TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3595 ~ 36
0
2
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4919
3595
Re
cei
v
ed O
c
t
ober 2
4
, 201
3; Revi
se
d Decem
b
e
r
16, 2013; Accept
ed Ja
nua
ry 6,
2014
Quality Abnormal Pattern Recognition of Dynamic
Process Based
on MSVM
Yumin Liu, Haofei Zho
u*, Shuai Zhang
Busin
e
ss Scho
ol of Z
hengz
ho
u Univ
ersit
y
, Z
hen
gzh
ou 45
0
001, He
na
n, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: seanzh
o
u
668
@16
3
.com
A
b
st
r
a
ct
T
he i
m
prov
e
m
ent of the effec
t
ive reco
gniti
on
of
qua
lity ab
n
o
rmal p
a
tterns
in dyn
a
m
ic
pro
c
ess has
seen i
n
cre
a
sin
g
de
ma
nds n
o
w
adays in the
real-ti
m
e
monit
o
r and
dia
g
n
o
s
e of auto
m
ati
c
ma
nufacturi
n
g
.
Based
on the
analys
is of the dyn
a
m
ic p
r
ocess of
qu
a
lity abn
or
mal
pattern, this p
aper pr
ese
n
ts
a
recog
n
itio
n mo
del of qu
al
ity abnor
mal patter
n
recog
n
it
io
n u
s
ing a Mu
lti-SV
M. Contrasting
w
i
th performan
c
e
of recogn
ition
mo
de
l base
d
on differe
nt kernel
functi
ons, suitab
le kern
el
functions w
e
re
selected for the
recog
n
itio
n mo
del. F
u
rther
mo
re, w
e
have contrasted th
e mo
de
l prop
ose
d
in this p
aper
w
i
th the mod
e
l
ado
pted
by Va
hid. Si
mulati
on
results sh
ow
that the rec
o
g
n
i
t
ion
mo
del
pro
pose
d
in th
is p
aper
has very
hig
h
recog
n
itio
n acc
u
racy for all p
a
tterns, and the
over
a
ll aver
ag
e recog
n
itio
n a
ccuracy is 97.7
8
%.
Ke
y
w
ords
:
pat
tern recog
n
itio
n, dyna
mic pr
o
c
ess, MSVM, kerne
l
functio
n
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Quality abno
rmal pattern reco
gnition of dynamic
p
r
o
c
ess plays a very importa
nt role in
monitori
ng b
o
th the ma
n
u
facturer
pro
c
e
ss
ru
nni
ng
in its inten
d
ed mo
de a
n
d
the p
r
e
s
en
ce of
abno
rmal
p
a
tterns and re
a
lizing online quality
diagn
ose
of autom
atic produ
ctio
n pro
c
e
s
s. Since
mode
rn in
du
strie
s
, such
as
petrol
e
u
m
, metallurg
y, machin
ery
and
othe
r i
ndu
strie
s
, ha
ve
become m
o
re large-scale
,
compl
e
x a
nd contin
u
o
u
s
, the mo
nitoring
and
di
agno
si
s of t
he
dynamic
pro
c
e
ss h
a
s
attracte
d many
sch
ola
r
s'
attention an
d n
o
w it be
com
e
s the resea
r
ch
hotsp
ot in the
field of qualit
y control [1-3
]. Si
nce a la
rge num
be
r of
dynamic dat
aflow h
a
s
be
en
gene
rated
through th
e p
r
o
c
e
s
s of auto
m
atic p
r
o
duction, the key
probl
em of
d
y
namic
pro
c
e
ss
quality cont
rol is ho
w to
monitor a
n
d
diagno
si
s th
e dynami
c
d
a
ta flow’
s
variation ten
d
e
n
cy
effec
t
ively.
Suppo
rt vector ma
chin
e (SVM) perfo
rms cl
as
sifica
tion tasks of
various
kin
d
s
quality
abno
rmal
dat
a by
con
s
tru
c
ting
optimal
se
parating
h
y
perpla
ne [4]
.
This
metho
d
can
effecti
v
ely
solv
e sev
e
r
a
l
pra
c
t
i
cal pr
o
b
lems,
su
ch as
t
he
small
sampl
e
p
r
ob
lem, nonlin
ea
r problem
an
d
high dime
nsi
onal pattern reco
gnition p
r
oblem an
d so
on [5].
Curre
n
tly, SVM ha
s wi
del
y applied to
quality
of the
indu
stry pro
c
e
ss
monito
ri
ng an
d
diagn
osi
s
. Th
e variation te
nden
cy of dynamic
dataf
lo
w con
s
ist
s
of
several patt
e
rn
s, incl
udin
g
trend patte
rn,
shift pattern and cy
clic pa
ttern. The
re
cognition of variation
tend
en
cy for dynami
c
dataflow is t
y
pical of mul
t
i-cla
ss
cla
ssi
ficati
on.Howe
ver, the basi
c
SVM deals with two-cla
ss
probl
em
s. Th
us, it
can
be
develop
ed fo
r multi-cla
s
s cl
assificatio
n
to
deal
with
dy
namic dataflo
w.
