TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5316 ~ 53
2
3
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.561
1
5316
Re
cei
v
ed
Jan
uary 9, 2014;
Re
vised Feb
r
uar
y 23, 201
4
;
Accepte
d
March 10, 201
4
Algorithm of Multi Sensor Data Fusion Based on BP
Neural Network and Multi-scale Model Predictive
Control
Guo Wa
ng*, Dong
Dai
Dep
a
rtment of computer sci
en
ce and tec
hno
l
o
g
y
, Hen
an Me
chan
ical a
nd El
ectrical En
gin
e
erin
g Col
l
eg
e,
Hen
an
Xin
x
ia
n
g
, 4530
03, Ch
i
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: edu
w
a
n
ggu
o
@
16
3.com
A
b
st
r
a
ct
Multi sensor d
a
ta fusion is the data fro
m
mu
lt
ipl
e
sens
o
r
s and infor
m
ation fro
m
th
e
rel
e
van
t
datab
ase
are
combi
ned,
w
h
i
c
h o
b
tai
ned
j
u
dg
me
nt a
nd d
e
scripti
on
th
at can not achi
ev
e
the
g
oal,
mo
re
accurate a
nd c
o
mpl
e
te by an
y singl
e sensor
. BP neural net
w
o
rk is a kind of artificial n
e
u
r
al netw
o
rk ba
se
d
on error b
a
ck-
prop
agati
on a
l
gorith
m
. It ado
pts addi
ng h
i
d
den
l
a
yer, to e
s
timate th
e err
o
r directly l
e
a
d
ing
layer
of
output
lay
e
r by
the
e
rror o
u
tput. T
h
e p
a
p
e
r pr
ese
n
ts Alg
o
rith
m
of multi
se
nsor
data
fusi
on
ba
se
d
on BP n
eur
al
netw
o
rk and
multi-scal
e
mo
d
e
l pr
edictiv
e c
ontrol. T
he
mu
lti-scale
mod
e
l
pred
ictive co
n
t
rol
can not o
n
ly ob
tain the prev
io
us infor
m
ati
on,
and in
cr
eas
e the flexi
b
il
ity in mo
de
lin
g an
d opti
m
a
l
phas
e.
Ke
y
w
ords
:
BP
neura
l
netw
o
rk, multi se
nsor,
data fusio
n
, multi-scal
e
mod
e
l
pred
ictive con
t
rol
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
Equation of
state, the tran
sfer fu
nctio
n
of
this kind of
traditional
m
odel can be
u
s
ed as a
predi
ctive mo
del. For line
a
r stable obj
ect
,
even the
step respon
se, i
m
pulse re
sp
o
n
se of this
ki
nd
of non
param
eter mo
del, a
l
so
can
be
di
rectly u
s
e
d
a
s
p
r
edi
ction
model. In a
d
d
ition, nonli
n
ear
system, a di
stributed p
a
ra
meter
system
model, a
s
lo
ng a
s
have t
he fun
c
tion, also
can
be i
n
of
the predi
ctive co
ntrol fo
r t
h
is
cl
a
s
s of
system
as a
model, the
r
ef
ore, the
p
r
ed
ictive control
to
brea
k the
stri
ct req
u
irem
e
n
ts on the m
odel struct
u
r
e
of conventio
nal co
ntrol, a
nd more focu
s on
the ba
sis of
the inform
ation a
c
cording
to the
fun
c
ti
onal req
u
ire
m
ents acco
rding
to
th
e most
conve
n
ient way of the model wa
s esta
bli
s
he
d.
BP neural n
e
twork i
s
a kin
d
of two or mo
re than two la
yers BP neu
ral netwo
rk ha
s, inter
layer ne
uro
n
s to achi
eve fu
lly conn
ecte
d, i.e. eac
h
neu
ron in
ea
ch n
euro
n
an
d th
e upp
er laye
r is
con
n
e
c
ted to
the right, a
nd ne
uro
n
la
yer witho
u
t conne
ction. A
typical BP netwo
rk i
s
th
ree
layered fee
d
forward hierarchi
c
al net
work, that
is: input layer, hidd
en layer an
d output layer.
Obtaine
d
fro
m
the sen
s
o
r
data to discri
minate analy
s
is, the p
u
rp
o
s
e is to b
e
tter classify
the data,
and
the ultimate
purp
o
se of th
is rese
a
r
ch is to better se
nso
r
m
ana
ge
ment [1]. Mul
t
i
s
e
ns
or
d
a
t
a
fu
s
i
on
r
e
fe
rs to
d
a
t
a fr
om
mu
ltip
le
s
e
nso
r
s ar
e mu
lti
le
ve
l, mu
lti as
p
e
c
t
an
d mu
lti
levels of p
r
o
c
essing, resulting in
significant ne
w information, and t
h
is info
rmatio
n is a
n
y singl
e
sen
s
o
r
can
n
o
t get. In o
r
d
e
r to
get the
accurate a
n
d
relia
ble
co
nclusio
n
an
d it
is redu
cin
g
t
h
e
potential in the information
processin
g
erro
rs,
often u
s
ed to dete
r
mine the obj
ect from different
angle
s
u
s
ing
multiple se
nsors.
Open l
oop
system is
sta
b
le and
ca
n
not gua
r
ant
ee the sta
b
ili
ty of the closed
-loo
p
system, but
it requires clo
s
ed
-loo
p
stability.
Open loo
p
st
ability is de
termine
d
by the
cha
r
a
c
teri
stics of p
r
o
c
e
s
s
and
cont
rolle
r of potentia
l.
So the multi
-
scale
m
odel predi
ctive co
ntrol
is the
use of
the expressio
n
of
the
sam
e
state
sp
ace
and
stan
dar
d mod
e
l p
r
ed
ictive co
ntrol
of
input and o
u
tput con
s
traints, in ad
dition to
all of the fea
s
ibility of th
e control st
ability
cha
r
a
c
teri
stics of time dom
ain ca
n be u
s
ed multi-scal
e model p
r
edi
ctive cont
rol.
Multi sen
s
or
data fu
sion
i
s
a
combi
nat
ion a
nd
a p
r
oce
s
s of
mul
t
i-sou
r
ce i
n
fo
rmation.
Thro
ugh
coordination,
com
b
inati
on,
com
p
lement
ea
ch
othe
r to
i
m
prove
the effect
iveness of
th
e
system to ob
tain informati
on from a va
riety of
sen
s
ors, p
e
rfo
r
ma
nce tha
n
a si
ngle sen
s
or
has
more g
ood.
