TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6127 ~ 6133
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.439
7
6127
Re
cei
v
ed Se
ptem
ber 15, 2013; Revi
se
d Ma
y 7, 201
4; Acce
pted
May 25, 20
14
Application Research Based on Artificial Fish-swarm
Neural Network in Sintering Process
Song Qiang*
1
,
Wang Ai-m
in
2
,
Li-Hua
3
1
Mechan
ical e
n
g
in
eeri
ng d
epa
rtment of
An
y
a
ng institute
of techn
o
lo
g
y
,
An
yan
g
cit
y
of
Hen
an prov
inc
e
, 4550
00
2
Computer sci
e
n
ce de
partme
n
t
of An
y
a
n
g
nor
mal Univ
ersit
y
,
An
yan
g
45
500
0
3
Sintering p
l
a
n
t
of An
y
a
n
g
Steel & Iron Cor
p
oratio
n,
An
yan
g
cit
y
of
Hen
an prov
inc
e
, 4550
04
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: songq
ia
ng0
1
@
12
6.com
A
b
st
r
a
ct
Sinter tu
mbler
strength
is a
n
i
m
p
o
rtant
para
m
eter
in
the
si
nterin
g pr
oces
s, and
has
a
n
i
m
p
o
rtan
t
influ
ence
on th
e perfor
m
ance
of finish
ed si
nter. Artificial fis
h
sw
arm a
l
g
o
ri
thm h
a
ve
go
od
abil
i
ty to acq
u
i
r
e
the gl
ob
al
pe
rforma
nce, th
e ne
ura
l
n
e
tw
ork has
strong
no
nli
n
e
a
r
abi
lity a
nd
l
o
cal
opti
m
i
z
a
t
ion
perfor
m
a
n
ce,; AF
SA+
BP algo
rithm c
o
mbi
n
e
d
w
i
th artificial
fish sw
arm a
l
g
o
rith
m an
d BP
alg
o
rith
m, rea
l
i
z
e
s
the co
mpl
e
me
ntary artificia
l
fish sw
arm
alg
o
rith
m gl
ob
al searc
h
cap
abil
i
ty an
d BP
algor
ith
m
'
s
lo
cal
opti
m
i
z
at
ion c
o
mbi
natio
n of perfor
m
a
n
ce, an artifici
al
fis
h
sw
arm ne
ur
al resu
lt
s sho
w
that the netw
o
rk
combi
natio
n a
l
gorith
m
, it
is
show
n
that c
o
mp
arin
g w
i
th t
he tra
d
itio
nal
BP ne
ural
net
w
o
rk forecasti
ng
met
h
od
，
the p
r
esente
d
forec
a
sting
meth
od
has better ad
aptive a
b
il
ity a
nd can
give b
e
tter forecasti
n
g
results.T
he art
i
ficial fis
h
—sw
arm
alg
o
rith
m netw
o
rk
is train
ed a
nd ch
e
cked w
i
th the
actual
prod
ucti
on
data
．
this al
go
rithm
has stro
ng g
ener
ali
z
a
t
i
on ca
pab
ility,
pred
ictive acc
u
racy i
m
prov
ed
signific
antly,
and
spee
d up th
e conver
genc
e r
a
te, provid
es a
n
effective
met
hod for stren
g
th pred
ictio
n
. W
h
ich be us
ed
for
off-line l
earn
i
n
g
and pr
ed
ictio
n
, a goo
d basis
for the onlin
e
app
licati
on.
Ke
y
w
ords
:
AF
SA, ANN, combin
ation
pr
ed
iction, tu
mbl
e
r strength
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
Artificial fish
swa
r
m al
go
rithm(AFSA),
prop
osed b
y
Dr. Li Xiaolei from Sh
ando
ng
University in 2002, is
a to
p-do
wn
ada
p
t
ion optim
ization algo
rithm
s
enlig
htene
d
by fish swarm
behavio
rs. A
c
cordi
ng to
Dr. Li, thi
s
a
l
gorithm i
s
a
pplied to th
e
followin
g
typical
beh
aviors:
foragin
g
be
h
a
vior, swa
r
m
behavio
r an
d followi
ng b
ehavior. AFS
A
, as a ne
w
efficient ad
ap
tion
optimizatio
n
algorith
m
, ha
s the
adva
n
tage
s--con
curren
cy, sim
p
li
city, quick co
nverge
nce, h
i
gh
optimizatio
n and fast e
s
ca
ping from a l
o
cal o
p
timu
m
.
