TELKOM
NIKA
, Vol.11, No
.2, Februa
ry 2013, pp. 63
0
~
63
6
ISSN: 2302-4
046
630
Re
cei
v
ed Au
gust 20, 20
12
; Revi
sed
De
cem
ber 2
2
, 2012; Accepte
d
Jan
uary 11,
2013
Discrimination of Chinese Her
b
al Med
i
cine by Machin
e
Olfaction
Deha
n Luo*, Ya
w
e
n Shao
Schoo
l of Information En
gi
ne
erin
g, Guang
d
ong U
n
ivers
i
t
y
of
T
e
chnol
og
y,
Guangzh
ou 5
100
06, P.R Chi
na.
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: deha
nl
uo@
g
dut.edu.c
n
A
b
st
r
a
ct
“Sm
a
ll Sample Si
z
e
”
(SSS)
problem
would occur wh
ile using linear dis
c
ri
minant analysis (LDA
)
alg
o
rith
m w
i
th
traditio
nal
F
i
sh
er criteri
o
n
if t
he w
i
th
i
n
-class
scatter
matrix
is s
i
ng
ul
ar. T
he c
o
mbi
nati
o
n o
f
m
a
x
i
m
u
m
sc
atter difference (
M
SD) criterion and LDA
algorithm for solve SSS prob
lem
is descr
ibed. It is
empl
oyed to d
e
tect three kin
d
s of Chin
ese
herba
l m
edic
i
nes fro
m
differ
ent grow
ing
ar
eas by
mach
in
e
olfactio
n. Co
mpare
d
w
i
th PCA or PCA +
L
D
A alg
o
ri
th
m, the classific
a
ti
on resu
lt w
a
s
enh
anc
ed. It
works
out that on
ly a
few
sampl
e
s of
Anhu
i Atractyl
odes
are cl
assi
fied i
n
correctly,
how
ever, the
classificati
on r
a
te
reach
e
s 97.8
%
.
Ke
y
w
ords
:
li
n
ear discri
m
in
a
n
t analys
is; maxi
mu
m scatte
r difference cri
t
erion; Ch
ines
e herb
a
l
med
i
cine;
mac
h
i
ne olfacti
o
n
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Atractylode
s i
s
an a
s
teraceae me
dicin
e
with sp
ec
ial
smell, an
d their qu
ality is affected
by pla
c
e
of o
r
igin, h
a
rve
s
t
time, bre
ed
a
nd oth
e
r fa
ct
ors,
amo
ng t
he o
r
igin
fact
ors is on
e of
the
most impo
rta
n
t criteri
a
in judgin
g
the quality.
With people’
s incre
a
sin
g
quality requi
rem
ents of
Chin
ese herb
a
l medici
ne, the i
dentification of medici
n
a
l herb
s
is p
a
r
ticula
rly imp
o
rtant.
Re
sea
r
ch of
electroni
c no
se b
ega
n in t
he 19
90
s,
it is a p
a
rt of th
e sp
ecifi
c
ity with the
comp
ositio
n of the gas sensor a
rray
and patte
rn recognitio
n
system is co
mposed of the
approp
riate i
n
strum
ents,
m
a
inly u
s
ed
to i
dentify sim
p
le
and
compl
e
x odo
rs [1]. Th
ere
are a
lot
of
resea
r
che
s
a
nd so
cial a
p
p
licatio
ns in
the food ind
u
stry [2-4], medical diag
nosti
cs [5
-7], and
environ
menta
l
monitori
ng
[8-10] at h
o
m
e an
d ab
road, but in
the fiel
d of
Chin
ese he
rbal
medici
ne
s are rarely repo
rted in the current.
