Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4133
~
4147
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
4133
-
41
47
4133
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Intellig
en
t Sensing
U
sin
g Metal
Oxid
e Se
micondu
ctor
B
as
ed
-
on
Su
pp
or
t Ve
ctor
Machin
e for
Odor Cl
assif
icat
i
on
Nya
yu
L
at
i
fah Hu
s
ni
1
, S
iti
Nu
rm
aini
2
, I
r
sy
adi
Yani
3
,
Ade Sil
via
4
1
,4
El
e
ct
ri
ca
l
Dep
art
m
ent
,
Polit
ek
nik
Nege
r
i
Sriwi
jay
a
,
Indon
esia
2
Inte
lligen
t
S
y
s
t
em Re
sea
r
ch
Gr
oup,
Facu
lty
of C
om
pute
r
Scie
n
ce
,
Univer
si
ta
s S
riwij
a
y
a
,
Indone
sia
3
Mec
hanica
l
En
gine
er
ing
Dep
artm
ent
,
Fa
cul
t
y
of
Engi
n
ee
ring
,
U
nive
rsit
as
Sriwij
a
y
a
,
Indone
si
a
Art
ic
le
In
f
o
ABSTR
A
C
T
Art
ic
le
history:
Re
cei
ved
N
ov
21
, 201
7
Re
vised
Jun
1
0
, 201
8
Accepte
d
Aug
4
, 2
01
8
Cla
ss
if
y
ing
odor
in
rea
l
expe
rim
ent
pre
sents
some
cha
llenges,
es
pec
i
al
l
y
th
e
unce
rt
ai
nt
y
of
t
he
odor
con
ce
n
tra
ti
on
and
disp
ersion
that
c
an
le
ad
to
a
diffi
cu
lty
in
obt
ai
ning
an
a
cc
ur
at
e
da
ta
sets.
In
thi
s
study
,
to
e
nhanc
e
th
e
ac
cur
acy
,
d
at
as
e
ts
arr
ange
m
ent
base
d
on
MO
S
sensors
par
am
et
ers
using
SV
M
appr
oac
h
f
or
odor
c
la
ss
ifi
c
at
ion
is
propose
d.
The
s
ensors
are
te
sted
t
o
det
ermine
the
sensors
'
ti
m
e
r
esponse,
sensors
'
pea
k
dur
at
io
n,
sensors
'
sensiti
vity
,
and
sensors
'
stabi
lit
y
when
applie
d
to
the
var
iou
s
sourc
es
at
diffe
ren
t
ran
ge
.
Thre
e
source
s
were
used
in
expe
riment
al
t
est,
name
l
y
:
et
hano
l,
m
e
tha
n
ol,
and
ac
e
tone.
The
g
as
sensors
cha
ra
cteri
sti
cs
ar
e
an
aly
z
ed
in
open
sam
pli
ng
m
et
hod
to
see
the
sensors
'
per
form
anc
e
in
real
situa
ti
on
.
The
se
per
form
a
nce
s
are
co
nside
red
as
the
base
of
choosing
the
positi
on
in
col
l
ec
t
ing
the
da
ta
sets.
Th
e
sensors
in
d
y
namic
expe
riment
h
ave
ave
rag
e
of
pre
ci
sion
of
9
3.
8
-
97.
0
%
,
the
ac
cur
acy
93.
3
-
96.
7%,
and
the
reca
l
l
93.
3
-
96.
7
%
.
Th
i
s
val
ues
ind
icates
tha
t
the
col
l
ect
ed
dataset
s
ca
n
support
the
SV
M
in
impr
oving
the
in
telli
gen
t
sensing
when
conduc
ti
ng
odor
cl
assifi
ca
t
ion
wo
rk.
Ke
yw
or
d:
In
te
ll
igent
sens
ing
MOS se
nsor
Odor cl
assifi
ca
ti
on
Suppor
t
v
ect
or
m
achine
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
:
Sit
i Nurm
ai
ni
,
In
te
ll
igent
Syst
e
m
Resea
rch
G
rou
p,
F
acult
y o
f
Com
pu
te
r
Sci
ence,
Un
i
ver
sit
as
Sr
i
wij
ay
a
,
J
al
an
R
ay
a Pr
a
bu
m
ulih
Inder
al
ay
a,
Palem
ban
g, S
outh
Su
m
at
era,
I
ndonesi
a
.
Em
a
il
:
siti_n
ur
m
ai
ni@u
nsri.a
c.id
1.
INTROD
U
CTION
Sensors
as
on
e
of
the
im
po
rtant
dev
ic
es
in
el
ect
r
on
ic
s
yst
e
m
s
hav
e
r
ecei
ved
pe
ople
’s
at
te
ntio
n.
En
or
m
ou
s cap
abili
ty
o
f
sensors
as inputs o
f
i
nfor
m
at
ion
p
r
oc
essing has b
e
en
de
velo
ped
within the
el
ect
ronics
industry
[1]
-
[4]
.
Lig
ht
se
ns
ors
a
nd
gas
se
nsors
are
tw
o
of
e
xisti
ng
se
nsors
that
ha
ve
been
m
os
t
act
ively
stud
ie
d.
Gas
sens
or
s
a
re
def
i
ned
as
a
dev
ic
e
t
hat
can
s
ubs
ti
tute
hu
m
an
olfacti
on.
A
ir
qual
ity
m
on
it
or
ing
[5]
-
[
7]
,
gas
le
a
ka
ge
local
i
zat
ion
[8]
,
[9]
,
forest
m
on
it
or
ing
[10
]
,
[11]
,
m
i
li
ta
ry
[12]
,
[
13]
,
are
so
m
e
exam
ples
of
ga
s
sensor
ap
plica
ti
on
s.
Gas
sen
s
or
s
c
onve
rt
the
env
ir
onm
ent
ph
ysi
cal
phen
om
eno
n
(in
this
case
po
ll
uta
nt
con
c
entrati
on)
into
el
ect
rical
s
ign
al
s.
They
re
present
pa
rt
of
the
interface
be
tween
the
ph
ysi
cal
world
an
d
t
he
w
or
l
d
of
el
ec
tric
al
dev
ic
es
,
su
c
h
as
com
pu
te
rs
[
6]
.
T
he
y
can
be
i
nteg
rated
i
n
t
wo
m
od
es,
m
ob
il
e
and
sta
ti
c,
howe
ver,
t
he
la
te
r
has
li
m
it
a
ti
on
in
co
ve
rin
g
a
la
r
ge
ar
ea.
Stat
ic
sens
or
s
can
be
br
ought
by
worker t
o detec
t t
he
surr
ound
ing
;
unfortu
nat
el
y, app
ly
in
g
t
hem
to
dangerou
s
and
poi
son
ou
s
area
can
h
a
rm
the
worker
.
Mo
bil
e
robo
ts
util
iz
at
ion
can
co
pe
the
pro
blem
s.
They
giv
e
m
or
e
adv
a
ntages
c
on
ce
r
ning
safe
ty
and
secur
it
y
[14]
,
s
uch
as:
i)
bein
g
able
to
be
de
plo
ye
d
t
o
the
d
ang
e
r
ou
s
a
rea
without
bei
ng
afr
ai
d
of
t
o
be
kill
ed;
ii
)
al
so
bei
ng
able
to
be
us
e
d
to
m
on
it
or
t
he
en
vir
onm
ental
ai
r
qu
al
it
y
con
ti
nu
ously
without
bei
ng
ti
red
or
giv
in
g
i
nacc
urat
e inform
at
ion
; and ii
i) offe
ri
ng ch
ea
pe
r
c
ost
in
trai
ni
ng the
m
[15]
.
Me
ta
l
Ox
ide
S
e
m
ic
on
duct
or
(
MOS)
sens
ors
are
one
of
the
m
os
t
inv
est
igate
d
groups
of
ga
s
sens
ors
.
Ther
e
we
re
a
l
ot
of
re
searc
he
s
us
e
d
them
as
their
se
nsors
[
16
]
-
[
20
]
.
T
he
r
easo
ns
of
c
hoosi
ng
t
his
se
ns
or
ar
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4133
-
4147
4134
du
e t
o: i)
lo
w
c
os
t an
d
fle
xib
il
it
y associa
te
d
to their
pro
duct
ion
; i
i) sim
plicity
o
f
their
us
e;
ii
i) larg
e num
ber
of
detect
able
gase
s/po
ssi
ble
ap
pl
ic
at
ion
[
21
]
;
a
nd
i
v)
rob
us
t,
l
igh
t
wei
gh
t,
fa
st
respo
ns
e
[
22]
.
