
o
p
t
i
m
i
z
e
d
?
)
Y
N
N
N
Y
Y
Figure
1.
The
process
of clas
sifyi
ng
a
sphyxi
at
ed
in
fan
t c
ry
The ker
nel
fun
ct
ion
ca
n be e
xpress
ed
as:
)
(
)
(
,
j
T
i
j
i
K
(2)
The
c
omm
on
kernel
us
e
d
is
ra
dial basis
f
un
ct
ion
(RBF)
whi
ch
ca
n be e
xpr
essed
a
s:
2
j
i
j
i
K
e
x
p
,
(3)
wh
e
re
γ
is t
he
kernel
width.
In
SV
M
a
ppr
oa
ch,
the
12
1
f
eat
ur
e
set
s
e
xtracted
f
r
om
th
e
MFC
C
analy
sis
wer
e
pass
ed
th
rou
gh
SV
M
with
RB
F
ke
r
nels.
I
n
t
he
sec
ond
m
eth
od
cal
le
d
PC
A
-
S
VM,
once
the
op
ti
m
al
featur
e
set
wa
s
obta
ined,
PCA
was
em
plo
ye
d
to
the
optim
iz
ed
featur
e
set
.
I
n
PC
A
-
SV
M
,
EOC,
CPV,
an
d
SC
REE
al
go
rit
hm
s
wer
e
app
li
ed
to
the
op
ti
m
al
featur
e
set
.
In
a
no
t
her
m
et
ho
d,
OL
S
was
a
ppli
ed
t
o
the
optim
iz
ed
featu
re
set
an
d
t
his
appr
oach is cal
le
d
O
LS
-
SV
M
.
In
orde
r
to
obta
in
a
n
op
ti
m
al
m
od
el
of
SV
M
with
t
he
po
ly
nom
ia
l
kernel,
C
wa
s
va
ried
from
0.000
001
t
o
0.
01
a
nd
d
wa
s
var
ie
d
f
ro
m
two
to
fou
r.
F
or
SV
M
with
RB
F
ke
rn
el
,
C
a
nd
γ
wer
e
va
ried
f
ro
m
0.01
to
100
a
nd
0.001
to
0.0
5
re
sp
ect
ively
.
The
range
of
C
,
d
,
an
d
γ
was
sel
ect
ed
ba
sed
on
the
previo
us
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
47
52
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
1
,
Jan
ua
ry
201
8
:
1
39
–
1
4
5
142
stud
ie
s
a
nd
al
s
o
f
ro
m
the
cur
r
ent
ex
per
im
ents.
For
reli
abili
ty
resu
lt
s,10
-
cr
os
s
-
validat
io
n
was
use
d
to
se
par
at
e
betwee
n
trai
nin
g an
d
te
st
dat
aset
s
.
3.
RESU
LT
S
A
ND AN
ALYSIS
Figure
2
s
how
s
the
var
ia
ti
on
of
cl
assifi
cat
ion
acc
ur
acy
of
SV
M
with
R
BF
kernel
(C
=
1
an
d
γ
=
0.009
)
obta
ine
d
wh
e
n
t
he
num
ber
s
of
c
oe
ffi
ci
ent
and
filt
er
ba
nk
wer
e
va
ried
from
te
n
to
20
an
d
20
to
30
resp
ect
ively
.
It
is
obse
rv
e
d
t
ha
t
feature
set
t
hat
co
ns
ist
s
of
20
c
oeffici
ent
s
gi
ves
poor
cl
assifi
cat
ion
ac
cur
acy
wh
ic
h
sho
ws
there
a
re
m
any
rep
et
it
ive
an
d
f
ewer
si
gn
ific
a
nc
e
feat
ur
es
.
T
he
hi
gh
est
cl
as
sific
at
ion
accu
racy
is
93.84%
wh
e
n
te
n
coeffic
ie
nt
s
with
22
an
d
25
filt
er
ba
nk
s
a
re
us
ed
.
So
,
featu
re
set
gen
e
rated
fro
m
te
n
coeffic
ie
nts
an
d
22
filt
er
ba
nks
is
chosen
as
the
best
crit
erion
for
de
velo
pi
ng
the
opti
m
al
featur
e
set
sin
ce
this
com
bin
at
ion
produces
sm
all f
eat
ur
es.
Figure
2
.
