TELKOM
NIKA
, Vol. 11, No. 5, May 2013, pp. 2377 ~ 2380
ISSN: 2302-4
046
2377
Re
cei
v
ed
Jan
uary 8, 2013;
Re
vised Ma
rc
h 4, 2013; Accepte
d
March
14, 2013
APSO-RBF Nonlinear Calibration Method in Carbon
Anode Baking Temperature Measurement
Li Xiaobin
1
, Sun
Haiy
an
2
, Han Co
ngd
ao
1
, Zhang J
i
e
3
1
School of Elec
trical an
d Elect
r
onic En
gin
eer
i
ng, Shan
gh
ai Institute of T
e
chno
log
y
,
Shan
gh
ai 20
02
35, Chi
n
a
2
School of Ecol
ogic
a
l T
e
chnol
og
y an
d Eng
i
n
eer
i
ng, Sha
ngh
ai Institute of T
e
chn
o
lo
g
y
,
Shan
gh
ai 20
02
35, Chi
n
a
3
CECEP W
i
nd-
Po
w
e
r (
X
i
n
ji
an
g)Corp
orati
on, Urumchi 8
3
0
0
0
2
, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: xbl
i
@sit.e
du.
cn
,
hai
ya
nsu
n
@
sit.edu.cn, zj
_j6
036
_cn
@
si
na.com
A
b
st
r
a
ct
A correcti
ng
n
onli
n
e
a
r
errors
of the
ther
mo
coup
le s
ens
or
base
d
on
Ra
di
al B
a
sis F
u
ncti
on
Neur
al
Netw
ork usi
n
g
partic
l
e sw
ar
m
opti
m
i
z
a
t
io
n
are
intro
duc
e
d
. It solves
t
he s
hortco
m
i
n
g of T
h
er
moc
o
upl
e
Sensor
’
s
app
lic
ation o
n
lar
ge
data. T
he resu
l
t
of ex
peri
m
ent
show
s that th
e non
lin
ear ca
li
bratio
n bas
ed
on
APSO-RBF ha
s hi
gher
pr
ecisi
on th
an
the
me
thod
bas
ed
on
RBF an
d ANFI
S. Then, APS
O-RBF is us
ed
to
test fire path temp
eratur
e in th
e ano
de b
a
kin
g
. It is proved that the metho
d
is effective.
Ke
y
w
ords
: car
bon a
n
o
de te
mperatur
e, partic
l
e sw
arm o
p
ti
mi
z
a
t
i
o
n
, radi
al b
a
sis functio
n
n
eura
l
netw
o
rk, N
thermoco
upl
e sensor, no
nli
n
e
a
r cali
bratio
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
Carbon a
nod
e bakin
g pro
c
ess of the anode ca
rb
on bl
ock temperature, flue temp
eratu
r
e,
furna
c
e n
egat
ive pre
s
sure, and the flu
e
g
a
s tem
per
atu
r
e an
d oth
e
r
physi
cal q
uan
tity are the ke
y
para
m
eters e
n
su
ring
ano
d
e
ca
rbo
n
blo
ck
ba
king
q
u
a
lity, wherei
n
the accu
rate
measureme
n
t
and cont
rol of
ano
de
b
a
k
ing
tempe
r
a
t
ure (fire pat
h
an
d the
ca
rbon
blo
c
k)
dire
ctly affect
the
quality perfo
rman
ce p
a
ra
meters incl
u
d
ing the po
rosity, spe
c
if
ic re
sista
n
ce
, compressiv
e
stren
g
th, ele
c
trolysi
s
, oxi
dation
rate
a
nd b
a
ck EM
F of the
ano
de
carbon. A
nd the
a
c
curate
negative p
r
e
s
sure me
asurement an
d control of fu
rn
ace
ch
ambe
r is the
cru
c
ia
l step to e
n
sure
the full
com
bustio
n
of th
e fuel
and
volatile, asp
halt overflo
w
, volatile de
comp
ositio
n
and
evaporation,
con
c
e
n
tration
,
while the
te
mperature
co
ntrol of flue
g
a
s e
m
ission
i
s
the m
a
in fa
ctor
of energy sa
ving and con
s
umptio
n re
d
u
cin
g
. Ther
efore, Establi
s
hment of intelligent dete
c
tion
and
co
rrectio
n
metho
d
fo
r the
s
e
nonli
near phy
sica
l qua
ntities i
s
the
p
r
emi
s
e conditio
n
f
o
r
reali
z
ing a
c
cu
rate co
ntrol of
key para
m
et
ers in a
nod
e baki
ng proce
ss.