Althou
gh
t
his
ty
pe
of
se
nsors
s
howed
th
ei
r
ad
va
ntages
,
howe
ve
r,
the
de
pl
oym
ent
of
gas
se
ns
ors
in
real
env
i
ronm
ent
faces
so
m
e
oth
e
r
pro
blem
s
,
su
c
h
as
the
phen
om
eno
n
of
patches
a
nd
e
ddie
s
that
occ
ur
du
e
to
th
e
tur
bu
le
nt
ai
rf
lo
w
of
th
e
wind
[
23
]
,
[24]
.
The
c
on
c
entr
at
ion
dec
rease
s
wh
e
n
m
olecules
m
ov
e
awa
y
fr
om
the
so
urce,
he
nce
m
olecular
diffusi
on
an
d
tur
bule
nt
diffus
ion
proce
s
ses
hav
e
the
m
ai
n
ro
le
i
n
determ
i
ning
t
he
s
ha
pe
of
plu
m
e.
Mol
ecular
diffusi
on
ca
use
s
rand
om
m
o
ti
on
of
th
e
m
olecules
to
m
ove
gr
a
dual
ly
ap
art,
w
hile
tur
bule
nt
diffusio
n
te
ars
apar
t
the
cl
oud
of
m
olecules
ph
ysi
cal
ly
by
ai
r
tur
bu
le
nce
[
25
]
.
M
olecular
diffusi
on
e
ff
e
ct
on
t
he
pl
ume
sh
a
pe
can
be
ne
glect
ed
[24]
du
e
to
it
s
sm
a
ll
eff
ect
on
t
he
plu
m
e
sh
a
pe.
It
is
c
ontra
dicti
ve
with
t
urbu
le
nt
dif
fu
si
on
that
can
change
th
e
sh
a
pe
of
the
pl
um
e.
The
turbule
nt
diffusio
n
that
do
m
inate
s
th
e
disp
ersi
on
of
od
or
m
olecules
beco
m
es
cru
ci
al
par
am
et
er
in
odor
res
ear
ch
.
So
m
e
research
ers
have
in
vestigat
ed
the
odor
char
act
e
risti
cs
in
ai
rf
l
ow
e
nv
iro
nm
ent
[26]
-
[28]
an
d
in
t
urb
ulence
e
nvir
on
m
ent
[29],
[
30]
.
Most
of
t
he
gas
sens
or
s'
w
orks
wer
e
do
ne
in
si
m
ulati
on
[
31]
-
[33]
.
S
om
e
research
e
rs
w
ho
had
a
great
desire
t
o
work
in
rea
l
exp
e
rim
ent
us
ua
ll
y
set
so
m
e
lim
i
ta
ti
on
s
on
their
e
nvir
on
m
ent
(in
co
nd
it
io
ned
ex
pe
rim
ent
and
s
cal
able
env
i
ronm
ent)
[
24
]
,
[34]
a
nd g
a
ve
m
or
e c
on
si
de
rati
on on a
n
i
ntell
igent se
ns
i
ng as t
heir
im
pr
ovem
ent.
In
te
ll
igent
sen
s
ing
in
odor
cl
a
ssific
at
ion
ref
e
rs
to
a
n
intel
li
ge
nt
a
ppro
ac
h
that
s
upplied
to
the
s
ens
ors
to
inc
rease
th
ei
r
abili
ti
es
in
detect
ing
an
d
cl
assify
ing
s
ub
sta
nces
.
I
n
a
ncient
resear
ch,
t
he
c
onve
ntion
a
l
te
chn
iq
ue
was
app
li
ed
t
o
the
s
ens
or
s
to
e
nh
a
nce
sens
ors
ca
pab
il
it
ie
s.
The
char
act
e
risti
cs
of
this
re
searc
h
wer
e
base
d
on the sen
sor supe
rio
riti
es
platfo
rm
.
T
he
sens
ors w
e
r
e equ
ip
pe
d
wit
h
va
rio
us
pow
e
rs,
i.e. a
ve
ry sp
eci
al
sens
or
with
a
pro
per
a
bili
ty
that
was
pro
du
ced
on
ly
to
de
te
ct
a
par
ti
cula
r
s
ub
sta
nce.
T
his
ap
proac
h
ga
ve
a
si
m
plici
t
y con
c
ern
i
ng
data pro
cessi
ng
a
nd com
pu
ta
ti
on
d
ue t
o
the senso
rs had
a
good sel
ect
ivit
y;
h
ow
e
ver
t
he
sens
or
s
im
pr
ovem
ents
need
e
d
m
or
e
tim
e
a
nd
m
on
ey
[35]
.
Ar
ti
fici
al
In
te
ll
igent
(A
I
)
off
ered
a
go
od
s
ol
ution
to
overc
om
e
t
he
prob
le
m
s
.
It
can
b
uild
i
deal
pe
rfor
m
ance
[
36]
an
d
gav
e
a
hi
gh
a
ccur
acy
[37]
.
Neural
Netw
ork
(
NN)
[38]
-
[
41]
,
k
-
ne
arest
[31]
,
P
rincipal
com
pone
nt
analy
sis
(PC
A)
,
cl
us
te
rin
g
analy
ses
(C
A)
[
31]
,
and
S
upport
V
ect
or
Ma
c
hie
(
SV
M)
[
42
]
-
[
45]
are
so
m
e
al
go
rithm
s
that
usual
ly
be
us
ed
a
s
the
odor
cl
as
sifie
rs.
Sar
kar
hav
e
pro
posed
S
N
N
(S
pi
king
Ne
ur
al
Network)
to
cl
assify
blac
k
te
a
odor
[
38]
,
[39]
.
The
propose
d
m
et
ho
d
was
s
uccess
fu
l
to
c
la
ssify
the
te
a
odor
data;
howe
ve
r,
it
s
le
arn
i
ng
al
gorithm
need
ed
a
bigg
e
r
arch
it
ect
ure
s
o
that
it
m
ade
the
c
om
pu
ta
ti
on
beco
m
e
slowe
r,
easi
ly
to
be
trap
ped
in
l
oc
al
m
ini
m
a,
coul
d
not
run
well
if
the
data
was
inc
om
ple
te
,
and
ha
d
ba
d
predici
ti
on
[
46
]
.
Y
u
[
47]
ha
d
c
om
par
ed
th
ree
al
go
rithm
s
(k
-
NN,
BP
N,
a
nd
S
VM)
a
nd
fou
nd
t
hat
SV
M
produce
d
m
or
e
accu
r
at
e
cl
assifi
cat
i
on.
S
VM
ha
d
so
m
e
su
pe
rio
riti
es
[48]
,
su
c
h
as:
i)
SV
M
is
f
or
m
ulate
d
as
a
qu
a
drat
ic
pr
og
ram
ming
pro
blem
w
it
h
no
local
m
i
nim
a,
ii
)
The
arc
hitec
ture
of
the
m
od
el
(num
ber
of
s
uppor
t
vecto
rs)
is
autom
at
ic
ally
sel
ect
ed
du
rin
g
t
he
op
ti
m
iz
ation
proces
s,
ii
i)
SVM
can
be
ap
plied
f
or
m
ulti
-
class
pro
blem
s
,
i
v)
T
hey
can
ge
ner
al
iz
e
well
f
ro
m
a
restr
ic
te
d
am
ount
of
trai
ning
data.
T
his
is
pa
rtic
ularly
inter
est
ing
i
n
t
he
ga
s
sen
sor
a
rea
w
her
e
it
is
ve
ry
costly
and tim
e
-
consu
m
ing
to o
btain
a large
r
el
ia
ble
and
represe
ntati
ve
set
of e
xa
m
ples.
SV
M
ap
proac
h
f
or
od
or
cl
as
sific
at
ion
ha
ve
been
a
pp
li
ed
i
n
m
any
wo
rks
[44],[
49]
,[50
]
.
Ma
rco
[
44]
inv
est
igate
d
th
e
cl
assifi
cat
ion
in
c
on
ti
nuous
m
on
it
or
in
g
for
obta
ini
ng
an
acc
ur
at
e
da
ta
set
s.
It
was
reall
y
diff
ic
ult
to
get
the
ste
a
dy
sta
te
of
the
sens
ors
du
e
to
the
s
ens
or
s
we
re
not
co
ntin
uous
l
y
expose
d
t
o
t
he
odor
so
urces
f
or
e
noug
h
ti
m
e.