Later on th
e basi
s
of AFSA, the survival
and
com
petition me
ch
ani
sm we
re i
n
tro
duced to im
p
r
ove AFSA,
makin
g
it a
more
su
cce
s
sful
swarm i
n
tellig
ence alg
o
rith
m. Based
on
the anim
a
l a
u
tonomo
u
s a
gent, AFSA h
a
s b
o
th
striki
ng
advantag
es
and di
sa
dva
n
tage
s. On
the one
sid
e
, it has
su
ch m
e
rits a
s
high
sea
r
ching
efficien
cy, good rob
u
stn
e
ss, good glo
b
a
l conve
r
ge
n
c
e,
less sen
s
itivity to the initial value and
small erro
rs
of inversio
n result
s. But on the ot
her h
and, there ex
ists lo
w optim
ization a
c
curacy,
low
conve
r
ge
nce
sp
eed in
the later p
e
ri
od an
d othe
r deficie
nci
e
s.
Artificial ne
u
r
al net
wo
rk i
s
a
mathemati
c
al
model fo
r th
e brain a
nd it
s a
c
tivities
a
s
well
as
a ma
thematical
ab
stra
ction fo
rm
ed
by the interconne
ction
of a larg
e nu
mb
er of p
r
o
c
e
s
sing unit
s
. Beside
s, it is al
so a l
a
rg
e-scale
nonlin
ear ad
a
p
tive mod
e
l.
Artificial n
e
u
r
al net
wo
rk
is featured
by h
i
gh co
m
putin
g po
we
r,
stro
ng
self- l
e
a
r
ning
ability, adap
tive cap
a
city, nonlin
ea
r m
appin
g
ability
and
goo
d f
ault toleran
c
e.
Therefore, it
has be
en
succe
ssfully u
s
ed i
n
patte
rn re
co
gnition
, image p
r
o
c
essing,
sign
al
pro
c
e
ssi
ng,
system optimi
z
ation, intelli
g
ent cont
rol
a
nd ma
ny oth
e
r field
s
. By
applying AFS
A
to
the stru
ctu
r
e
optimizatio
n
and feat
ure
sele
ction of
neural networks,
this
stu
d
y built stum
bler
s
t
r
e
ng
th
o
p
t
imiz
a
t
io
n mode
l. T
h
is
mode
l h
a
s
n
o
t
o
n
l
y red
u
ce
d th
e comp
utatio
n of the
syst
em,
greatly impro
v
ed predi
ctio
n accu
racy a
nd co
nverg
e
n
t spee
d, but
also obviou
s
ly improved the
gene
rali
zatio
n
of the syst
em. As a go
od re
sul
t, it has a
c
hi
eved
compl
e
ment
ary between
the
global
sea
r
ch
ing ability of AFSA and the local o
p
timization of BP algorith
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 612
7 –
6133
6128
2. Combined
Prediction Model for Ar
tificial Neur
a
l
Net
w
o
r
ks
2.1. Principle of AFS
A
Unli
ke hum
a
n
being
s, fish
don’t have such adva
n
ce
d intelligen
ce
as logical re
aso
n
ing
and syntheti
c
judgment ca
pabilitie
s.
They achieve or express their
aims throu
g
h
the simple
act
of individual o
r
gro
u
p
s
, whi
c
h can be d
e
s
cribe
d
as th
e followin
g
four beh
aviors.
2.1.1. Foraging behav
i
or
This i
s
the m
o
st ba
si
c and
primitive beh
avior of
artificial fish. Besi
d
e
s, it is a b
e
h
a
vior of
food tropi
sm
for fish, whi
c
h ma
ke
s its sele
ction
s
b
y
using the
sight o
r
smel
l to detect the
physi
cal qu
alities or
con
c
e
n
tration in the
water.
Artificial fish
sea
r
che
s
for
food in th
e water throug
h
vision o
r
sm
ell, and
swim
rapi
dly
toward
regio
n
s
with m
o
re food. In
th
e optimi
z
atio
n-o
r
iente
d
p
r
oce
s
s, ba
se
d on
its p
r
e
s
ent
loc
a
tion, artific
i
al fish
s
e
arc
h
es
for more optimal
l
o
cation withi
n
v
i
sual
after fini
te try-num
be
r. If
not foun
d, it will pe
rform random
walk
behavio
r. Th
erefo
r
e, fora
ging
behavio
r is the
artificia
l
fish’s be
havi
o
r of
se
archi
ng for mo
re
optimal
lo
cati
on b
a
sed
on
its lo
catio
n
and
ca
pabilit
y,
spe
c
ifically, the pro
c
e
s
s of searchi
ng for local an
d indi
vidual optimu
m
.
Behavior d
e
scriptio
n: set the cu
rrent
sta
t
e of artificial fish as Xi, sel
e
ct a state at rand
om a
s
i
X
.
()
.
Rand
Visual
X
X
i
j
(
1)
Whe
r
e
Ran
d
( ) re
pre
s
e
n
ts
a ran
dom n
u
m
ber
betwee
n
0 and
1. T
hen
step forward toward that
dire
ction.
()
.
.
