TELKOM
NIKA
, Vol.16, No
.4, August 20
18, pp. 1457
~14
6
7
ISSN: 1693-6
930,
accredited First Grade by
Kemenristekdikt
i
, Decree No: 21/
E/KPT/2018
DOI
:
10.12928/TELKOMNIKA.v16i4.7127
1457
Re
cei
v
ed Au
gust 26, 20
17
; Revi
sed Ma
y 8, 2018; Accepte
d
May 2
9
, 2018
Design of Electronic Nose System Using Gas
Chromatography Principle and Surface Acoustic
Wave S
e
nsor
Anifa
t
ul Fari
cha*
1
, Su
w
i
to
2
, M. Ri
v
a
i
3
, M. A. Nanda
4
, Djoko Purw
a
n
to
5
, Rizki
Anhar
R. P.
6
1,2,
3,5
Department of Electrical
Engi
neer
in
g, Inst
itut T
e
knolog
i
Sepul
uh
N
ope
mber (IT
S
),
601
11, Sura
ba
ya, Ind
o
n
e
sia
4
Departme
n
t of Mechan
ical a
n
d
Bio-s
y
stem E
ngi
neer
in
g, Bogor Agric
u
ltura
l
Universit
y
,
166
80, Bog
o
r, Indon
esi
a
6
Departme
n
t of Electrical En
gi
neer
ing, PGRI Adi
Bua
na U
n
i
v
ersit
y
, 6
024
5, Surab
a
y
a, Indo
nesi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: anifatulfar
ich
a
@gma
il.com
1
, muhamma
d_ri
v
ai@e
e.its.ac.i
d
3
Abstr
a
c
t
Most gases are odorl
e
ss, col
o
rless an
d als
o
ha
z
a
rd to be s
ense
d
by the h
u
man olf
a
ctory
system.
Henc
e, an electronic nos
e system
is re
quir
ed for the gas
classification
pr
ocess. This study presents the
desi
gn of elect
r
onic nos
e system usi
ng a c
o
mbi
natio
n of Gas Chro
mato
grap
hy Col
u
mn and a Surfa
c
e
Acoustic W
a
ve
(SAW
). T
he Gas Chro
matogr
aphy
Col
u
mn i
s
a techn
i
qu
e
base
d
on th
e c
o
mpo
und
partit
i
o
n
at a c
e
rtai
n te
mp
eratur
e. W
h
ereas,
th
e SA
W
sensor
w
o
rks bas
ed
on
th
e res
o
n
ant fre
que
ncy c
han
g
e
. I
n
this study, gas
samples
including
methanol, acetonitrile, and ben
z
ene
ar
e used for system
perfor
m
ance
me
asur
e
m
ent.
Each
g
a
s s
a
mp
le
ge
ner
ate
s
a s
pec
ific
a
c
oustic s
i
gn
al
data
in
the
for
m
of a
freq
ue
ncy
chan
ge r
e
cord
ed by
the SA
W
sensor. T
h
e
n
, the ac
oustic
sign
al
data
is
ana
ly
z
e
d
to o
b
tain t
he
acou
sti
c
features, i.e. the p
eak
ampl
i
t
ude, the
neg
a
t
ive slo
pe, the
posit
iv
e slo
p
e
,
and the
le
ng
th. T
he Sup
p
o
r
t
Vector Mach
in
e (SVM) meth
od us
ing
the
a
c
oustic featur
e
as its in
put p
a
ra
meters
are
app
lie
d to cl
as
sify
the gas s
a
mpl
e
. Radi
al B
a
si
s F
unction is
u
s
ed to b
u
il
d th
e opti
m
al hy
pe
rpla
ne
mo
del
w
h
ich dev
ide
d
into
tw
o processes
i.e., the traini
ng proc
ess a
n
d
the exte
rn
al
valid
atio
n pro
c
ess. Accordin
g to the resu
l
t
perfor
m
a
n
ce, the trai
nin
g
pro
c
ess has th
e a
ccuracy of
9
8
.7% a
nd the
ex
ternal v
a
li
datio
n proc
ess has
th
e
accuracy
of 93
.3%. Our elect
r
onic
nose sys
tem h
a
s t
he
a
v
erag
e sens
itiv
ity of
51.43
H
z
/mL to s
ense t
h
e
gas sa
mp
les
.
Ke
y
w
ords
: ac
oustic featur
es
, gas chro
matogra
phy (GC),
hyper
pla
ne, s
upp
ort vector
mac
h
i
ne (SVM
),
surface acoustic wave (SAW)
Copy
right
©
2018 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Gas i
s
a
matter which ha
s
an ind
epen
de
nt sha
pe b
u
t tends to
expa
nd ind
e
finitely. Most
gases a
r
e co
lorle
ss a
nd o
dorle
ss whi
c
h difficult to be se
nsed b
y
the naked
eye and hum
an
olfactory
syst
em. In additi
on, the ga
se
s which re
sult in toxic odo
r are forbidd
en to be
se
n
s
e
d
usin
g the h
u
man no
se
dire
ctly [1].
Therefor
e, a
n
electroni
c device is
required for
gas
recognitio
n
.
Over th
e la
st
de
cad
e
s, th
e ele
c
tr
o
n
ic
nose d
e
vice
has exten
s
ively bee
n u
s
e
d
in
indu
stry for t
he q
uality m
onitorin
g
syst
em, ga
s
i
den
tification, che
m
ical
analy
s
i
s
, et
c. Elect
r
onic
nose techn
o
l
ogy refers to the capa
bility of t
he human olfaction u
s
ing a sen
s
o
r
config
uratio
n and
a pattern recognition alg
o
rithm [2,3].
In the ele
c
tro
n
ic
no
se
syst
em, a
sen
s
o
r
ar
ray is requ
ired to
sen
s
e
the od
or. T
h
e Metal-
Oxide-Semi
condu
ctor (M
OS)
sen
s
o
r
such
a
s
T
agu
chi
Ga
s Se
nsor
(TGS
) b
e
comes the
type of
sen
s
o
r
wi
del
y used fo
r ga
s sen
s
ing
ap
plicatio
ns
du
e to its sim
p
li
city [4,5]. However, it ha
s the
low
sen
s
itivity whi
c
h g
ene
rally re
qui
re
s the hig
h
sa
mple of
co
ncentration, i.e.