Anothe
r
researc
h
cond
ucted
by
Am
y
Lou
tfi
[49]
w
ho
us
e
d
t
he
tran
sie
nt
respon
s
e
of
the
sens
or
s
.
I
n
that
researc
h,
the
sens
or
s
w
ere
assum
ed
to
be
in
a
sta
te
of
tra
ns
it
ion
and
th
e
input
to
the
cl
ass
ific
at
ion
a
lgorit
hm
was
ta
ken
from
the
com
par
ison
be
tween
a
baseli
ne
an
d
ste
a
dy
sta
te
;
ho
we
ve
r,
it
sti
l
l
was
not
easy
to
get
go
od
dataset
s.
Os
un
a
[
50
]
sta
te
d
that
in
the
com
pr
es
sion
of
the
data,
the
sens
or
tr
ansien
t
respo
ns
e
did
not
im
pr
ov
e
the
accu
racy
pre
di
ct
ion
but
im
pair
the
acc
urac
y.
It
c
ou
l
d
be
con
cl
ud
e
d
t
hat
the
database
was
over
fitt
in
g
[50]
and
the
c
on
cl
us
io
n
ab
out
the
best
database
s
for
the
odor
c
la
ssific
at
ion
w
as
sti
ll
vague.
In
t
his
resea
rc
h
S
VM
ap
pro
ach
ba
sed
on
the
MOS
perf
or
m
ance
was
pro
po
se
d
to
overc
om
e
the
pro
blem
of
col
le
ct
ing
the
dat
abases
f
or
a
tr
ai
nin
g
dataset
s.
T
he
c
har
act
e
r
ist
ic
of
t
he
M
OS
sens
ors'
re
sp
onse
tim
e,
sensors'
pea
k
re
spo
nse
durati
on,
the
se
ns
it
ivit
y
an
d
the
sta
bi
li
t
y
response
of
the
s
ens
or
we
re
inv
est
igate
d.
T
he
m
a
in
go
al
of
this
researc
h
is
to
analy
ze
t
he
act
ive
cov
e
rag
e
area
of
ga
s
sensors.
By
hav
i
ng
the
se
ns
or'
s
pe
rfor
m
ance
data
,
a
plan
ning
of
fin
ding
a
n
a
ppr
opriat
e
locat
io
n
for
ta
ki
ng
da
ta
set
s
of
S
VM
cou
l
d
be
set
up.
T
o
the
aut
hor'
s
know
le
dg
e
,
ther
e
is
no
one
discuss
e
d
ab
out
the
perform
ance
of
gas
se
ns
ors
for
the
odor
cl
assifi
cat
ion
.
It
is
hoped
that
the
MOS
sens
or
s'
perfor
m
ance
can
hel
p
to
fin
d
a
go
od
dataset
s
f
or
SV
M.
This
pa
per
c
onsist
s
of
5
pa
rts,
i.e.:
pa
rt
1
desc
ribes
t
he
backgro
und
of
the
resea
rc
h,
par
t
2
i
ntrodu
ce
s
par
am
et
ers
in
odor
cl
assi
ficat
ion
ta
sks,
pa
rt
3
desc
ribes
t
he
exp
e
rim
ental
set
up,
pa
rt
4
di
sp
la
ys
the
res
ul
t
and
the d
isc
us
si
on
of the e
xperim
ent, a
nd p
a
rt
5 i
s the c
on
cl
us
i
on of t
he
re
sea
rch.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
In
te
ll
igent
Se
nsi
ng U
si
ng Met
al O
xi
de Semic
onduct
or
Ba
se
d
-
on
.
.
..
(
Ny
ay
u
La
ti
fa
h Husni)
4135
2.
PARA
METE
RS
I
N OD
OR
CLASSIFI
C
A
TION T
AS
K
S
2.1.
Ga
s
Sens
ors P
erfo
rm
an
ce
s
The
MO
S
sens
or
s
a
re
su
it
a
ble
for
rec
ognizing
ei
the
r
re
du
ci
ng
or
ox
i
dizing
gases
by
or
cond
uctiv
e
m
easur
em
ents
[51]
.
Senso
r'
s
perform
ance
is
on
e
of
the
im
p
or
ta
nt
p
arts
of
the
sens
or
a
pp
l
ic
at
ion
.
By
knowi
ng
the
pe
rfo
rm
ance,
it
can
be
e
asi
ly
app
li
ed
t
o
a
n
a
ppr
opria
te
app
li
cat
io
n
with
ce
rtai
n
li
m
it
a
ti
on
a
nd
s
cop
e
.
Accor
ding
to
V.
E.
B
oc
henk
ov
[
52]
,
S
om
e
of
par
am
et
ers
sho
uld
be
pai
d
at
te
ntio
n
i
n
order
to
cha
ra
ct
erize
sens
or
pe
rfo
r
m
ance,
nam
ely:
sensiti
vity
,
sel
ect
ivit
y,
st
abili
ty
,
detect
i
on
li
m
it
,
dyna
m
ic
ran
ge,
li
near
it
y,
reso
l
ution,
res
pons
e
tim
e,
re
cov
e
ry
ti
m
e,
work
i
ng
te
m
per
at
ur
e
,
hyste
r
esi
s,
an
d
li
fe
-
cy
cl
e.
Sens
it
iv
it
y
is
a
change
of
m
ea
su
re
d
sig
nal
pe
r
analy
te
concentrat
io
n
unit
[52]
.
Xin
Z
ho
u
in
[
53
]
an
al
yz
ed
the
se
ns
it
ivit
y
of
the g
as
sens
or
base
d
on Z
nF
e
2
O
4
sphere
s and Z
nF
e
2
O
4
na
nopa
rtic
le
s.
The
g
as se
nsor'
s r
e
sp
onse t
o
the
30 ppm
and
100
pp
m
acet
on
e
wer
e
r
e
corde
d.
T
he
ga
s
sens
or
'
s
respon
s
e
to
the
ace
ton
e
var
ie
d
with
the
cha
nge
of
the
tem
per
at
ur
e
.
Gas
se
nsor
ga
ve
lo
w
res
ponse
to
t
he
acet
on
e
at
lo
w
te
m
per
at
ur
e
(
20
0
o
C)
due
to
a
cet
on
e
m
olecules
cann
ot
ef
fecti
vely
react
with
the
su
r
face
abs
orb
ed
oxyge
n
sp
e
ci
es.
The
res
pons
es
of
por
ous
ZnF
e
2
O4 sphe
res we
re
good at
high
er tem
per
at
ur
e
at
o
pe
rati
ng te
m
per
at
ur
e 20
0
o
C an
d 237.5
o
C.
Sele
ct
ivit
y
ref
ers
to
cha
racteri
sti
cs
that
determ
ine
wh
et
her
a
sen
sor
can
res
pond
to
a
gro
up
of
analy
te
s
sel
ect
i
vely
or
eve
n
to
a
sing
le
analy
te
sp
eci
fical
ly
.
Sele
ct
ivit
y
is
on
e
of
esse
ntial
par
am
et
ers
in
odor
identific
at
ion
[
54
]
.
Sele
ct
ivit
y
will
be
easy
if
the
odor
s
to
be
id
entifi
e
d
ar
e
re
al
ly
diff
ere
nt.
It
is
con
tra
dicti
ve
wh
e
n
the
od
ors
are
qu
ie
t
sim
il
ar
as
in
pa
pe
r
[31]
.
T
he
sel
e
ct
ivit
y
of
the
s
ens
or
s
will
be
bette
r
by
a
ddit
ion
al
m
et
ho
ds
or
te
c
hn
i
qu
e
s,
s
uc
h
as
the
i
ntegr
at
i
on
of
PC
A,
L
DA,
NN,
S
V
M,
et
c.
Sele
ct
ivit
y
has
a
ti
gh
t
relat
ion
with
the
sta
bili
ty
that
ref
e
rs
t
o
the
a
bili
ty
of
a
se
ns
or
t
o
prov
i
de
reprod
uc
ible
res
ults
f
or
a
certai
n
pe
ri
od
of
tim
e.
It
include
s
retai
ning
th
e
sensiti
vity
,
sel
ect
ivit
y,
respon
s
e,
a
nd
rec
overy
ti
m
e.