No
wad
a
ys,
MSVMs a
r
e
gainin
g
ap
pli
c
ation i
n
the
area
of control ch
art p
a
ttern recognitio
n
an
d
fault diag
nose in in
du
strial
pro
c
e
s
se
s [6
-7]. Jia
ng
(2
0
09)
esta
blish
e
s
a MSVM
model
ba
sed
on
four
SVM cla
ssifie
r
s, all
of
whi
c
h have cho
s
e
n
G
a
u
s
sian
kern
el fu
nction
to dia
g
nose the fa
ult of
blast fu
rna
c
e
[8]. Vahid
(2
010) p
r
e
s
ent
s a
MSVM m
odel
usin
g th
e “o
ne
-agai
n
s
t-all
”
meth
o
d
to
recogni
ze the
quality pattern for all six
pattern
s [9]. Wu (2
010
) prese
n
ts a MS
VM model, and
four SVM cla
ssifie
r
s a
r
e e
s
tabli
s
he
d to recogni
ze
tre
nd pattern,
shift pa
ttern, cyclic pattern
and
mixed pattern [10]. Xiao (201
0) p
r
e
s
ents a lea
s
t
squa
re MS
VM model to re
cog
n
ize
three
different
shift pattern
s of t
he TE p
r
o
c
e
ss
and
sim
u
l
a
tion expe
rim
ent indi
cate
s
that MSVM can
recogni
ze all
the quality patterns effectiv
ely [11
]. As
mentione
d ab
ove, many studie
s
of MSVM
recognitio
n
model are
re
stri
cted
to so
me
spe
c
ial
i
ndu
strial pro
c
e
ss and
al
ways ado
pt
"one-
again
s
t-all" m
e
thod
whi
c
h i
s
on
e kernel f
unctio
n
to dia
gno
se differe
nt quality abn
ormal
pattern
s.
Thus the
recognition
effect of this MSV
M
re
co
gnition
mod
e
l
will b
e
g
r
eatly
wea
k
en
ed. O
w
in
g t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3595 – 36
02
3596
the cha
r
a
c
ters of
dynami
c
dataflo
w,
su
ch
as n
online
a
r, n
on-no
rm
al. The
r
e
are
large
differe
nce
on variation
tenden
cy o
f
quality abnorm
a
l patte
rns.
Con
s
e
q
uently, the large
s
t probl
ems
encounte
r
ed
in setting u
p
the MSVM
model
are h
o
w to
sele
ct the
ke
rnel
functio
n
s an
d
its
para
m
eter va
lues, when we using MSV
M
model to reco
gni
ze the
quality patterns of dynam
ic
p
r
oc
es
s
.
Based
on the
analysi
s
of t
he dynami
c
pro
c
e
ss
of q
uality abno
rm
al pattern, thi
s
pa
per
pre
s
ent
s a
re
cog
n
ition fra
m
ewo
r
k of
qu
ality abno
rma
l
pattern
reco
gnition
by u
s
ing a
Multi-SV
M
model. Throu
gh the simul
a
tion experi
m
ent, a MSVM reco
gniti
on model of
quality abno
rmal
pattern in
dynamic
pro
c
e
s
s is p
r
o
p
o
s
e
d
in this p
a
p
e
r after
com
parin
g an
d a
nalyzin
g different
kernel fun
c
tions’ a
c
curacy. Furthermo
re,a qua
lity abno
rmal re
cog
n
ition mo
del of dynamic
pro
c
e
ss
ba
se
d on MSVM is esta
blishe
d, whic
h pro
v
es the effectiv
eness achi
eved by qual
ity
abno
rmal p
a
ttern of dynami
c
pro
c
e
s
s.
2. Qualit
y
Patter
n
of Dy
n
a
mic Proces
s
Thro
ugh
the
contin
uou
s
p
r
odu
ction
p
r
o
c
e
s
s o
r
a
u
to
matic
equip
m
ent tre
a
ting
p
r
ocess,
we o
b
taine
d
a mount
of d
y
namic d
a
taflow
reflectin
g
the ope
ratin
g
co
ndition
which
effects t
he
quality and chang
es of pro
duct
s
und
er a
continu
o
u
s
p
r
odu
ction p
r
o
c
e
ss.
Figure 1. Qua
lity Abnormal Pattern of Dynamic Process
The
pro
d
u
c
tive process i
s
u
nde
r
a n
o
rm
al
runnin
g
statu
s
wh
en dyn
a
mi
c
dataflo
w
fluctuating
around
the
de
signi
ng ta
rge
t
nume
r
ical v
a
lue
ran
doml
y
. The follo
w variation
tre
n
d
sho
w
n
by Fig
u
re
1(a
)
i
s
th
e no
rmal
pattern
(NR) of d
y
namic
pro
c
e
s
s ba
sic qu
ality pattern. Whe
n
the dynamic
dataflow tend
ing to trend, shift or
cycl
e, there are a
b
norm
a
l factors leadin
g
to a
quality pro
b
l
e
m amo
ng p
r
odu
ctive pro
c
e
ss. Th
e
a
bnormal con
d
ition of dyn
a
mic p
r
od
uct
i
ve
quality contai
ns th
ree
type
s: tren
d, shift and
cy
clic.