The N type therm
o
couple
is comp
ared
with
the orig
inal K, S type thermo
co
u
p
le. The
nonlin
ear e
r
ror
accou
n
ted
for 1
300
E
M
F in
0.4%
,
and th
e n
onli
near e
rro
r i
s
also
far le
ss than
that of K type thermo
co
upl
e in the
scop
e of 20
-40
0
. Beca
use N type therm
o
co
uple h
a
s sh
orter
usa
ge hi
sto
r
y, and its
measurement
prin
cipl
e
is sam
e
a
s
o
t
her the
r
mo
couple, n
onlin
ear
corre
c
tion
sh
ould al
so be
done in
N type thermo
co
u
p
le. At prese
n
t the extensi
v
e research
can
be mainly cla
ssifie
d
into two dire
ction
s
, one is
pi
ecewise line
a
ri
zati
on of nonline
a
r anal
og ci
rcuit
fitting method
, which is m
a
inly used fo
r anal
og
m
e
ters
and te
mp
eratu
r
e tra
n
smitters. Anot
her
adopt
s loo
k
-up table
met
hod by mi
cro
p
ro
ce
ssor to
reali
z
e
n
onlin
ear co
rrectio
n
and col
d
e
n
d
temperature
comp
en
satio
n
. But these
method
s hav
e the fitting and loo
k
-up ta
ble erro
r, as
well
as th
e sen
s
o
r
pe
riph
ery a
n
tioxidant p
r
otecti
on
layer and
othe
r reason
s, in th
e ano
de
ba
ki
ng
unde
r the sp
e
c
ial environm
ent of
actual
have som
e
problem
s.
Curre
n
t corre
c
tion metho
d
s
incl
ude pie
c
e
w
ise linea
rization of no
nlinea
r anal
o
g
circuit
method, a
s
well a
s
the
microprocessor (or
singl
e
comp
uter) lo
ok-up t
abl
e method;
but these
method
s due
to the exte
rnal temp
era
t
ure ra
nge,
therm
o
couple,
thermal re
si
stan
ce of the
cha
nge
s, e
s
p
e
cially in the t
herm
o
couple
to take
certai
n oxidation p
r
otection m
e
a
s
ures
and
oth
e
r
rea
s
on
s, will
cau
s
e th
e therm
o
couple
or the
r
ma
l
resi
stan
ce the
output
characteri
stics of
the
cha
nge,
whe
n
bein
g
u
s
e
d
in hig
h
p
r
eci
s
ion te
mp
eratu
r
e m
e
a
s
urem
ent an
d
cont
rol
syst
em,
can
not meet the actu
al me
asu
r
em
ent an
d control re
qu
ireme
n
ts
[1-4]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2377 – 238
0
2378
2. The Basic
Principle of Thermoco
u
p
l
e Nonlinear
Corre
ction
Gene
ral the
r
moco
uple te
mperature m
eas
ure
m
ent
model can be
expresse
d a
s
:
E =
f
(
T)
(1)
Whe
r
e i
n
E repre
s
e
n
ts th
e
therm
o
coupl
e out
put volta
ge, T d
enote
s
the
r
mo
cou
p
l
e input
(mea
su
red te
mperature
)
.