On
e
of
t
he
ga
s
s
e
nsor'
s
sta
bili
ty
was
cond
ucted
by
Z
hen W
e
n
[55]
. Th
ey
te
ste
d
se
ver
al
g
as
se
nso
r'
s
par
am
et
ers
in
their
resea
rc
h,
s
uc
h
as:
sens
it
ivit
y,
detect
ion
lim
i
t,
w
orkin
g
te
m
per
at
ur
e,
res
po
ns
e/
rec
ov
e
ry
ki
netic
s,
sel
ect
ivit
y
,
an
d
sta
bil
it
y
of
the
sens
or.
I
n
that
researc
h,
they
got
that
Rho
m
bic
Co
3
O
4
nanoro
d
(
NR)
arr
ay
-
base
d
ga
s
sens
or
had
a
good
sta
bili
ty
o
ver the
3
m
on
ths
test
.
Wor
king
te
m
p
eratur
e
is
t
he
tem
per
at
ure
tha
t
cor
re
spo
nd
s
t
o
m
axi
m
u
m
se
ns
it
ivit
y.
Po
r
ous
Z
nF
e
2
O
4
sp
he
res
we
re
good
at
operati
ng
te
m
per
at
ur
e
200
o
C
a
nd
23
7.5
o
C;
it
m
eans
that
the
wor
king
te
m
per
at
ur
e
range
of
t
hat
sens
or
was
200
o
C
unt
il
23
7.5
o
C.
It
is
due
to
the
se
ns
i
ti
vity
of
the
sens
or
bec
ome
qu
it
e
high
in
that
range.
Anothe
r
exam
ple
of
work
i
ng
tem
per
at
ur
e
ca
n
be
seen
from
the
researc
h
of
Zhe
n
W
e
n
[
55]
that
analy
zed s
om
e
p
aram
et
ers
in
gas
se
nsor'
s p
e
rfor
m
ances.
Et
hanol a
naly
te
w
as c
ho
se
n
as
the g
as
source
in their
researc
h.
T
he
op
ti
m
al
wo
rk
i
ng
te
m
per
at
ur
e
of
Rh
om
bic
C
o
3
O
4
na
nor
od
arr
ay
base
d
ga
s
sens
or
f
or
m
axim
u
m
sensiti
vity
was
at
160
o
C.
T
he
respo
ns
e
of
the
sens
or
inc
reased
with
t
he
op
e
rati
ng
te
m
per
at
ur
e
an
d
then
decr
ease
d
with
a
f
ur
t
her
rise
of
the
operati
ng
te
m
per
at
ure.
T
he
ph
e
no
m
ena
we
re
e
xpla
ined
by
Zh
en
Wen
us
in
g
t
he
a
ds
orptio
n
a
nd
des
orptio
n
kin
et
ic
s
on
t
he
s
urfac
e
of
the
sem
ico
nd
uc
ti
ng
m
etal
ox
i
des.
When
th
e
work
i
ng te
m
per
at
ur
e
was
sm
a
ll
(
below
150
o
C), th
e chem
ical act
ivati
on
was al
so
sm
all;
t
her
e
fore the
re
sp
on
s
e
was
al
so
sm
all
(b
el
ow
10)
.
The
ads
orbe
d
gas
m
olecules
escaped
befo
re
their
reacti
on
if
the
op
e
r
at
ing
tem
per
at
ur
e
w
as
rea
ll
y
hi
gh
(abo
ve
200
o
C)
.
T
he
res
ponse
w
ou
l
d
al
s
o
decr
ease
(
belo
w
10).
The
w
orki
ng
tem
per
at
ur
e
was bet
ween 1
50
o
C u
ntil
200
o
C, w
it
h t
he hig
hest se
ns
it
ivit
y at
160
o
C.
The othe
r
MO
S p
e
rfor
m
ances w
ere
desc
rib
ed
as foll
ows:
Det
ec
tion
li
mi
t
is t
he
lo
west
con
ce
n
trat
io
n
of
the
a
naly
te
t
hat
can
be
dete
ct
ed
by
the
sen
so
r
unde
r
giv
e
n
co
nd
it
io
ns
,
pa
rtic
ularly
at
a
giv
e
n
tem
per
a
ture.
On
e
of
t
he
sel
ect
ive
detect
io
n
researc
hes
w
as
offe
red
by
Qian
qian
[
56
]
who
obser
ve
d
Zn
O
gas
se
nso
r
to
the
acet
on
e
s
ource
s.
T
he
resea
rc
h
ob
ta
ine
d
t
hat
det
ect
ion
lim
i
t
for
the
acet
one
was
0.25
ppm
.
Dy
n
amic
range
ref
e
rs
to
the
a
naly
te
con
c
ent
rati
on
ra
ng
e
be
tween
t
he
detect
ion
li
m
i
t
and
the
hi
gh
e
st
lim
i
ti
ng
c
on
ce
nt
rati
on
.
Li
nearity
re
fe
rs
to
the
relat
ive
de
viati
on
of
an
e
xperim
e
ntall
y
determ
i
ned
cal
i
br
at
io
n
gr
a
ph
f
ro
m
an
ideal
strai
gh
t
li
ne.
Reso
lu
tio
n
m
eans
the
lo
we
st
con
ce
ntrati
on
dif
fer
e
nce
t
hat
can
be
dis
ti
ng
uis
he
d
by
sens
or
.
Resp
on
se
tim
e
is
the
tim
e
req
ui
red
f
or
se
ns
or
to
res
pond
to
a
ste
p
con
cent
rati
on
c
ha
ng
e
f
r
om
zero
to
a
certai
n
co
nce
nt
rati
on
value
.
Qian
qian
[56]
analy
zed
the
re
sp
onse
ti
m
e
of
the
Z
nO
se
nso
r
to
th
e
ac
et
on
e.
The
respo
ns
e
tim
e
of
the
se
nsor
was
as
sho
rt
as
3
s.
Rec
over
y
time
is
the
tim
e
i
t
ta
kes
for
the
sens
or
sign
al
to
return
to
it
s
ini
ti
al
value
after
a
ste
p
co
ncen
t
r
at
ion
cha
nge
f
r
om
a
certai
n
value
to
zer
o.
H
ys
tere
sis
m
ean
s
the
m
axi
m
u
m
diff
eren
ce
in
o
ut
put
wh
e
n
the
va
lue
is
app
r
oa
ched
with
an
i
ncr
easi
ng
a
nd
a
decr
easi
ng
a
naly
te
con
ce
ntrati
on
r
ang
e
.
Li
fe c
ycl
e
is t
he peri
od
of tim
e o
ver w
hich
t
he
se
nsor
w
il
l co
ntin
uousl
y op
e
rate.
2.2.
Dispersio
n
M
od
el
The
gas
can
m
ov
e
easi
ly
fr
om
on
e
place
to
an
oth
e
r
pla
ce
du
e
t
o
the
wind
or
the
di
ff
e
ren
ce
of
con
ce
ntrati
on
in
on
e
place
.
The
longer
the
distance
of
the
gas
fro
m
the
so
ur
ce
is,
the
s
m
a
ller
the
con
ce
ntrati
on
will
be.
I
n
ot
he
r
w
ord,
t
he
c
on
ce
ntrati
on
ne
ar
the
s
ource
will
be
highe
r
than
the
c
once
ntrati
on
of
t
he
gas
fa
r
f
ro
m
the
source
.
The
tur
bule
nc
e
an
d
the
dif
fusion
t
hat
can
c
ause
the
gas
to
m
ov
e
are
in
flu
ence
d
by
the
en
vir
onm
ental
con
diti
on,
especial
ly
the
wi
nd
c
har
ac
te
risti
cs.
The
m
ov
e
m
ent
of
the
gas
from
i
ts
so
urce
to
the
area
ar
ound
it
will
pro
du
ce
a
c
oncent
ra
ti
on
pa
tt
ern
or
al
way
s
cal
le
d
as
th
e
pl
um
e
disp
e
rsion.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4133
-
4147
4136
The
c
hange
of
the
co
ncen
t
rati
on
patte
rn
will
con
ti
nuously
happe
n
in
acc
orda
nce
with
t
he
occ
urre
nce
of
th
e
wind that m
oves in t
he
sam
e d
irect
io
n or
i
n t
he
di
ff
e
ren
t
di
recti
on.
The
plu
m
e
disp
ersi
on
can
be
m
od
el
ed
us
i
ng
m
at
he
m
at
ic
equ
at
io
n.
I
n
ge
ner
al
,
the
m
od
el
s
ca
n
be
div
ide
d
into
tw
o
form
s,
i.e.,
i)
Ba
sic
Mod
el
;
and
ii
)
Dif
f
us
ion
Mo
del.