The trend
of
dataflow incl
ude
s two typ
e
s
whi
c
h a
r
e in
crea
sing t
r
end
(IT) an
d de
crea
sing t
r
end
(DT
)
(sh
o
wn
by Figure 1(b) an
d 1(d))
as
well a
s
the
shift pattern
contain
s
up
wa
rd
shift
(US
)
and d
o
wnward shift (DS)
(sho
wn
by Fig
u
re
1(c) a
nd 1
(
e
)
). The
r
efore, the dynami
c
pro
c
e
ss quali
t
y
abnorm
a
l model can b
e
illustrated as
increasing trend, de
creasi
ng trend, upward shift, downward sh
ift and cycli
c
pattern (CC). Figure
1 sho
w
s deta
ils of dynami
c
proces
s ba
sic q
uality model which in
clud
es n
o
rm
a
l
pattern an
d five
abno
rmal p
a
ttern
s.
Duri
ng th
e a
c
tual
co
ntinu
ous produ
cti
v
e pr
oc
ess
o
r
au
to
ma
tic tr
e
a
t
in
g pr
oc
es
s
,
the
abno
rmal ch
ange
s of dataflow will le
ad to the
problem
s of produ
cts on varying de
gre
e
s.
Therefore, in
orde
r to re
d
u
ce the
pro
b
l
e
ms in p
r
od
uc
tive
p
r
oc
ess, it is
n
e
c
essa
r
y
to
r
e
c
o
gniz
e
effectively the quality abnormal model of
dynamic d
a
ta
flow.
3. Establish
m
ent th
e Re
cognition M
odel Bas
e
d
on MSVM
In this se
cti
on, we e
s
ta
blish
ed th
e
quality patte
rns
re
co
gnitio
n
mo
del
of
dynamic
pattern. We combine
d
the roug
h SVM cl
assifier an
d subdivisi
on SVM classifie
r
to establi
s
h th
e
recognitio
n
m
odel
of dyna
mic p
a
ttern. I
n
orde
r to
provide the
o
reti
cal
ba
sis for
kernel fu
nctio
n
sele
ction
of
singl
e
sup
porting vecto
r
m
a
chi
ne i
n
the
re
cog
n
ition
model, th
ree
kin
d
s of
kerne
l
function
s, in
cludi
ng Li
ne
ar,Polynomial
and
RBF,
whi
c
h a
r
e ty
pically u
s
e
d
to SVM wil
l
be
introduced in this section as well.
0
50
10
0
15
0
20
0
25
0
30
0
35
0
-4
-2
0
2
4
（）
aN
R
（）
bI
T
（）
cU
S
（）
aN
R
（）
dD
T
（）
eD
S
（）
fC
C
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Quality Abn
o
rm
al Pattern Reco
gnition of
Dynam
ic Pro
c
e
ss Ba
sed o
n
MSVM (Yum
in Liu)
3597
3.1. Recog
n
ition Model Based on MS
VM
A
n
S
V
M
pe
rf
orm
s
cla
ssif
i
c
a
t
i
on t
a
sk
s
b
y
co
n
s
tructin
g
optimal
se
paratin
g hyp
e
r
plane
s
(OSHs) [12]. There a
r
e
six basi
c
q
uality pattern
s
in t
he dynami
c
d
a
taflow to
re
cogni
ze, so M
u
lti-
SVM model i
s
ne
ce
ssary.
There are two
widel
y u
s
ed metho
d
s t
o
extend SVM to Multi-cl
ass
probl
em
s [9]. One of them is calle
d the “on
e
-a
gai
nst-all
”
(OAA
) method. An
other meth
od
is
calle
d “on
e
-a
gain
s
t-on
e” (OAO) meth
o
d
. The two m
e
thod
s all ha
ve good re
co
gnition effect.
In term of trend, shift an
d cycli
c
thre
e pattern
s of
dynamic d
a
taflow, we first sho
u
l
d
adopt
OAA
method
and
try to
re
co
gnize the
s
e
three
quality
abn
orm
a
l p
a
tterns
with
the
establi
s
hm
en
t of SVM
1,
SV
M
2
and SVM
3
classifie
r
s. And then, we esta
blish
e
d
SVM
4
and SVM
5
cla
ssifie
r
s an
d re
co
gni
ze i
n
crea
sing
tre
nd o
r
d
e
crea
sing t
r
en
d an
d up
wa
rd
shi
ft or do
wn
wa
rd
shift. In order to cl
assify the
six basi
c
quality
patterns
of dynami
c
proc
ess,
we constructed a
model of qual
ity abnormal
pattern recog
n
ition ba
sed
on MSVM. The structu
r
e
of the recogni
tion
Model is
sho
w
n in Figu
re
2.