1
Rand
Step
X
X
X
X
X
X
t
i
j
t
i
j
t
i
t
i
(
2)
Otherwise, re
sele
ct the
ra
ndom
state
a
t
rand
om
to
determi
ne whether
the co
ndition of
goi
ng
forward i
s
satisfied. If the condition can not
be m
e
t after several
repeated try
-
num
ber, it will
move forward
at random.
()
.
1
Rand
Visual
X
X
t
i
t
i
(
3)
2.1.2. S
w
a
r
m
Behav
i
or
Fish
will natu
r
ally gathe
r in
group
s du
rin
g
sw
im
ming,
and the artifi
cial fish swam
can b
e
viewed a
s
se
veral group
s
of clu
s
ter
cen
t
er. The
s
e
living habit
s
a
r
e
formed to e
n
sure the
su
rvival
of grou
ps
an
d avoid n
a
tural ha
zards. T
he form
ation
of fish swa
r
m
is al
so a vivi
d life exampl
e. It
is g
ene
rally
consi
dered fi
sh do
es not n
e
ed a
lead
er
.
Only if ea
ch
membe
r
of
a
grou
p follo
ws the
local inte
ra
cti
on rul
e
, the swarm p
hen
omeno
n will
stand o
u
t as a whol
e mo
del or th
rou
g
h
individual l
o
cal interaction
.
Fish
swarm follows th
ree rule
s: se
paratio
n rule
-- try
not to
be
overcro
w
di
ng
with neigh
bo
ring p
a
rtne
rs; alignment
rul
e
-- try to mat
c
h the ave
r
ag
e dire
ction
with
neigh
bori
ng p
a
rters; co
he
si
on rule
-- try to move towa
rd the cente
r
o
f
neighbo
ring
partne
r
s.
Behavior d
e
scriptio
n: Fish
in nature
will naturally
gath
e
r in group
s, mainly to prot
ect their g
r
ou
ps
from dan
gers and to su
rvive. In AFSA, r
u
les ove
r
arti
f
i
cial fish a
r
e a
s
follows: 1)
To move toward
the cente
r
of neigh
bori
ng p
a
rters;
2) To
Avoid overcro
w
din
g
.
Set the
cu
rre
nt sid
e
of
arti
ficial fish a
s
Xi, sea
r
ch fo
r the
numb
e
r
of pa
rtners nf
within
visual(dij
≦
Visual),
and
the
ce
nter lo
cati
on X
c
. If
i
c
c
Y
n
Y
it is a
d
e
mon
s
tration that
the
r
e
are
enou
gh food
and spa
c
e in
the cente
r
of the parters
. Then
step forward towa
rd
the dire
ction
of
the parte
rs.
()
.
.
1
Rand
Step
X
X
X
X
X
X
t
i
c
t
i
c
t
i
t
i
(
4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Applicatio
n Rese
arch Ba
se
d on Artificial
Fi
sh
-s
wa
rm
Neu
r
al N
e
two
r
k in…
(Song
Qiang
)
6129
2.1.3. Rear
-e
nd Beh
a
v
i
or
Duri
ng th
e swimmin
g
of t
he fish
swa
r
m, whe
n
on
e
or
seve
ral of
them find
ce
rtain le
ss
cro
w
d
re
gion
with mo
re fo
od, the pa
rtn
e
rs
nea
rby wi
ll follow to re
ach th
e re
gio
n
. If the artificial
fish finds the partner i
n
the optim
um
location
withi
n
perception,
it w
ill move step forward;
otherwise, it
will perform
foragi
ng behavior.
Rea
r
-end be
havior will al
ways a
c
cele
rate artificial fish to move toward the more optima
l
positio
n. Rea
r-e
nd
beh
avior i
s
a
n
inte
rpretation to
be
qui
cker,
stronge
r
an
d faste
r
,
whi
c
h
ensure
s
the g
l
obal optimal
solutio
n
and
conv
e
r
ge
nce and ra
pidity of the algorith
m
.
Behavioral d
e
scriptio
n: Rear-en
d
be
ha
vior is
the
arti
ficial fish’
s
b
e
havior of follo
wing it
s
partne
r
with h
i
ghe
st fitness nearby. Optimization
algo
rithm can b
e
unde
rsto
od a
s
the p
r
o
c
e
s
s of
moving toward the optimu
m
partne
r
ne
arby. Set the
c
u
rrent
s
t
ate of artific
i
al fis
h
i as
Xi, the
partner with maximum
as Xj
within
vi
sual(dij
≦
Vi
sual
). If, it shows t
hat
there are
enough food
in
the cente
r
of Xj which i
s
no
t cro
w
ed. The
n
step forwa
r
d toward the dire
ction of Xj.
()
.
.