I
n
ba
sic
m
od
el
,
the
gas
on
ly
m
ov
es
du
e
to
the
ai
r
flow
sp
ee
d.
Th
e
con
ce
ntrati
on
us
ing
this
m
od
el
is
consi
der
e
d
to
be
con
sta
nt.
T
hus,
the
con
ce
ntrati
on
in
one
place
in
on
e
ti
m
e
is
the
sa
m
e
with
the
con
ce
ntrati
on
in
ano
t
her
place
in
diff
e
ren
t
tim
e.
The basic
m
odel
u
ses t
he
e
quat
ion
as
r
e
pr
es
ented
i
n
e
quat
ion (
1)
:
∂
C
∂
t
+
u
∂
C
∂
x
=
0
(1)
wh
e
re
is
gas
con
ce
ntrati
on
(kg/m
3
)
,
de
note
s
the
ai
r
fl
ow
sp
ee
d
(m
/s),
sh
ows
th
e
ti
m
e
(s)
,
an
d
i
s
coor
din
at
e
(m
).
F
or
the
dif
fu
si
on
m
od
el
s
,
there
wer
e
a
lot
of
re
sea
rch
e
rs
offe
re
d
so
m
e
so
luti
on
s.
In
pap
e
r
[
57
]
,
G
aussian
m
od
el
s
an
d
Far
rel'
s
filam
entou
s
m
od
el
we
re
dis
cusse
d.
The
se
m
od
el
s
assu
m
ed
the
m
et
eor
ologica
l
conditi
on
a
nd
plu
m
e
e
m
issio
n
a
re
sta
ti
onary.
The
gas
c
on
ce
ntrati
on
c
an
be
sta
te
d
i
n
tw
o
form
s,
i.e. 2 an
d 3 dim
ension
s
that can
be
see
n
in
the
f
ollow
i
ng equati
ons:
(
,
,
)
=
4
√
(
−
(
−
0
−
)
2
4
−
(
−
0
−
)
2
4
)
(2)
(
,
,
,
)
=
(
4
)
3
/
2
√
(
−
(
−
0
−
)
2
4
−
(
−
0
−
)
2
4
−
(
−
0
−
)
2
4
)
(
3)
3.
MA
TE
RIA
L
S
AND
METH
OD
3.1.
Experim
en
ta
l
Setu
p
Thr
ee
Me
ta
l
Ox
i
des
ga
s
se
ns
ors
(TGS
2620,
T
GS
26
02,
an
d
T
GS
2600)
wer
e
us
e
d
in
this
exp
e
rim
ent
.
These
sens
ors
a
re
c
om
bin
ed
t
og
et
her
in
ord
er
to
m
ake
an
arr
ay
se
nsor
.
T
he
se
ns
i
ng
el
em
ent
of
the
gas
se
nsor
s
is
m
ade
of
a
m
et
al
ox
ide
se
m
ic
on
duct
or
la
ye
r
form
ed
on
the
al
um
ina
su
bs
trat
e
of
a
se
ns
in
g
chip
to
gethe
r
with
an
integ
r
at
ed
heater.
T
he
res
pons
e
pa
tt
ern
of
the
se
ns
ors'
arr
ay
w
as
analy
zed
in
the
real
env
i
ronm
ent
i
n
or
der
to
see
the
cha
racteri
s
ti
cs
and
perfor
m
ance
of
eac
h
sens
or
(T
he
use
of
m
or
e
tha
n
on
e
sens
or
i
n
odor
cl
assifi
cat
io
n
is
ve
ry
im
po
rtant
du
e
to
t
he
ou
t
pu
t
of
on
e
se
ns
or
ca
n
ref
e
r
to
dif
fer
e
nt
con
ce
ntrati
on
of v
a
rio
us
a
nal
yt
es
)
[
58
]
.
In
t
his
researc
h,
t
he
data
of
the
se
ns
or
perform
ance
will
be
a
naly
zed
i
n
fou
r
cat
e
go
ries,
na
m
ely:
sens
or
s'
tim
e
respon
se
,
se
nso
rs'
peak
du
rati
on,
se
nsors'
se
ns
it
ivit
y,
an
d
s
ens
or
s'
sta
bili
ty
.
The
ex
pe
r
im
ent
wa
s
done
in
a
roo
m
of
4
m
x
10
m
.
The
resp
onses
of
the
sen
so
rs
wer
e
m
ea
su
re
d
us
i
ng
2
scenari
os
.
F
or
check
i
ng
the
sens
ors'
resp
onse
t
im
e
an
d
sens
ors'
peak
durati
on,
t
he
so
urce
wa
s
ex
po
s
ur
e
to
the
e
nv
i
ronm
ent
fo
r
20
s
.
The
distances
of
the
sens
ors
to
t
he
s
ource
s
we
re
var
ie
d,
i.e.
by
m
ov
in
g
th
e
r
obots
60
cm
from
it
s
init
ia
l
po
sit
io
n.
The
s
ource
was
s
witc
hed
i
nto
on
a
nd
off
c
onditi
on
to
se
e
t
he
se
ns
it
ivit
y
of
the
sens
or
s
.
T
he
s
i
m
ple
blo
c
k
dia
gr
am
o
f
t
he
ga
s sens
or
de
velo
pm
ent is sh
own
i
n
F
igure 1
(a). T
he
te
m
per
at
ur
e
of
t
he
se
ns
or
w
il
l no
t
be
disc
us
se
d
in
detai
l
du
e
t
o
t
he
data
got
fro
m
the
first
ste
p
in
this
re
searc
h
w
ould
be
s
upplied
as
the
da
ta
set
s
of
od
or
cl
assif
ic
at
ion
wh
ic
h
was
done
in
t
he
real
en
vir
onm
ent
(the
e
nvir
on
m
ental
te
m
per
at
ur
e
was
ta
ke
n
arou
nd 28
0
C
unti
l 31
0
C).
(a)
(b)
Figure
1. (a
)
T
he bloc
k diag
r
a
m
o
f gas
sens
or
s
d
e
velo
pm
e
nt, (b
)
T
he gas
so
urces
setu
p
S
o
u
r
c
e
s
G
a
s
s
e
n
s
o
r
s
A
r
r
a
y
M
i
c
r
o
c
o
n
t
r
o
l
l
e
r
O
u
t
p
u
t
D
a
t
a
W
i
c
k
T
u
b
e
I
n
l
e
t
f
o
r
f
r
e
s
h
a
i
r
A
i
r
C
o
m
p
r
e
s
s
o
r
L
i
q
u
i
d
c
o
m
p
o
u
n
d
s
(
e
t
h
a
n
o
l
o
r
m
e
t
h
a
n
o
l
o
r
i
s
o
-
b
u
t
a
n
o
l
)
V
a
p
o
r
C
o
m
p
o
u
n
d
(
e
t
h
a
n
o
l
o
r
m
e
t
h
a
n
o
l
o
r
i
s
o
-
b
u
t
a
n
o
l
)
o
u
t
p
u
t
i
n
p
u
t
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
In
te
ll
igent
Se
nsi
ng U
si
ng Met
al O
xi
de Semic
onduct
or
Ba
se
d
-
on
.
.
..
(
Ny
ay
u
La
ti
fa
h Husni)
4137
Figure
1(a)
sho
ws
the
blo
c
k
di
agr
am
of
gas
sens
or
s
de
velo
pm
ent
.
The
che
m
ic
al
so
ur
ce
s
detect
ed
by
sens
or
s
ar
ray
wer
e
bein
g
c
onve
rted
i
nto
physi
cal
par
am
et
er
.
It
was
t
he
n
processe
d
i
n
m
ic
ro
con
t
ro
l
le
r
to
ob
ta
in
t
he
se
nsor
res
pons
e.
I
n
this
researc
h,
the
outp
ut
was
al
so
sent
to
t
he
com
pu
te
r
by
m
eans
of
wire
le
ss
com
m
un
ic
at
ion
.
T
he
AD
C
data
of
the
se
ns
ors
wer
e
ob
serv
e
d
a
nd
pr
ocesse
d
to
get
the
value
of
sens
or
'
s
respo
ns
e.
To
validat
e
ga
s
sens
ors'
perf
or
m
ance,
the
e
xp
e
rim
ents
we
re
te
ste
d
in
op
en
e
nv
i
ronm
ent.