The Chi
n
e
s
e herb
a
l medi
ci
ne Atractylod
es is the obj
ect in this pa
per, and d
e
te
cted by
electroni
c no
se. In pattern re
cog
n
ition
with the el
ectro
n
ic
no
se, the prin
ci
pal co
mpo
n
e
n
t
Analysis
(Pri
ncip
al Co
mp
onent Analy
s
is,
PCA) a
n
d
LDA h
a
s
been
widely
use
d
[8]. The
outstan
ding f
eature
of LDA is it can e
n
su
re t
hat af
ter the proje
c
tion, mod
e
l sampl
e
ha
s the
smalle
st
wit
h
in-cl
a
s
s
di
st
a
n
ce a
nd max
i
mum
betwe
en-cla
ss dist
ance
in
the new spa
c
e,
that
model ha
s th
e best sepa
ra
bility in the space. Ho
wev
e
r, there i
s
al
so not ap
plicable in the "small
sampl
e
probl
em" and othe
r sh
ortcomin
g
s
. In respon
se to this sh
ort
c
omin
g, man
y
schol
ars ha
ve
use
d
a meth
od of combi
n
ation with PCA and L
D
A
[9],
the advantage
s of the PCA and L
D
A
together fully integratio
n, a
nd it
can
not
only solve
th
e problem
of
PCA algo
rith
m is
not sen
s
itive
to the different training sa
mple data problem,
but al
so L
D
A algo
rithm wh
en the within
-cl
a
ss
scatter matri
x
is singul
ar,
and obtain
a better cla
s
sificatio
n
re
sults. In this pape
r, maximum
scatter
difference
crite
r
ion
and
L
D
A
wil
l
be
integ
r
ate
d
togeth
e
r, it
solve
d
th
e p
r
oble
m
of
sm
all
sampl
e
s, an
d
there is a bet
ter cla
s
sificati
on re
sult than
PCA and PCA + LDA.
2. Rese
arch
Metho
d
2.1. Electron
i
c nose (E-n
ose)
Experiment
s
were pe
rform
ed with
a co
mmercia
l E-n
o
se
(PEN3
)
. It is provid
ed
by WMA
AIRSENSE Analyse
n
techni
k Gm
bH
(Schwe
rin, Ge
rm
any). Tabl
e 1
.
summa
ri
zes the se
nsitivity
of different se
nso
r
s in PEN3.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Discrim
i
natio
n of Chine
s
e
Herbal Me
di
ci
ne by Ma
chi
n
e Olfaction
(Dehan L
uo)
631
PEN3 in
clu
d
ed a
n
a
r
ray
of 10
differe
nt MOS
se
n
s
ors,
and
th
e sen
s
o
r
respon
se i
s
defined a
s
th
e ratio of co
n
ducta
nce: G/G0. Whe
r
e,
G rep
r
e
s
ent
s the resi
stan
ce of each
se
nso
r
in the ch
amb
e
r after
expo
sing to the ta
rget ga
s a
n
d
G0 re
pre
s
e
n
t
s t
he re
si
st
a
n
ce
while
ea
ch
sen
s
o
r
i
s
exp
o
se
d to th
e
zero
ga
s filtere
d
by a
c
tive
carbo
n
. Th
e el
ectro
n
ic no
se
co
nsi
s
ts mai
n
ly
of the followin
g
se
ction
s
: co
mputer
、、
sam
p
ling chann
el
sen
s
o
r
ch
a
nnel, as
sho
w
ed in Figu
re 1
.
Table 1. The
sen
s
itivity list
of 10 sen
s
o
r
s in PEN3
Number in a
rra
y
Sensor name
Sensitive to
S1 W1C
Aromatic
components
S2
W5S
Nitrogen o
x
ides, ver
y
sensitive
S3
W3C
Ammonia and ar
omatic components
S4
W6S
Mainly
h
y
d
r
ogen,
selective
l
y
,
(bre
ath gases)
S5
W5C
Alkanes and aro
m
atic components
S6 W1S
Propane
S7
W1W
Sulfur organic compounds
S8 W2S
Ethanol
S9 W2W
Aromatic
component
s and org
anic-sulfides
S10 W3S
Propane
(selective
sometimes)
Figure 1. Dia
g
ramm
atic La
yout of Electronic
No
se
2.2. Experimental sample
This medi
cin
e
sampl
e
i
s
suppo
rted
by
Guan
gzhou
University
of Ch
in
ese Me
di
cine.
They
are i
r
regul
ar clum
ps of
hypertr
ophy,
ga
s fra
g
ra
n
c
e,
sweet a
nd
slightly a
c
rid. Atractyl
ode
s
sampl
e
s
we
re provid
ed from thre
e ki
n
d
s of O
r
igin:
Baoding
of Heb
e
i provin
ce, Hao
z
ho
u
of
Anhui province, Shaoxing o
f
Zhejiang p
r
o
v
ince.