But sensors with limited reso
ur
ce
s, and ho
w mu
ch sen
s
or m
anag
ement
has
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Algorithm
of
Multi Senso
r
Data Fu
sion
Based o
n
BP Neu
r
al Netwo
r
k an
d… (Gu
o
Wan
g
)
5317
become the
key data fusio
n
system
perf
o
rma
n
ce.
Th
e pape
r p
r
e
s
ents Algo
rith
m of multi se
nso
r
data fusio
n
b
a
se
d on BP neural n
e
two
r
k and multi-scale model p
r
e
d
ictive co
ntrol
.
2. Using Mul
t
i-sc
ale Mod
e
l Predictiv
e
Cont
rol to Multi Sensor
Data F
u
sion
Method
s
Multi sensor information fusion i
s
mainly
refers
to mul
t
iple informati
on so
urce o
r
sen
s
o
r
informatio
n a
c
qui
sition,
proce
s
sing, i
n
tegrate
d
p
r
o
c
ess, it can
b
e
tter u
nde
rst
and th
e o
b
se
rved
obje
c
t. In recent yea
r
s,
wit
h
the
ra
pid
d
e
velopm
e
n
t o
f
com
puter te
chn
o
logy
and
man
u
factu
r
in
g,
multi-sen
s
or
system for t
he co
mplex
appli
c
at
ion b
a
ckgroun
d of
the emerge
nce of a la
rge
numbe
r of. The com
p
lexity of the military and ci
vili
an fields g
r
o
w
ing, urgent
need to u
s
e
new
techn
o
logy fo
r
comp
reh
e
n
s
ive tre
a
tmen
t of too m
u
ch
inform
ation, i
n
terp
retation
and
evaluatio
n,
whi
c
h ma
ke
s the multise
n
s
or
data fu
si
on theo
ry
to obtain the
co
nsid
era
b
le d
e
velopme
n
t, the
techn
o
logy h
a
s be
en wi
del
y used in ma
ny fields.
Two lin
ea
r q
uadratic o
p
timal re
gulato
r
pro
b
le
m
s
in
optimal control theory
occupie
s
an
importa
nt po
sition, it belong
s to the open loo
p
optimal co
ntroller. T
he so
-call
ed
regul
ator
probl
em, i
s
when th
e
syste
m
of the
cont
rolled
out
p
u
t
deviates from
the e
quilib
riu
m
point, h
o
w
to
desi
gn a pro
per control input sign
al is applie
d to
the system, so
that the output tends towa
rds
equilibrium. If the control
signal
to make a perform
ance index function is m
a
ximum, sai
d
the
regul
ator for t
he optimal re
gulator.
Therefore,
in
predi
ctive
co
ntrol, o
p
timization i
s
n
o
t a
n
off-line,
but
repe
atedly o
n
line; this
is the
mea
n
i
ng of the
rolling optimi
z
ati
on. Subo
ptim
al limitation
s
of this finite ti
me optimi
z
ati
on
goal to ma
ke
it in the ideal ca
se
can
only get
the global o
p
timization
solutio
n
, but the rol
ling
impleme
n
tation can
be e
s
timated
due
to the mo
de
l mismat
ch, ti
me varying, i
n
terferen
ce a
n
d
other
un
ce
rtai
nties, timely
remedy. Alwa
ys put th
e
new
op
timiz
a
tion
b
a
s
e
d
on
pr
a
c
tic
a
l
b
a
s
i
s
,
to
maintain
opti
m
al control
i
n
fact. Fo
r th
e co
mplex in
dustri
a
l p
r
o
c
e
ss, th
e mo
de
l mismat
ch, ti
me
varying, the i
n
terferen
ce
caused by
u
n
certainty is in
e
v
itable, so th
e esta
blishm
ent of the roll
ing
optimizatio
n strategy in finite perio
d but more effe
ctive.
Four threat sensor differe
nt
systems a
dopts di
strib
u
t
ed,
network detectio
n
method
ca
n
effectively de
al with th
e m
odern
war [2
]. In
the
di
stributed netwo
rk det
e
c
tion system,
u
s
u
a
lly
take the foll
o
w
ing m
ann
er:
firstly by ea
ch dete
c
tor to
form lo
cal tra
c
ks; an
d then
the tra
ck
dat
a
repo
rt to
sup
e
rio
r
intelli
ge
nce
cente
r
; finally In
form
ation
Cente
r
o
n
the
lo
cal t
r
acks is a
fusi
on
sub
optimal trajecto
ry tracking by data fu
sion. Mu
lti se
nso
r
targ
et associ
ation me
thod mainly h
a
s
two
kind
of
schem
e, o
ne i
s
m
e
a
s
urem
ent sch
e
me
of a track
asso
ciation;
an
other is to tra
c
k a
track a
s
sociat
ion schem
e.
In the multi
rate si
gnal
proce
s
sing,
often r
equi
re m
u
ltiple sen
s
o
r
s simultan
eo
usly
o
n
different
scal
es of th
e ph
enome
non
or pro
c
e
s
s ob
servation.
Ho
w will
the
se
nso
r
s of diffe
rent
types, differe
nt scale
s
of i
n
formatio
n o
b
tained
by the effective comprehe
nsiv
e is the
co
m
m
on
con
c
e
r
n at work, the multi
-
scale a
nalysis and m
u
lti-scale mod
e
li
ng is an im
p
o
rtant re
se
arch
probl
em.
Feedb
ack forms are diverse, ca
n gua
rantee
a predi
ctive model b
a
se
d on inva
riant, for
the future to predi
ct and
compen
sate th
e error; it
can
also modify the pre
d
ictio
n
model a
c
cord
ing
to the princi
p
l
e of on-line i
dentificatio
n. Whatev
e
r
the
corre
c
tion fo
rm, predi
ctive
control b
a
sed
optimizatio
n is esta
blished
in the actual
system,
an
d tries to ma
ke
more a
c
curate predi
ction
s
on
the future sy
stem dynamics in the optim
iz
ation, a
s
is
sho
w
n by Eq
uation (1
).
n
i
i
n
x
f
x
x
x
L
L
1
2
1
)
;
(
)
;
,
,
,
(
)
(
(1)
Whe
r
e L(x
1
,x
2
) is
c
o
rrec
tion form func
tion, f(x
i
)
is
Mu
lti s
e
ns
or
da
ta
w
h
ic
h p
r
o
v
id
e
s
a
powerful mul
t
i-sou
r
ce dat
a
fusi
on processing
tool
s, it is the
dat
a from
multi
p
le
sen
s
o
r
s
and
informatio
n from the
relev
ant data
base
are
co
mbin
e
d
, whi
c
h
obta
i
ned u
s
in
g a
n
y
single
sen
s
or
whi
c
h
can
no
t be achieved
, the target i
s
more
accu
ra
te and
com
p
l
e
te de
scriptio
ns of ju
dgm
e
n
t
and.
