TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5284 ~ 52
9
2
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.470
8
5284
Re
cei
v
ed O
c
t
ober 1
1
, 201
3; Revi
se
d Febru
a
ry 15, 2
014; Accepte
d
March 12, 2
014
Adaptive Neural Network Approach for a Class of
Uncertain Non-affine Nonlinear Systems
Hui H
U
*
1
, Wa
ng Yingjun
2
, Xilong
Qu
3
1
Dept of Electri
c
al an
d Information En
gi
neer
i
ng, Hun
an Insti
t
ute of Engin
e
e
r
i
ng, Hu
nan
Xi
angta
n
, Chi
na,
2
School of Infor
m
ation & Eng
i
n
eeri
ng, Hen
an
Instit
ute of Science a
nd T
e
chnol
og
y, Xin
x
i
a
n
,
China,
3
Dept of Comp
uter Scienc
e, Hun
an Instit
ute
of Engine
eri
n
g
,
Hunan
Xi
an
gtan, Chi
n
a
*Corres
p
o
n
id
n
g
author, e-ma
i
l
: onl
ym
yh
ui@
126.com
1
, w
a
ngy
i
n
g
j
un
@
s
in
a.co
m
2
, quxil
on
g
@
12
6.com
3
A
b
st
r
a
ct
T
he p
a
p
e
r pr
opos
es a
n
e
w
outp
u
t fee
d
b
a
ck a
daptiv
e t
r
ackin
g
co
ntrol
sche
m
e us
in
g n
eura
l
network for a class of unc
ertain no
n-affine
nonlinear systems that only th
e system
output
variables c
an
b
e
me
asur
ed. The
sche
m
e
ad
opt
s low
-
pass filte
r
to transfor
m
non-
affine
non
l
i
ne
ar syste
m
s i
n
to affine
in th
e
pseu
do-i
n
p
u
t d
y
na
mics. No s
t
ate observ
e
r i
s
empl
oy
ed
a
nd few
ada
pti
ng p
a
ra
meters
to be tune
d a
n
d
Lipsc
h
i
z
ass
u
mpti
on, SPR c
ond
ition
is not
requir
ed.
Onl
y
the outp
u
t error is use
d
in
control l
a
w
s
an
d
weights update laws which
make the syst
em
construct
sim
p
le. Boundedness for the
output tracking
error
and
all
states
in the
clos
e
d
-lo
op syste
m
are g
uar
ante
ed, an
d si
mul
a
tion r
e
sults
have v
e
rifie
d
the
effectiveness
o
f
the propos
ed
appr
oach.
Ke
y
w
ords
:
ne
ural n
e
tw
ork, non-affin
e
no
nli
near syste
m
s, uncerta
in, outp
u
t feedb
ack
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
In the p
a
st
d
e
ca
de, the
a
daptive
control
ha
s see
n
rapid
an
d sig
n
ificant devel
opment
leadin
g
to global stability and asymptoti
c
trackin
g
re
sults for larg
e cla
s
ses of un
certai
n nonli
n
ear
sy
st
em
s.
I
n
r
e
ce
nt
y
ear
s,
f
u
z
z
y
logi
c
co
nt
rol [
1
-8] a
n
d
ad
aptive n
e
u
ral
network [
9
-12] th
at mo
del
the fun
c
tional
mechani
sm
of the h
u
man
brai
n t
hat
ca
n coop
erate
with h
u
man
e
x
pert kno
w
le
dge
have bee
n succe
ssfully a
pplied to ma
ny cont
rol problem
s be
ca
use they ne
ed no a
c
curate
mathemati
c
al
model
s of th
e syste
m
un
d
e
r
control.