Figure 2. Re
cognition Mo
d
e
l Based o
n
MSVM
In the Figure 2, trend, shift and cycli
c
th
ree pattern
s of dynamic dat
aflow thro
ugh
SVM
1
,
SVM
2
and SVM
3
classifie
r
s be recogni
zed firstly. W
hen the cu
rrent dataflo
w be
identified as
trend patte
rn
through SV
M
1
, the datafl
o
w will be
cl
assified by SVM
4
which
can re
cog
n
ize
th
e
increa
sing trend a
nd de
crea
sing t
r
end
two qua
lity
abno
rmal p
a
tterns.Su
bseq
uently, upward
shift and
do
wn
ward shift two qu
ality abno
rmal
pat
tern
with SVM
2
and SVM
5
clas
sif
i
e
r
s
be
cla
ssifie
d
a
s
t
he
sam
e
me
chani
sm.
Whe
n
SVM
3
ide
n
tified
the datafl
o
w
i
s
not an cycli
c
al pattern
after SVM
1
and SVM
2
clas
sif
i
er
s,
S
V
M
6
will und
erta
ke
the recogniti
on for no
rmal
pattern.
3.2. Kernel F
unction Sele
ction of MSV
M
The ba
si
c pri
n
cipl
e of SVM is to solve
linearly sep
a
rabl
e proble
m
s through fi
nding
a
linear
hype
rpl
ane. In the li
nearly
sep
a
rable
con
d
itio
n, a set of (x
i
,y
i
), i=1,2,…
…
n,
xR
d
as
training
sam
p
le of dynamic dataflow whi
c
h
have
kno
w
n the value
of catego
ry. The y
i
is the l
abel
value of traini
ng sam
p
le an
d sele
cting th
e value from1
or -1.
The pu
rpo
s
e
of training SV
M model i
s
to fi
nd an optim
al hyperplane
that divides the two
abno
rmal
pattern
s
so th
at it ca
n turn the
error
of traini
ng into
ze
ro
while
maximu
m margin. Th
us,
the probl
em finding thi
s
hyperpl
ane tran
sform
ed this
optimizatio
n probl
em:
2
11
mi
n
(
)
22
T
J
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3595 – 36
02
3598
s.t
:
(
)
1
,
1
2
n
T
ii
yx
b
i
，
，
…,
(1)
It can
ado
pt
Lagrnag
e o
p
timize
metho
d
to tre
a
t the
primal
proble
m
a
s
o
ne
si
mple d
ual
probl
em. The
optimal deci
s
ion functio
n
(ODF
) is then
given by:
**
1
(
)
s
gn[
(
,
)
]
n
ii
i
i
f
xa
y
x
x
b
(2)
Whe
r
e the
a
*
are o
p
timal
Lagrang
e mu
ltipliers,
whe
n
*
0
i
a
the trainin
g
sampl
e
is SV
(su
ppo
rt vect
or)
，
an
d
*
b
are
cla
ssifi
cation
threshold. A
s
the sample
data of q
ualit
y abno
rmal
pattern all a
r
e linea
rly inseparable, SV
Ms can e
ffici
ently perform
non-li
nea
r cl
assificatio
n
u
s
ing
kernel fun
c
tio
n
, implicitly m
appin
g
the
s
e
linearly in
se
p
a
rabl
e sampl
e
data into
hi
gh-di
men
s
ion
a
l
feature spa
c
es so that ca
n solve linea
rly inse
p
a
ra
bl
e probl
em
s. Thus the p
r
o
b
lem to find an
optimal hype
rplane tra
n
sfo
r
med anoth
e
r
optimal probl
em as follo
w
：
1
1
mi
n
(
)
+
C
2
n
T
i
i
J
s.
t
(
)
1
-
,
1
2
n
T
ii
i
yx
b
i
：，
，
…,
(3)
i
is a sl
ack variable an
d
0
i
,c d
e
fined a
s
pen
alty coefficien
t (c>0). The
g
r
eate
r
the
C is, the heavier puni
shment fo
r
misclassifi
cation will be. Afte
r adding the kernel function
k(
,
)
i
x
x
，
the c
l
ass
i
fy func
tion turn to
：
**
(
)
sgn[
a
(
,
)
]
ii
i
SV
f
xy
k
x
x
b
(4)
Acco
rdi
ng to above-mentio
ned analy
s
is,
it shoul
d be
cho
o
se different kernel fu
nction
s
for different
SVM classifi
ers. Th
us, it
is ess
ential
to sele
ct a suitable
ke
rnel functio
n
for
improvin
g th
e pe
rforman
c
e of SVM
cl
assifiers.
Th
e
r
e are
three
basi
c
ke
rnel
function
s
i
n
t
he
appli
c
ation of
SVM. These mathemati
c
e
x
pressio
n
of
kernel fun
c
tio
n
s are sh
own
in Table 1 [13].
Table 1. Basi
c Kern
el Fun
c
tion
Name
Expression
Linear
(,
)
(
)
ii
Kx
x
x
x
Poly
nomial
(,
)
(
1
)
,
1
,
2
,
n
q
ii
Kx
x
x
x
q
……
，
RBF
22
(,
)
e
x
p
2
ii
Kx
x
（
x-
x
）
Curre
n
tly, it has
no g
ene
ra
l rule
to
sele
ct
kern
el fun
c
tion. The
r
efore, it need
s
consi
d
e
r
the a
c
tual id
entify obje
c
t to sele
ct kernel
fun
c
tio
n
by existin
g
expe
rien
ce
and
sim
u
lat
i
on
experim
ents.