2.3. Experiment pro
cedur
e
The expe
rim
ents were carrie
d out
in an air-co
nditione
d la
borato
r
y wh
ere the
temperature
wa
s ke
pt at 25±1 an
d the
humidit
y at 54±2%. Static head
sp
ace sampling meth
od
wa
s used be
cause of its ac
ce
ssi
bility and stability [10].
The sample
s with different
origin
were put into four
bea
kers (500
ml) label
ed Hebei, Anhui a
n
d
Zhejian
g
, re
spectively. Th
e amou
nt of e
a
ch
sam
p
le i
n
the be
aker
wa
s 10
0g. Th
en thre
e be
akers
were h
e
rm
etically
cap
ped
with
pla
s
tic wrap fo
r
70
minute
s
in
orde
r to
ge
n
e
rate
a
stea
dy
head
sp
ace re
spe
c
tively. The samplin
g time for e
a
ch
sampl
e
is
60
se
con
d
s,
whi
c
h i
s
en
oug
h
for
each se
nso
r
to reach a st
able value. T
he rin
s
ing tim
e
is set as 1
10 se
co
nds,
durin
g whi
c
h
th
e
sen
s
o
r
s a
r
e ri
nse
d
with ch
arcoal filtere
d
to force
the
sign
als of se
nso
r
s to b
a
se
line. The interva
l
for data
colle
ction
wa
s on
e se
co
nd. On
e mea
s
u
r
eme
n
t cycle
wo
ul
d last fo
r abo
ut three mi
nu
tes.
Whe
n
the m
easure
m
ent
wa
s co
mplet
ed, the obt
ai
ned data
wa
s sto
r
ed in a
comp
uter for later
analysi
s
. Th
e
head
sp
ace
gas
of ea
ch
bea
ker of Atractylode
s
sa
mple
wa
s me
asu
r
ed
30 ti
mes
Computer
Sample volatile
Built-in pump
Sensor arr
a
y
Clean air
Clean air
Built-in pump
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 630 – 636
632
respe
c
tively. Thus 90
data
set
s
were
co
llected
fo
r all three group
s of
Atractylod
e
s
sampl
e
s. T
he
90 sa
mple
s
were divided
into two gro
ups: 45
sam
p
les
(15
sa
mples
of each gro
up) fo
r
the
training
set a
nd the re
st 45
sample
s (15
sampl
e
s of e
a
ch g
r
ou
p) fo
r the testing
set.
2.4. Patter
n
recognitio
n
LDA is
one
of the wid
e
l
y
used
cla
ssification te
ch
nique
s. However, wh
en t
he total
number of s
a
mples
is
s
m
all or
the number of selec
t
ed features
is
large, SSS problem would
occur
while
u
s
ing LDA alg
o
rithm with
traditional Fish
er
crite
r
ion if
the within
-cla
ss scatter ma
trix
is
singul
ar. T
herefo
r
e, a
n
optimize
d
di
scrimi
nant crit
erion
called maximum scatter
differen
c
e
(MSD) criteri
on wa
s ad
opt
ed [11].
Suppo
se th
e
numbe
r
of kn
own
pattern
classe
s i
s
N a
s
1
G
,
2
G
,
,
N
G
, pattern
d
x
R
is d
-
dime
nsi
o
nal real ve
cto
r
,
i
N
is th
e nu
m
ber
of traini
n
g
sa
mple
s in
i
t
h
cla
ss,
i
m
is t
he me
an
feature ve
cto
r
of trai
ning
sample
s in
i
th
c
l
as
s
,b
e
t
w
een
-
c
la
ss
sc
atter matrix is
b
S
, within-c
lass
s
c
atter matrix
is
W
S
,and they defined a
s
fol
l
owin
g re
spe
c
tively:
Mean of sa
m
p
les
i
m
:
1
i
i
xG
i
mx
N
,
1,
2
,
,
iN
(1)
within-c
lass
sc
atter mat
r
ix
S
:
1
i
N
T
ii
ix
G
Sx
x
(2)
1,
2
,
,
iN
betwe
en-cla
s
s scatter mat
r
ix
b
S
:
1
()
()
N
T
bi
i
i
Sm
m
m
m
(3)
among,
1
1
N
i
i
mm
N
(4)
Fishe
r
crite
r
i
on i
s
that th
e choi
ce m
a
ke
s the
maxi
mum of th
e
gene
rali
zed
Rayleig
h
quotient a
s
the proje
c
tion d
i
rectio
n vecto
r
T
b
F
T
w
S
J
S
(5)
The b
a
si
c id
ea of MSD
crite
r
ion i
s
tr
y to find an
optimal p
r
oje
c
tion ve
ctors
.It is
different from
Fishe
r
criterion be
cau
s
e
in MS
D, the
differen
c
e of
betwee
n
-cla
ss
scatter a
n
d
within-cla
ss
scatter is emp
l
oyed
a
s
discrimina
n
t crite
r
ion
rath
er th
an thei
r
ratio.