Likewi
se, for
a class of no
nlin
ear
contin
uou
s-
time system
s,
adaptive di
re
ct and
i
ndirect control
usi
n
g fuzzy logi
c have be
en p
r
opo
sed in [3,
4]
by using “do
m
inate input
s” con
c
e
p
t. Co
ntrolle
rs in
[3,
4] using a st
ate feedba
ck approa
ch is v
a
lid
if all of the
system
state
s
a
r
e
availa
ble fo
r me
asurem
ent. In
pra
c
tice,
ho
wever, th
e
state
feedba
ck co
n
t
rol doe
s not
alway
s
hold
becau
se sy
st
em state
s
are not alway
s
available. Ba
sed
on [3, 4], ref
e
renc
es
[5, 6] pres
ent a
d
a
p
tive output
control
algo
ri
thms ba
sed o
n
state ob
server
and e
r
ror o
b
server. Mo
st o
f
them deal
with the co
ntrol
probl
em of t
he affine n
onl
inear
syste
m
s.
Ho
wever,
in
pra
c
tice,
the
control m
e
th
ods of
affine nonlin
ear
sy
stems do not alway
s
h
o
ld and
the co
ntrol
m
e
thod
s of the
non
-affine n
online
a
r
syst
ems
are
ne
cessary. And
few results a
r
e
available fo
r
non-affine n
o
n
linea
r
syste
m
s in
whi
c
h
the control
in
put app
ea
rs
in a n
onlin
ea
r
fashio
n. In [7] authors a
d
d
r
essed th
e ind
i
rect a
daptive
fuzzy
control
probl
em of S
I
SO non-affine
nonlin
ear
systems. The ap
proa
ch i
s
ba
sed on t
he ap
p
r
oximation of
the nonlin
ear
plant dynami
cs
by
a f
u
zzy
sy
st
em a
nd t
h
e
n
t
he
co
nt
rol
act
i
on
is
c
o
m
put
ed
ba
sed
on lo
cal
inv
e
r
s
ion
of
t
h
e
f
u
zzy
model. In [8], an indirect a
daptive fuzzy controller
i
s
prop
osed, wit
h
in this ap
proach, the SISO
non-affine no
nlinea
r syste
m
is firstly transfo
rme
d
into an affine form by con
s
i
derin
g a Taylor
seri
es exp
a
n
s
ion a
r
ou
nd
an ope
rating
trajecto
ry
. Ho
wever, the in
dire
ct adaptiv
e approa
ch h
a
s
the drawb
a
ck of the contro
ller
singul
arit
y proble
m
, i.
e., division by
zero may o
ccur in th
e cont
rol
law. In [9], an obs
e
rver-based dire
ct ad
a
p
tive fuzzy-n
eural
control
scheme
is p
r
ese
n
ted fo
r
n
o
n
-
affine nonli
n
e
a
r
system
s. By using im
p
licit func
tio
n
theorem
a
nd Taylor seri
es
expan
sion a
nd
SPR Lyapu
n
o
v theory, th
e sta
b
ility of the clo
s
e
-
l
o
o
p
sy
stem i
s
verified. Recently, in [10] an
output fee
d
b
a
ck-b
ased
ad
aptive ne
ural
co
ntrolle
r
ha
s b
een
p
r
e
s
e
n
ted fo
r a
cl
a
s
s of
un
certa
i
n
non-affine no
nlinea
r sy
ste
m
s
with unm
odelle
d dyna
mi
cs whi
c
h
redu
ce the
co
mplexity of control
desi
gn. But i
n
the
scheme
,
a low-pa
ss f
ilter is
de
sign
ed to ma
ke
the e
s
timation
error
dynami
c
s
satisfy the strictly positive-real
(SPR) co
ndition so that they can use Meyer-K
al
mon-Ya
ku
bo
vitz
(MKY) lemm
a, which m
a
ke
s the st
ability anal
ysis of the close
d
-lo
op system and real
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Neu
r
al Net
w
o
r
k A
ppro
a
ch for a
Class
of Un
certain Non
-
affine No
nlinea
r… (Hui
HU)
5285
impleme
n
tation very
com
p
licate
d
. And
the pa
ram
e
ters
of filter a
r
e h
a
rd
to be
cho
s
e
n
. In [11],
output feedb
ack tra
cki
ng
control sch
e
me is i
n
vestigated for a
cla
ss of u
n
c
ertai
n
no
nli
near
system
s. Th
e
distin
gui
she
d
a
s
pe
ct of
the al
gor
ith
m
is that
no
Lip
s
chitz
assu
m
p
tion a
nd SP
R
con
d
ition a
r
e
employed
which m
a
kes t
he syste
m
construct
simp
le. But the obse
r
ver m
u
st
be
employed.