In the
fram
e
w
ork of q
uali
t
y abno
rmal
pa
ttern
re
co
g
n
ition, it is
si
gnifica
nt that
to
cho
o
se a
sui
t
able kernel
function to
i
m
prove th
e
quality pattern re
co
gnition
perfo
rma
n
ce
of
MSVM
cla
s
sifiers.In ou
r pa
per, we comp
ered the pe
rf
orma
nce of SVM classi
fier
with usin
g four
different
ke
rn
el fun
c
tion
s, i
n
clu
d
ing
Lin
e
a
r, Po
lyno
mi
al, RBF
an
d
Sigmoid, fo
r
different
quali
t
y
pattern
s. The
simulation ex
perim
ent is a
s
follow.
4. Results a
nd Analy
s
is
In this se
ction, in order to
ensu
r
e the
perfo
rman
ce
of reco
gnitio
n
model, we
analyze
perfo
rman
ce
of SVM cl
assifiers
ba
sed
on
differe
nt ke
rn
el fun
c
t
i
ons for the
different
qual
ity
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Quality Abn
o
rm
al Pattern Reco
gnition of
Dynam
ic Pro
c
e
ss Ba
sed o
n
MSVM (Yum
in Liu)
3599
abno
rmal pat
terns. First,
we use
Mo
nte-Carl
o
meth
od to
simul
a
te the d
a
ta of
dynamic proce
ss
pattern
s. The
n
, we
com
pare the a
c
curacy of each
SVM cla
s
sify mo
del ba
se
d on
different
kern
el
function,
so
t
hat it
can
offe
r
some
evide
n
ce
s
fo
r
ch
oo
sing
pa
ram
e
ters of SVM
m
odel.
Comp
ared
with the re
co
gnition
mo
del
prop
osed by
Vahid, the model p
r
op
ose
d
in this pa
pe
r wa
s an
alyzed.
Furthe
rmo
r
e,
it is si
gnificance that
thi
s
re
co
g
n
ition
mod
e
l b
e
p
r
oved
for qu
ality diagn
ose of
dynamic p
r
o
c
ess.
4.1. Data
De
scription
In parti
cula
r,
the con
s
ide
r
e
d
patterns refer to t
he follo
wing
cl
asse
s:
normal
(NR), cycli
c
(CC), in
crea
sing tre
nd
(IT),
decre
asi
ng t
r
end
(DT)
, up
ward shift
(US),
and do
wn
ward shift
(DS)
.
These
patterns
of all
the
s
e different types were
g
e
n
e
rated
u
s
ing
the Mo
nte-Ca
rlo m
e
thod.
T
h
is
equatio
n as f
o
llows:
()
()
()
x
td
t
r
t
(5)
In the Equation (1
), x(t) is
the data of q
ua
lity pattern
in dynami
c
proce
s
s whi
c
h
need
s to
be si
mulate
d.
μ
rep
r
e
s
e
n
ts the val
ue o
f
desig
n targ
et, containi
ng
three
pa
rts.
To si
mplify the
simulatio
n
d
a
t
a, this pa
pe
r set
s
μ
to
ze
ro. r(t) is the
data
chan
ge
cau
s
e
d
by th
e existed
cau
s
al
factors. He
re
white noi
se is
adopted to repre
s
e
n
t the data ch
ange
r(t) in ou
r exp
e
rime
nt. d(t) is
the data ch
a
nge re
sulte
d
by t
he abnormal factors. It causes
five
abnormal m
ode
s, includi
ng
increasing trend, decreasing tr
end, upward
shift, downward
sh
if
t
and cycl
e patterns. Dat
a
simulatio
n
formula of each qua
lity pattern is
as
follows
:
1) Normal patterns
:
xt
r
t
(6)
2) Increa
sin
g
trend patte
rn
s:
xt
r
t
+
k
t
(7)
3) De
crea
sin
g
trend patte
rns:
xt
r
t
k
t
(8)
4) Upward
s
h
ift patterns
:
x t
r
t
b
s
(9)
5) Do
wn
wa
rd
sh
ift patterns:
x t
r
t
-
b
s
(10)
6) Cycli
c
patt
e
rn
s:
xt
r
t
+
a
s
i
n
2
t
/
T
(11)
Here, r(t) i
s
a fun
c
tion
that ge
nerates
ran
dom
numb
e
rs
n
o
rmally
distributed(
rt
(
0
1
)
N
，
), k i
s
the
gra
d
ient of in
cr
e
a
sin
g
o
r
de
creasi
ng trend
pattern
trend
pattern
(set in
the
rang
e 0.3 to 0.5), b indicates the shift position in an
upward
shift pattern an
d a
down
w
a
r
d shift
pattern
(b
=0
before
the
shi
ft and b
=
1 a
t
the shift a
n
d
there
a
fter),
s is the
mag
n
itude of th
e
shi
f
t
(set bet
wee
n
1
and 2), a is the amplitud
e of cyclic
va
riation
s
in acyclic pattern (set in the ran
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3595 – 36
02
3600
0.8 to 2), T is the peri
od of
the cycle
(se
t
betwee
n
15
and 30
), t is the di
screte ti
me at whi
c
h t
he
monitored
proce
s
s vari
abl
e is sample
d
(set
within
th
e range
0
to2
9
), a
nd x(t
)
i
s
the valu
e of
the
dynamic data
point at time
t. 360 a
bove
-
mentio
ned
p
a
tterns, 60 of
ea
ch
type, were
p
r
eviou
s
ly
gene
rated. F
o
rme
r
20 g
r
o
ups a
r
e trai
n sampl
e
s, whil
e the other 4
0
grou
ps a
r
e
testing.