Thu
s
we
ca
n
define maxim
u
m scatter dif
f
eren
ce criteri
on functio
n
a
s
belo
w
:
()
T
b
M
T
SC
S
J
(6)
Whe
r
e, C
i
s
a
con
s
tant,
f
o
r co
nvenien
ce,
thi
s
articl
e
is set
to
1, to
bala
n
ce
m
a
ximizing
the
betwe
en-cla
ss scatter and
minimi
ze
the divergen
ce bet
wee
n
cl
asse
s.
b
SC
S
is call
ed matri
x
of generali
z
ed diverg
en
ce differen
c
e a
s
paramete
r
s
for the
C
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Discrim
i
natio
n of Chine
s
e
Herbal Me
di
ci
ne by Ma
chi
n
e Olfaction
(Dehan L
uo)
633
It can p
r
oved
that the opti
m
al proje
c
tio
n
dire
ction
is to make the maximum
scatter
differen
c
e cri
t
erion
fu
nctio
n
()
M
J
to take th
e maximum
value of th
e
sol
u
tion,
wh
ich th
e
followin
g
gen
erali
z
ed ei
ge
nvalue proble
m
is solved:
()
b
SC
S
(7)
So, maximum scatte
r difference crite
r
io
n
can
be attrib
uted to the sa
ke of eige
nve
c
tor
probl
em of the gene
rali
zed
divergen
ce d
i
fference matrix
b
SC
S
.
3. Results a
nd Analy
s
is
3.1. Sensors
respons
e
Figure 2 sho
w
s the typical
resp
on
se cu
rves of
10 se
nso
r
s to the three
sele
cted
sampl
e
grou
ps. T
he
hori
z
ontal
axis is the
sam
p
ling time,
a
nd the ve
rtical axis i
s
the
sen
s
o
r
re
sp
onse
value.
It shows
rapi
d ch
ang
e at t
he be
ginni
ng
of t
he sampl
i
ng time
whil
e the respon
se valu
es
rea
c
h to th
e
steady
state
soo
n
. After a
pproxim
ately 60 second
s a
l
most all th
e
sen
s
o
r
s re
ached
to stable resp
onse value
s
. This Fig
u
re cl
early s
hows
different re
sp
onse sig
nal
s of sen
s
o
r
s a
r
ray
to Atractylod
es
sampl
e
s
with differe
nt gro
w
ing
a
r
e
a
s. Each sen
s
or
ha
s re
sp
onse to different
varieties of Chine
s
e he
rbal
medicin
e
s.
Figure 2. The
resp
on
se curves of Atractylode
s sam
p
le
s
3.2. Featur
e selectio
n
Feature
sele
ction i
s
of g
r
eat im
porta
nce,
whi
c
h requires
the conve
r
si
on o
f
sampl
e
feature
s
to
pattern
s that
have
con
d
ense r
epresentation
s
, id
eally co
ntain
i
ng only
ma
in
information.
In this
study,
initially eight
different
su
b
-
f
eatures were sele
cted a
s
the
ori
g
inal
feature
vector fro
m
the sen
s
o
r
re
sp
onse sig
nal
s:
10
40
60
,,,
,
m
a
x
,
v
a
r
,
,
T
f
f
f
avg
s
t
d
diff
(8)
whe
r
e
i
f
re
pre
s
ent
s the
re
spo
n
se value
at i se
con
d
of sen
s
o
r
a
rray
(i=1
0, 4
0
, 60);
avg
rep
r
e
s
ent
s the averag
e va
lues of ea
ch
res
pon
se
cu
rve for the du
ration of 60
se
con
d
s;
max
0
20
40
60
0
0.
5
1
1.
5
2
2.
5
3
3.
5
4
4.
5
A
n
hui
A
t
rac
t
y
l
od
es
S
a
mp
li
n
g
t
i
me
(
s
)
R
e
s
pons
e v
a
l
ue[
G
/
G
0
(
G
0/
G
)
]
0
20
40
60
0
0.