In orde
r to
si
mplify the d
e
sig
n
of
co
n
t
roller, i
n
[12
], an outp
u
t
feedba
ck-b
ased
adaptive ne
u
r
al controll
er
has b
een p
r
o
posed for
a
cla
ss of un
ce
rtain
nonli
n
e
a
r sy
stems.
No
state observe
r wa
s employ
ed in the algo
rithm and onl
y the output e
rro
r wa
s used
in control la
ws
and weight
s update la
ws.
Based
on
the
above
ob
se
rvation, a n
o
vel sy
stematic
desi
gn
pro
c
e
dure
is devel
oped
for
non-affine no
nlinea
r syste
m
s witho
u
t state obse
r
ve
r to simplify the de
sign of
control
syst
em.
First, a
lo
w-p
a
ss filter i
s
e
m
ployed to
transfo
rm th
e n
o
rmal
form
n
on-affine
no
nl
inear sy
stem
into
affine in the
pse
udo
-inp
ut dynamics.
No
stat
e ob
server i
s
e
m
pl
oyed an
d th
e neu
ral
wei
ghts
update
la
ws i
s
tun
ed
acco
rding to
only t
he o
u
tput
tra
cki
ng
error.
T
he
stability a
nalysi
s
d
epe
nds
heavily on the unive
rsal
function a
p
p
roximatio
n
prop
erty, onl
y one RBF
N
is empl
oye
d
to
approximate t
he lump
ed u
n
ce
rtain n
onli
near fu
ncti
o
n
. There are n
o
re
stri
ctive condition
s on t
he
desi
gn con
s
tants. The
pro
posed sch
e
m
e
has fe
w ad
apting pa
ram
e
ters to
be tu
ned an
d Lip
s
chi
z
assumptio
n
, SPR con
d
itio
n are not req
u
ired.
The pap
er i
s
org
ani
zed
as follows. First,
the problem is formulated in
Section II.
Adaptive neural net
work
controlle
r design i
s
given in III. In Se
ction IV, stability analysis is
inclu
ded. Sim
u
lation results are present
ed to co
nfirm
the effectiveness a
nd ap
plica
b
ility of the
prop
osed met
hod in Sectio
n V. Finally, concl
u
si
on
s are inclu
ded.
2. Problem Formulation
The following notations and def
initions
will be used
extensiv
ely throughout this paper.
Let
R
be the re
al numbe
r, a
nd
n
R
r
e
pr
es
en
t th
e
r
e
a
l
n
-vec
tors
.
k
den
ote
s
the usual E
u
clid
ean
norm of a vector
k
. In c
a
s
e
where
k
is a scalar,
k
den
otes
its absolute value.
We co
nsi
der
the followin
g
non-affine no
nlinea
r syste
m
:
1
1
1
,
,
1
(
,
)
ii
n
x
xi
n
xf
x
u
yx
(1)
Whe
r
e
yR
,
uR
are the outputs a
nd input
of the system a
nd
1
[,
,
]
Tn
n
x
xx
R
is the
system state
vector. The smooth functio
n
()
f
is unkno
wn
. The states
are not mea
s
urabl
e, only
y
is available fo
r cont
rol de
si
gn.
For the controllability issue, the following assumption must be made.
Assu
pmtion
1
: T
h
e va
lu
e
o
f
f
u
is no
nzero.
Witho
u
t lo
ss
of gen
erality, we
assu
me t
hat
for all
n
x
R
,
0
u
f
f
u
.
The control o
b
jective i
s
to
desi
gn a
n
a
daptiv
e ne
ural network co
ntrolle
r for a
cla
ss
of
non-affine no
nlinea
r syste
m
s (1
)
su
ch that the syste
m
output
y
follows a de
sired
trajecto
ry
d
y
,
while all
sign
als in the cl
osed-lo
op sy
ste
m
are bo
und
ed.