, in the
ran
g
e
of
parts
pe
r mill
ion (p
pm) l
e
vel [6]. Another commo
n gas se
nsor
i
s
qu
artz crystal
microbala
n
ce
(QCM),
which is able to
sense the odor at very
low
concentrations, i.e., single
parts
per milli
on
(ppm
) or eve
n
parts p
e
r bi
llion (pp
b
) [7,8]. To obtain a sen
s
itive gas sen
s
ing, a
n
array of Q
C
M
sen
s
o
r
s is u
s
ed in
the
ele
c
tronic no
se
[9
,10]. Ho
weve
r, the m
a
in
problem
s of th
e
s
e
se
nsors
can
lead to
com
p
l
e
xity and inte
rfere
n
ce. The
r
efore, in
thi
s
study, we
co
nstru
c
ted
the
electroni
c no
se
system
whi
c
h ha
s th
e
simple
confi
guratio
n
wi
th
high
se
nsiti
v
ity and go
od repe
atabi
lity.
A Surfac
e Acous
tic
Wave
(SAW)
s
e
ns
or was
s
e
le
cte
d
as the det
ector. Prin
cip
a
lly, both SAW
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 145
8-1467
1458
and Q
C
M
work
ba
sed
o
n
the re
so
na
nt freque
nc
y
cha
nge. In
the analytical approximati
on,
Sauerbrey’
s
formula
p
r
esented i
n
Eq
uation
1 i
s
widely used to
dete
r
mine
the cha
nge
o
f
resona
nt freq
uen
cy affecte
d
by the abso
r
bed m
a
ss on
the crystal’
s
surfa
c
e
s
[11]:
2
0
2
F
F
m
A
(1)
w
h
er
e
∆
F
i
s
the chan
ge
o
f
re
son
ant fre
quen
cy
(Hz),
F
0
is the
re
sonant f
r
eq
ue
ncy
(Hz),
∆
m
is
the
ma
ss ch
a
nge (g),
A
is
t
he ac
tive cr
ys
tal ar
ea
(c
m
2
),
ρ
is the
c
r
ys
tal dens
i
ty
(g/c
m
3
), and
μ
is
the she
a
r mo
dulu
s
of the crystal (g/
c
m
s
2
).
In 2014, Hari
Agus Sujo
n
o
et al.,
appl
ied QCM se
nso
r
array
s
for vapo
r ide
n
tification
system which
has the re
so
nant fr
equ
en
cy of 20 MHz [9]. This typ
e
of sensor a
rray indu
ce
s the
compl
e
x configuration and the interference i
s
sues
. T
herefore, onl
y
a singl
e sensor of SAW
will
be u
s
ed
to
se
nse
the o
dor
whi
c
h h
a
s th
e re
so
nant
freque
ncy of
3
4
MHz. Th
e
SAW sen
s
or
use
d
in this expe
ri
ment op
erate
s
at
a hi
ghe
r re
son
ant fre
quen
cy. Hen
c
e it
affects
the in
cre
a
se
in
sen
s
itivity becau
se the chang
e
of the reso
nant fre
quen
cy (
∆
F
) to sense th
e mass ch
a
nge
absorb
ed
by
cry
s
tal a
r
ea f
o
r b
o
th
sen
s
ors a
r
e d
epe
ndent
on thei
r resonant f
r
eque
ncy
(
F
0
) as
explained by
Sauerbrey’
s
formul
a.
To a
c
hieve
a go
od
sel
e
ctivity in the
ele
c
troni
c n
o
se
sy
stem,
we
ap
plied
a G
a
s
Chromato
gra
phy (G
C) pri
n
ciple fo
r the
compou
nd a
n
a
l
ysis. Th
e G
C
is a te
ch
niqu
e ba
sed
on th
e
comp
oun
d p
a
r
tition at a
ce
rtain tem
pera
t
ure
whi
c
h
i
n
volves the
two ph
ases, i.e.
, the statio
na
r
y
pha
se a
nd
the mobil
e
gas pha
se.
The
statio
nary p
h
a
s
e
materi
al is located in
the
chromato
gra
phy column
a
s
the
pa
rtitio
n mate
rial,
where
a
s the
m
obile
ga
s p
h
a
s
e
co
nsi
s
ts o
f
a
sampl
e
carri
ed by d
r
y ai
r into the
pa
rtition colum
n
[12]. Each
sam
p
le ha
s different el
u
t
ion
strength because of the
polarity suitability with
the stationary phase mate
rial in the partition
colum
n
[13].
In 2016, th
e ele
c
troni
c
nose sy
st
em
by integrating G
C
an
d
TGS se
nsor
wa
s
con
d
u
c
ted by Radi et al. [1
2]. Howeve
r, the TG
S sen
s
or h
a
s a lo
w sen
s
itivity w
h
ich requi
re
s the
high am
ount
of co
ncentration for th
e
mea
s
ureme
n
t.
Therefo
r
e
,
in this
stud
y, a combi
n
a
t
ion
betwe
en G
C
and SAW sen
s
or in the el
e
c
troni
c
sy
ste
m
is expecte
d to overco
m
e
the issue
s
.
For the reco
gnition pa
rt in the elect
r
o
n
ic sy
stem, we u
s
ed a l
earni
ng alg
o
rithm of
Suppo
rt Vector Ma
chine
(SVM) for the classifi
catio
n
pro
c
e
ss. T
he SVM is propo
se
d as
an
effective tech
nique fo
r dat
a cla
s
sificatio
n
. It is derive
d
from
statisti
cal le
arni
ng t
heory int
r
odu
ced
by Vladimir
Vapnik
et al. [14]. Basica
lly, another
comp
etitive learni
ng al
gorithm is Artificial
Neu
r
al
Net
w
ork (A
NN), b
o
th of them
are i
n
cl
uded
in the
sup
e
rv
ised l
e
a
r
ning
cla
ssifie
r
[1
5].
Ho
wever, m
any re
sea
r
ch
ers
re
porte
d
that SVM
classifier ofte
n outpe
rform
s
than th
e
ANN
cla
ssifie
r
[16]
. The A
N
N cl
assifier
achie
v
es a
n
optim
a
l local
sol
u
tion, while the SVM cl
assifi
er
obtain
s
an
o
p
timal glob
al
solutio
n
. It is not su
rp
ri
si
n
g
that the
sol
u
tion of
the ANN cla
s
sifier
is
different for
each traini
ng
pro
c
e
ss
whi
c
h result
s in
a different
optimal soluti
on, whe
r
e
a
s
the
solutio
n
offered by SVM
cla
ssifie
r
i
s
same fo
r eve
r
y runni
ng.
Hence, it ge
ne
rates the
sa
me
optimal
solut
i
on [17
-
19].