I
n
m
os
t
work
s,
the
dat
a
was
ob
ta
ine
d
us
in
g
a
cha
m
ber
[58]
-
[60]
wh
ere
the
sensors
we
re
pla
ced
in
and
in
je
ct
ed
so
urces
thr
ough
a
ti
ny
tu
be.
I
n
this
resea
r
ch,
t
he
e
xperi
m
ental
so
urce
was
set
as
show
n
i
n
Fi
gure
1
(b).
The
s
ource
set
-
up
im
it
a
te
d
the
odor
s
ources
i
ntr
oduce
d
by
T
ho
m
as
Loch
m
at
te
r
[15]
.
The
so
urces use
d
in
this
exp
e
rim
ents
wer
e
et
ha
nol
(C
2
H
6
O
),
m
eth
an
ol
(CH
3
O
H)
,
a
nd
acet
one
(C
4
H
10
O)
.
Liqu
i
d
source
s
wer
e
trans
form
ed
to
be
gas
us
in
g
t
he
help
of
t
he
w
ic
k
an
d
t
he
fr
e
sh
ai
r
s
u
ppli
ed
to
the
sourc
es'
cham
ber
.
T
he
wick
and
a
piece
of
t
ube
(for
t
he
fr
es
h
ai
r
in
le
t)
wer
e
us
e
d
to
increa
se
the
ai
r
-
et
ha
nol
inter
face
s
urface.
Eva
porated
et
ha
no
l
was
m
ixed
with
t
he
ai
r
o
n
the
to
p
par
t
of
th
e
c
ham
ber
.
The
m
ixed
ga
s
was
the
n
pu
m
ped
towa
rd
the
gas
outl
et
.
T
he
ai
r
pum
p
us
e
d
as
the
e
xperim
ental
so
u
rces
ha
d
capa
bili
ty
to
exh
a
us
t
14.
0
L
it
tre
li
qu
id
s
ource
i
n
a
m
inu
te
.
It
has
si
x
outp
uts
that
ena
ble
th
e
de
plo
ym
ents
of
t
he
s
ources
to
any
possibl
e
an
d
desire
d
po
i
nts.
The
press
ur
e
of
the
pum
p
is
m
or
e
than
0.
016
MP
a
(
≈
m
or
e
than
2.3
2
ps
i)
with
50
-
60
Hz
fre
qu
e
ncies.
T
he
ty
pe
of
th
e
pu
m
p
is
a
diap
hr
a
gm
pum
p
that
allows
the
outp
ut
(
in
this
case
e
than
ol
con
ce
ntrati
on
)
to
be
c
on
tr
olle
d.
I
n
a
dd
it
io
n,
it
is
equ
i
pped
with
a
kn
ob
th
at
can
be
t
urne
d
a
rou
nd
f
ro
m
the
lowest
pr
e
ssur
e to the
h
i
ghest
one that
offe
re
d
a
n
e
asi
ness o
f
c
on
tr
olli
ng th
e co
ncen
t
rati
on.
3.2.
SVM
Ap
pr
oa
ch
In
t
his
resea
r
ch,
Ard
uino
Me
ga
an
d
Ra
sp
be
r
ry
wer
e
us
e
d
as
in
te
ll
igent
se
ns
in
g
con
t
ro
ll
ers
.
The
si
gn
al
sen
sed
by
the
sen
s
or
ar
ray
was
se
nt
to
t
he
Ard
uin
o
Me
ga
.
T
he
sign
al
was
t
he
n
c
onve
rted
t
o
dig
it
al
sign
al
in
the
A
rduin
o.
T
he
dig
it
al
sign
al
wa
s
then
se
nt
to
Ra
sp
be
rr
y.
In
t
his
co
ntr
oller,
the
SV
M
proc
ess
was
cond
ucted.
A
n
al
go
rithm
us
ing
one
ve
rsus
oth
er
s
te
ch
niq
ue
was
us
ed
to
identify
a
nd
cl
assify
th
e
gas.
The
pr
ocess
of
the
cl
assifi
cat
ion
ca
n
be
di
vid
ed
i
nto
tw
o
gro
ups,
i.e.
Trainin
g
a
nd
Test
ing
as
s
hown
i
n
Figure
2.
Figure
2. Cl
as
sific
at
ion
P
r
oc
ess
Thr
ee
eq
uatio
ns
wer
e
us
e
d
i
n
trai
ning
an
d
te
sti
ng
process:
i
)
Ke
r
nel
Ra
dia
l
Ba
sis
Functi
on
(RBF
)
as
sh
ow
n
i
n
e
qu
a
ti
on
(4)
was
use
d
to
m
ap
the
data
f
ro
m
inpu
t
sp
ace
t
o
featur
e
sp
ace
in
th
e
trai
ning
a
nd
te
sti
ng
process
of this
researc
h;
(
⃗
,
⃗
)
=
(
−
‖
⃗
−
⃗
‖
2
)
(4)
ii
)
qu
a
d
rati
ng
pro
gr
am
m
ing
i
n
e
qu
at
io
n
(5)
was
us
e
d
to
de
te
rm
ine
the
s
uppo
rt
vect
or
va
lue
≠
0
that
was
ob
ta
ine
d by c
ountin
g
t
he valu
e of
1
,
2
,...
.
S
V
M
T
R
A
I
N
I
N
G
·
D
e
t
e
r
m
i
n
e
t
h
e
n
u
m
b
e
r
o
f
c
l
a
s
s
e
s
i
n
S
V
M
p
r
o
c
e
s
s
·
M
a
p
t
h
e
d
a
t
a
f
r
o
m
i
n
p
u
t
s
p
a
c
e
t
o
f
e
a
t
u
r
e
s
p
a
c
e
u
s
i
n
g
K
e
r
n
e
l
R
B
F
·
D
e
t
e
r
m
i
n
e
S
V
v
a
l
u
e
T
E
S
T
I
N
G
·
D
e
t
e
r
m
i
n
e
t
h
e
n
u
m
b
e
r
o
f
c
l
a
s
s
e
s
i
n
S
V
M
p
r
o
c
e
s
s
·
M
a
p
t
h
e
d
a
t
a
f
r
o
m
i
n
p
u
t
s
p
a
c
e
t
o
f
e
a
t
u
r
e
s
p
a
c
e
u
s
i
n
g
K
e
r
n
e
l
R
B
F
·
C
o
u
n
t
t
h
e
d
e
c
i
s
i
o
n
f
u
n
c
t
i
o
n
C
l
a
s
s
i
f
i
c
a
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4133
-
4147
4138
(
)
=
∑
=
1
−
1
2
∑
∑
=
1
=
1
〈
,
〉
(
5)
The
c
orrelat
io
n
data
that
co
r
relat
es
to
≠
0
as
s
uppo
rt
vect
or
ca
n
be
ac
hieve
d
by
us
i
ng
this p
rogr
am
m
ing
;
ii
i) eq
uatio
n (
6) wa
s
us
ed
to
c
ount the
d
eci
s
on
functi
on.
(
)
=
(
∑
∈
(
,
)
+
2
)
(
6)
4.
RESU
LT
S
A
ND D
I
SCUS
S
ION
4.1. Se
nsor P
erfo
rm
an
ce
s
The
se
nsors'
pe
rfor
m
ances
th
at
wer
e
m
easur
ed
wer
e
f
ocus
ed
on
t
he
sen
s
or
s'
tim
e
resp
onse,
se
nsor
s'
peak
re
spo
ns
e
du
rati
on,
the
sensiti
vity
and
the
sta
bili
t
y
respon
se
of
the
sensor.
The
s
peed
of
the
se
ns
or'
s
respo
ns
e
m
eas
ur
e
d
in
di
ff
e
re
nt
posit
ion
wa
s
sh
ow
n
in
F
igure
3.
Fig
ur
e
3(
a)
a
nd
(b)
sh
ows
that
the
gas
sens
or
'
s
res
ponse
tim
e
to
et
han
ol
a
nd
m
et
ha
no
l
a
re
al
m
os
t
the
sam
e.
Fo
r
40
cm
un
ti
l
24
0
cm
distance,
they
need
e
d
on
ly
a
bout
20
s
to
re
sp
on
d
the
gas
so
urce.
For
et
ha
no
l
re
spo
n
se,
there
wa
s
a
ch
ang
e
at
32
0
-
72
0
c
m
wh
e
re
it
need
e
d
lo
nger
res
po
nd
tim
e
than
m
et
hanol,
a
rou
nd
50
-
60
s;
w
hil
e
f
or
m
et
hanol
on
ly
need
e
d
40
-
50
s
.