1
Rand
Step
X
X
X
X
X
X
t
i
j
t
i
j
t
i
t
i
(
5)
2.1.4. Impro
v
e
d AFS
A
-ju
m
p Behav
i
or
Strictly sp
ea
king, the th
re
e ba
si
c be
ha
vior of
artific
i
al fis
h
swarm belong to
the loc
a
l
optimizatio
n
pro
c
e
ss. If the predi
ction
accuracy d
oes
not cha
nge,
it indicates the ite
r
ative
pro
c
e
s
s ha
s f
a
llen into
lo
cal extrem
e. Since
there i
s
no p
o
int of e
x
ecuting
the i
t
eration, it mi
ght
as well p
e
rfo
r
m jump beha
vior. This stu
d
y attempts
to add the jum
p
behavio
r
for the redu
ction
o
f
predi
ction
a
c
curacy
so
as to obtain th
e iteratio
n p
r
oce
s
s out
of the lo
cal ex
treme. Thi
s
will
undo
ubtedly
incre
a
se th
e po
ssibility of rea
c
hin
g
glob
al o
p
timization
a
n
d
sp
eed
up
the
conve
r
ge
nce
spe
ed a
s
wel
l
. This seemi
ngly negligi
b
l
e
jump b
e
hav
ior can
save t
he artifici
al fish
deep in
cri
s
is.
2.1.5. Contr
o
l
Parameter
Selections
Despite artifi
cial fish’
s
sensit
ivity to
the initial value, it
is still necessary to set control
para
m
eters.
AFSA parameters in
cl
ude the nu
m
ber of attempts (t
ry-n
umbe
r), se
n
s
ing
rang
e(vi
sual
), step
(step), t
he
cong
estio
n
facto
r
(
δ
)
and
th
e
nu
mb
er o
f
a
r
tific
i
a
l
fis
h
(
N
)
.
AF
SA is
tolerant to the
range of pa
rameter valu
e and al
so le
ss
stri
ct in the initial value of the algo
rithm
s
.
In short, the
characteristi
c
of artifici
al fish
swarm
is each of them
will
select the optimum
orientatio
n af
ter compa
r
in
g the results of re
a
r
-end,
swarm
and
foragin
g
. Re
ar-end
beh
avior
focu
se
s o
n
enha
nci
ng th
e rapidity a
n
d
glo
bal
su
p
e
rio
r
ity of al
gorithm
conv
erge
nce,
swarm
behavio
r enh
ances the glo
bal su
peri
o
rit
y
of al
gorithm conve
r
ge
n
c
e on the ea
rly stage and
the
stability on th
e late stag
e. Yet the fora
ging be
havio
r is the
co
re
and foun
dati
on of the wh
ole
algorith
m
, playing a vital role in gu
aranteein
g
the
spee
d, stab
ility and con
v
ergen
ce
of the
algorith
m
and
effectively avoiding the al
g
o
rithm falling
into local extremum.
2.2. Principle of BP Neur
al Net
w
o
r
k
BP neural n
e
twork,
sho
r
t fo
r the
erro
r b
a
c
k p
r
opa
g
atio
n ne
ural
net
work,
con
s
ist
s
of one
input layer, o
ne or more hi
dden l
a
yers a
nd an
output
l
a
yer. Each la
yer is
com
p
o
s
ed
of a nu
m
ber
of neurons.
Just like the n
e
rve cells of
human b
e
ing
s
, these n
e
u
r
ons a
r
e
corre
l
ated with ea
ch
other. The
structure is
sho
w
n in Figu
re
1:
Figure 1. BP
Neu
r
al Netwo
r
k Mo
del
In
p
ut la
y
er
Hidde
n la
y
er
Out
p
ut la
y
er
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 612
7 –
6133
6130
The tra
n
smi
s
sion
of biolo
g
ical n
e
u
r
on
sign
als i
s
a
compl
e
x ele
c
tro
c
he
mical
pro
c
e
s
s
passe
d syna
pse. As fo
r artificial
neu
ral
netwo
rks, this pro
c
e
ss i
s
si
mplified an
d simulate
d a
s
th
e
contin
ued
ch
ange
s
and
u
pdate
s
of
a
set of digita
l
si
gnal
s throug
h
ce
rtain l
e
a
r
ni
ng rule
s. Th
e
s
e
digital si
gnal
s are
sto
c
kpile
d in the
wei
g
ht con
n
e
c
tion
betwe
en n
e
u
r
on
s. The
net
work i
nput lay
e
r
simulate
s
se
nso
r
y neu
ron
s
of the ne
uron sy
stem,
receivin
g the i
nput sample
sign
als. Sign
als
input via the
input laye
r a
nd outp
u
t fro
m
the out
p
u
t layer afte
r the compl
e
x cal
c
ulatio
n in
th
e
hidde
n layer.