The
co
ntents of this p
a
p
e
r
are
o
r
ga
n
i
zed
a
s
follo
ws:
se
ction
2
discu
s
ses th
e expe
riment
al de
sign
of
the el
ec
tro
n
ic
no
se
sy
stem, featu
r
e extra
c
tion,
and
elabo
rate the
SVM cla
ssifi
er te
chniq
ue.
Furthe
rmo
r
e
,
sectio
n 3 d
e
mon
s
trate
s
the re
sult
s a
n
d
verification a
nalysi
s
. Finall
y
, we pre
s
ent
our co
ncl
u
si
on in se
ction
4.
2. Researc
h
Method
2.1. The Experimental De
sign
In this study, the experime
n
tal desig
n of the el
ectro
n
ic no
se syste
m
includ
es fo
ur mai
n
parts, i.e., a gas
sampl
e
, a GC colum
n
, a det
ecto
r, and data a
nalysi
s
. Thre
e types of gas
sampl
e
s
were used in this study, i.e., methanol
, acetonitrile, and b
enzene. The
chromato
gra
phy
colum
n
co
nsi
s
ted of The
r
mon-300
0 an
d ShinCa
rb
o
n
as the statio
nary pha
se m
a
terial. The S
A
W
device
whi
c
h
has the resonant freq
ue
ncy of 34 M
H
z
wa
s used
as the dete
c
tor to re
co
rd the
freque
ncy ch
ange of the a
c
ou
stic
sign
al
generated by
each g
a
s
sa
mple.
2.2. The Experimental Pr
ocedur
e
Figure 1
p
r
e
s
ents th
e d
e
si
gn of
ou
r el
e
c
troni
c
no
se
system.
T
h
e
experim
ental setup
is
depi
cted in
F
i
gure
1a,
wh
erea
s th
e schematic layo
ut is
sho
w
n i
n
Figu
re 1
b
.
There a
r
e th
ree
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Des
i
gn of Elec
tronic
Nose Sy
s
t
em Us
ing
Gas
Chromatography
....
(Anifatul
Faric
h
a)
1459
main p
a
rt
s of
our
de
sign, i.
e. the
carrie
r
gas
, th
e
chro
matography
column, a
nd th
e SAW
se
nso
r
.
The d
r
y ai
r i
s
used to
carry the ga
s
sa
mples of A
(methanol
), B
(acetonitril
e),
and
C
(be
n
ze
ne).
The ch
rom
a
togra
phy col
u
mn is mad
e
o
f
the gla
ss m
a
terial with th
e diamete
r
of 3.2 mm and the
length of
1.6
m. It has the
operat
ing te
mperature
va
lue ab
ove 7
0
o
C. The
ch
a
m
ber as the
oven
whe
r
ein the
chrom
a
tograp
hy column lo
cated i
s
ma
d
e
of the aluminum with the
geomet
rical size
of
40 cmx
7
.
5
cmx
8
cm.
(a)
(b)
Figure 1. The
desig
n of ele
c
troni
c no
se
system: (a
) th
e experim
ent
al setup,
(b) the
schem
atic layout
Before sta
r
tin
g
the mea
s
urement, the chrom
a
tograp
hy colum
n
is
need
ed to be
clean
ed
for 30
minute
s
u
s
ing
the
carri
er
ga
s pu
she
d
by the
air p
u
mp. T
h
e amo
unts o
dor vol
u
me
s
of 20
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 145
8-1467
1460
mL ga
s samp
le are i
n
tro
d
u
c
ed i
n
to the i
n
jectio
n po
rt. Then the
ga
s sampl
e
is t
r
a
n
sp
orted
by the
carrie
r g
a
s i
n
to the chrom
a
togra
phy
col
u
mn
whi
c
h i
s
locate
d in
a
cham
be
r op
e
r
ated
und
er t
h
e
controlled t
e
mperature
of
80
0
C. Intera
ction
s
bet
we
en statio
na
ry pha
se m
a
te
rial a
nd the
gas
sampl
e
comp
ound gen
erat
e
a seri
es
of fraction
s wh
i
c
h is
co
nverte
d by the SA
W
sen
s
o
r
a
s
the
aco
u
sti
c
freq
uen
cy ch
ang
e data. Fu
rth
e
rmo
r
e, t
he
aco
u
sti
c
si
gn
al data a
r
e transmitted to
the
comp
uter through
the F
r
e
quen
cy
Coun
ter (FC) d
e
vice fo
r the
da
ta analy
s
is.
Acco
rdi
ng to
the
measurement
, the SAW se
nso
r
records
the freq
uen
cy
value of a
b
o
u
t 34 M
H
z
at
initial co
nditio
n
before inj
e
cti
ng the ga
s sa
mple. The fre
quen
cy ch
an
ge is de
scrib
ed as
()
.
re
f
f
ft
f
(2)
whe
r
e
∆
f
is the freq
uen
cy chan
ge,
f
ref
is the initial freque
ncy of 3
4
MHz, and
f
(
t
)
is
de
te
c
t
ed
freque
ncy aft
e
r inje
cting t
he ga
s
sam
p
le. Furth
e
rmore, to
coll
ect the a
c
o
u
s
tic
sign
al d
a
ta
prod
uced by
each ga
s sa
mple ne
ed
s 5
00 se
co
nd
s.
Figure 2 sho
w
s the
se
nso
r
re
spo
n
se to
th
e
acetonitrile.
Figure 2. The
sen
s
or
re
spo
n
se to the ga
s sa
mple of a
c
etonitril
e
2.3. Featur
e Extrac
tion of Acous
tic Signal Proces
sing
In this study, the acousti
c sign
al dat
a w
ould b
e
pro
c
e
s
sed to
obtain the aco
u
sti
c
feature
s
. Fig
u
re 3
de
scrib
e
s t
he
param
eters i
n
cl
ude
d to determin
e
the a
c
ou
stic features
usi
n
g
the thre
sh
old
of -100
Hz
value. The
four
acou
stic
feature
s
in
clu
d
ing th
e p
e
a
k
a
m
plitude
A
p
,
the neg
ative slop
e
S
(-)
, the positive
slop
e
S
(+)
, and th
e length
L
a
r
e
use
d
in thi
s
study determi
n
e
d
in Equation 3,
4,5 and 6 re
spectively:
Figure 3. Aco
u
stic featu
r
e
para
m
eters
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Des
i
gn of Elec
tronic
Nose Sy
s
t
em Us
ing
Gas
Chromatography
....