Usi
ng no
n p
a
ram
e
tric
mo
del of the ob
ject (the im
p
u
lse
re
spo
n
se and
step
resp
on
se
model
) for predictio
n, and
the introdu
ction of m
anag
ement in are
a
s such as the long
-sta
n
d
ing
rolling o
p
timization ide
a
, the modeli
ng errors
and u
n
ce
rtaintie
s a
r
e timely feedba
ck
corre
c
t
i
on
usin
g the tre
e
clu
s
te
r erro
r current, forming a
non
- para
m
etri
c
model p
r
e
d
ict
i
ve cont
rol, which
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5316 – 53
23
5318
laid the
foun
dation fo
r th
e devel
opme
n
t of p
r
edi
cti
v
e co
ntrol. B
e
ca
use the
para
m
eters
are
redu
nda
nt, non pa
ramet
r
i
c
mod
e
l ro
bu
st co
ntrol,
bu
t the online computation;
according to
the
curre
n
t outpu
t erro
r on future outp
u
t feedba
ck co
rre
c
tion while th
e cal
c
ulatio
n is sim
p
le, but
the
corre
c
tion a
c
t
i
on co.
Multiscale data fusion of multi-sensor
case
was
sig
n
ificantly different from th
e single
sen
s
o
r
, the latter is mainly according to the ord
e
r
of time by measu
r
ing the blo
c
k in the value
of
the state
block up
date; whi
l
e the
form
er
state blo
c
k e
s
timation i
n
g
eneral can
b
e
ca
rri
ed o
u
t
at
two levels
: the firs
t level is es
timated at
differ
ent times a
c
cordi
n
g
to the time
seq
uen
ce of
the
system i
s
up
dated, the se
con
d
le
vel is
at the same ti
me acco
rdin
g
to the senso
r
mea
s
ureme
n
ts
comin
g
order
state blo
c
ks for the time
up
date, as is
sh
own by Figu
re1.
Figure 1. Usi
ng Multi-scal
e Model
Pre
d
i
ctive Control to Multi Sensor Data F
u
si
o
n
Method
s
Aiming at the
con
s
trai
ned
multivariable
DMC
co
ntroll
er p
r
op
osed
an onlin
e pa
rameter
tuning
strateg
y
: acco
rdin
g to the
predi
ctive output and
t
he paramet
ers
of co
ntroll
er obj
ect
s
(su
c
h
as
weight
coe
fficient, the cl
ose
d
-lo
op e
r
ror contro
l co
e
fficient) ap
pro
x
imate linear
relation
shi
p
, at
each samplin
g time acco
rding to the d
e
viation
of a
d
justin
g the
controlle
r pa
rameters o
n
line
referen
c
e tra
c
e outp
u
t and the expect
ed obje
c
t,
and then use t
he weig
ht co
efficient the new
rolling optimi
z
ation
a new round of,
so that t
he
response tracking refe
rence trajectory the
desi
r
ed cont
roller.
Con
s
trai
nt condition
s a
r
e
impo
sed;
predict
ive cont
rol
a
m
ount o
f
cal
c
ulation
will
be
greatly in
crea
sed [3]. Early
method
s of t
w
o pl
annin
g
are a
s
a
n
onli
ne optimi
z
ati
on st
rategy.
Afte
r
the shift limit
algorith
m
, ca
n effectively
solve
som
e
speci
a
l control
pro
b
lem, the
basi
c
id
ea i
s
to
predi
ct the ti
me after
seve
ral
step
s to li
mit the
output
, which i
s
eq
u
a
l to the
set seque
nce valu
es,
to improve th
e pe
rform
a
n
c
e of the
syst
em. For
thi
s
kind
of probl
em, only eq
u
a
lity con
s
trai
nts,
though the
perfo
rman
ce
index is th
e two ty
pe, the Lagran
ge multiplie
r method to
the
uncon
strain
e
d
optimizatio
n probl
em is
obtaine
d by analytical meth
od, cont
rol strategy.
These
th
ree model
s will survey
three
t
y
pes and
th
ree kind
s
of g
eometri
c obje
c
ts up.
If
the
minimum geomet
ric distance definiti
on
for ea
ch
of
the two g
e
o
m
etric
obje
c
t
s
an
d a
s
ked f
o
r a
geomet
ric obj
ect of
any
po
int to a
nothe
r geo
metri
c
o
b
ject to
ok th
e minim
u
m
o
ne p
o
int of
the
segm
ent le
ng
th, the minim
u
m ge
ometri
c dista
n
ce
can
be
used a
s
measurement
of a
s
soci
atio
n
measure a
measurement
of the.
Me
asu
r
e
the
si
milarity de
gree i
s
to
det
ermin
e
the
t
w
o
measurement
or two mea
s
urem
ent set
s
are fro
m
the same ta
rget.
This i
s
a real-time algorithm
provides the possibility of. Wh
ile the other al
gorithm
s only a
piece of m
e
a
s
ureme
n
t an
d blo
c
k the
relation
ship
b
e
twee
n the
st
ates, thu
s
re
al-time
estim
a
tion
of the difficult
y of system. In this
cha
p
ter, we u
s
e MS
BKF algorith
m
to deal
with multi rate
multi
sen
s
o
r
d
a
ta
fusion, in
o
r
de
r to ea
se
in wr
iting, without conf
usin
g situ
ation, we
use
and
rep
r
e
s
ent
s the estimate an
d the estimati
on error
cova
rian
ce, and d
on't use a
nd t
o
rep
r
e
s
ent.