Cl
assifi
cat
ion
acc
uracy
o
f
S
VM w
i
th RBF
kernel
for vari
ou
s
co
e
ff
ic
ie
nts a
nd f
il
te
r
ba
nk
nu
m
ber
s
In
PCA
-
S
VM
ap
proac
h
us
i
ng
CP
V
,
54
pr
i
ncipal
c
ompone
nts
a
re
r
et
ai
ned
as
th
e
pri
ncipal
com
po
ne
nts
ha
ve
a
bout
80
%
of
t
he
c
umulat
ive
per
ce
nt
of
the
var
ia
nc
e.
U
sin
g
CP
V,
the
dim
ension
of
the
input
featu
re
is
reduce
d
f
rom
[1
240
x
600]
to
[54
x
600]
.
Wh
en
t
he
fea
tures
ar
e
sel
ect
ed
by
E
OC,
on
ly
23
pr
i
ncipal
com
pone
nts
are
re
ta
ined
since
th
ey
hav
e
ei
ge
nval
ue
gr
eat
er
t
han
on
e
.
N
ote
that
with
EO
C,
the
or
i
gin
al
dim
ension
is
re
du
ce
d
fr
om
[1
24
0
x
600]
to
[
23
x
600].
In
Sc
ree
te
st,
on
ly
three principal
com
ponen
ts
wer
e
retai
ne
d
as
they
ha
ve
the
la
r
ge
ga
p
be
tween
eac
h
ot
her
as
s
how
n
in
Fig
ur
e
3.
W
it
h
this
m
eth
od,
t
he
dim
ension
of
i
nput
featur
e
is
reduce
d
f
r
om
[
1240
x 600] t
o
[3 x 6
00]
.
Figure
3. Scree
p
lot
of in
put f
eat
ur
es
f
or
SVM
with RBF
kernel
90
91
92
93
94
20
21
22
23
24
25
26
27
28
29
30
C
l
a
s
s
i
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
N
u
m
b
e
r
o
f
f
i
l
t
e
r
b
a
n
k
1
0
c
o
e
f
f
i
c
i
e
n
t
s
1
2
c
o
e
f
f
i
c
i
e
n
t
s
1
4
c
o
e
f
f
c
i
e
n
t
s
1
6
c
o
e
f
f
c
i
e
n
t
s
1
8
c
o
e
f
f
c
i
e
n
t
s
2
0
c
o
e
f
f
i
c
i
e
n
t
s
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
Princip
al Co
mpo
nen
t
An
aly
sis a
nd O
rt
hogonal Le
as
t
Square
on… (
R.
Sahak
)
143
Figure
4
s
how
s
cl
assifi
cat
ion
accur
acy
of
P
CA
-
S
VM
with
RB
F
ke
rn
el
w
hen
C
inc
rease
s
from
0.
01
to 1
00 for
γ
eq
ual to 0
.02
5.
As shown
in
the
fig
ur
e, the hig
hest cla
ssific
at
ion
accu
racy i
s o
btaine
d
when
C
=
1
us
in
g
E
OC
al
gorithm
.
The
rat
e
of
inc
rem
ent
in
cl
ass
ific
at
i
on
acc
ur
a
cy
is
influ
e
nce
d
by
the
dim
ension
of
the
input feat
ur
e
vec
tor
.
Figure
4. Ef
fec
t of va
ryi
ng C
on the classi
ficat
ion
acc
ur
acy
of SV
M
w
it
h
RB
F k
e
rn
el
c
om
bin
ed
with
E
OC,
CPV,
a
nd SCR
EE
Chan
ges
in
the
cl
assifi
cat
ion
accuracy
of
E
OC,
CPV
a
n
d
SCR
EE,
w
hen
com
bin
ed
with
SV
M
with
RB
F k
er
nel as
γ
is var
ie
d
at
th
e o
ptim
al
C
(w
hich
is
on
e
)
is
sh
ow
n
in Fi
gur
e 5
. Note that i
ncr
easi
ng
γ
do
es n
ot
yi
el
d
good
cl
a
ssific
at
ion
acc
ur
acy
for
E
OC
,
CPV
,
a
nd
S
CR
EE.
The
op
tim
a
l
γ
is
obta
ined
at
0.0
25
s
ince
it
pro
du
ces
the
highest
cl
assifi
cat
ion
accura
cy
(9
4.8
4%).
This
optim
al
value
is
ob
ta
i
ned
wh
e
n
the
EOC
sel
ect
ion
is em
plo
ye
d.