In the foll
owi
ngs,
we
will a
dopt lo
w-pa
ss filter
to tran
sform
(1) into
affine i
n
the
pse
udo
-
input dynami
cs
[13]
. The overall scheme
is illustrated i
n
Figure 1.
The tran
sfe
r
functio
n
of the low-p
a
ss filter is:
()
Ls
s
(2)
Whe
r
e
is
a p
o
sitive d
e
sig
n
co
nsta
nt. Th
en, altho
ugh t
he p
s
e
udo
-co
n
trol
p
u
sho
w
s a
ch
at
t
e
ring
phen
omen
on
due to a swi
t
ching fun
c
tio
n
, the actual
control input
u
applie
d to the real pla
n
t is
smooth b
e
ca
use
u
is made
by low pa
ss fi
ltering of
p
u
,
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: 5284 – 52
92
5286
p
uu
u
(3)
Figure 1. Basic Idea for S
m
oothing
Co
ntrol
We d
e
fine the au
gmented
state variable as
(1
)
(
)
12
1
=
,
,,
,
,
,,
,
T
T
nn
m
nn
x
xx
x
y
y
y
y
R
and
=,
T
m
x
uR
with
1
mn
. Then:
1
1
1
1
1
1
1,
2
,
,
1
(,
)
:
(,
)
=
ii
nn
n
n
i
i
i
n
ip
i
i
xx
i
n
xf
x
u
x
xf
x
u
ff
xu
xu
f
ff
x
uu
x
uu
(4)
If we define the functio
n
s
()
a
and
()
b
as
:
1
1
()
n
i
i
i
f
f
ax
u
x
u
()
f
b
u
(5)
Whe
r
e
()
b
is n
o
n
ze
ro
and
p
o
sitive a
c
cording to A
s
su
mption 1. T
hus th
ere ex
ist po
sitive
con
s
tant
0
b
su
ch that
0
()
bb
for all
m
R
.And we ca
n see that the origin
al
n
th-order n
on-
affine nonlin
e
a
r sy
stem be
come
s the
m
th-o
rde
r
affine
in the pseu
d
o
-inp
ut nonlin
ear sy
stem:
1
1,
2
,
,
()
(
)
ii
mp
x
xi
n
xa
b
u
(6)
3. Adap
tiv
e
Neur
al Net
w
ork Con
t
roll
er Desig
n
Define t
he
re
feren
c
e ve
cto
r
()
[]
nT
m
dd
d
d
yy
y
y
R
.
The refe
re
nce
si
gnal
d
y
an
d
its time
deriv
ative are
a
s
sumed to
be
smooth a
nd
b
ound
ed.
We
also
defin
e th
e tra
c
king
error a
s
d
ey
y
and co
rresp
ondin
g
e
rro
r vector as
()
[,
,
]
nT
m
d
ey
e
e
e
R
. A filtered t
r
ackin
g
error is d
e
fine
d as:
1
1
m
T
d
s
ee
e
dt
(7)
Whe
r
e
0
is a
desi
gn
con
s
tant.
12
,(
1
)
,
,
(
1
)
T
mm
mm
,
1
T
. The time
derivative of
s
is de
rived a
s
:
out
put
C
ontr
o
l
input
y
u
p
u
s
Real
plant
Pseudo-
cont
rol
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Neu
r
al Net
w
o
r
k A
ppro
a
ch for a
Class
of Un
certain Non
-
affine No
nlinea
r… (Hui
HU)
5287
()
()
1
()
11
11
()
()
()
(
)
Tm
m
d
TT
m
dd
p
T
p
se
y
y
y
ya
b
u
ab
u
v
(8)
W
h
er
e
1
0
T
T
,
()
11
mT
dd
vy
y
.
F
o
r
t
he
sy
st
em
(1)
s
a
t
i
sf
ie
s
Assu
mption
1, if the id
e
a
l
control is de
si
gned a
s
:
*1
()
()
()
m
p
av
uk
t
e
b
(9)
Whe
r
e
1
()
2
kt
is a
desi
gn pa
ra
meter,
12
()
(
)
()
TT
aa
b
k
,
12
1
2
,,
vv
v
v
b
v
,
22
T
d
vk
y
,
2
2
0
,
(1
)
,
(1
)
,
1
T
m
mm
, Then,
s
converg
e
s to zero.