To
study a
n
d
sai
d
from th
e a
s
pe
cts of
func
tion,
stru
cture
a
nd m
a
thematical
m
odel to
informatio
n fusio
n
model
[4]. Functio
n
model
fro
m
the fusion
process of informatio
n fusion
whi
c
h, descri
bes the mai
n
function, d
a
taba
se,
as
well as the i
n
tera
ction inf
o
rmatio
n fusi
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Algorithm
of
Multi Senso
r
Data Fu
sion
Based o
n
BP Neu
r
al Netwo
r
k an
d… (Gu
o
Wan
g
)
5319
system b
e
tween the vari
ous
com
pon
ents of the
pro
c
e
ss;
stru
ct
ure m
odel
from inform
a
t
ion
fusion, inform
ation fusion
system
soft, hardware,
data flow, system and
external environment
interface; mathematics mo
del is the in
fo
rmation fu
sio
n
and integ
r
at
ed logi
c.
For a sta
b
le
pro
c
e
ss, th
e choi
ce of
an app
rop
r
ia
te control an
d predi
ction
hori
z
on,
makin
g
the a
s
sumption th
at the 1 and 2 set up.
An unsta
ble process, peopl
e need to prove
to
the current st
ate value possible
stability. This i
s
a transform an
d
makes the
control vari
ables to
achi
eve the d
e
sired target value, so t
hat
the unco
n
trol
led mode b
e
comes
stable.
Becau
s
e of t
he com
p
lexity of the degrees of fre
edo
m and the o
p
timal pro
b
le
m is the
numbe
r of in
put variabl
es,
if the cont
rol
time co
ns
tra
i
nts in a finit
e
time a
s
sh
ort a
s
po
ssib
le,
whi
c
h redu
ce
s the am
ount
of comp
utation in the ve
ry
great d
egree
. We ca
n al
so to expre
s
s the
control ve
ctor by a n
u
mbe
r
of pa
ramete
rs, so in
th
e fa
ce of
ce
rtain
variable
s
i
s
li
mited when t
he
control time
is infinite. P
eople
ca
n d
egen
erate
to
pre
d
ict th
ese re
stri
ction
s
will
cau
s
e t
he
perfo
rman
ce
or even
ma
ke the sy
stem
unsta
ble.
To
cont
rol the ti
me length
ch
oice to
a la
rg
e
extent by the desig
n de
ci
sion
s, on
ce i
t
's fixed is maintaine
d
as
a con
s
tant. In recent yea
r
s,
resea
r
chers
have pro
p
o
s
e
d
chan
ge
s of MPC in the
time domain, the main is to
ensu
r
e that the
adju
s
tment time length so as to en
sure the stability of the system.
Perform
a
n
c
e
measureme
n
t system in
MPC or
performan
ce i
n
d
e
x can
not be
fixed in
advan
ce. In
stead of th
e
co
mpen
sation, i
t
is in
ord
e
r to en
su
re th
e
maximizin
g
o
r
minimi
zin
g
t
he
perfo
rman
ce
index. Before
the o
p
timization p
r
o
c
e
s
s i
s
com
p
leted,
the pe
rforma
nce
ind
e
x val
u
e
is un
kn
own. In the u
n
con
s
traine
d MP
C
due to th
i
s
so
lution can b
e
expre
s
sed
as a cl
osed fo
rm,
then we ca
n
according to
the system
para
m
eter
s a
nd weighting
in the obje
c
tive function
to
rep
r
e
s
ent the
s
e p
r
op
ertie
s
. Thus the p
e
rform
a
n
c
e t
heory al
so
can be a
d
ju
sted well. You
can
cha
nge the di
stributio
n of arbitra
r
y pro
p
e
r
ties to chang
e the weig
hting value.
In the model
predi
ctive co
ntrol sta
b
ility in t
he literatu
r
e only pe
opl
e in a time d
o
main i
s
given, the sta
t
e and the ou
tput fr
om a gi
ven initial value to a pred
etermin
ed target value ne
ar.
Whe
n
thi
s
time be
com
e
s larg
er
and
a
pproxim
at
e i
n
finite, then t
he deviatio
n
of the state
and
output to the target value i
s
approxim
ate to zero.
The multi-scale mod
e
l predictive control of
clo
s
ed
-l
oop sta
b
ility is very simil
a
r to the
traditional
mo
del p
r
edi
ctive
co
ntrol
[5]. We
ca
n
u
s
e
Lyapun
ov the
o
ry to
prove
t
he a
pproxim
ate
stability.
The
multi-scale model pre
d
ictive
contro
l d
oes not n
e
e
d
to limit a
n
y
to en
sure
the
stability, so th
e algo
rithm o
n
ly need
s the
final time poi
nt to refe
ren
c
e line
can
be.
This vie
w
ca
n
be
con
s
ide
r
e
d
a
com
p
re
ssion
metho
d
. The
refe
ren
c
e
path i
s
a
filter, whi
c
h
define
s
the t
r
ue
optimal ope
n
-
loop
state will be close to this rang
e of values, so
that the loop is approximate
stability (time
domain
at the en
d poi
nt of N, but
p
e
ople
kno
w
th
ere
exists
on
the input of
the
solutio
n
, the approximate t
he true ta
rget
value, and
th
e sol
u
tion to the equ
ation,
as i
s
sh
own
by
Equation (2).
2
11
cn
ik
i
k
ik
Ld
(2)
Whe
r
e
L is data fusi
on
occu
rs fro
m
t
he existi
ng rese
arch
in the dete
c
tion a
n
d
estimation, u
ik
is po
sition e
s
timation a
n
d
d
ik
is
attribute es
timates
t
hat the three
part. In fac
t,
all
kind
s of fu
sio
n
can b
e
real
ized t
ogethe
r in a
unifi
ed f
u
sio
n
sy
stem
, it is very n
a
t
ural. Th
ere
a
r
e
three mai
n
types of st
ructu
r
e mod
e
l of multi sen
s
o
r
data fusio
n
: centrali
zed, di
stribute
d
, hybrid.
Centralized
structu
r
e i
s
a f
u
sio
n
of
ori
g
inal o
b
se
rvati
on d
a
ta, the
sen
s
o
r
data t
r
an
smitted to
the
fusion
center,
in the
fusi
on
data
calib
rati
on, data
a
s
so
ciation, t
r
a
c
k
/ trace fu
sion,
predictio
n a
n
d
tracking cent
er
executive.