(Anifatul
Faric
h
a)
1461
pp
A
y
(3)
()
p
f
p
r
yy
S
tt
(4)
()
rp
rp
yy
S
tt
(5)
rf
L
tt
(6)
whe
r
e
t
f
is
the fall time,
y
f
is the fall amplitude,
t
p
is the pea
k time,
y
p
is the peak amplitud
e,
t
r
is
the rise time,
and
y
r
i
s
the rise amplitu
d
e
.
2.4. Support
Vector Ma
ch
ine (SVM) Cl
assifier
In this study, we used the
SVM classifier
to identify the gas type
whi
c
h in
clude
d four
aco
u
sti
c
feat
ure
s
a
s
the i
nput pa
ram
e
ters. T
he g
a
s types are di
vided into three cl
asse
s.
To
unde
rsta
nd t
he ba
sic
prin
ciple of SVM
classifie
r
,
the simpl
e
line
a
r sepa
rabl
e
ca
se is
sh
own in
Figure 4. In
the ori
g
inal
space, a lin
ea
r hype
rpla
ne
f
(
x
) is u
s
ed
to se
pa
rate
the data
poi
nts
according to
the sup
port
vector po
siti
on. Hen
c
e, t
he linea
r hy
perpl
ane
f
(
x
) grou
ps the
data
points i
n
to two
cla
s
ses,
i.e., class
+1
and
cla
s
s -1 whi
c
h
are
con
s
trained
by
f
(
x
)
+1 and
f
(
x
)
+1, re
sp
ectively [20]. Ho
wever, m
a
ny real
ca
se
s contai
n noi
se or
outlier d
a
ta point
s wh
ich
are
non
-lin
ea
rly sepa
rabl
e
[21]. Thu
s
, t
he mai
n
o
b
je
ctive of the
S
V
M cla
s
sifier
is to
obtain
the
optimal hype
rplane mo
del that can maxi
mize the ma
rgin (
M
) of the c
l
ass
e
s
[22].
Figure 4. The
Linear
sep
a
rable case usi
ng SVM method
The SVM classifier in
clu
des the kernel
s to opti
m
ize the hyperpl
ane mo
del for a
nonlin
ear
se
para
b
le
ca
se
. The kernel
s allo
w
tran
sformi
ng the
data poi
nts into a hi
gh
er
dimen
s
ion
a
l spa
c
e calle
d
feature spa
c
e
to
obt
ai
n
a linea
r hyp
e
rpla
ne alth
o
ugh o
c
casi
on
ally
result in a n
online
a
r hyp
e
rpla
ne in th
e origin
al sp
ace. In this
study, we u
s
ed Ra
dial B
a
si
s
Functio
n
(RBF) as the
ke
rnel fun
c
tion.
The
RBF kernel i
s
deriv
ed in Equati
on 7. Then the
hyperpl
ane m
odel
f
(
x
d
) is d
e
termin
ed in
Equation 8 [2
3,24].
2
(
,
)
e
xp(
)
id
i
d
Hx
x
x
x
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 145
8-1467
1462
1
()
(
,
)
n
di
i
i
d
i
f
yb
xH
x
x
(8)
w
h
er
e
x
d
is
th
e
da
ta
po
in
t,
α
is lagrange multiplier,
y
i
is memb
ership
of
the ga
s sam
p
le cla
ss,
γ
is gamm
a
,
x
i
is su
ppo
rt vector,
b
is intercept, and
i =
1, 2, 3, . . . ,
n
.
It is broadly
mentione
d that the SVM
cla
ssi
fie
r
usi
ng RBF kern
el requi
re
s the be
st
combi
nation
of two hype
rpa
r
amete
r
s, i.e., gamma (
γ
) a
nd
cost (
c
) to
build the optimal
hyperpl
ane m
odel. The ga
mma explain
s
how
signifi
c
ant the influence of ea
ch
data point in the
training
set. For
example,
a hig
h
e
r
val
ue of g
a
mma
lead
s the
ov
er-fitting
pro
b
l
em be
ca
use
it
tries to fit ex
actly ea
ch
d
a
t
a point
in th
e trai
ning
set. Whe
r
e
a
s, t
he
co
st i
s
u
s
ed to
control
the
trade
-off betwee
n
sm
ooth
deci
s
ion b
o
u
ndary a
nd cl
assifying the data point
s in
the training
set
c
o
rrec
tly [25,26].
In this study
, the identification g
a
s
syste
m usin
g
SVM classifier co
nsi
s
ted
of two
p
r
oc
es
se
s
,
i.e
.
th
e
tra
i
n
i
ng
pr
oc
es
s a
n
d
th
e e
x
te
r
nal va
lid
a
t
io
n pr
o
c
e
s
s
.
Th
e tr
a
i
n
i
ng
pr
oc
ess
wa
s used to
build the hyp
e
rpla
ne mo
d
e
l whi
c
h in
clu
ded the a
c
o
u
s
tic
signal
da
ta from the to
tal
numbe
rs of 150 ga
s sa
m
p
les. Since the external
v
a
lidation p
r
o
c
ess wa
s u
s
e
d
to assess the
SVM perfo
rm
ance. It used
the acou
stic sign
al dat
a
obtaine
d fro
m
the total n
u
mbe
r
s
of 30
gas
sampl
e
mea
s
urem
ents.
To de
scri
be t
he p
e
rfo
r
man
c
e
analy
s
is o
f
the
SVM
classifier, the
3x3 confu
s
io
n matrix
table
wa
s ap
plied fo
r thi
s
study, sho
w
n
in Ta
ble
1
[2
7]. In the
con
f
usion
matrix
table, the
act
ual
result is the data base
d
on
the obse
r
vation (re
ality).
It is con
s
iste
d of three cla
sses i.e., class A,
B, and C, wherea
s the pre
d
icted result is the ident
ification re
sult a
s
sesse
d
by the SVM classifier
whi
c
h also consi
s
ted of three
cla
s
ses.
The ca
ses a
r
e divided int
o
nine value
s
:
TA
,
FA1
,
FA
2
,
FB1
,
TB
,
FB
2
,
FC1
,
FC
2
, and
TC
.
F
i
nally, the
a
c
cura
cy
(
AC
) used
to a
s
se
ss the
SVM
perfo
rman
ce
to classify ga
s sa
mple is d
e
termin
ed in
Equation 9 [2
8].