Make a
co
mpari
s
o
n
bet
wee
n
the i
n
p
u
t sign
al an
d
the expe
cte
d
output, if there
exists e
r
ror, l
e
t the erro
r
sign
al counte
r-p
rop
agate
s
from the
ou
tput layer to
the input lay
e
r
throug
h the
p
r
ocess of the
hidd
en l
a
yer. In this pr
ocess, the
erro
r is
allo
cated
to all u
n
ites
of
each layer vi
a gra
d
ient de
scent algo
rith
m. T
hen the
error
signal
o
f
each unit
can be obtai
n
e
d
.
Revise the
weig
ht of each
unit ba
sed
on the
error
sign
al, thus the n
e
twork
weig
ht i
s
redi
strib
u
ted.
Whe
n
thi
s
p
r
oce
s
s is finished, t
he i
nput
sign
al
will en
ter the
network a
gain th
rou
g
h
the input layer to repe
at the above pro
c
e
ss.
Thi
s
adju
s
tment
process of
positive sig
nal
prop
agatio
n
and e
r
ror b
a
ck-p
ro
pag
ation
among
wei
g
hts in e
a
ch l
a
yer will
carry
out rep
eatedl
y,
until the network outp
u
t error is redu
ced to t
he acceptabl
e level, or the pre-set numbe
r of
learni
ng i
s
re
ach
ed. Th
e
continuo
us we
ight adj
ustme
n
t pro
c
e
s
s i
s
just th
e n
e
twork’
s le
arni
ng
p
r
oc
es
s
.
2.3. Combined Arti
ficial Fish Neur
al Foreca
sting
Model
Set the pre
d
iction value
o
f
artificial fish swarm a
s
f1, the predi
ction value
o
f
neura
l
netwo
rk
as f2
, the pre
d
ictio
n
value of o
p
t
imal co
m
b
in
ation a
s
fc, p
r
edictio
n erro
r as e
1
, e2, a
n
d
ec, re
spe
c
tively. Define w1, w2 as the
rele
vant wei
g
h
t
coefficient, and w1+W2
=
1, then:
2
2
1
1
f
f
f
c
So the erro
r a
nd varian
ce a
r
e:
2
2
1
1
e
e
e
c
Var(
c
e
)=
)
,
cov(
2
)
(
)
(
2
1
2
1
2
2
1
1
2
1
e
e
e
Var
e
Var
Evaluate Var(ec)with rega
rd to
ω
1, then we have:
)
,
cov(
2
)
(
)
(
)
,
cov(
)
(
2
1
2
1
2
1
2
1
1
1
e
e
e
Var
e
Var
e
e
e
Var
Als
o
ω
2=
1-
ω
1. Since
f1 a
nd f2 a
r
e
predictio
n mo
d
e
ls in
dep
end
ent of ea
ch
other, o
b
viou
sly,
cov(e
1
, e2)=0, note Var(e
1)=
δ
11 a
nd
Var(e
2)=
δ
1
1
, then the wei
ght coeffici
en
t of combinat
ion
predi
ction a
r
e
resp
ectively.
22
11
22
1
22
11
11
2
Acco
rdi
ng to
the minim
u
m
statisti
cal
error th
eory, thi
s
com
b
inatio
n meth
od,
su
perio
r to
a singl
e on
e
,
can yield t
he final com
b
ined p
r
e
d
ict
i
on mod
e
l of artificial fish
swa
r
m n
eural
netwo
rk al
go
rithm.
3. Applicatio
n of Ar
tificial Fish
S
w
a
r
m
Neural Net
w
ork Algorith
m
The stability of sinter tum
b
ler
strength
is
a perform
a
nce index th
at has been put more
empha
si
s by
the enterpri
s
e. It is the
key to
maint
a
in a go
od
run for the
e
n
tire iron-ma
king
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Applicatio
n Rese
arch Ba
se
d on Artificial
Fish
-swa
rm
Neu
r
al Netwo
r
k in…
(Song
Qiang
)
6131
system. Th
e
existing in
sp
e
c
tion met
hod
and me
ch
ani
ca
l e
quipm
en
ts for
sinter tumble
r st
ren
g
th
detectio
n
are
too outdate
d
to meet th
e dema
nd
of
large
-
scale
prod
uctio
n
proce
s
s. Du
e to
equipm
ent problem
s, the i
n
sp
ectio
n
cy
cle is lon
g
e
r
than 12 h
o
u
r
s.
The seri
ou
s lag of inspe
c
tion
results
ha
s g
r
eatly ham
pe
red th
e d
e
vel
opment
of
si
ntering
produ
ction. In p
a
rti
c
ula
r
, when
the
prod
uctio
n
q
uality is found abno
rmal, the sint
er m
a
ster-control room
can’t g
e
t the feedback
timely, unabl
e to adju
s
t sinteri
ng p
r
odu
ction ti
mely nor g
u
ide furna
c
e
prod
uctio
n
. The
investigatio
n
sho
w
s that the majority
of dom
e
s
tic enterp
r
ises
have simil
a
r probl
ems.