Whe
n
the d
e
tection
and
m
u
ltiple sen
s
ors si
mu
ltane
o
u
sly for m
u
ltiple target tra
cki
ng a
n
d
identificatio
n, reso
urce sche
duling p
r
oblem
mu
st
be solved
between
sensor an
d target
detectio
n
, an
d trackin
g
an
d re
cognitio
n
, which is ho
w under
ce
rtai
n con
d
ition
s
, make full u
s
e
of
the se
nsor
reso
urce
s to
meet the o
p
timal sy
st
em perfo
rman
ce
requi
rem
ents,
becau
se of
the
obje
c
tive env
ironm
ental re
stri
ctions ma
ke
som
e
sen
s
ors
ca
n not
fully its fun
c
tion, for
so
me
purp
o
ses on
spe
c
ific con
s
traints,
sen
s
or usi
n
g
the
propo
sed
in
ad
dition, sen
s
o
r
re
so
urce
s
are
limited.
The multi-scale model p
r
edi
ctive con
t
ro
l (MSMPC) and cl
assi
cal mod
e
l predictiv
e
control (MPC) shari
ng
a l
o
t of theoretical
cha
r
a
c
teristics.
The
r
ef
ore
MSMP
C can be dire
ct
ly
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5316 – 53
23
5320
derived di
re
ctly from the MPC,
it and the time domai
n MPC. This
make
s the sy
stem theo
ry; all
of the MPC can be ad
opte
d
.
3. Rese
arch
of Multi Sen
s
or Da
ta Fu
s
i
on Base
d on BP Neur
al Net
w
o
r
k
BP neural n
e
twork
algo
rith
m: the net
wo
rk i
s
comp
ose
d
of
n
euron
s and neu
ron
weights,
namely: in
put
nod
es an
d
o
u
tput no
de
s
of hidd
en
l
a
yer
nod
es, th
ree p
a
rt
s; ea
ch ne
uron
wei
ghts
are
m
u
tually con
n
e
c
ted strength.
Th
e n
eural
network throu
gh trai
n
i
ng, learning
kno
w
le
dge from
the sam
p
le,
and the
kno
w
led
ge sto
r
e
d
in the con
nectin
g
wei
g
hts value in.
The cl
assification
pro
c
e
ss n
e
u
r
al netwo
rk i
s
divided into two pa
rts,
first
learnin
g
net
work weight
s, weight
s to obtain
the data mod
e
l usin
g the
data for trai
ni
ng the
net
wo
rk a
r
e
kno
w
n
;
then the un
kno
w
n
sam
p
l
e
categ
o
rie
s
on
the basi
s
of the existing n
e
twork
stru
cture an
d wei
g
ht param
eters [6].
Neu
r
al
n
e
two
r
k pre
d
ictive control
is use
d
as
th
e ne
ural network id
entification m
odel to
prod
uce pre
d
i
ctive signal,
and then u
s
i
ng the cont
ro
l vector opti
m
ization te
ch
nique, so as to
reali
z
e the
co
ntrol, predi
ction of
no
nline
a
r
system fu
rther, get the
optimal
control traje
c
tory,
can
also b
e
train
ed to anothe
r as the ne
ura
l
netwo
rk
con
t
roller, so a
s
to approa
ch the time co
ntrol
function, afte
r the
end
of
the traini
ng, to directly
co
ntrol the
co
ntrolled
obje
c
t.
Artificial n
e
u
r
al
netwo
rk i
s
a
pplied i
n
the
followin
g
three a
s
pe
cts: sign
al
processing
and pattern re
cog
n
ition,
kno
w
le
dge e
ngine
erin
g an
d expert sy
stem and p
r
o
c
e
ss
cont
rol, as is sho
w
n by
Equation (3).
mid
j
i
t
j
i
t
j
i
t
j
i
j
i
mid
j
i
sp
p
r
p
g
p
p
p
,
,
1
,
,
4
,
,
)
1
(
)
(
4
1
(3)
W
h
er
e P
i,j
is the BP m
o
d
e
l of the
I/O
probl
em fo
r
a set of
sam
p
les into
a
n
online
a
r
optimizatio
n
probl
em, r i
s
the most
com
m
on optim
i
z
a
t
ion gra
d
ient
desce
nt meth
od. If the neu
ral
netwo
rk
as th
e input to o
u
tput mappi
ng,
this ma
p
p
ing
is a hi
ghly no
nlinea
r map
p
i
ng. The d
e
si
g
n
of a neural n
e
twork expe
rt system focuse
s on t
he
stru
cture and
learnin
g
alg
o
rithm of mo
de
l
sele
ction. Ge
nerally spea
king,
the stru
cture is dete
r
mined a
c
cord
ing to the re
sea
r
ch field and
the pro
b
lem
s
to be solved
. Throug
h th
e analysi
s
of
a large
amo
unt of histo
r
ical data
and
the
data of the
a
n
terio
r
ne
ural
netwo
rk the
o
ry devel
op
ment level, the e
s
tabli
s
h
m
ent of a
su
itable
model, and t
he co
rrespon
ding lea
r
nin
g
algorithm
b
a
se
d on the
model, the n
e
twork lea
r
ni
ng
pro
c
e
ss,
co
nstantly adju
s
t the network
pa
ra
met
e
rs, u
n
til the output re
sults m
eet
the
requi
rem
ents.
Based o
n
th
e com
p
re
hen
sive treatme
nt of all
sensor data, to o
b
tain the fina
l data is
more a
c
cu
rat
e
. Acco
rdin
g
to the differe
nt resear
ch
q
uestio
n
s a
n
d
it is data fusion for differen
t
purp
o
ses. G
enerally spe
a
k
ing, data fu
sion can
be
broa
dly divided into two types. A class of
probl
em
s is to study cha
r
a
c
teri
stics of data throug
h d
a
ta fusion; a feature of the
object of stu
d
y
wa
s rep
eated
determin
a
tio
n
. In this kind
of observatio
n
, each
sen
s
or ca
n be co
nsid
ere
d
simi
lar;
to solve
this
kind of
proble
m
s
ca
n often
be d
a
ta
fu
sio
n
by m
ean
s
o
f
mutual
su
pp
ort b
e
twe
en t
h
e
s
e
ns
or
s
.
In the
co
ntrol
use
d
in
mo
re
BP netwo
rk
a
nd RBF network,
BP
net
work
a
nd RBF
netwo
rk
are ve
ry sh
all
o
w bi
ologi
cal
backg
rou
nd a
nd cl
ose
to th
e perso
n with
ability of non
linear fu
nctio
n
.