4.2. Experiment Settings
In orde
r to compa
r
e the reco
gnition a
c
curacy of SVM in different
kernel fun
c
tion, the
penalty factor C of the SVM identifier n
eed
s to
be set. C is the tradeoff betwe
en identify errors
and
algo
rithm
com
p
lexity. If the value
of
C i
s
to
o hig
h
,
the alg
o
rith
m will
be
com
e
very
com
p
l
e
x.
And the
lo
wer th
e valu
e
of C,
the
st
rong
er t
he
m
odel’
s
g
ene
ralizatio
n a
b
ility, but the e
rro
r
become
s
big
ger which wa
s pr
o
p
o
s
ed b
y
Cherka
ssky
[14].
ma
x
x
+
3
x
-
3
xx
C
（
,
）
(12)
x
is the m
ean v
a
lue of the t
r
aining
sam
p
l
e
s in th
e dyn
a
mic p
r
o
c
e
ss quality patterns,
σ
x
is the stan
dard deviation of
the training sample
s.
In the simul
a
tion experi
m
ent, the value of
c is sele
cted a
s
1.2 accordi
ng to the
Che
r
kassky's empirical formula (11
)
for Linea
r and P
o
ly kern
el function mod
e
l, while
gr
id
s
ear
c
h
method is a
d
opted for RB
F kernel fun
c
tion model.
4.3. Compari
s
on of th
e M
odels’ Perfor
mance
After analyzi
ng the effect
of different ker
nel fun
c
tions o
n
pe
rfo
r
man
c
e, the
suitabl
e
kernel fu
ncti
ons were
se
lected
for re
cog
n
ition
m
o
del p
r
op
osed
in thi
s
p
a
p
e
r. Acco
rdin
g to
above-mentio
ned d
a
ta,
60
grou
ps
of dat
a of ea
ch q
u
a
lity pattern wil
l
be me
rge
d
i
n
to 360
gro
u
p
s
of data, a
s
th
e sample
dat
a of SVM
1
, SVM
2
, SVM
3
and SVM
6
pa
ramete
r o
p
timization
ba
sed o
n
RBF ke
rnel f
unctio
n
. 60 group
s of data of two
quality patterns
will be merged in
to 120 gro
u
p
s
of
data, as the
sampl
e
d
a
ta
of SVM
4
and
SVM
5
para
m
eter o
p
timi
zation
ba
sed
on
RBF
kernel
function. T
h
e
re
sults
of p
a
ram
e
ters o
p
timization
with grid
se
arch m
e
thod f
o
r SVM
1
whic
h
recogni
ze
tre
nd p
a
ttern
a
nd SVM
4
wh
ich
re
cog
n
ize in
cre
a
si
ng
and
de
crea
sing
pattern
are
s
h
ow
n
in
F
i
gu
r
e
3
.
(a)
(b)
Figure 3. (a)
Paramete
r op
timization for
trend SVM
1
, (b) Para
meter
optimizatio
n for SVM
4
Thro
ugh
sim
u
lation expe
riment, the
accuraci
es
o
f
reco
gnition
based on
different kernel
function
s are sho
w
n in Ta
b
l
e 2.
-5
0
5
-5
0
5
30
40
50
60
70
80
90
100
l
o
g2c
l
og2g
A
c
c
u
ra
c
y
(%
)
-5
0
5
-5
0
5
30
40
50
60
70
80
90
100
lo
g
2c
lo
g
2
g
A
ccu
r
a
cy(
%
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Quality Abn
o
rm
al Pattern Reco
gnition of
Dynam
ic Pro
c
e
ss Ba
sed o
n
MSVM (Yum
in Liu)
3601
Table 2. The
Accu
ra
cy of SVM on Different Ke
rn
el F
unctio
n
Un
de
r Optimized P
a
ram
e
ters
Kernel function
RA
Parameters an
d
recognition accuracy
SVM
1
SVM
2
SVM
3
SVM
6
SVM
4
SVM
5
Linear
85.21%
C=1.2
64.58%
66.67%
96.25%
83.75%
100%
100%
Polynomial 95.83%
C=1.2
q=2
q=2
q=2 q=2 q=3
q=3
92.08%
97.92%
98.33%
86.67%
100%
100%
RBF
98.4%
c=1.74
c=3.03
c=0.189 c=3.03 c=0.004
c=0.004
g=0.02
g=0.02
g=0.012
g=0.02
g=0.004
g=0.004
98.33%
97.92%
100%
94.17%
100%
100%
Indicate
d by
recognitio
n
result
s sho
w
n
in
Table
2, the re
cog
n
ition pe
rform
a
n
c
e
s
are
quite differe
nt betwee
n
the
different qu
ali
t
y patte
rns a
n
d
their dive
rse ke
rnel fun
c
t
i
ons. Th
e RB
F
function
have
excelle
nt pe
rforma
nce for all t
he
patte
rns, and average RA
re
ach
98.4%,
while
Linea
r an
d Polynomial fun
c
tion
s al
so a
c
hieve hi
gh
classificatio
n
a
c
cura
cy which equal to
10
0%
for increa
sin
g
trend
pattern, decrea
s
in
g
trend p
a
ttern
, upwa
r
d
shift pattern a
nd
downward
sh
ift
pattern.