This
situation
ha
s se
riou
sly co
nstrai
ned
si
n
t
ering
pro
d
u
c
tion and
cau
s
ed
non
-n
egl
igible lo
sse
s
to
ironm
aki
ng
p
r
odu
ction, whi
c
h ha
s
be
en a
bottlen
ec
k f
o
r the
devel
o
p
ment of
cu
rrent iro
n
ma
kin
g
prod
uctio
n
. T
herefo
r
e, it
is
an u
r
g
ent n
e
ed
for
sinte
r
in
g plants in China to develop the predi
ct
ion
model of
sint
er tumbl
e
r
strength. Only b
y
approx
im
ating or re
achin
g
the inte
rnat
ional a
d
van
c
ed
level in the same i
ndu
stry soon
can t
hese
sinte
r
in
g plants
cre
a
te huge e
c
onomi
c
ben
e
f
its,
redu
ce u
nne
cessary
slag a
nd wa
ste so as to
experi
e
nce hi
gh-te
ch
benefits from
energy savin
g
.
Sinter tumble
r strength i
s
one of the im
portant
in
dica
tors to eval
u
a
te the si
nter quality
and al
so the
reflectio
n
of the sinter’
s
mech
ani
cal
strength, havi
ng a great i
n
fluen
ce on
the
techni
cal
-
e
c
o
nomic in
dica
tor of blast-furna
c
e p
r
o
c
ess. Therefo
r
e sinte
r
tu
mbler
stren
g
t
h
p
r
e
d
i
c
t
io
n is
ve
r
y
imp
o
r
t
an
t. Sin
c
e th
e s
i
n
t
er
in
g
p
r
o
c
e
s
s ha
s
su
ch characte
ri
stics a
s
lo
ng ti
me
delay, stron
g
cou
p
ling a
nd
nonlin
ear, ad
opting conve
n
tional alg
o
rit
h
ms i
s
hard to achi
eve. Even
some intelligence al
gorithm
s, including neural
net
works and
su
pport vector m
a
chi
ne
algorithm,
have formida
b
le sh
ortcomi
ngs. Neural n
e
twor
ks have
both incom
p
arabl
e advan
tages a
nd fatal
disadvantages. On the one si
de, neural
network
s features t
he capab
ility of high-speed
operation, se
lf-learning,
self-ada
ption,
nonlin
ear m
a
pping a
nd e
r
ror-corre
c
tion
. However, they
are al
so ea
si
ly trapped int
o
a local min
i
mum and
ca
nnot extricat
e themselve
s
. Besides, th
eir
weig
hts an
d
threshold
s
a
r
e hard to ide
n
tify.
These
have contain
ed the ap
plication of neu
ral
netwo
rks. Th
e use of
su
pport ve
ctor machine
al
gorithm to
d
e
termin
e kernel fun
c
tion
and
regul
ari
z
ation
function i
s
al
so time-co
n
suming.
The
a
pplication of
ASFA to optimize the n
e
u
r
al
netwo
rk
will greatly improv
e the gl
obal search capa
bil
i
ty of combin
ation algo
rith
m as well a
s
the
local sea
r
ch cap
ability. Furtherm
o
re, the gene
rali
zati
on and ro
bu
stne
ss of the
algorithm al
so
performs
well.
Predi
ction
pa
ramete
rs of A
S
FA are
a
s
f
o
llows:
p
opul
ation size
of artificial swa
r
m
is 50,
sen
s
in
g rang
e 0.8, m
a
ximum movin
g
step 0.56
an
d
co
nge
stion
degree fa
cto
r
3.28. Stru
ct
ure
para
m
eters a
nd perfo
rma
n
c
e pa
ramete
rs of fish sw
arm neural n
e
tworks are: th
e input layer,10;
neuron
s n
u
m
bers
of the
hi
dden
layer,
1
7
; the
output l
a
yer,1; the
st
ructu
r
e
of
arti
ficial fish
swa
r
m
neural networks a
s
10
× 1
7
× 1 simila
r
to the empiri
cal value; init
ial learni
ng rate is 0.3, wh
ich
varies
dynam
ically with th
e further le
arning in
BP networks; the action fun
c
tio
n
slop
e of the
hidde
n layer is 0.5. Through the opti
m
izati
on d
e
si
gn of ASFA, BP neural netwo
rk furth
e
r
cal
c
ulate
s
error
ba
ck prop
agation
for 1
000 tim
e
s.
Used
a
s
p
r
e
d
i
c
tive net
wo
rk for
data te
sti
ng,
the single
sa
mple pre
d
icti
on time is no
more than
1
2
ms. Iterative curve of art
i
ficial fish swarm
neural network is sho
w
n in
Figure1, and
tumb
ler
stren
g
th predi
ction
in Figure 2.