5
1
1.
5
2
2.
5
3
3.
5
4
4.
5
Hebe
i
A
t
rac
t
y
l
o
des
S
a
m
p
l
i
ng t
i
m
e
(s
)
R
e
s
pons
e v
a
l
ue[
G
/
G
0
(
G
0/
G
)
]
0
20
40
60
0
0.
5
1
1.
5
2
2.
5
3
3.
5
4
4.
5
Z
h
ej
i
a
n
g
A
t
rac
t
y
l
od
es
S
a
mp
li
n
g
t
i
me
(
s
)
R
e
s
pons
e v
a
l
ue[
G
/
G
0
(
G
0/
G
)
]
W1
C
W5
S
W3
C
W6
S
W5
C
W1
S
W1
W
W2
S
W2
W
W3
S
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 630 – 636
634
rep
r
e
s
ent
s th
e maximum
values
of ea
ch re
spo
n
se curve for th
e
duratio
n of 6
0
se
co
nd
s;
va
r
rep
r
e
s
ent
s th
e varian
ce
of respon
se
da
ta for the du
ration of 60
seco
nd
s;
st
d
repres
ents
the
stand
ard
dev
iations
of th
e re
sp
on
se
sign
als,
p
r
e
s
enting the
fluctuatio
n a
r
o
und the
average
values of ea
ch respon
se
curve;
diff
rep
r
esents the differentiati
on of the respon
se
si
gnal
s.
3.3. Discriminant cla
ssifi
cation
There is
15
sampling tim
e
s for Atractyl
ode
s tr
aini
ng
sam
p
les of
each growi
n
g
are
a
, so
the total of Atractylod
es t
r
a
i
ning
sampl
e
s with
three dif
f
erent g
r
o
w
in
g are
a
s is
45,
and PEN3 h
a
s
10 sen
s
o
r
s,
e
a
ch
se
nsor m
easure
m
ent
s are extrac
t
e
d eight
c
h
arac
teris
t
ic
parameters
, thus the
total charact
e
risti
cs ve
cto
r
dimen
s
ion
are 80
-d
im
e
n
sio
nal, then
, clearly the total number of
training
sam
p
les a
r
e le
ss than the feature vect
o
r
dimen
s
ion, a
"small sa
mp
le" probl
em that
arise, at thi
s
time LDA
alg
o
rithm
can
n
o
t pro
c
e
ed at
this time. Fi
gure
3 i
s
the
PCA and
PCA +
LDA analy
s
is
cha
r
t of three
Atractylode
s training
sam
p
les.
Figure 3. Analytic result of three g
r
o
u
p
s
It can be see
n
from Figure 3 (a) that the cla
ssi
fi
cati
on re
sults of three sets of training
sampl
e
s
with
a sepa
rate
PCA algorit
hm are not
satisfa
c
to
ry, the batch of
sample p
o
i
n
ts
intertwin
ed, a
nd indistin
gui
sha
b
le. The reason is
whe
n
the differen
c
e of sam
p
le quality grad
e is
small, the
r
e i
s
a bi
g ove
r
l
ap of info
rma
t
ion or
releva
nce i
n
the dif
f
eren
ce
s in t
he sample
th
at
reflect by el
e
c
troni
c n
o
se
sen
s
o
r
, PCA algorithm t
o
find only th
e data di
strib
u
tion of spin
dle
orientatio
n [1
2], retaine
d
a
fter dimen
s
io
nality red
u
cti
on by the i
n
formatio
n
is n
o
t necessa
ril
y
the
mos
t
effec
t
ive for
c
l
as
s
i
fication .
Figu
re
3 (b
)
can
be
see
n
that th
e distin
gui
sh
result of PCA
+
LDA m
e
thod
is b
e
tter th
an u
s
ing
PCA algorith
m
alone
betwe
en three
gro
ups of trai
ni
ng
sampl
e
s, a
n
d
interspe
rse
d
with the
origin
al traini
ng sa
mple
points h
a
ve all been
cle
a
rly
sep
a
rate
d. This is be
ca
u
s
e the main
idea of LD
A
algorithm is to minimize
the within-cl
a
ss
distrib
u
tion a
nd maximize the spread b
e
t
ween
cla
s
se
s.