Acco
rdi
ng to
the result
s
of experime
n
t,
the recog
n
i
tion model
sho
u
ld choo
se RBF
function for S
V
M
1
, SVM
2
and SVM
3
; Linear, Polynom
ial and RBF f
unctio
n
is be
st for SVM
5
and
SVM
6
. The Result of ch
oo
sing
kernel fu
nction for MS
VM model are sho
w
n in T
able 3.
Table 3. Sele
ction of Kern
el Functio
n
for MSVM
NO.
Ty
p
e
SVM
1
RBF
SVM
2
RBF
SVM
3
RBF
SVM
4
RBF
SVM
5
Linea
r
、
RBF
、
Po
l
y
no
m
i
a
l
SVM
6
Linea
r
、
RBF
、
Po
l
y
no
m
i
a
l
Afterwa
r
ds,
compa
r
ing th
e
re
cog
n
ition
accu
racy
with
two mo
del
s
based o
n
RBF ke
rnel
function, vali
dity of quality abno
rmal
re
cog
n
iti
on m
o
del for dyna
mic p
r
o
c
e
s
s
prop
osed
by this
pape
r
wa
s p
r
oved. The
M
SVM model
p
r
opo
se
d by V
ahid b
e
sele
cted a
s
the
co
ntrast
mod
e
l. We
set above
-
m
entione
d MSVM model a
s
model
Ⅰ
, while the mo
del pro
p
o
s
ed
in this pape
r as
model
Ⅱ
. In o
r
de
r to com
p
are the recog
n
ition accu
ra
cy of these t
w
o mo
del, RBF kernel fun
c
tion
is ado
pted b
y
kernel fun
c
tion of these
two SVM model
s. Acco
rding to abov
e quality pattern
sampl
e
data of
dynami
c
p
r
ocess
by si
mulation,
th
e
optimal
pa
ra
meter
co
mbi
nation
ba
sed
on
RBF wa
s fou
nd with gri
d
-sea
rch meth
od. The re
co
gnition re
sult
s of two mod
e
l are sho
w
n
in
Table 4.
Table 4. The
Re
cog
n
ition
Accu
ra
cy of Tw
o Mod
e
ls u
nder the
Optimized Pa
ram
e
ters
Ty
p
e
R
A
The recognition
accurac
y
unde
r t
he optimized par
ameters
IT DT
US
DS
CC
NR
Model
Ⅰ
96.6%
98.33%
92.5%
97.5%
97.08%
100%
94.17%
Model
Ⅱ
97.78%
98.33%
98.33%
97.92%
97.92%
100%
94.17%
Comp
ared
wi
th the MSVM
model
p
r
op
o
s
ed
by Va
hid
,
the MSVM
model
ado
pte
d
in thi
s
pape
r ha
s hi
gher
re
cog
n
ition accu
ra
cy
and the
ov
erall ave
r
ag
e
recognition
accuracy rea
c
h
97.78%. The
validity of recognition mo
d
e
l pro
p
o
s
ed b
y
this pape
r for qu
ality abn
ormal p
a
ttern
of
dynamic p
r
o
c
ess be verifie
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3595 – 36
02
3602
5. Conclusio
n
Thro
ugh th
e
simulatio
n
ex
perim
ent, a
MSVM re
cog
n
ition mod
e
l
of quality ab
norm
a
l
pattern i
n
dy
namic p
r
o
c
e
s
s i
s
p
r
op
ose
d
in thi
s
pap
er afte
r
com
parin
g diffe
re
nt accu
ra
cie
s
of
several
kerne
l
functio
n
s. F
u
rthe
rmo
r
e, t
h
is
pape
r
co
n
t
rast the
met
hod
we
propo
sed
in thi
s
p
a
per
with the methods commo
nl
y adopted.Th
e simulatio
n
result
s indicate that the proposed algo
rit
h
m
has ve
ry hig
h
re
co
gnition
accu
ra
cy. This hi
gh efficiency i
s
a
c
hi
eved with
qu
ality abno
rm
al
pattern
s of d
y
namic p
r
o
c
ess. Besid
e
s, it offe
rs a novel tech
niq
ue and tho
u
ght for re
al-ti
m
e
quality monitoring a
nd dia
gno
sis in dyn
a
mic p
r
o
c
e
s
s.