Table 1. The
3x3 confu
s
io
n matrix
Actual
Result
A
B
C
Predicted
Result
A
TA FA1
FA2
B
FB1
TB
FB2
C
FC1
FC2
TC
1
00%
12
1
2
1
2
T
AC
TF
A
F
A
F
B
F
B
F
C
F
C
(9)
whe
r
e
TT
A
T
B
T
C
,
TA
is t
he co
rr
ect
l
y
cla
ssif
i
e
d
cla
ss A
,
TB
is th
e co
rrectly cl
assified
cla
ss
B,
TC
is t
he
cor
r
e
c
t
l
y
cla
s
sif
i
ed cl
as
s
C
,
FA1
is t
he cla
ss B
clas
s
i
f
i
ed int
o
cla
s
s A
,
FA2
is the
cla
s
s
C
cla
s
s
i
f
i
ed int
o
cl
as
s A
,
FB
1
i
s
t
he
cla
s
s A
cl
assified i
n
to
cla
s
s B,
FB2
is
t
he
cla
s
s
C
cla
ssif
i
e
d
int
o
cla
ss B
,
FC1
is t
he
cla
s
s
A
clas
sif
i
ed i
n
t
o
cla
s
s C,
FC2
i
s
t
he
cl
as
s B
cla
s
sif
i
ed
int
o
cla
ss
C.
3. Results a
nd Analy
s
is
In this
study,
we
de
sig
ned
the el
ect
r
oni
c n
o
s
e
sy
st
e
m
w
h
ich
con
c
er
ns
int
o
t
w
o t
e
rm
s,
i.e., having hi
gh
sen
s
itivity and
go
od
re
peatabilit
y. Fi
gure
5
sho
w
s the
me
asurement
re
sult
of
the sign
al acousti
c data re
corded by the
electro
n
ic n
o
s
e sy
stem. Base
d on the Equation 2, e
a
ch
gas
sa
mple
h
a
s the
spe
c
ific fre
que
ncy
chang
e (
∆
f
)
curve affe
cted
by ea
ch g
a
s sam
p
le'
s
ma
ss
absorb
ed by cry
s
tal are
a
of SAW sen
s
ors. In the
term of sen
s
itivity, according
to Figure 5
a
by
usin
g 5
mL
o
dor volume, t
he a
m
plitude
pea
k
dete
r
m
i
ned i
n
e
quati
on 3
from
g
a
s
sam
p
le A,
B,
and C a
r
e -20
00 Hz, -100
0 Hz, an
d -85
Hz, respectiv
e
ly.
Furthe
rmo
r
e,
by a
pplying
20 m
L
o
dor v
o
lu
me
s, the
gas sampl
e
A, B, and
C reach the
amplitude
pe
ak
of -2
800
Hz,
-17
50
Hz, and
-85
0
Hz, re
sp
ectivel
y
. It means t
hat the el
ect
r
onic
nose system
coul
d sen
s
e t
he odo
r from methanol,
a
c
etonitrile, and
benze
ne wit
h
the sen
s
itivity
of
53.3 Hz/m
L,
50 Hz/m
L,
and 51 Hz/
m
L,
re
spe
c
ti
vely. Hen
c
e th
e average
se
nsitivity is 51
.4
3
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Des
i
gn of Elec
tronic
Nose Sy
s
t
em Us
ing
Gas
Chromatography
....
(Anifatul
Faric
h
a)
1463
Hz/mL.
Rivai
et al. [29,30] have desig
ned the
ele
c
t
r
oni
c no
se u
s
ing g
a
s
ch
romatog
r
ap
hy and
QCM
sen
s
or.
According
t
o
the
re
sult
perfo
rman
ce
,
it ha
s th
e a
pproxim
ate
sensitivity of
6.5
Hz/mL
to
sen
s
e th
e od
or
o
f
ethanol. An
other
re
sea
r
ch pap
er sh
o
w
s that the f
r
eque
ncy
cha
nge
curve resulte
d
by specifi
c
odor, frag
ra
nce,
and g
a
s in the range
of hundred
s herzt [31]. Our
system
offers highe
r o
pera
t
ing re
so
nant
frequ
en
cy. H
ence, it is a
b
l
e
to ge
nerate
the spe
c
ific
or
distin
ctive
a
c
ousti
c sign
al comin
g
from each
g
a
s
sa
mple in
the
ra
nge
of thou
sa
nds he
rtz,
sh
own
in Figure 5.
(a)
(b)
Figure 5. The
acou
stic
sign
al data affect
ed by
the odo
r volume of (a
) 5 mL, and (b) 20 mL
In the term of repeatability
,
each ga
s samp
le compo
und produ
ce
s a spe
c
ific
aco
u
sti
c
sign
al data
becau
se it has p
a
rticula
r
intera
ct
ion
s
with station
a
ry pha
se
material in t
h
e
chromato
gra
phy colu
mn.
The mai
n
differen
c
e
of ea
ch g
a
s
sam
p
le cu
rve is p
o
inted out by
its
pea
k amplitu
de
A
p
. For
example, in Fi
g
u
re
5a, the hi
ghe
st pea
k a
m
plitude i
s
a
c
hieve
d
by g
a
s
sampl
e
C. T
he ga
s
sam
p
le A ha
s the
lowe
st value
of amplitud
e
pea
k an
d th
ese t
r
en
ds
a
r
e
repe
ated i
n
Figure 5
b
when
usi
ng th
e am
ount
s o
dor volume
s of 20
mL.
Furthe
rmo
r
e,
the
aco
u
sti
c
sign
al data
gen
erated by e
a
ch
ga
s
samp
l
e
are
proce
s
se
d to obtai
n th
e four a
c
ou
st
ic
feature
s
.
Figure 6
pre
s
ents th
e di
stribution of fo
u
r
a
c
ou
stic fe
ature
s
i.e., th
e pea
k
ampli
t
ude
A
p
,
the ne
gative
slop
e
S
(-)
, the pos
i
tive s
l
ope
S
(+)
, and th
e len
g
th
L
which
refe
r to
Equation
3, 4
,
5,
and 6
re
sp
ect
i
vely. The dist
ribution
re
sult
given
by Fig
u
re
6 in
clud
e
s
50
mea
s
u
r
e
m
ents
usi
ng
20
mL odo
r volu
mes. The di
st
ribution
re
sult
s of the pea
k amplitude
A
p
, the negative slope
S
(-)
, the
positive slo
p
e
S
(+)
, and the length
L
ca
n be
seen
in
Figure 6
a
, 6b
, 6c,
and
6d
seq
uentially.
As
sho
w
n in Fig
u
re 6a, 6b,
6c, and 6
d
, we can co
n
c
lude that the
distributio
n of the acou
stic
feature
s
from
gas sample
A, B, and C a
r
e cla
s
sified i
n
to the non-li
nearly sepa
ra
ble ca
se.
(a)
(b)
Figure 6
.
The
acou
stic feat
ure
s
of gas
sample: (a
)
the pea
k amplit
ude, (b
) the n
egative slo
p
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 145
8-1467
1464
(c
)
(d)
Figure 6
.