To avoid the
small sampl
e
proble
m
, we
use L
D
A alg
o
rithm ba
se
d
on maximum
scatter
differen
c
e crit
erion. Th
e re
sults
sho
w
n i
n
Figure 4:
As can b
e
se
en from Figu
re 4 that the d
i
sti
ngui
sh result of LDA algorithm is b
e
tter than
PCA an
d P
C
A + L
D
A algo
rithm.
While
He
bei
Atractylod
e
s
sampl
e
p
o
ints
are
m
a
inly
con
c
e
n
trated
in the lower h
a
lf of the feature sp
a
c
e, Zhejian
g
Atractylodes sa
mp
le points mai
n
ly
in the upp
er l
e
ft part and
Anhui Atra
ctylode
s samp
le
points
are
concentrate
d i
n
the up
per
ri
ght
part. Vari
ou
s
training
sa
mp
le point
s can
be cl
ear
ly
distinguished, a
nd compa
r
e
d
to PCA +
LDA
method, th
e
distrib
u
tion
o
f
sam
p
le
poi
nts
with
in
a
cla
s
s is eve
n
mo
re
con
c
entrated,
mo
re
obviou
s
the interface between the cl
asses.
(a) Analytic
result of three
grou
ps by
P
CA
(b) Analytic
result of three
grou
ps by
PCA+
LDA
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Discrim
i
natio
n of Chine
s
e
Herbal Me
di
ci
ne by Ma
chi
n
e Olfaction
(Dehan L
uo)
635
Figure 4. Analytic result of three
g
r
o
u
p
s
by LDA base
d
on MSD crit
erion
3.4. Discriminant cla
ssifi
cation
Table 1
sho
w
s the
cla
s
sification re
sul
t
of each test sam
p
le in
the two-dim
ensi
ona
l
feature spa
c
e, and the accura
cy rate
wa
s cal
c
ul
at
ed by the ratio of the number of correctly
predi
cted
sa
mples a
nd th
e numbe
r of total testing sample
s.
Table2. Pred
icted re
sult
s of three testin
g sets
The
re
sults
shows th
at, fo
r the
45
sam
p
les teste
d
, t
here
i
s
o
n
ly
an e
r
ror to b
e
carried
out to d
e
termine, the
re
cog
n
ition
rat
e
of
An
hui Atractylode
s wa
s
9
3
.3%,while
Heb
e
i
and
Zhejian
g
Atra
ctylode
s reco
gnition
rate
wa
s 10
0%. T
he di
scrimin
a
n
t re
sults re
a
c
he
d 97.8%
o
f
c
o
rrec
t c
l
ass
i
fic
a
tion rate for all tes
t
s
a
mples
.
4. Conclusio
n
More an
d mo
re studi
es ha
ve shown tha
t
the
use of electro
n
ic n
o
se technol
ogy for odor
analysi
s
i
s
no
t only obj
ecti
ve and
a
c
curate, but
also
rep
r
od
uci
b
le
and
co
nvenie
n
t. In this pa
p
e
r,
PEN3 ele
c
tro
n
ic n
o
se u
s
e
d
to test Atra
ctylode
s
sam
p
les of
thre
e gro
w
ing area
s,
data analy
s
is
method
usi
n
g
LDA al
gorith
m
ba
sed
on
MSD criteri
o
n to solve th
e problem
of small
sa
mpl
e
s,
also
distin
gui
sh
with three
Atractylode
s from
three
different g
r
o
w
ing
are
a
s correctly, and
the
corre
c
t re
cog
n
ition rate of
all testing sample
s
rea
c
hes 9
7
.8%, furthe
rmo
r
e, the cla
s
sificati
on
results
clea
rly supe
rio
r
to the use of PCA or PC
A +
LD
A a
l
g
o
r
i
th
m. T
h
is
pr
o
v
id
es
th
e
as
su
ra
nc
e
for the quality of Chine
s
e h
e
rbal m
edi
cin
e
an effective
way.
Referen
ces
[1]
Gardner JW
, Bartlett PN. A Br
ief Histor
y
of Electron
ic Nos
e
.
Sensors a
n
d
Actuators B
. 199
4; 15(1
8
):
211-
220.
[2]
Ghasemi-V
a
rn
amkhasti M, Mohtase
b
i SS.
Meat Qualit
y
Assessment b
y
El
ectronic N
o
se.
Sensors
.
200
9; 9(8): 605
8-60
83.