This
co
ndit
ion
co
ntin
ue
d
unti
l
the
longest
dis
ta
nce
ab
ov
e 8
00
cm
,
wh
ere
th
e
respo
ns
e
tim
e
to
et
hanol
be
ca
m
e
slow
e
r
tha
n
m
et
hanol.
F
or
Fi
gure
3(c)
,
the
respo
ns
e
ti
m
e
of
acet
on
e
was
faster
tha
n
et
hanol
an
d
m
eth
an
ol
.
These
res
ponse
tim
e
diff
erenc
es
wer
e
due
to
so
m
e
factor
s,
su
ch
as
te
m
p
eratur
e
,
hum
id
it
y,
and
al
so
the
ty
pe
of the
odors t
hem
sel
ves.
In
the
resea
rc
h,
the
te
m
per
at
ur
e
an
d
the
hum
idity
w
er
e
m
ade
as
ideal
as
possible.
T
her
e
fore,
th
e
cause
of
the
di
ff
ere
nces
was
due
to
the
ty
pe
of
t
he
odor
substances
on
ly
.
Aceto
ne
ha
s
hea
vier
m
olecular
weig
ht
tha
n
et
hanol
a
nd
m
eth
an
ol,
t
hus
it
cou
l
d
r
each
the
gas
se
nsor
m
or
e
quic
kly
tha
n
two
oth
e
r
s
ubs
ta
nces.
The
m
olecular
weig
hts
a
re
46.06
844
g/m
ol
for
et
ha
nol,
32.04
g/m
ol
fo
r
m
et
hanol
,
and
58.
08
g/m
ol
fo
r
acet
on
e.
T
hey
can
e
va
porate
quic
kly
due
to
thei
r
lo
w
bo
il
ing
point
(
78
0
C,
64
0
C,
56
0
C
res
pecti
vely
).
T
he
heav
y
m
olecular
weig
ht
of
ac
et
on
e
m
ade
it
no
t
be
ea
sil
y
to
be
diffuse
d
by
the
wind
o
r
to
re
apar
t
by
tu
r
bu
le
nt.
Th
us
, th
e s
ub
st
ances could
m
ov
e strai
gh
t t
o t
he
gas
se
ns
ors
w
it
h
high con
centrati
on and can
b
e
detect
ed
m
or
e
qu
ic
kly t
ha
n
t
wo o
t
her subst
ances.
(a)
(
b)
(c)
Figure
3. Se
nsor'
s t
i
m
e
respo
ns
es t
o
s
om
e so
urces
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
In
te
ll
igent
Se
nsi
ng U
si
ng Met
al O
xi
de Semic
onduct
or
Ba
se
d
-
on
.
.
..
(
Ny
ay
u
La
ti
fa
h Husni)
4139
The
se
nsors'
pe
ak
res
ponse
durati
o
n
was
re
pr
ese
nted
i
n
Fi
gure
4.
T
he
pe
ak
res
ponse
du
rati
on
t
o
the
three
s
ources
,
et
hanol,
m
et
han
ol,
an
d
ace
t
on
e
show
ed
dif
fe
ren
t
tre
nds.
H
oweve
r,
al
l
sour
ces
ha
ve
al
m
os
t
the
sam
e
char
act
erist
ic
s.
T
he
nea
rest
the
s
ource
s
to
the
ga
s
se
ns
ors,
t
he
lo
ngest
the
res
ponse
durati
on
of
the
ga
s
sens
or
be.
T
he
peak
respo
ns
e
durati
on
of
ga
s
sens
or
to
t
he
et
hanol
and
m
et
han
ol
at
40
-
400
cm
decr
eased
from
3
5
s
into
10 s.
(a)
(b)
(c)
Figure
4. Se
nsor'
s
pea
k d
ur
at
i
on r
es
ponse
s
f
or 3 dif
fer
e
nt s
ources
The
res
pons
e
seem
ed
to
be
const
ant
to
the
val
ue
of
10
s
econd
f
or
the
distance
a
bove
of
400
cm
.
Wh
il
e
f
or
th
e
acet
on
e,
t
he
durati
on
ti
m
e
was
hi
gh
e
r
th
an
et
ha
no
l
a
nd
m
et
han
ol;
it
has
val
ue
of
arou
nd
30
-
40
s
f
or
the
dista
nce
of
40
-
40
0
cm
.
It
only
exp
e
rience
d
sm
a
ll
chan
ges
un
ti
l
the
distan
ce
of
760
cm
.
Thi
s
ph
e
nom
eno
n
was
a
gain
cau
s
ed
by
the
m
ole
cular
weig
ht
of
th
os
e
three
odor
s
ources
.
T
he
hea
vier
m
olecular
weig
ht
of
t
he
acet
on
e
m
ade
the
acet
one
c
ould
sta
y
l
onge
r
ar
ound
the
ga
s
sen
sor.
The
li
gh
te
r
weig
ht
of
et
hanol and m
et
hanol m
ade them
to
be
e
asi
ly
to be
disp
e
rse
d by the
wind.
The
se
ns
it
ivit
y
an
d
sta
bili
ty
respon
se
of
the sen
s
or
s w
as d
i
sp
la
ye
d
in
Fig
ur
e 5
a
nd
Fig
ure
6.
Eac
h
of
the
sen
sors
in resp
on
se
ti
m
e
t
est
ing
s
howe
d
that
the
res
ponse
of
them
beco
m
e
slow
er d
ue
to
the
l
onger of
t
he
distance
of
the
so
urces
to
rea
ch
the
se
nsors.
It
was
quit
e
the
sam
e
with
t
he
sen
sors'
peak
res
pons
e
dur
at
io
n
wh
e
re
the
re
spon
s
e
w
ou
l
d
be
sh
ort
er
wh
e
n
the
distance
betwee
n
the
s
o
urce
an
d
the
gas
se
ns
ors
be
com
e
longer
.
T
he
se
ns
it
ivit
y
of
the
sens
ors
is
sho
wn
in
Fig
ur
e
5
.
It
in
dicat
ed
that
the
TGS
s
ens
or
was
se
nsi
ti
ve
enou
gh
to
th
e
change
of
th
e
env
ir
onm
ent
.
The
od
or
s
ources
wer
e
e
xpose
d
a
nd
un
e
xpos
e
d
to
the
rob
ots
intercha
ngeabl
y
f
or
obta
inin
g
the
data
.
T
he
so
urce
was
s
w
it
ched
on
for
120
s
a
nd
was
s
witc
hed
off
120
s.
The dat
a
was
a
chieve
d
in
the
sta
ti
c p
os
it
io
n wh
e
re t
he
r
obot
s w
ere
p
la
ce
d i
n
the
strai
ght f
ace t
o face
po
s
it
ion.
The
co
nce
ntrat
ion
of
the
acet
on
e
t
hat
the
robo
ts
m
eas
ur
ed
was
the
hi
gh
e
st
on
e
(a
bove
900
of
A
DC
va
lue)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4133
-
4147
4140
wh
il
e
the
m
eth
an
ol
was
t
he
s
m
al
le
s
t
on
e
(abo
ve
500
of
AD
C
val
ue).
The
detect
in
g
values
of
eac
h
gas
sens
or
s
wer
e
no
t
di
rectl
y
chan
ge
wh
e
n
the
pr
oc
esses
of
on
a
nd
off
ha
pp
e
ne
d.
All
of
the
sensors
w
ai
te
d
arou
nd 30
-
50 s
econds
unti
l t
he
y chan
ge
d
th
e
ir r
ea
ding
valu
e b
ase
d o
n
the
conditi
on of t
he
g
as
sou
rces.
It
was
due
to
the
sens
ors
ne
eded
m
or
e
tim
e
to
be
bac
k
to
it
s
init
ia
l
con
diti
on.
T
he
s
ens
or
s
sti
ll
sense
d
the
gas
that
was
le
ft
in
their
s
urrou
nd
i
ng
(alt
hou
gh
the
s
ource
ha
s
been
switc
he
d
off
);
he
nce,
the
sens
or
s
sti
ll
read
the
gas
c
on
centrati
on
as
hi
gh
value
for
a
bout
30
-
50
sec
onds
a
fter
the
gas
s
ource
off.
To
ov
e
rc
om
e
and
m
ini
m
iz
e
this
transiti
on
c
ondi
ti
on
,
a
fa
n
was
us
e
d.