The
acou
stic feat
ure
s
of gas
sample:
(c) the
positive slo
p
e
, (d) the len
g
th
The
re
cognit
i
on alg
o
rith
m of SVM
cla
ssifie
r
wa
s u
s
e
d
to
solve the n
o
n
-line
a
rly
sep
a
ra
ble
ca
se. In thi
s
study, the total
numb
e
rs of
150 g
a
s sam
p
les
we
re
used for traini
n
g
pro
c
e
ss to b
u
ild the optim
al hyperplane
model
u
s
ing
RBF ke
rnel.
The RBF
ke
rnel re
quires t
h
e
best
com
b
ina
t
ion of two hy
perp
a
ramete
rs of
gam
ma
and
co
st. In t
he hyp
e
rp
ara
m
eters tuni
ng
,
we set the interval of gam
ma starte
d from 2
-15
to 2
2
, whe
r
ea
s the
co
st has th
e lowe
r limit of 2
-15
and the u
ppe
r limit of 0.25. Figur
e
7 pre
s
ent
s the det
ail distri
butio
n
of the hyperp
a
ram
e
ter tuni
ng
perfo
rman
ce.
The dark blue colo
r (left
)
and the
da
rk green
colo
r (rig
h
t) have
the lowest a
nd
highe
st accu
racy, respe
c
tively. According to t
he tra
i
ning p
r
o
c
e
s
s, the best combinatio
ns
o
f
gamma an
d co
st are 1 an
d 0.2, resp
ectively, wh
ich have the accura
cy of 98.7%. It means the
total numbe
rs of 150 o
b
s
ervatio
n
s
u
s
ed fo
r the
training
pro
c
ess: 148 d
a
ta are
co
rre
ctly
cla
ssifie
d
and
the others ar
e inco
rr
ect
l
y
cla
ssif
i
e
d
.
Figure 7. The
hyperpa
ram
e
ters tu
ning p
e
rform
a
n
c
e
The extern
al
validation
wa
s used f
o
r the
t
e
st
in
g t
o
as
se
ss
t
he S
V
M
cla
ssif
i
e
r
perfo
rman
ce.
It included the acou
stic signal data fro
m
30 gas
sa
mples
(10 d
a
t
a for each g
a
s
sampl
e
). Tabl
e 2 sho
w
s th
e confu
s
ion
matrix result
from extern
al validation pro
c
e
ss. The result
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Des
i
gn of Elec
tronic
Nose Sy
s
t
em Us
ing
Gas
Chromatography
....
(Anifatul
Faric
h
a)
1465
of case
TA
,
FA1
,
FA2
,
FB1
,
TB
,
FB2
,
FC1
,
FC2
, and
TC
are 9, 1, 0, 1, 9, 0, 0, 0, and 10,
respe
c
tively. The SVM cla
ssifie
r
ha
s th
e accu
ra
cy of 93.3% to identify the gas sample, wh
ich
mean
s from t
he total n
u
m
bers of
30 o
b
se
rvation
s
u
s
ed
for th
e e
x
ternal valid
a
t
ion process:
28
cor
r
e
c
t
l
y
cla
s
sif
i
ed an
d 2 i
n
co
rr
ect
l
y
cla
ssif
i
e
d
.
The
s
e re
sult
s in
di
cat
e
t
hat
t
h
e
S
V
M
clas
sif
i
er
can
be
categ
o
rized a
s
the
robu
st alg
o
ri
thm whi
c
h
ca
n be integ
r
at
ed with th
e e
l
ectro
n
ic
no
se
sy
st
em.
Table 2. The
3x3 confu
s
io
n matrix re
sul
t
Actual
Result
A
B
C
Predicted
Result
A 9
1
0
B 1
9
0
C 0
0
10
4. Conclusio
n
The de
sig
n
o
f
the electro
n
i
c
no
se
syste
m
with goo
d repeata
b
ility and high
se
nsi
t
ivity by
integratin
g th
e ga
s chrom
a
togra
phy p
r
i
n
cipl
e
an
d Surface Acou
stic Wave
(SA
W)
se
nsor
was
su
cc
es
sf
ully
demon
st
r
a
t
e
d
.
Three
ga
s
sampl
e
s
we
re used fo
r th
e mea
s
u
r
em
ent process,
i.e.,
methanol,
acetonitrile, a
n
d
ben
ze
ne. I
n
the p
r
ev
iou
s
research, t
he ele
c
troni
c nose u
s
in
g
gas
chromato
gra
phy and QCM sen
s
o
r
ha
s only the ap
pr
oximate se
nsitivity of 6.5 Hz/mL to sense
the odo
r. An
other
re
sea
r
ch also sho
w
s that t
he freq
uen
cy ch
ang
e cu
rve resul
t
ed by sp
ecif
ic
odor, fra
g
ran
c
e, and
ga
s only in the ra
nge of hu
ndr
eds h
e
rzt. Based o
n
the re
sult analy
s
is,
our
electroni
c no
se sy
stem h
a
s t
he avera
ge se
nsitivity of 51.43
Hz/mL and al
so offers hig
her
operating
re
sonant frequ
e
n
cy. Hen
c
e, i
t
is a
b
le to
ge
nerate
mo
re
spe
c
ific or
di
stinctive
acou
stic
sign
al comi
ng
from each ga
s sa
mple.
The re
peata
b
ility performa
n
ce i
s
shown by the
dist
inctive acou
stic sign
al cu
rve from
each ga
s sa
mple du
e to the specifi
c
in
teractio
ns
bet
wee
n
the od
or an
d the m
a
terial lo
cate
d in
the ch
romato
grap
hy colu
m
n
. In this stud
y, Support Vector Ma
chi
ne
usin
g Ra
dial
Basis F
u
n
c
tio
n
wa
s a
pplie
d t
o
recogni
ze
the o
d
o
r
. The
four
acou
stic
feature
s
obtai
ned f
r
om
a
c
o
u
stic si
gnal
d
a
ta
were u
s
e
d
fo
r input
param
eters in
the
cl
assifier,
i.e., t
he am
plitude
pea
k, the
neg
ative slo
pe, t
h
e
positive sl
ope
, and the leng
th. The cla
s
si
fication u
s
ing
Support Ve
ct
or Ma
chin
e was divide
d int
o
two p
r
o
c
e
s
ses, i.e., the
training
p
r
ocess a
nd
th
e
external
val
i
dation
pro
c
e
ss.