A nonlinea
r relation
ship b
e
twee
n outpu
t and netwo
rk
conne
ction o
f
the former, whi
c
h ma
ke
s the
learni
ng alg
o
r
ithm mu
st use the no
nlin
ear meth
od,
thus inevitabl
y exist the proble
m
of local
minimum. F
o
r the RBF
netwo
rk, al
though h
a
s
a uniqu
e op
timal approximation poi
nts,
con
n
e
c
ting th
e network p
o
w
er an
d the
output cha
r
a
c
te
ri
stic
of linear
relationship so that it
can
use line
a
r opt
imization al
g
o
rithm can g
uara
n
tee t
he
global conve
r
gen
ce, but its center p
o
int set
sele
ction i
s
n
o
t easy, as is
sho
w
n by Eq
uation (4
).
2
)
2(
2
2
2
2
L
B
L
B
L
B
i
Z
n
Z
Z
n
Z
Z
n
Z
Z
(4)
Whe
r
e Z
i
is
nonlin
ear fu
n
c
tion, n i
s
th
e dist
ribute
d
stru
ctu
r
e n
o
des fi
rst in
the lo
cal
sen
s
o
r
local pro
c
e
ssi
ng of
the obse
r
vation informat
io
n, L is the local processin
g
results which
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Algorithm
of
Multi Senso
r
Data Fu
sion
Based o
n
BP Neu
r
al Netwo
r
k an
d… (Gu
o
Wan
g
)
5321
are tran
smitte
d to the d
a
ta fusio
n
center,
forming
th
e final glo
bal e
s
ti
mate in the fu
sion
ce
nter. I
n
the dist
ribute
d
structu
r
e, a
s
in the
data f
u
sio
n
c
enter is the ve
ctor
d
a
ta, so the
co
mputation lo
a
d
of the fusion center is greatly
reduced. Ability of
distributed
structure not only has the l
o
cal
tracking,
and
the sy
stem o
v
erhea
d i
s
no
t expensive
a
nd go
od
stabi
lity, so t
he
structure i
s
wi
d
e
ly
use
d
in engi
n
eerin
g [7], as is sh
own by Equation (5).
(5)
Whe
r
e E is t
he neu
ral net
work mo
del, p
i
is neu
ral n
e
twork n
onlin
ear o
b
ject
t
o
sele
ct
different ne
ural network m
odel
s and
structure det
e
r
mination p
r
o
b
lems; a
nd from the no
nli
near
identificatio
n
asp
e
ct, the
r
e
is
suffici
ent
incentiv
e, p
a
ram
e
ter ide
n
tification,
wi
th the n
o
ise
of
s
y
s
t
em identific
a
tion, identific
a
tion algorit
h
m is
fast an
d the conve
r
g
ence pro
b
lem
.
For these two
probl
em
s, an
d it is to the
a
pplication of
nonlin
ear
th
e
o
ry and
optim
ization m
e
tho
d
for the
exist
i
ng
and d
e
velop
m
ent. The
modelin
g alg
o
rithm a
nd t
he control system an
d
neural net
work
conve
r
ge
nce and stability,
as
the co
ntrol
l
ed obj
ect a
n
d
neu
ral n
e
twork are no
nlin
ear, it is
difficul
t
to s
o
lve.
The BP algorithm not only has the inpu
t laye
r, output layer nodes, also can ha
ve 1 or
more
nod
es
of the hidd
en
layer. Fo
r th
e input
sign
al
, the first forward p
r
op
ag
ated to the
n
ode
s
of the hidden
layer, the fun
c
tion, the hid
den no
de
out
put sign
al to prop
agate to
the output no
de,
finally, the output re
sults [8]. The rol
e
of in
centive functio
n
n
ode
s u
s
ually
sele
cted S
type
function
s, su
ch a
s
1 f (x) = Sigm
oid
+ e - x
/Q
parameters of t
he 1 type
of Q to adj
ust
the
excitation fu
nction form. The lea
r
ning
process
of the algorithm
is by forwa
r
d and reverse
transmissio
n
comp
one
nt. In the fo
rward
pro
pag
ation
pro
c
e
s
s inp
u
t inform
ation f
r
om th
e inp
u
t
layer, the hi
d
den laye
r by
layer p
r
o
c
e
s
sing, and
tran
smitted to th
e output l
a
ye
r. Each
layer of
neuron
s in the next layer state affect
s o
n
ly the states of the neuro
n
s.
In the trainin
g
pro
c
e
ss of
BP neural ne
twor
k, the ne
ed for so
me
training p
a
ra
meter is
set, in o
r
de
r to sp
eed
u
p
the p
r
o
c
e
s
s of n
e
two
r
k trainin
g
, im
prove th
e st
ability of net
work
training,
try t
o
a
c
hieve
th
e be
st
effect
of trai
ni
ng. T
o
compl
e
te t
he training
p
a
ram
e
ters,
which
can b
e
call
ed
the training functio
n
is
u
s
e
d
to train the BP neural net
work.
Thro
ugh
data
fusion
and it
is to dete
r
m
i
ne the
ch
a
r
a
c
teri
stics of o
b
ject o
w
n
e
rship. In
su
ch p
r
obl
e
m
s, the
sen
s
or i
s
often
different,
di
fferent chara
c
teri
stic
s of
each sen
s
or is
respe
c
tively studie
d
were
dete
r
mine
d.
This ki
nd
of probl
em due
to
the cha
r
a
c
teristics
of
e
a
c
h
sen
s
o
r
wa
s n
o
t the same,
so th
e u
s
e
of
the mutual
su
pport
and
un
satisfa
c
to
ry, but al
so m
a
y be
due to the ne
glect of a sen
s
or i
s
the deg
ree of su
ppo
rt is relatively low, thus lo
sin
g
some u
n
iqu
e
prop
ertie
s
of the se
nsor by the determi
na
tion of the object.
4. Algorithm
of Multi Sensor Data F
u
sion Ba
sed
on BP Neu
r
al Net
w
o
r
k
and Multi-sc
ale
Model Predictiv
e
Control
The data fusi
on task is the
value of the stat
e, determi
nation, re
sea
r
ch obj
ect o
w
nership.
Each
sen
s
o
r
on the d
e
termined
re
spe
c
tively differe
nt cha
r
a
c
teri
stics of the
study, in orde
r to
determi
ne th
e ch
aracte
rist
ic of the
re
se
arch o
b
je
ct t
h
rou
gh the
sensor, in
pra
c
tical
appli
c
at
ions
often ch
oo
se inde
pen
de
nt cha
r
a
c
teri
zation
re
sea
r
ch
obje
c
t
chara
c
te
risti
c
s indexe
s
were
observed, fro
m
the viewpo
int of proba
bil
i
ty t
heory is regarded a
s
i
ndep
ende
nt of each oth
e
r, at
least rel
a
tive degree is ve
ry low.