It
c
ou
l
d
help
to
cl
ea
n
t
he
gas
re
sid
ue
that
sta
cked to t
he
s
ens
or
s'
area.
(a)
(b)
(c)
Figure
5. Se
nsor'
s sensi
ti
vity
f
or
3 diff
e
re
nt
so
urces
The
sta
bili
ty
resp
onse
of
the
s
ens
or
s
wa
s
represente
d
in
Fig
ur
e
6.
This
sta
bili
ty
was
needed
in
orde
r
to
m
easur
e
the
tim
e
occu
pie
d
by
the
sensor
s
in
detect
ing
the
source
in
a
sta
ble
conditi
on.
By
knowin
g
the
sta
bili
ty
resp
onse,
t
he
m
ob
il
e
rob
ots
intel
li
gen
ces
w
her
e
the
gas
se
nsors
wer
e
p
la
c
ed
can
be
des
ign
e
d
pro
per
ly
.
I
n
t
his
resear
ch
,
the
sta
bili
ty
data
go
t
by
c
ollec
ti
ng
the
se
nsor
s'
respon
se
t
o
the
change
of
sit
ua
ti
on.
The
a
bili
ty
of
the
se
ns
or
t
o
r
each
it
s
init
ia
l
conditi
on
w
as
m
easur
ed
a
nd
recorde
d.
F
r
om
Figu
re
6,
it
can
be
sta
te
d
that
the
gas
sen
sors
us
ed
in
this
e
xpe
rim
ents
wer
e
sta
ble
enou
gh.
The
co
nce
ntrat
ion
m
easur
ed
by
the
sens
or
s
was
al
m
os
t
the
sa
m
e
from
the
first
s
a
m
pling
un
ti
l
t
he
fi
fteenth
sa
m
pl
ing
.
Fig
ur
e
6
(
a
)
-
6
(
c
)
sho
ws
the
sta
bili
ty
of
the
sensors
to
the
et
han
ol,
m
et
han
ol,
a
nd
acet
one.
Fig
ur
e
6(
d
)
rep
rese
nts
the
sensors'
sta
bil
it
y
of
the
sens
ors
to
al
l
of
the
s
our
ces
us
e
d
in
th
e
exp
e
rim
ent.
The
ga
s
sen
sors
us
e
d
in
t
his
exp
e
rim
ent
has
the
sta
bili
ty
r
espon
se to eac
h
s
our
ces ar
ound
34
seco
nd un
ti
l 4
5 sec
on
d
f
or eac
h
se
nsors
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
In
te
ll
igent
Se
nsi
ng U
si
ng Met
al O
xi
de Semic
onduct
or
Ba
se
d
-
on
.
.
..
(
Ny
ay
u
La
ti
fa
h Husni)
4141
(a)
(b)
(c)
(d)
Figure
6. Se
nsor'
s stabil
it
y for
3 diff
e
ren
t
sources
4.2.
Co
ll
ectin
g D
ataset
s
The
trai
ning
da
ta
set
s
wer
e
got
f
ro
m
the
ga
s
sens
ors
of
t
he
rob
ots
that
wer
e
place
d
in
a
fix
place
appr
ox
im
at
e
60
cm
fr
om
the
so
urce.
The
c
hoic
e
of
the
60
c
m
away
from
the
sour
ce
as
the
trai
ning
da
ta
sets
was
base
d
on
the
cha
racteri
s
ti
c
of
t
he
se
nsor
res
pons
e
ti
m
e,
sens
or
pe
ak
respo
ns
e
durati
on,
sta
bili
ty
and
sensiti
vity
.
At t
his
po
sit
io
n,
ga
s sen
s
ors
s
how
ed
thei
r goo
d p
erfor
m
ance.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4133
-
4147
4142
The
ex
pe
rim
en
ta
l
env
iro
nm
ent
was
then
e
xpos
e
d
to
the
ga
s
so
urce
f
or
20
seco
nd
s
.
The
n,
the
data
ca
m
e
fr
om
the
3
sens
ors
of
each
r
obot
were
plo
tt
ed
in o
ne
gr
a
phic
s
as
sh
own
in
Fi
gure
7.
These
data
be
ca
m
e
the
trai
ning
da
ta
set
s
fo
r
the
cl
assifi
cat
ion
.
The
data
w
as
gro
up
e
d
into
t
heir
cl
asses.
Wh
e
n
e
xpos
e
d
to
the
et
hanol
source
,
the
data
that
was
sense
d
by
TGS
26
00,
TGS
26
02,
an
d
TGS
26
20
as
et
hanol
cl
asses
were
recorde
d,
as
w
el
l
as
the
data
of
the
ot
her
s
ource
,
i.e.
m
et
han
ol
an
d
acet
one.
Fig
ur
e
7(b)
and
7(c)
s
how
s
the
respo
ns
e
of the
g
as
sen
s
ors
for
m
et
han
ol clas
s and t
he
acet
one cla
ss
res
pec
ti
vely
.
(a)
(b)
(c)
Figure
7. Cl
as
ses u
se
d as
data trai
ning
for g
as cla
ssific
at
io
n
4.3.
Odo
r
Cl
as
sific
at
i
on
u
sing
Suppor
t Vect
or Mac
hine
In
t
his
resea
rc
h,
tw
o
m
od
es
of
e
xperim
ental
set
up
will
be
us
ed
i
n
de
te
ct
ing
the
odor
s
ource,
nam
e
l
y:
static
and
m
ob
il
e
sensors.
Stat
ic
sens
or
is
a
syste
m
that
on
ly
do
e
s
the
detect
ion
process
w
it
ho
ut
changin
g
th
e
po
sit
io
n,
w
hile
m
ob
il
e
sensor
has
the
op
posit
e
char
act
eris
ti
cs,
i.e.
by
chan
gi
ng
the
se
nsors
po
sit
io
n
to
the
odor
s
ources
.
Ba
sed
on
the
detect
ion
res
ul
t,
the
robo
ts
s
hould
be
able
to
cl
assify
the
odor
ty
pe.
T
he
pro
cess
of
t
he
cl
assifi
cat
ion
usi
ng
SV
M
ap
p
ro
ac
h
was
est
ablishe
d
in
ra
sp
be
r
ry
w
hile
th
e
nav
i
gation
of
t
he
m
ob
il
e r
obot
s w
as
proces
s
ed
in
Ar
du
i
ono.
4.3.1.
St
ati
c se
nso
r
Stat
ic
sen
sors
data we
re obtai
ning b
y
dep
l
oying
the
r
obots that w
e
re equip
ped
with
gas
s
ens
or
s T
G
S
2600,
T
GS
26
02,
an
d
T
GS
2620
to
t
he
e
xperim
ental
env
iron
m
ent.
The
r
obots
wer
e
pla
ced
in
fro
nt
of
the
gas
so
urce
a
nd
m
ov
ed
them
to
a
certai
n
place
m
anu
al
ly
to
ge
t
furthe
r
f
r
om
t
he
s
ource
(in
s
ca
le
of
0.5
m
)
in
each
m
ov
e
m
ent.
4
po
sit
io
ns
(0.5,
1.0,
1.5,
a
nd
2.0
m
)
wer
e
te
ste
d
to
see
the
s
up
e
rio
rity
of
t
he
r
obot
s
in
determ
ining
th
e
odor
s
ource.
The
odor
cl
as
sific
at
ion
te
sti
ng
was
c
ondu
c
te
d
10
ti
m
es
for
eac
h
sou
rc
e.
The
data
was
rec
orded
in
Ta
ble
1.
Most
of
them
wer
e
su
cce
ssf
ul
in
sp
eci
fyi
ng
the
source
.
F
or
the
ra
nge
of
0.
5
un
ti
l
2.0
m
,
al
l
the
ro
bots
co
ul
d
determ
ine
th
e
et
han
ol
an
d
acet
on
e
co
rr
ec
tl
y
with
the
su
ccess
rate
90
%.
Fr
om
10
ti
m
es
test
in
g,
all
o
f
the ro
bo
ts co
ul
d
dete
rm
ine the so
urce cor
rectl
y for
9
tim
es. Th
ere w
as o
ne
ti
m
e
wh
e
re
the ro
bo
t
got 1
00% s
uccess.
G1
-
2 i
n t
he 0.
5 m
d
ist
an
ce co
uld
sp
eci
fy m
et
hanol s
ource c
orrectl
y for
10
tim
es.
Evaluation Warning : The document was created with Spire.PDF for Python.