The t
r
ain
i
ng
pro
c
e
ss
and
the external
validation p
r
ocess
h
a
ve
the high a
c
curacy of 98
.7% and 93.
3%,
respe
c
tively. These results indi
cate th
at the cl
a
ssif
i
er can b
e
a
pplied to the
electroni
c n
o
se
system. Final
ly, to
achieve the comprehensive re
sult performance,
the fu
ture work
will concern
on a deep i
n
vestigatio
n of sen
s
itivity and rep
eat
a
b
ility perform
ances at the
electro
n
ic n
o
se
system by va
rying the tem
peratu
r
e of th
e cham
be
r, the pre
s
su
re
of the air pu
mp, and diffe
ren
t
type kern
el function
s in the
Support Vect
or Ma
chin
e al
gorithm.
Ackn
o
w
l
e
dg
ment
This
re
sea
r
ch
wa
s carried
out with fina
n
c
ial
ai
d supp
ort from th
e
Ministry of
Rese
arch,
Tech
nolo
g
y and Hig
her Educatio
n
of
The
Repu
b
lic of
Indo
nesi
a
(Keme
n
riste
k
di
kti RI).
The autho
rs also
woul
d like to ackno
w
le
dge LPPM ITS for the intense di
scu
ssi
o
n
.
Referen
ces
[1]
NS Nat
h
, I Bh
attachar
ya, A
G
T
u
ck, DI Schlip
al
ius, PR
Ebert. Mech
an
isms of P
hosp
h
in
e T
o
xic
i
t
y
.
Journ
a
l of T
o
xi
colo
gy
. 201
1; 2011: 1-9.
[2]
K Prasa
d
, E
Basse
y, P S
a
llis.
El
ectron
ic
nos
e for
ga
s sens
ing
a
p
p
licati
ons
. Co
nferenc
e on
Optoelectro
n
ic
and Micr
oel
ectronic Mater
i
als
& Devices. Per
t
h. 2014: 11
3-1
16.
[3]
W
Lu, W
Yu, C Gan, Q Liu, J Li.
Appl
icati
on
of Electron
ic N
o
se T
e
ch
nol
og
y in the D
e
tecti
on of W
h
e
a
t
Quality
. Interna
t
iona
l Co
nferen
ce on Intel
lig
en
t
T
r
ansportatio
n
, Big Data
an
d Smart Cit
y
.
Halo
ng B
a
y.
201
5: 133-
136.
[4]
Radi, M Riv
ai,
MH Purnom
o. Stud
y
on El
ect
r
onic-N
o
se-B
a
s
ed Qua
lit
y
M
o
nitori
ng S
y
ste
m
for Coffee
U
n
de
r R
o
a
s
ting
.
Journa
l of Ci
rcuits, Systems
,
and Co
mp
ute
r
s
. 2016; 25(
10
): 16501
16.1-
1
650
11
6.19.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 145
8-1467
1466
[5]
AA Alkand
ari, S Moein. Impl
em
entati
on of
Monitori
ng S
y
s
t
em for
Air Qualit
y
usi
ng Ra
spberr
y
PI:
Exp
e
rim
ental
Stud
y
.
Indo
ne
sian J
our
nal
o
f
Electrical
En
gin
eeri
ng
an
d
Co
mp
uter Sc
ienc
e
. 20
18
;
10(1): 43-
49.
[6
]
T
Itoh
, T
Akama
t
su
,
A Tsuru
t
a
,
W
Shin.
Se
lecti
v
e
Detectio
n of
T
a
rge
t
Vo
latile Organic
Compound
s in
Contamina
t
ed
Humid Air U
s
i
ng a Sens
or Array
w
i
th Principal Compone
nt Analy
s
is.
Senso
r
s
. 2017
; 17(7
)
: 1-14
.
[7]
K Z
hang, G F
an, R Hu, G Li.
Enha
nce
d
Dib
ut
y
l
Ph
t
hal
ate Sensi
ng Perfor
mance of a Qu
artz Cr
y
s
ta
l
Microba
la
nce
Coate
d
w
i
th A
u
-Dec
orated Z
n
O Porous Mi
crosph
eres.
Sensors.
20
15;
15: 211
53-
211
68.
[8]
M Procek, A
Stolarcz
yk, T
Pusteln
y
, E
Maciak. A
Stud
y
of a QC
M Sens
or Ba
sed
on T
i
O2
Nan
o
structures
for the D
e
tection
of
NO2 a
n
d
E
x
pl
osives
V
apo
urs in
Air.
Sensors.
2
0
1
5
;
15: 95
63-
958
1.
[9]
HA Su
jon
o
, M
Rivai. V
a
p
o
r Id
entificati
on
S
y
s
t
em
Usin
g Qu
a
r
tz Reso
nator
Sensor
Arra
y
a
nd S
u
p
port
Vector Machine.
ARPN Journ
a
l of Engi
ne
eri
ng an
d App
lie
d
Sciences
. 20
1
4
; 9(12): 24
26-
243
0.
[10]
M Rivai, T
a
sripan. F
u
e
l
Qu
a
lificati
on
Usin
g Quartz Sen
s
ors.
ARPN J
ourn
a
l of En
gi
neer
ing
an
d
Appl
ied Sci
enc
es
. 2015; 1
0
(1
6): 6737-
67
43.
[11]
KMM Kabir, SJ Ippolit
o, YM Sabri, CJ Harris
on, G Matthe
w
s, SK Bhargav
a.
A comp
ariso
n
of Surfac
e
Acoustic W
a
ve
(SAW
) and Quart
z
Crystal
Microba
la
nce (
Q
CM) based s
ensors for p
o
rtabl
e, onl
i
n
e
m
e
rc
ury vapour sensing
. CHE
MECA. Perth, West
ern Australia. 2
0
1
4
.
[12]
Radi, S Cipto
had
ijo
yo, W
S
Litan
a
n
da, M Rivai
, MH Pur
nomo. Electro
n
ic nos
e base
d
on partiti
o
n
colum
n
i
n
tegr
a
t
ed
w
i
t
h
g
a
s s
e
nsor for
fruit i
d
entificati
on
an
d
classific
a
tio
n
.
Co
mp
uters a
n
d
El
ectronics
in Agric
u
lture
.
201
6; 121: 42
9
-
435.
[13]
G Greco,
T
Letzel. Main Inte
ractions a
nd I
n
fl
ue
nces of th
e Chrom
a
togr
aph
ic Param
e
ters in HILIC
Separ
atio
ns.
J Chro
mat
ogr Sc
i
. 2013; 5
1
(7): 684-
693.
[14] V
Vapnik.