Usually the
comp
en
sated
output p
c
an be
drawn wh
en o
u
tput E und
ergoe
s a
comp
en
satio
n
model g ( E
).
p = g
(
E)
(2)
If for any E o
u
tput, can be
found on inp
u
t output cha
r
acte
ri
stic curve corre
s
po
n
d
ing to
the input
T, that the
com
p
ensated
o
u
tp
ut with d
e
si
re
d prope
rties,
P = T. Ap
parently g ( E
)
also
is n
onlin
ear,
and th
e a
c
tu
al process i
s
very
compli
cated, it i
s
d
i
fficult to u
s
e
the a
nalytical
expressi
on, therefore, we use APSO-RBF method to
carry on the study, then
replace the g(
E ).
3. Correc
t
io
n Method o
f
APSO-RBF Metho
d
The first through the A
PSO algorith
m
for RBF
netwo
rk
nee
ds to dete
r
mine the
para
m
eters for fast global
optimization,
so we
ca
n define an opt
imized
sea
r
ch spa
c
e, at this
time, the A
PSO algo
rith
m of
RBF
netwo
rk o
p
timization
results a
s
the
initial value
s
of
para
m
eters, and then pl
a
y
the RBF network local s
earch ability, high preci
s
io
n, using g
r
ad
ient
method i
s
o
p
timized fu
rt
her, the
r
eby
to achi
ev
e th
e req
u
ired a
c
cura
cy of
correctio
n
mo
del,
namely APSO-RBF method. Us
ing AP
SO to train
t
he RBF neural network
, tak
e
s
the following
as an in
dicator of fitness functio
n
:
2
,,
11
1
Nm
d
ji
j
i
ij
Jy
y
N
(3)
Among the
m
, the trai
ning
sample
set i
s
D
= {
(yi , xi )| i = 1 ,
2 , , N},
N i
s
the
training
set sampl
e
s;
m is the
num
ber
of hidd
en
layer u
n
its.
For the
RBF
netwo
rk
,
be
cause of its i
n
put
vector compo
nents to a
di
stan
ce fu
ncti
on
com
b
inat
i
on, so the
n
o
r
mali
zed
inp
u
t
data i
s
crucial.
Here, u
s
ing
Matlab tool
bo
x of prem
nmx
functi
o
n
, the
input data
is
norm
a
lized to
(-
1 , 1
)
.
,
d
j
i
y
is
the i sampl
e
s of the j network
output no
de output
s;
,
j
i
y
is a sample
of the j output n
ode
s of the
actual o
u
tput value. The al
gorithm for th
e basi
c
ste
p
s
can b
e
expre
s
sed a
s
follo
ws:
(1)
Initialization:
initial sea
r
ch
point posit
i
on and
spe
ed is u
s
uall
y
in the ran
ge allo
wed
rand
omly ge
nerate
d
PSO
algorithm. T
o
deter
mine
the maximu
m numb
e
r of
iteration
s
of
Tmax = 10
0;
popul
ation
si
ze
30; hi
dde
n
layer cent
er numbe
r
i
s
8; weig
ht
facto
r
C1 = C2
=
2; the weight
function
w fro
m
0.9 lineari
z
ation red
u
ced
to 0.4.
(2)
Evaluation of adaptability: according to
the
training
sample set an
d type (3), evaluation of
every pa
rticl
e
's fitne
s
s, if g
ood
on th
e p
a
rticle
current
fitness valu
e
,
the pb
est i
s
set to
the
positio
n of the parti
cle is u
pdated, an
d the indivi
du
al fitness valu
e. If all the particle
s
of the
individual fitn
ess valu
e of
the be
st b
e
tter tha
n
the
current all
pa
rticle
s of
optim
al ad
aptive
value, the gb
est is
set to the positio
n of t
he particl
e, the particl
e numb
e
r re
cording, a
n
d
update
s
the b
e
st fitness value.