Wh
e
n SCR
EE is
us
ed
, th
e
w
or
st cl
as
sific
at
ion
acc
uracy
is obtai
ned
.
Figure
5. Cl
assifi
cat
ion
acc
uracy
o
f
S
VM w
i
th RBF
kernel
wh
e
n
γ
is
var
ie
d usin
g
E
OC,
CPV,
a
nd
SCR
EE
The
te
n
c
oeffici
ents
r
a
nk
e
d
w
it
h
the
us
e
of
OLS
al
gorithm
is
sho
wn
in
T
able
1.
C
oeffici
ent
with
the
highest
ERR
is
ar
range
d
at
t
he
to
p
of
the
li
st
whereas
coe
f
fici
ent
wit
h
th
e
le
ast
ERR
is
placed
at
the
bott
om
of the
li
st. T
he fo
ur
t
h,
first,
thi
rd, se
ven
t
h,
fift
h
a
nd tenth
c
oe
ff
ic
ie
nts c
onta
in si
gn
ific
a
nt in
form
ation
.
In
the
optim
iz
a
ti
on
proce
ss,
th
e
op
ti
m
al
C
is
on
e
for
al
l
m
eth
ods.
H
ow
e
ve
r
,
dif
fer
e
nt
opti
m
al
values
of
γ
are
found
for
SV
M
(
0.0
09),
PC
A
-
S
V
M
(0
.
025)
an
d
OLS
-
SV
M
(
0.0
07).
T
he
hi
ghest
cl
assifi
cat
ion
accuracy
(
94.84%)
is
obta
ine
d
from
PCA
-
SV
M
(EO
C
),
f
ol
lowed
by
SVM
(9
3.8
4%
)
as
sh
ow
n
in
Figure
6.
Howe
ver,
OL
S
-
S
VM
(eig
ht
coeffic
ie
nts)
prov
i
des
co
m
par
able
accuracy
(
93.
34%)
to
S
VM
since
th
e
diff
e
re
nce
between
their
cl
as
sific
at
ion
acc
uraci
es
is
a
bout
0.5
03%
on
ly
.
It
can
be
co
ncl
ud
e
d
t
hat
the
f
eat
ur
e
sel
ect
ion
al
gor
it
h
m
s
aff
ect
the
cl
assifi
cat
ion
accu
racy
ba
sed
on
the
f
act
that
the
l
ow
est
sup
port
vect
or
pro
du
ces
the
hi
gh
e
st cl
assifi
ca
ti
on
acc
ur
acy
.
50
60
70
80
90
100
0
.
0
1
0
.
1
1
10
100
C
l
a
ss
i
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
C
EO
C
C
P
V
S
C
R
EE
γ
=
0
.
0
2
5
82
84
86
88
90
92
94
96
0
.
0
0
1
0
.
0
1
0
8
0
.
0
2
0
6
0
.
0
3
0
4
0
.
0
4
0
2
0
.
0
5
C
l
a
ss
i
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
γ
EO
C
C
P
V
S
C
R
EE
C
=
1
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
47
52
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
1
,
Jan
ua
ry
201
8
:
1
39
–
1
4
5
144
Ta
ble 1. Ra
nk
i
ng of c
oeffi
ci
ents u
si
ng OLS
Ran
k
Co
ef
f
icien
t Nu
m
b
er
ERR
1
4
0
.17
8
2
1
0
.16
1
3
3
0
.15
5
4
7
0
.02
0
5
5
0
.02
0
6
10
0
.01
1
7
6
0
.01
1
8
2
0
.00
9
9
8
0
.00
7
10
9
0
.00
5
Figure
6. Cl
assifi
cat
ion
acc
uracy
o
f
S
VM, PC
A
-
S
VM, a
nd
OLS
-
S
VM w
it
h
RB
F
kernel f
or
var
i
ou
s
γ
Table
2
sho
ws
the
accuracy
and
c
om
pu
ta
tio
n
ti
m
e
between
SV
M
,
PC
A
-
SV
M
,
an
d
O
LS
-
S
VM
with
RB
F
ker
ne
l
in
cl
assify
ing
cry
patte
rn
s.
T
he
PCA
-
SV
M
ta
ke
s
on
ly
1.9
s
w
hich
is
the
sho
rtest
tim
e
to
classify
the inf
a
nt cr
ie
s
. Th
e c
om
pu
ta
ti
on
ti
m
e o
f
OL
S
-
S
VM is short
er th
an SVM.