Proof:
Con
s
i
der th
e Lya
p
u
nov fun
c
tion
2
1
2
s
Vs
. Taki
ng th
e time de
rivative of
s
V
along
(8) y
i
eld
s
:
11
1
11
1
22
2
()
()
()
()
()
(
)
()
()
(
)
(
)
()
(
)
(
)
T
sp
Tn
mT
T
d
Vs
s
s
a
b
uv
av
s
ab
k
t
e
v
b
sb
k
t
e
b
k
b
k
y
bk
t
s
(10
)
Acco
rdi
ng to the Lyapu
nov theor
e
m
, the results impli
e
s that
li
m
0
t
s
.
Ho
wev
e
r,
()
a
()
b
are
unkn
o
wn in ideal controll
e
r
(9
),
an
d the state vecto
r
can n
o
t
be mea
s
u
r
ed.
*
p
u
is not availa
ble. The idea
l controll
er (9) can be
rewrit
tened a
s
:
*1
*
()
m
ad
uk
t
e
u
(11)
Whe
r
e
*
()
()
ad
av
u
b
is an unkno
wn fun
c
tion.
In this paper,
a radial ba
si
s function
(RBF) neur
al n
e
twork (NN) is use
d
to ca
pture the
unkno
wn no
nlinea
rity
*
ad
u
in (11). In ge
neral, the ou
tput
of the
multiple-i
nput
-sin
gle
-
output
R
B
F
N
N
is
des
cr
ib
ed
b
y
:
()
()
T
hW
(12)
Whe
r
e
()
hR
is
the RBFN output,
L
WR
is the adjust
able pa
ram
e
ter vecto
r
,
1
()
:
nL
RR
is a nonli
nea
r vector fun
c
tion of the inpu
ts with
L
bein
g
the numb
e
r of
RBFs. The
i
t
h
element of
W
,
,1
,
,
i
iL
, is the syna
ptic wei
ght be
tween the
i
th
neuron in the
hidde
n layer
and outp
u
t
neuron an
d
()
i
is a Gau
s
sian
function in th
e form of:
2
()
e
x
p
2
i
i
i
(13)
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: 5284 – 52
92
5288
Whe
r
e
i
is a m-dime
nsi
o
n
a
l vector represe
n
ting the
cente
r
of the
i
th basi
s
functi
on and
i
is
the varian
ce repre
s
e
n
ting the sp
rea
d
of the ba
sis fun
c
tion.
The key adv
antage
of RBFN
i
s
th
at it
has
the
capability to
a
pproxim
ate n
online
a
r
mappin
g
s to
any degree of
accura
cy. So:
**
()
()
()
T
ad
av
uW
b
(14)
Whe
r
e app
ro
ximation
erro
r
sat
i
sf
y
0
,
11
1
2
(
)
,(
)
,
,(
(
1
)
)
,
(
)
,
(
)
T
yt
y
t
d
y
t
m
d
v
t
v
t
is
the inp
u
t vector to th
e
RB
FNN a
nd
1
0
d
is
a po
sitive tim
e
del
ay .
*
W
is
an id
eal
para
m
eter
vector
which minimizes the
function
an
d
be defined a
s
:
**
arg
m
i
n
s
u
p
(
)
T
ad
W
WW
u
(15)
Whe
r
e
|
WW
,
0
is the desi
gn
con
s
tant. So the neural
network ou
tput
feedba
ck co
n
t
roller
can b
e
descri
bed a
s
:
1
ˆ
()
m
pa
d
uk
e
u
(16)
Whe
r
e
ˆ
ˆ
()
()
T
ad
uW
is the
output of RB
FNN,
ˆ
W
is the e
s
timated valu
e of the optimal wei
ght
*
W
.