T
he
Nature of Stati
s
tical Le
arni
ng
T
heory.
Ne
w
Y
o
rk: Spring
er-V
erla
g. 199
5.
[15]
T
Saha
y
,
A
Aggar
w
a
l, A
Bansa
l
, M C
h
andr
a.
SVM a
nd A
NN: A c
o
mpar
ative
ev
alu
a
tion
. 1st
Internatio
na
l
C
onfere
n
ce on Ne
xt
G
ener
ati
on C
o
mputi
n
g
T
e
chnolo
g
ies
(NGCT
)
.
Dehradu
n. 20
15
;
960-
964.
[16]
K Pasup
a
, W
Sunh
em.
A co
mp
ariso
n
b
e
tw
een s
hal
low
a
nd d
e
e
p
arch
itecture cl
assifie
r
s on s
m
a
l
l
dataset
. Information T
e
chn
o
lo
g
y
and El
ectrical En
g
i
n
eeri
ng (ICIT
E
E) 2016 8th
Internation
a
l
Confer
ence. Y
o
g
y
ak
arta, Ind
ones
ia. 20
16: 1-6.
[17]
HF
Kuo, A
F
a
richa. Artific
i
al
Ne
ural
N
e
t
w
o
r
k for D
i
ffraction B
a
se
d Ove
r
la
y M
easur
em
ent.
In IEEE
Access
. 2016;
4: 7479-
74
86.
[18]
HR Kh
oei, M Z
o
lfag
hari. N
e
w
Mode
l Refer
e
n
c
e Ada
p
tive S
ystem Spee
d O
b
server for F
i
el
d-Oriente
d
Contro
l Inducti
on Motor
Driv
es Usin
g N
e
u
r
al Net
w
o
r
ks.
Bull
etin of E
l
ectrical E
n
g
i
n
eeri
ng a
n
d
Informatics (BEEI)
. 2016; 5(1
)
: 25-36.
[19]
SS Salve, SP
Narote.
Iris re
cogn
ition
usi
n
g SVM an
d A
NN.
Internati
o
nal
Confer
enc
e on W
i
r
e
les
s
Commun
i
cati
o
n
s, Signa
l Proc
essin
g
an
d Net
w
o
r
kin
g
(W
iSPNET
). Chenn
ai
. 2016: 47
4-47
8.
[20]
SHSA Ub
aid
ill
ah, R Sa
ll
ehu
ddi
n, NA Ali.
Canc
er Detect
ion
Usin
g Ariti
f
ical N
eura
l
N
e
t
w
o
r
k a
n
d
Supp
ort Vector Machin
e:
A Compar
ative Stud
y.
Jurna
l
T
e
knol
ogi (Sci
en
ces & Engin
e
e
r
ing)
.
2013;
65(1): 73-
81.
[21]
W
F
F
uentes,
DH Bal
b
u
ena,
JCR Quiñ
on
ez
, DO Ugald
e
, F
F
G
Navarro, O Sergi
y
enk
o,
MR Lóp
ez,
FNM Rico, V
M
Kartashov.
Outlier
mini
n
g
of
a vis
i
on
sens
ing
d
a
ta
base
for SV
M regress
i
o
n
improve
m
ent
.
IECON 2015 -
41st A
nnual
C
onfer
ence of
the IEEE Industrial E
l
ectronics S
o
ciet
y
.
Yokoh
a
ma. 20
15: 000
20
8-00
021
3.
[22]
MKA Ghan
i,
DH Suta
nto.
Im
provin
g C
l
a
ssificatio
n
Acc
u
rac
y
for
No
n-Commu
nic
a
b
l
e D
i
se
ase
Predicti
on Mo
del B
a
sed
on
Supp
ort Vecto
r
Machin
e.
Jur
nal T
e
k
nol
ogi
(Scienc
es & E
ngi
neer
in
g)
.
201
5; 77(1
8
): 29-36.
[23]
X Z
h
an
g, Z
Yan. Ap
plic
ation
of su
p
port vector
machi
ne to r
e
lia
bi
lit
y a
nal
ysis of e
ngi
n
e
s
y
stems.
T
E
LK
OMNIKA Indon
esia
n Journ
a
l o
f
Electrical Eng
i
ne
erin
g
. 201
3; 11(7): 355
2-3
560.
[24]
MA Nand
a, KB Seminar, D
Nan
d
ika, A M
add
u.
A Comp
ariso
n
Stud
y o
f
Kernel F
unct
i
ons i
n
th
e
Supp
ort Vector
Machin
e an
d Its Applicati
on for T
e
rmite Detection.
Inform
ation
. 201
8;
9
(1)
:
1-14.
[25]
W
ang
R, H
u
Z
,
Che
n
L, Xi
on
g
J.
A
n
Ap
pro
a
ch
on
F
eatur
e Se
lecti
on
of
Casca
ded
Su
p
port Vect
or
Machi
nes w
i
th Particle S
w
arm Optimi
zation Al
gorith
m
. Advanc
es
in Electronic
Engin
eeri
ng,
Commun
i
cati
o
n
and M
ana
ge
ment Vol2: Pro
c
eed
ings
of
20
11 Intern
ation
a
l
Confer
enc
e o
n
Electron
ic
Engi
neer
in
g, Communic
a
tio
n
and Ma
na
gem
ent (EEC
M 20
11). Beij
ing, C
h
in
a. 201
1: 2: 689-
700.
[26]
H Dua
n
, X L
i
u.
Low
er C li
mits in s
upp
o
r
t vector mac
h
in
es w
i
th rad
i
al b
a
sis funct
i
on ker
nels
.
Internatio
na
l
S
y
mp
osi
u
m o
n
Informatio
n
T
e
chnolo
g
ies
in Me
dici
ne
and
Educ
ati
o
n
.
Hok
o
d
a
te,
Hokkai
do. 2
0
1
2
: 768-7
71.
[27]
NF
Shamsu
di
n, H Bas
i
ron,
Z
Saa
y
a, A
F
NA Rahm
an
, MH Zakaria
,
N Hassim.
Sentime
n
t
Classific
a
tio
n
of Unstructur
e
d
D
a
ta Us
ing
Le
xic
a
l B
a
se
d
T
e
chniques.
J
u
rnal Teknologi (Sciences
&
Engi
neer
in
g
). 2015; 77(
18): 11
3-12
0.
[28]
KM
T
i
ng. Conf
usio
n Matri
x
. Enc
y
c
l
o
ped
ia of
Machi
ne Le
arn
i
ng. Boston: Sp
ring
er US. 201
0.
Evaluation Warning : The document was created with Spire.PDF for Python.