It can be concl
ud
e
d
that c
ombini
ng
featur
e
selec
t
io
n
al
go
rithm
s w
it
h
SV
M
im
pr
ov
es t
he
cl
assifi
cat
ion
acc
ur
ac
y and com
pu
ta
ti
on
ti
m
e.
Table
2.
Cl
assi
ficat
ion
acc
ura
cy
an
d com
pu
t
at
ion
Tim
e o
f SVM,
PCA
-
S
VM,
and OL
S
-
SV
M
w
it
h
RB
F
K
e
r
nel
Metho
d
s
Clas
sif
icatio
n
Accuracy
Co
m
p
u
tatio
n
T
i
m
e
(
s)
SVM
9
3
.84
%
6
2
.87
7
PCA
-
SV
M
9
3
.84
%
1
.98
2
OLS
-
S
VM
9
3
.34
%
3
6
.89
4
4.
CONCL
US
I
O
N
An
i
nv
est
igati
on
i
nto
the
cl
assifi
cat
ion
pe
rfor
m
ance
of
t
hr
ee
a
ppr
oaches,
SV
M,
PCA
-
S
VM
an
d
OLS
-
S
VM
in
cl
assify
ing
as
phyxia
te
d
in
fan
t
cries
with
SVM
com
bin
ed
with
featu
re
se
le
ct
ion
te
chn
i
ques
ha
s
been
disc
us
se
d.
RB
F
ke
r
nel
was
em
plo
ye
d
an
d
op
ti
m
izati
on
pr
oc
ess
was
car
ried
out
to
obta
in
optim
a
l
par
am
et
ers
.
Fo
r
t
he
first
a
ppr
oach,
w
hich
is
SV
M,
the
op
ti
m
al
feature
set
was
gen
e
rated
f
ro
m
te
n
coef
fici
e
nts
with
22
filt
er
banks
.
T
he
highest
cl
assifi
cat
ion
acc
uracy
is
93.
84
%
f
or
RB
F
ke
rn
el
.
In
t
he
sec
ond
a
ppr
oac
h,
wh
ic
h
is
PCA
-
SV
M,
the
high
est
cl
assifi
cat
i
on
acc
ur
acy
f
or
RB
F
kernel
is
94.84%
w
he
n
EOC
sel
ect
ion
was
e
m
plo
ye
d on t
he op
ti
m
iz
ed
featur
e set
.
In
OLS
-
S
VM
appr
oach,
the
highest
cl
assif
ic
at
ion
accu
ra
cy
achieved
is
93.34%
a
nd
the
highest
cl
assifi
cat
ion
was
ac
hieve
d
wh
e
n
ei
gh
t
c
oe
ff
ic
ie
nts
t
hat
sel
ect
ed
f
ro
m
OLS
analy
sis
was
em
plo
ye
d.
For
t
he
50
60
70
80
90
100
0
.
0
0
1
0
.
0
1
0
8
0
.
0
2
0
6
0
.
0
3
0
4
0
.
0
4
0
2
0
.
0
5
C
l
a
s
s
i
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
γ
S
V
M
P
C
A
-
S
V
M
O
L
S
-
S
V
M
C
=
1
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
Princip
al Co
mpo
nen
t
An
aly
sis a
nd O
rt
hogonal Le
as
t
Square
on… (
R.
Sahak
)
145
resu
lt
s
ob
ta
ine
d
in
this
st
ud
y,
it
can
be
s
hown
t
hat
the
cl
a
ssific
at
ion
acc
ur
acy
is
obta
in
ed
withi
n
the
s
hortes
t
com
pu
ta
ti
on
ti
m
e
after
OLS
and
PC
A
wer
e
e
m
plo
ye
d.
The
RB
F
kernel
is
su
it
able
for
cl
a
ssifyi
ng
asp
hy
xiate
d
infa
nt cr
y.
ACKN
OWLE
DGE
MENTS
T
he
a
uthors
w
ou
l
d
li
ke
to
th
ank
t
he
Mi
nist
ry
of
Scie
nce
,
Tech
no
l
og
y,
and
I
nnovat
io
n
(e
Scie
nce
Gr
a
nt
06
-
01
-
01
-
S
F02
10),
M
al
ay
sia
,
and
U
niv
e
rsiti
Teknolo
gi
Ma
ra,
M
al
ay
sia
fo
r
t
he
facil
it
ie
s
pr
ov
ided
t
o
carry
out
this
r
esearch
pro
j
ect
.