The
stand
ard
model
pre
d
i
c
tive time len
g
th co
ntrol i
s
use
d
a
s
the
para
m
eters,
becau
se
the output constrai
nt limits the time choi
ce, if t
he inappropriate
choi
ce of
words
will not produce
the feasi
b
ility probl
em (input constrai
nt is al
ways feasible). Contro
l in each link is the choi
ce
of
time domain
algorithm in
finite simulati
on ma
ke
s the output con
s
traint
wa
s a
l
most meet t
he
predi
ctive m
u
ltiscal
e
mo
de
l. On the
cha
nge
s
of tim
e
and th
e
reference p
a
th tra
cki
ng
MSMPC
set it in time domain fo
r the infinite hori
z
on
ca
se.
The
stand
ard
deviation
as the pe
rcent
age of
ea
ch
numbe
r, it can not
only
variation
betwe
en u
n
its of
com
p
a
r
ison value
s
of
different
va
ria
b
les,
but al
so
ca
n
comp
are the
differen
c
e
betwe
en the
mean va
riatio
n between va
riable val
u
e
s
. Nee
d
to poi
n
t
out, the first
point: variabili
ty
index variatio
n values o
r
disp
ersion, o
ften wi
th position index a
v
erage m
e
th
od, the varia
b
le
value centrali
zed
lo
cation
and dispersio
n
de
gre
e
[9]. The se
con
d
point:
Althou
gh
the
varia
b
ility
i
L
i
i
p
p
E
2
1
0
log
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5316 – 53
23
5322
indicator
spe
c
ie
s, but any
variation in
de
x, its val
ue is greate
r
, that variation i
s
greater, the mo
re
seri
ou
s the
nume
r
ical
stagge
r; said t
he
smalle
r v
a
riation
is small, Equatio
n (6
) i
s
sh
o
w
n
con
c
e
n
trated.
n
i
i
i
n
i
i
b
ax
y
b
a
F
0
2
0
2
)
(
)
,
(
(6)
Whe
r
e F
(
a, b
)
is the mo
del
of BP neural
netwo
rk p
a
ra
meters on th
e netwo
rk, E
i
need to
initialize, y
i
is variability indi
cator. Sin
c
e the syste
m
is
nonlin
ear, the
initial value for lea
r
nin
g
ca
n
achi
eve the local minim
u
m and ca
n conve
r
ge
nce
result
s gre
a
t
relationship
.
An important
requi
rem
ent i
s
: the initial
weig
hts
so th
at each
ne
uron state
valu
es
clo
s
e to
zero i
n
the in
put
accumul
a
tion
, weig
ht is in
(-1, 1
)
ran
d
o
m
num
be
r b
e
t
ween, to
rel
a
tively small.
Also
ho
pe t
hat
input sampl
e
wa
s normali
zed, so
that relatively
larg
e inp
u
t can
still fall in the
transfe
r fu
nct
i
on
gradi
ent.
The
perfo
rma
n
ce
of n
e
two
r
k traini
ng
an
d the t
r
aini
ng
sa
mple
s
are
clo
s
ely li
nke
d
. The
desi
gn
of a
g
ood t
r
ainin
g
set sh
ould
not
e that th
e
sa
mple
si
ze, b
u
t
also
sho
u
ld
pay attention
to
the quality of
the sample.
Becau
s
e
the
road
informat
ion a
c
qui
sitio
n
may be i
n
consi
s
tent in
the
unit, highe
r value a
nd so o
n
, therefo
r
e b
e
fore BP ne
u
r
al net
wo
rk
predictio
n, we
must ta
ke so
me
data pro
c
e
s
si
ng method
s
of normali
zati
on pret
reatm
ent on roa
d
traffic flow da
ta, making th
e
pro
c
e
s
sed d
a
t
a can fall o
n
the (0, 1
)
, to accele
ra
te t
he convergen
ce of trai
ning
netwo
rk,
whi
c
h
facilitates the
BP neural net
work mo
del.
The ro
bu
st stability of MSMPC, or is th
e
system mo
deling e
r
ror o
r
the ability to handl
e
the ch
ang
e o
f
param
eters
in other
wo
rd
s, we
use the
MPC meth
o
d
to deal
with
. The obj
ect
at
different scal
es of model
mismat
ch, using a diffe
ren
t
calcul
ation
stru
cture ca
n
only improve
the
robu
stne
ss, so the op
en
-lo
op optim
al lin
k is to
get m
o
re i
n
form
ation ab
out the
system th
rou
g
h
different
scal
es o
r
freque
n
c
y ban
dwi
d
th
. The expe
ri
ment can
sel
e
ct pa
ram
e
te
rs
adju
s
tment
o
f
MSMPC, ma
ke
s it possibl
e to accu
rate
ly obtain simi
lar to MPC a
s
re
sult
s, so
that it can ha
ve
the same p
e
rf
orma
nce.
This i
s
a dy
n
a
mic
sy
st
em;
we a
ssu
me
t
hat
all States are co
mple
tely obse
r
vab
l
e. No
modelin
g an
d
measureme
n
t error a
nd
external di
stu
r
ban
ce i
n
he
re, the refe
ren
c
e p
a
th follo
wed
by the dynamic mod
e
l is
from a sta
r
tin
g
point to
the
target point
pro
c
e
ss. In o
r
de
r to verify the
prop
osed al
g
o
rithm a
nd it
is sho
w
th
at the
propo
sed
algo
rith
m is effe
ctive and
practi
cal.
Example is a
c
cordi
ng to th
e 4 sensor
da
ta to
determi
n
e
an obj
ect of
study, and fi
nally determi
ne
the resea
r
ch obje
c
t.
Figure 2. De
sign of Multi Sensor Data F
u
sio
n
Base
d on BP Neural
Netwo
r
k a
n
d
Multi-scale
Model
De
sign
of M
u
l
t
i Sensor Data Fu
sio
n
b
a
s
ed o
n
BP
Ne
ural
Netwo
r
k
and
Multi-sca
l
e Mo
del
is sho
w
n by Figure
2.