Several i
s
su
es
sho
u
ld
be
explored fu
rther.
Fo
r exa
m
ple, it ha
s
very high
re
cognition
accuracy fo
r
singl
e qu
ality pattern
with
RBF
ker
nel f
unctio
n
. Ho
wever, ho
w to
recogni
ze
mi
xed
pattern
s with
RBF ke
rnel fu
nction n
eed
s
further
study.
Ackn
o
w
l
e
dg
ements
This research
was
sup
port
ed by the Nat
i
onal Scie
nce
Foundatio
n o
f
China un
der grants
7127
2207 an
d
6127
114
6.
Referen
ces
[1]
Che
n
N, Zhou
S. Delecta
b
il
it
y stud
y f
o
r sta
t
istical mo
nitori
ng of
multivariate dy
namic pr
ocesses.
II
E
T
r
ansactio
n
s
. 200
9; 41(7): 59
3–6
04.
[2]
Shi J, Z
hou S. Qualit
y
c
ontro
l and im
pr
ove
m
ent for multis
tage s
y
stems: A surve
y
.
IIE T
r
ansactio
n
s.
200
9; 41(9): 74
4–7
53.
[3]
Qiang L, T
i
an
yo
u C. Progre
ss of data-driv
en an
d kno
w
l
edg
e-dr
iv
en pr
ocess mon
i
tori
ng an
d faul
t
dia
gnos
is for in
dustr
y
process.
Control a
nd D
e
cisio
n
.
20
10; 25(6): 80
1–
807
.
[4]
Burges
C. A tutorial
on support ve
ctor machi
nes for p
a
ttern recognition.
Da
ta
Mi
ni
n
g
a
nd Kn
o
w
l
e
dg
e
Discovery.
1
9
9
8
; 2(1): 21-67.
[5]
Shi
y
u
an Y,
De
hui
W
.
Contro
l
Chart P
a
ttern
Reco
gniti
on
B
a
sed
on
PCA
and
SVM.
Jo
ur
nal
of Syste
m
Simulation
. 20
06; 18(5): 1
314
–13
18.
[6]
Xi
ao
xia Z
.
S
u
pport V
e
ctor
Machi
ne
w
i
th
Gauss Ker
n
e
l
Cl
assifie
d
M
e
thod
a
n
d
Mo
del
Sel
e
ctio
n,
Co
mp
uter Engi
neer
ing a
nd Ap
plicati
ons
. 2
0
0
6
; (1): 76-79.
[7]
Xi
ao-k
a
i G. Ine
r
tial n
a
vig
a
tio
n
cond
ition
dat
a
ana
l
y
sis and
e
s
tablis
hment of
dia
g
n
o
stic
mo
del bas
ed o
n
SVM.
Systems Engin
eeri
ng T
heory & Practic
e
.
2012; 2(
32): 405-
410.
[8]
Shao
hu
a J. Monitor
i
ng mo
de
l base
d
on ke
rnel pri
n
ci
pal
compo
nent a
n
a
l
y
sis an
d mul
t
iple su
pport
vector machi
n
e
s
and its ap
plic
ation.
Syste
m
s
Engin
eeri
ng T
heory & Practic
e
.
2009; 2
9
(9): 153-
159.
[9]
Vahi
d R, At
a E
Rez
a
G. Ap
pli
c
ation
of th
e P
S
O-
SVM mode
l for r
e
cog
n
iti
o
n of c
ontro
l ch
art patterns
.
ISA T
r
ansactio
n
s
. 2010; (4
9): 577-
586.
[10]
Cha
ngku
n
W
.
T
he pattern recogn
ition
of
control chart b
a
s
ed on
w
a
ve
let
anal
ys
is an
d SVM
. Chin
a
Mecha
n
ica
l
En
gin
eeri
ng.
20
1
0
; 21(13): 1
572
-157
6.
[11]
Ying
w
a
n
g
X. P
r
ocess M
onitor
i
ng
Base
d
on
W
a
velet T
r
ansform Ker
nel
Pri
n
cip
a
l
Comp
on
ent An
al
ys
i
s
and
Multi
p
l
e
S
upp
ort Vector
Machi
nes.
J
o
u
r
nal
of
Chi
nes
e co
mputer
sy
stems.
201
1; 32(7): 13
48-
135
3.
[12]
Cortes C, Vap
n
ic V. Supp
ort vector net
w
o
rk.
Machin
e Le
arnin
g
.
19
95; (20)
: 1-25.
[13]
Huaz
hon
g W
,
Jinsh
ou Y. St
u
d
y
o
n
the
Ke
r
nel-B
ased
Met
hods
an
d Its
Mode
l Se
lectio
n.
Jour
nal
of
Souther
n Yan
g
t
z
e
U
n
ivers
i
ty
. 200
6; 5(4): 500
-504.
[14]
Cherk
a
ssk
y V,
Ma Y. Practica
l
selecti
on
of S
V
M param
eter
s and
n
o
ise
est
i
matio
n
for SV
M regress
i
on.
Neur
al Netw
or
ks
. 2004; 17(
1)
: 113-12
6.
Evaluation Warning : The document was created with Spire.PDF for Python.