(3)
The parti
cle update: acco
rding
to
APS
O
theo
ry
fo
r ea
ch
pa
rticl
e
velo
city an
d po
sition
updatin
g.
(4)
Output de
ci
si
on: if the cu
rre
nt iterat
io
n numb
e
r
re
ach
e
s the
predetermine
d maximum
numbe
r, then
stop the itera
t
i
on, the output optimal sol
u
tion,
otherwi
se go to ste
p
2.
(5)
The
RBF network
learning: acc
o
rding to
APSO algorithm
pos
i
tioning
of an optimiz
ed
s
e
ar
ch
sp
ac
e, g
r
a
d
i
en
t d
e
s
c
e
n
t
me
th
od to
t
he netwo
rk
paramete
r
s a
r
e mo
difie
d
iterative
algorithm is as follows:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
APSO-RBF Nonlinear Calibration Method in Ca
rbon Anode Baki
ng T
e
m
per
ature…
(Li Xiaobin)
2379
1
12
()
()
(
(
)
(
)
)
((
)
(
)
)
d
j
jj
jj
kk
y
k
y
k
h
kk
(4)
2
2
((
)
(
)
)
j
d
jj
j
j
j
XC
yk
y
k
h
(5)
()
(
1
)
((
1
)
(
2
)
)
j
jj
jj
kk
kk
(6)
2
()
(
(
)
(
)
)
d
j
j
jj
j
x
c
ck
y
k
y
k
(7)
()
(
1
)
((
1
)
(
2
)
)
j
jj
jj
ck
ck
c
ck
ck
(8)
2
2
e
x
p
(
1
,
2
,
...,
)
2
j
j
j
XC
hj
m
(9)
Whe
r
e in
is t
he lea
r
nin
g
speed, i
s
momentum fa
ctor. Here set
= 0. 01,
=
0. 05.
j is the numb
e
r of hidde
n layer units.
(6) E
nd ju
dgi
ng: test
wh
ether th
e m
a
ximum n
u
mb
e
r
of iterations ( or minimu
m
error
), if
achi
eved, will
cea
s
e
op
era
t
ions, if not,
go to
st
ep fif
t
h. The maxi
mum nu
mbe
r
of iteration
s
is
100, the mini
mum error
re
quire
ment is
0.001.
4. APSO-RBF Method Si
mulation an
d Applicatio
n
Becau
s
e AP
SO-RBF i
s
di
recte
d
to a l
a
rge
sa
mple
of the sol
u
tio
n
to the p
r
ob
lem, so
according to
the indexin
g table, fro
m
1-10
00
℃
at 1
℃
take a sam
p
le
, calcul
ate the
corre
s
p
ondin
g
thermo
ele
c
tric pote
n
tial, and to
the
thermal p
o
tential tempe
r
ature a
s
inp
u
t,
output, 1000
grou
ps
of training
sam
p
l
e
s. Use APS
O-RBF meth
od to und
ert
a
ke trainin
g
, and
with the
RBF
,
ANFIS traini
ng meth
od to
co
mpa
r
e. By
simul
a
tion
a
nd
comp
ari
s
o
n
re
sult
s
sho
w
that APSO-RBF method
convergence is faster
, the training preci
s
ion reached 0.001 to
8
gene
ration
(Figure 1
)
, a
nd
RBF
rea
c
h
0.1 to
40
gen
eratio
n,
ANFIS re
ach
ed 0.0
1
to
400
gene
ration.
APSO-RBF network
training
from1.5~
1000.5chec
k
each
1
o
C
sel
e
ct
a sam
p
le,
result
s
as shown in
Figure 2, the error of calibration stability in 0 to0.04
o
C.
The APSO-RBF
c
o
rrec
tion method
is
app
lied to the anode
roas
ting c
hamber
temperature
measurement
,
the
in situ measur
ement
of ne
arly 4
00 g
r
ou
ps u
s
ing
PSO-RBF
method fo
r
the co
rrectio
n
of data
with the
international stan
dard N
type
thermo
co
up
le
measurement
value, erro
r correctio
n
, as
sho
w
n in Fig
u
re 3.