Figure 2. Iterative Curve of
Artific
i
al Fish
Sw
a
r
m N
e
ur
al
N
e
tw
or
k
Figure 3.
T
u
mbler Stre
ngt
h Predi
ction for
Sinter T
u
m
b
ler
Strength
0
1
2
3
4
5
6
7
8
10
-8
10
-6
10
-4
10
-2
10
0
10
2
10
4
8
E
poc
hs
T
r
ai
n
i
ng
-
B
l
u
e
G
o
al
-
B
l
a
c
k
P
e
r
f
or
m
anc
e i
s
1.
3
888e
-
008
,
Goal
i
s
1e-
007
0
50
100
150
200
250
300
55.
5
56
56.
5
57
57.
5
58
58.
5
59
59.
5
T
he out
p
u
t
r
e
s
u
l
t
s
of
t
r
ai
n
i
ng net
w
o
r
k
ti
m
e
s
D
r
um
s
t
r
e
ngt
h
Labor
at
or
y
v
a
l
u
e
E
s
ti
m
a
te
d
v
a
l
u
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 612
7 –
6133
6132
As sh
own in
Figure2, squ
a
re e
r
ror
con
v
ergen
ce
of artificial fish
netwo
rks ai
s
3.85, and
the predi
ctio
n value
of si
nter tum
b
ler
stren
g
th in
Fi
gure
3
i
s
fit to the a
c
tual
values,
with
only
several bi
g a
b
sol
u
te e
rro
rs at th
e poi
n
t
s. The
s
e
ha
ve bee
n en
o
ugh to
meet
the dem
and
s of
sinteri
ng p
r
od
uction.
4. Conclusio
n
This
study se
amlessly co
mbined ASF
A with arti
fici
al neu
ral n
e
tworks to buil
d
artificia
l
fish swarm n
eural n
e
two
r
ks. Thro
ugh th
e applied re
search of sinte
r
tumbler
stre
ngth pre
d
ictio
n
,
the re
sult
sh
ows that thi
s
algo
rithm ca
n not onl
y
re
alize th
e glo
bal optimi
z
ati
on, but g
r
eat
ly
improve the
convergen
ce spe
ed an
d gene
rali
zat
i
on ability.
Specifically, this algo
rith
m is
summ
ari
z
ed as
follo
ws:
Seamle
ss integratio
n of A
S
FA and BP
neural n
e
two
r
ks; thi
s
h
a
s a
c
celerated th
e search
pro
c
e
ss of BP algorithm, ensure
d
the o
p
timal sele
ct
i
on of both no
des a
nd actio
n
function
s of the
hidde
n layer as well as the optimization of
net
work
weig
ht and thre
sh
old, re
sulting
in
compl
e
me
nta
r
y combin
ation of ASFA’
s global
sea
r
chi
ng ability
and BP al
gorithm’
s
local
optimization ability.
Applicatio
n of
artificial fish
swarm n
e
u
r
al
net
wo
rks to
sinter tum
b
le
r stren
g
th pre
d
iction.
This can not
only meet the accuracy
requi
reme
nt of tumbler stren
g
th pre
d
iction b
u
t the
deman
ds
of a fast co
nvergen
ce spee
d
and onli
ne
real-time
cont
rol. With g
o
o
d
relia
bility and
operability, this m
odel
h
a
s p
r
ovid
ed
a scientific
a
nd effective
method fo
r t
u
mble
r st
ren
g
th
predi
ction.
Ackno
w
l
e
d
g
ements
I would li
ke t
o
than
k my adviso
r
, Prof
WANG Ai
-min, for his
guida
nce, co
nce
r
n a
nd
advice in
all
matters.
Without him no
ne of th
is
could have
h
appe
ned. Thi
s
re
se
arch
wa
s
sup
porte
d in
part by a g
r
a
n
t provide
d
b
y
The Natio
n
a
l Natu
ral Sci
ence Fou
nda
tion (60
973
0
51)
and in pa
rt by a grant pro
v
ided by Major scientif
ic a
nd tech
nolo
g
ical proje
c
t in Hen
an Provi
n
ce
(102
102
210
4
24).
Referen
ces
[1]
Che
ng Yo
ngm
ing. On inte
lli
genc
e optim
iz
ation
alg
o
rith
m and its a
p
p
licati
on i
n
co
mmunicati
on.
Shan
do
ng Un
i
v
ersit
y
. 2
010.
[2]
Shan
Xia
o
j
u
a
n
. On the
ap
plicati
o
n
of i
n
tellig
ent c
o
mp
uting
in
net
w
o
rk o
p
timizati
o
n
. Sha
n
d
o
n
g
Univers
i
t
y
. 2
0
0
7
.