The
a
utho
rs
would
al
s
o
li
ke
to
tha
nk
D
r.
Ca
rlos
A.
Re
y
es
-
Ga
rcia,
Dr
.
Em
il
i
o
Ar
c
h
-
Tirad
o,
Dr
.
Ed
gar
M
.
Gar
ci
a
-
Tam
ayo
an
d
the
IN
R
-
Me
xico
gro
up
for
their
ded
ic
at
ion
in
c
ollec
ti
ng
the
Ba
by Chill
anto
D
at
ab
as
e as
w
el
l as In
sti
tut
o Tec
nolo
gico S
up
e
rio
r de
Atli
xco of Me
xico for thei
r
data
ba
se on
infa
nt
cry
sig
na
ls.
The
Ba
by
Chil
la
nto
Data
Ba
se
is
a
property
of
the
I
ns
ti
tuto
Nacio
nal
de
Ast
rofisi
ca
Op
ti
c
a
y El
ect
ro
nica
–
CON
AC
YT,
Me
xico.
REFERE
NCE
S
[1]
La
ber
g
e
M.
Infa
nc
y
through
Ado
le
sce
n
ce Gal
e
.
F
arm
ingt
on
Hill
s
,
MI:
Gal
e
Group
,
Thoms
on
Gal
e, 2006.
[2]
Le
der
m
an
D.
Au
tomati
c
Cla
ss
ific
at
ion
of
Inf
ant
s’
Cr
y
.
M.Sc
.
Degr
ee
,
Ben
-
Gurion
Univer
sit
y
of
th
e
Nege
v
,
2002
.
[3]
Băni
c
ă
I
,
Cucu
H,
Buzo
A
,
Bu
ril
e
anu
D,
Burileanu
C
.
Au
tomati
c
Me
thods
for
Infant
Cry
Cla
ss
if
ic
ati
on
,
201
6
Inte
rna
ti
ona
l
Co
nfe
re
nc
e
on
Co
m
m
unic
at
ions (
COM
M)
,
2016; 51
–
54
.
[4]
Sriji
ra
non
K
,
E
i
amkanit
ch
at
N.
Appl
ic
a
ti
on
o
f
Neuro
-
fuzzy
approache
s
to
re
co
gnit
ion
and
cl
as
sifi
cation
of
inf
a
nt
cry
,
IEEE
R
egi
o
n
10
Confer
ence
,
2014;
1
–
6.
[5]
Rosita
Y
D,
Junae
di
H
.
In
f
ant's
cry
sound
cl
ass
if
ic
a
ti
on
using
Me
l
-
F
reque
ncy
C
epstrum
Coef
fi
c
ients
fe
ature
extract
io
n
and
Bac
kprop
agati
on
Neural
Net
work
.
2nd
Int
ern
ational
Confe
re
nce
on
Scie
n
ce
and
Te
chnol
og
y
-
Com
pute
r
(ICST
)
,
2016
;
160
–
1
66.
[6]
Uyun
S.
Fe
at
ur
e
Selecti
on
M
amm
ogra
m
Based
on
Brea
st
C
ancer
Mining
,
Indo
nesian
Journal
of
Elec
tric
al
an
d
Computer
Engi
n
ee
ring
,
2018;8
(
1):
Acc
epted
for
publi
c
at
ion
.
[7]
Salmam
F
Z,
M
ada
ni
A,
Kiss
i
M.
Emotion
re
c
ognit
ion
from
fa
ci
a
l
expr
ession
base
d
on
fiduc
ial
point
s
det
ection
and
using
Neura
l
N
et
work,
Indon
esian
Journal
of
El
e
ct
rica
l
and
Computer
Enginee
ring
,
2018
;8
(1):
Acc
ept
ed
for
publi
c
at
ion
.
[8]
Fauzi
M
A,
Za
i
nal
Arifin
A,
Gos
ari
a
S
C.
Indone
sian
New
s
Cla
ss
ifi
c
at
ion
U
sing
Naïve
Ba
yes
and
Two
-
Phase
Feat
ure
Se
le
c
ti
o
n
Model.