From the graph 3, the APSO-RBF
met
hod fo
r the
correction of t
y
pe N therm
o
couple
measurement
erro
r within
-0.55,
has a
re
latively high accura
cy.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2377 – 238
0
2380
Figure 1. APSO-RBF training res
u
lts
Figure 2. APSO-RBF parit
y
error
Figure 3. APSO-RBF
result
s of application error
5. Conclusio
n
The re
se
arch
takes the a
node ba
kin
g
proc
ess N t
herm
o
couple
temperatu
r
e
sen
s
or
calib
ration fo
r the pre
c
i
s
ion
measuremen
t as the o
b
jective. Throug
h
the actu
al dat
a by colle
ctin
g
in sce
ne id
entify and optimizing the
measu
r
em
e
n
t model pa
ramete
rs, re
alize p
r
e
c
isi
o
n
measurement
of an
ode
b
a
kin
g
temp
erature.
F
o
r N type the
r
mo
cou
p
le tem
p
eratu
r
e
se
nsor
calib
ration
m
e
thod. Acco
rding to the
si
ze of the
sa
mple
size, A
PSO-RBF
N-t
y
pe therm
o
co
uple
corre
c
tion
m
e
thod
s a
r
e
p
r
esented. By
co
mpa
r
ison,
as well as
simulation and application in
temperature precisi
on
measurem
ent of anode
baking process show
s that APSO-RBF correction
methods
c
an s
o
lve to the prec
is
e meas
uremen
t problem. APSO-RBF c
o
rrec
tion methods
are
better than th
e origin
al met
hod.
Ackn
o
w
l
e
dg
ments
This
wo
rk i
s
p
a
rtially
sup
porte
d b
y
Shangh
ai
key
scienti
f
ic re
se
arch
proj
ect
No.11
510
502
700, and sci
ence and technolo
g
y i
nno
vation focus
of SHMEC No.12ZZ1
89, and
Scien
c
e Fo
un
dation of SIT (YJ2
011
-33
)
and (Y
J20
11-22).
Referen
ces
[1]
Z
H
ANG YH, LIU JH. Improve
m
ent of sens
or
’s
performa
nce
b
y
usin
g
BP
n
eutral net
w
o
rk and its
ne
w
algorithms.
Chi
nese Jo
urn
a
l o
f
Sensors an
d Actuators
. 200
2; 15(3): 18
5-1
88.
[2]
T
A
IB MN, NARAYANANAS
WAM Y R. M
u
ltichannel
c
a
libration technique for
optic
al fibre
c
hemic
al
sensor us
ing
a
r
tificial n
eura
l
n
e
t
w
o
r
k.
Sensor
s and Actuator
s B
. 1997; B39
(
123): 65-
37
0.
[3]
DEMPSEY GL, AL IG JS, OLSON B, et al. Control s
ens
or lineariz
a
tion us
ing artificial
neural net
w
o
rks.
Anal
og Integr
ated Circ
u
its and
Signa
l Proces
sing
. 19
97; 13(
3): 321-3
33.
[4]
W
A
NG XD. A
pplic
atio
n
of R
B
F
neur
al
net
w
o
rks i
n
s
ens
or ca
libr
a
tion.
Chin
ese
Jo
urn
a
l
of Scie
ntific
Instrum
e
nt
. 20
03; 24(1): 9
5
-9
8.
[5]
YANG YX, LIU D. Nonli
nea
r calibrati
on fo
r te
mperature sensors
b
a
sed
on ada
ptive
neur
al fuzz
y
inference s
y
stem (ANFIS).
Chinese Jo
urn
a
l o
f
Scientific Instrument
. 200
5; 2
6
(5): 511-
51
4.
Evaluation Warning : The document was created with Spire.PDF for Python.