[3]
Z
hou Jun
he. On DNA enco
d
in
g base
d
o
n
h
y
brid o
p
ti
mizatio
n
alg
o
ri
thm and AF
SA. Z
hengzho
u
Univers
i
t
y
of Li
ght Industr
y
.
2
007.
[4]
Li Z
h
i
w
u.Impro
v
ement
of AS
F
A
and
its
ap
p
lic
atio
n
in
w
i
r
e
l
e
ss se
nsor
co
verag
e
optimiz
ation
. H
u
n
an
Univers
i
ty
. 201
2.
[5]
Jian
g Ming
ya
n
,
Yuan Do
ngfe
ng. S
y
stem d
e
s
ign of
e
nerg
y
efficient-bas
e
d
w
i
rel
e
ss sen
s
or net
w
o
rks
.
Co
mp
uter Systems.
2
010; (1):
7-12.
[6]
Guo Qia
ng, Z
han
g C
hao, M
o
T
i
anshe
ng.
Appl
ic
atio
n of
ASF
A in h
o
t rolli
ng c
o
il
in
g temper
atu
r
e
pred
iction.
Sci
ence & T
e
ch
no
logy R
e
view
. 2010; 28(
1): 137
-138.
[7]
Niu
Don
g
x
i
ao,
Che
n
Z
h
i
y
e,
Xi
ng M
i
a
n
, etc. Combi
n
e
d
o
p
timum gr
a
y
n
eura
l
n
e
t
w
ork
mode
l of th
e
seaso
nal p
o
w
e
rloa
d forecasti
ng
w
i
th the d
o
ubl
e trendsl.
C
h
in
a Electric Pow
e
r Universit
y
. 2002; (1):
29- 32.
[8]
Li
Xia
o
l
e
i, Qian Ji
xi
n. Res
earch o
n
the
decomp
o
siti
o
n
-coor
din
a
tio
n
-
base
d
ASF
A.
Circuits a
n
d
System
s
. 20
03
. 8 (1): 1-6.
[9]
Ma Jian
w
e
i, Z
han
g Guoli,
Xie Ho
ng, etc. A
pplic
ation o
f
ASF
A on fe
ed-for
w
a
r
d ne
ural n
e
t
w
or
k
optimiz
ation.
C
o
mputer Ap
plic
ations
. 20
04; 2
4
(10): 21 -2
3.
[10]
Song Qi
an
g. On the sinter
chemic
al com
positi
on pr
edic
t
ion mo
del
ba
sed o
n
BP an
d Gre
y
ne
ur
a
l
net
w
o
rks.
Kun
m
i
ng U
n
ivers
i
ty of Science.
2
006.
[11]
Hua
ng
Lin, Z
hou J
i
a
hu. H
y
br
id Optim
i
za
tion
Al
gor
ithm
base
d
optim
al p
o
w
e
r
flo
w
calcu
l
ati
on.
Electricity Jour
nal.
20
11: (2): 111- 1
14.
[12]
Yang L
i
, Liu
Gaofeng, Ya
n
g
Z
h
iji
e, etc.
ASF
A based e
m
pt
y
d
e
tection
.
C
o
m
p
u
t
e
r
Eng
i
ne
e
r
in
g
and
Appl
icatio
ns
. 2
011; 47(
11): 30
- 33.
[13]
Gu Xi
hu
a, Xin
g
Mia
n
, Niu
D
ong
xiao. Mu
ltifactoria
l
Gre
y
n
eura
l
net
w
o
rk-
base
d
lo
ad for
e
castin
g.
East
Chin
a Electric
Pow
e
r
. 2006; (1): 6 -9.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Applicatio
n Rese
arch Ba
se
d on Artificial
Fish
-swa
rm
Neu
r
al Netwo
r
k in…
(Song
Qiang
)
6133
[14]
Jingj
in
g Y
ang,
Benz
he
n Gu
o, Ji
xi
ang
Go
u,
Xia
o
Z
h
ang
. Bion
ic Inte
lli
gent Optim
i
zat
i
on
Alg
o
rith
m
Based o
n
MMAS and F
i
shS
w
arm Alg
o
r
i
thm.
T
E
LKOMNIKA Indon
esia
n J
ourn
a
l
of Electrica
l
Engi
neer
in
g
. 2013; 11(
9): 551
7-55
22.
[15]
Li
w
e
i T
i
an, L
i
n
T
i
an. Lia
oni
n
g
Informati
on I
n
tegr
ati
on
T
e
chno
log
y
En
gin
eeri
ng Rese
ar
ch
Ce
nter of
Internet of things. China.
[16]
A H
y
brid
Clust
erin
g Alg
o
rith
m Based o
n
Improve
d
Artificial F
i
sh S
w
a
r
m.
T
E
LKOMNIKA Indon
esia
n
Journ
a
l of Elec
trical Eng
i
ne
eri
ng.
201
4; 12(5)
: 3786-3
7
9
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.