[3]
Brezmes J, F
r
u
c
tuoso M
LL. E
v
alu
a
tion
of A
n
Electro
n
ic
Nos
e
to Ass
e
ss F
r
uit Ri
pe
ness.
I
EEE Sens
ors
Journ
a
l.
20
05;
12
(8): 97
–1
08.
Testing sets
Correctl
y
predict
ed
Wrongl
y
p
r
edicted
Accuracy
Hebei Atract
y
l
od
es
15
15
0
100%
Anhui Atract
y
l
od
es
15
14
1
93.3%
Zhejiang Atract
y
l
odes
15
15
0
100%
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 630 – 636
636
[4]
W
ang YW
, W
ang J. Mon
i
torin
g
Storag
e T
i
me an
d
Qual
it
y
Attribute of Eg
g Base
d on E
l
ectronic N
o
se
.
Analytic
a Chi
m
ica Acta
. 200
9; 16(3): 183-
18
8.
[5]
T
haler ER, H
anso
n
CW
. M
edic
a
l A
ppl
icat
ions
of El
ectronic
Nos
e
T
e
chno
log
y
.
Expe
rt R
e
vie
w
of
Medic
a
l Dev
i
ce
s
. 2005; 2(5): 5
59-5
66.
[6]
Mazzone, PJ.
Analy
s
is of Volatile Organic
Co
mp
oun
ds in
the Exh
a
l
ed
Breath for the
Diag
nosis
of
Lun
g Ca
ncer.
Journ
a
l of T
hora
c
ic Oncolo
gy
. 200
8; 3(7): 774
-780
[7]
Kateb B, Ry
a
n
MA. Sniffing Ou
t Canc
er
Using
the JP
L
Electron
ic N
o
se: A pil
o
t stu
d
y
of a
nov
el
appr
oach to d
e
t
ection an
d diff
erenti
a
tion
of brain ca
ncer.
Ne
uroi
ma
ge
. 2
0
0
9
; 7(4): 5-9.
[8]
Szczurek A,
S
z
eco
w
k
a
PM.
Application
of
Sensor
Array
and Neural Net
w
orks
for Quantification
of
Organic So
lve
n
t Vapours i
n
Air.
Sensors a
nd Actuators B
. 1999; 8(5): 42
7-43
2.
[9]
Martin MA, Santos JP, Vasq
uez H, et al.
Stud
y
of the Interferenc
es o
f
NO and CO in Sol
i
d Stat
e
Commercial S
ensors.
Sens
or
s and Actuator
s B
. 1999; 8(5)
: 469-47
3.
[10]
Martinel
li E, Z
a
mpetti E. De
s
i
gn
and T
e
st o
f
An Electron
ic
Nose for M
oni
toring th
e Air
Qualit
y i
n
th
e
Internatio
na
l Space Statio
n.
Micrograv
ity Scienc
e an
d T
e
chno
l
og
y. 2
007;
19(8):60-
64.
[11]
AZ
Berna, J
La
mmert
y
n
S. El
ectronic
Nos
e
S
y
stem
s to Study
Shelf Life
and Cu
ltivar Effect on T
o
mato
Aroma Profil
e.
Sensors a
nd A
c
tuators B: Chemical
. 2
004; 9
(
7): 324-3
33.
[12]
A Martin
ez, A
Kak. PCA
V
e
rsus
LDA .
IEEE Transact
i
ons
on Patte
rn Analysis
and M
a
chine
Intelli
genc
e
. 20
01; 23(2): 2
28-
233.
[13]
MP Marti, R Boque O. Electro
n
ic Nos
e
s in
th
e Qualit
y Co
ntrol of Alcoh
o
lic
Bevera
ges.
T
r
ends i
n
Analytic
al Ch
e
m
istry
. 200
5; 2
4
(7): 57-6
6
.
[14]
F
X
Son
g
, D Z
hang JY. Ada
p
ti
ve Classific
a
ti
o
n
Algor
ithm Ba
sed on Ma
ximu
m Scatter Difference
Discrimi
nant C
r
iterion.
Acta Automa
tica Sinica
. 2006; 3
2
(4) : 541-5
49.
[15]
Scott SM, James D, Ali Z
.
Data ana
l
y
sis for e
l
ectron
ic nose
s
y
stems.
Microchim
i
ca Acta
. 2
006;
15(6):1
83-2
07.
Evaluation Warning : The document was created with Spire.PDF for Python.