Indon
esian
Journal
of
El
e
ct
rica
l
and
Computer
Engi
n
ee
ring
,
2017
,
8
(3):
Acc
ep
te
d
fo
r
publi
c
at
ion
.
[9]
Sahak
R,
Manso
r
W
,
Le
e
Y K,
Yass
in
A I
M,
Za
b
idi
A.
An
Or
thogonal
Least
Squa
re
Appro
ach
to
S
el
e
ct
F
eat
ures o
f
Infant
Cry
wit
h
Asphyx
ia
.
6th
In
te
rna
ti
ona
l
Col
lo
quium
on
Signal
Proce
ss
ing
and
Its
Applicati
ons
(CSP
A),
2010;
1
-
4.
[10]
Sahak
R,
Manso
r
W
,
Le
e
Y
K,
Yass
in
A
I
M,
Za
bidi
A.
Or
thog
onal
Least
Square
Based
Support
Vect
or
Mac
hin
e
for
the
C
lassif
i
cat
ion
o
f
Infan
t
Cry
wit
h
Asphyx
ia
.
Int
ern
a
ti
o
nal
Confer
ence
on
Biom
edi
c
al
Engi
ne
eri
ng
a
nd
Inform
at
ic
s,
201
0;
986
-
990.
[11]
Sahak
R,
Mansor
W
,
L
ee
Y
K,
Yass
in
A
I
M,
Z
abi
di
A.
Opt
imized
Support
Vect
or
Mac
hine
for
Classify
ing
In
fa
nt
Cries
wit
h
Asphyx
ia
us
ing
Or
thogonal
Least
Square
.
Inte
rn
ational
Confer
ence
on
Com
pute
r
Applic
ations
&
Industria
l
Elec
tr
onic
s,
2010
;
692
-
696.
[12]
Zhou
S,
W
u
L,
Yuan
X,
Ta
n
W
.
Paramete
rs
Selec
t
ion
of
SVM
f
or
Func
ti
on
App
roximati
on
Base
d
on
Diff
ere
n
ti
a
l
Ev
olution
.
Int
ern
at
ion
al
Conf
e
re
n
ce
on
Int
el
l
ige
nt
S
y
stems
and
Kn
owledge
Engi
n
e
eri
ng
,
2007;
7.
[13]
Vapnik
V.
An Overvi
ew
of
Sta
ti
s
ti
c
al
Learni
ng
T
heor
y
.
IEEE
Tr
ansacti
ons on
N
e
ural
Net
works.
1
999;
5:
988
-
99.
[14]
Chan
W
C,
Cheung
K
C,
Harri
s
C
J.
On
the
Mo
del
li
ng
of
Nonli
nea
r
D
y
n
amic
Sy
stem
Us
ing
Su
pport
Vec
tor
Ne
ura
l
Networks.
Eng
in
ee
ring
Appl
i
cat
i
ons of
Art
if
i
ci
al
Inte
lligen
ce,
vol
.
14,
105
-
113.
[15]
Orozc
o
J
&
C
A,
Re
y
es
-
Garc
i
a
C
A.
Dete
c
ti
ng
Pathol
ogie
s
fro
m
Infa
nt
Cr
y
A
ppl
y
ing
Scaled
Conjuga
te
Gr
adient
Neura
l
N
et
work
s.
Proceedi
ngs o
f
th
e
2003
Europ
ean
Symposium
on
Arti
f
icial
N
eu
ral Ne
twork
,
349
-
354
.
[16]
Re
y
es
-
Garc
i
a, O
F.
[onl
ine da
t
ab
ase
]
.
Available:
htt
p://ingenieria.
uatx
.
mx
/~orionfrg/c
ry/
.
2009
.
[17]
W
ang,
W
J,
Xu
Z
B,
Lu
W
Z.
D
et
ermina
ti
on
of
t
he
Sprea
d
Para
m
et
er
in
the
Gau
s
sian
Kerne
l
for
Cla
ss
ifi
c
at
ion
an
d
Regre
ss
ion.
N
eu
rocomputing.
20
03;
55:
643
-
663.
[18]
Zhu
G
Q,
Li
u
S
R
&
Yu
J
S.
Su
pport
Vec
tor
Ma
chi
ne
and
Its
Applicati
ons
to
Functi
on
Approxi
m
at
ion.
Journal
of
East
China
Univ
ersity
o
f
S
ci
en
ce
and
Technol
og
y.
2002;
5
:
555
-
55
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.