In
t
he con
s
tru
c
ti
on
p
r
o
c
e
s
s o
f
BP neural
netwo
rk, fo
r
any continuo
us
function in cl
ose
d
interval
can be ap
proximat
ed by BP network with one hid
den layer, an
d a
three laye
r BP neural n
e
twork i
s
to form deci
s
i
on
region
s of arb
i
trary co
mple
xity, arbitra
r
y n-
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Algorithm
of
Multi Senso
r
Data Fu
sion
Based o
n
BP Neu
r
al Netwo
r
k an
d… (Gu
o
Wan
g
)
5323
dimen
s
ion
a
l mappe
d to m
dimen
s
ion.
Therefore, thi
s
pap
er u
s
e
s
only a singl
e
hidden laye
r o
f
three layers
BP neural net
work st
ructu
r
e to c
onst
r
uct
BP neural ne
twork pre
d
icti
on model ba
sed
on time se
rie
s
modeli
ng.
The
pape
r
prese
n
ts Alg
o
rit
h
m of m
u
lti sensor
data
fu
sion
ba
se
d o
n
BP ne
ural
netwo
rk
and multi-scale model p
r
edictive co
ntrol. In
this paper, multi-scale domai
n model predi
ctive
control; a
nd t
hen
gives a
new represe
n
tation m
e
th
od a
nd it
s
ch
ara
c
teri
stics;
the p
r
oble
m
s of
traditional
m
e
thod
s, gives the cal
c
ul
ation metho
d
o
f
parallel. Fi
nally, it is the algo
rithm
and
s
i
mulation.
5. Conclusio
n
In the multi
sensor
data fu
sion
sy
stem,
esp
e
ci
ally in
the large m
u
l
t
i sen
s
o
r
d
a
ta fusi
on
system, the
existen
c
e of
a larg
e nu
mb
er of
ho
mog
eneo
us and hetero
gen
eo
us sen
s
o
r
,
which
can reflect th
e cha
r
a
c
teri
st
ics of the external
wo
rl
d space from different si
de
s. Ho
w much se
nso
r
manag
eme
n
t ha
s b
e
co
me
the
key dat
a fusi
on
syst
em pe
rforma
nce
?
Th
at is ho
w the
se
nso
r
resou
r
ces all
o
catio
n
, in order
to ma
ke the syste
m
achieve the be
st overall perfo
rman
ce.
BP neural n
e
twork (Ba
ck Propagatio
n
Neural Ne
t
w
ork) is a ki
nd of artificial neural
netwo
rk
ba
se
d on erro
r ba
ck-p
rop
agatio
n algo
rith
m. It adopts a
ddi
ng hidd
en lay
e
r, to estimat
e
the erro
r di
rectly lea
d
ing
layer
of out
put la
yer u
s
i
ng the
erro
r output, a
n
d
then th
e e
r
ror
estimation
error
of a layer, so a l
a
yer back
propagation under go to
, will
obtain the
error
estimate
s for all other lay
e
rs. Th
e mult
i-scale
m
odel
predi
ctive co
ntrol
can not
only obtain the
previou
s
info
rmation, and i
n
crea
se the
flexibility in modeling a
nd op
timal phase.
Referen
ces
[1]
Xi
ao
nin
g
D
u
,
Yuge
ng
Xi, Sh
ao
yu
an
Li. An
Efficient Co
nst
r
ain
ed Mo
de
l
Predictiv
e Co
n
t
rol Alg
o
rithm
Based
o
n
Ap
p
r
oximate
Com
putatio
n.
Jo
urn
a
l
of Syste
m
s
Engi
neer
in
g a
nd E
l
ectro
n
ics
. 20
02;
13(1
)
:
42-4
7
.
[2]
XR
Li. C
o
mpar
i
s
on of t
w
o m
easur
ement fu
sion m
e
thods
for Kalma
n
-F
ilter-Base
d
multi
s
ensor
da
t
a
fusion.
IEEE Transactions on Aerospace and Electronic Sys
t
em
s
. 20
01; 37
(1): 273-2
80.
[3]
Xi
ao
nin
g
D
u
,
Yuge
ng
Xi
an
d Sh
ao
yu
an
L
i
. A Com
putati
ona
ll
y Effici
ent
Aggr
egati
o
n
Optimizatio
n
Strateg
y
of Model Pre
d
ictive
Contro
l.
High t
e
chn
o
lo
gy lette
rs
. 2002; 8(2): 68-7
2
.
[4]
Yue
Hou, Y
u
emei M
a
i. C
h
aotic Pr
edicti
o
n for
T
r
affic F
l
o
w
of Im
proved
BP N
e
u
r
al N
e
t
w
ork
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(3): 1
682-
169
0.
[5]
Yan Y
i
n, Gu
o
Peng-fe
i. Ap
pli
c
ation
of th
e P
r
edict
i
on M
o
d
e
l
in
Ch
lori
ne
Re
sidu
al
Co
ncent
ration
Bas
e
d
on BP ne
ural n
e
t
w
o
r
k.
JDCTA
. 2012; 6(2
1
): 444-4
48.
[6]
A Ghazza
w
i
, A Nouh, E Zafiriou.
On
-l
in
e T
u
n
i
ng
Stra
te
gy
fo
r
Mod
e
l Pred
i
c
tive
Co
ntrol
l
er
s.
Journ
a
l of
Process Co
ntrol
. 200
1; 11(3):
265-2
84.
[7]
LiuB
in,
Xi Yu
g
eng. Mor
e
rel
a
xed c
ond
itio
n
s
of
model predictiv
e
contro
l w
i
t
h
g
uara
n
teed sta
b
il
it
y.
Journ
a
l of Co
ntrol T
heory a
nd
Appl
icatio
ns
. 2
005; 2: 18
9-19
4.
[8]
Che
ng Y
u
g
u
i.
Electric En
erg
y
Dema
nd F
o
recast of N
anc
han
g b
a
se
d o
n
Ce
llu
lar G
e
n
e
tic Alg
o
rit
h
m
and BP N
eura
l
Net
w
ork.
T
E
LKOMNIKA Indon
esia
n Jour
nal
of El
ectric
al Eng
i
n
eeri
n
g
. 2013; 11(
7)
:
382
1-38
25.
[9]
W
en Che
ngl
in,
Z
hou Don
ghu
a. Multiscal
e
stochasti
c s
y
ste
m
modeli
ng a
n
d
multiscal
e
R
e
cursiv
e Data
fusion estim
a
ti
on.
Chi
nes
e Jo
urna
l of Electro
n
ics
. 200
2; 11(
2): 192-1
95.
Evaluation Warning : The document was created with Spire.PDF for Python.