The ada
ptive law for the e
s
timated para
m
eter
s of the NN i
s
determi
ned a
s
the followin
g
:
ˆˆ
()
We
e
W
(17)
Whe
r
e
ada
p
t
ive gain
,0
an
d
the
e
-modification term is in
tro
d
u
c
ed
to im
prove the
robu
stne
ss of adaptive law in the presen
ce of
the app
roximation e
r
ror, and the
r
e
exists compa
c
t
s
e
t.
ˆ
ˆ
|
m
WW
(18
)
Whe
r
e
()
m
,
m
is
a cons
tant. If
ˆ
(0
)
W
, then
ˆ
()
,
0
Wt
t
.
Proof:
Co
nsi
der
the Lya
punov
fun
c
ti
on
1
ˆ
ˆ
2
T
VW
W
.
The time de
rivative of the
function
V
alon
g (19
)
is de
rived as:
2
2
1
ˆˆ
ˆ
ˆ
()
ˆ
ˆ
ˆˆ
ˆ
ˆ
(
)
TT
T
m
VW
W
W
e
e
W
We
e
W
We
e
W
eW
W
(19
)
Thus
, it follows
that if
ˆ
m
W
then
0
V
.
So
ˆ
()
,
0
Wt
t
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Neu
r
al Net
w
o
r
k A
ppro
a
ch for a
Class
of Un
certain Non
-
affine No
nlinea
r… (Hui
HU)
5289
From (8), the
time derivative of the
filtered tracking e
r
ror ca
n be de
ri
ved as:
11
**
11
*
11
1*
2
()
(
)
(
)
()
()
()
()
(
)
()
()
(
)
()
ˆ
()
(
)
()
(
)
T
T
ip
T
pa
d
a
d
T
pa
d
mT
T
T
sa
b
u
v
ab
u
b
u
b
u
v
av
ab
u
b
b
u
v
b
bk
e
W
W
k
e
bk
s
W
(20
)
Whe
r
e
*
ˆ
WW
W
.
4. Stabilit
y
Anal
y
s
is
We a
r
e no
w ready to pre
s
e
n
t our main th
eore
m
whi
c
h
is sum
m
ari
z
e
d
in Theo
rem
1.
Theorem 1:
Con
s
id
er
th
e pure
-
fee
dba
ck
n
online
a
r
system (1)
wit
h
the
cont
roll
er in
put
(16
)
and ad
a
p
tive law (17
)
. Then, all the signal
s in the clo
s
ed
-loo
p
system are b
ound
ed and t
h
e
state vecto
r
x
remains in:
0
()
|
(
)
2
,
1
,
2
,
,
0.5
ii
m
m
xi
b
x
te
t
i
m
k
,
tT
Whe
r
e
*
m
bW
.
Proof:
Let th
e Lyapu
nov functio
n
2
1
2
s
Vs
. Taking the time
d
e
rivative of
s
V
according
to (20), we get:
2
2
0
2
0
()
(
)
()
()
()
()
()
()
()
()
T
s
Vs
s
b
k
s
W
s
bk
s
b
W
s
b
s
bk
s
b
W
s
b
s
bk
s
b
s
W
(21
)
Since
ˆ
W
is boun
ded a
s
sh
own in (18
)
, it follows that
Wb
,
*
m
bW
, th
en:
2
0
()
()
sm
Vb
k
s
b
s
b
(22)
From the ine
quality
22
()
2
, it
follows
that:
22
2
0
2
2
0
2
0
()
0
.
5
(
)
(
)
()
()
(
0
.
5
)
2
()
2
(
)
(
0.5
)
4(
0.5
)
sm
m
m
s
Vb
k
s
b
b
s
b
bk
s
b
bk
V
k
(23
)
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: 5284 – 52
92
5290
Let
2
0
()
4(
0
.
5
)
m
ss
b
VV
k
an
d n
o
te
2
0
()
4(
0.5
)
m
b
k
is
co
nstant. Usin
g t
he comp
ari
s
on
princ
i
ple, it follows
that:
0
2(
0
.
5
)
(
(
)
)
(0
)
t
kb
d
ss
VV
e
(
2
4
)
Therefore,
0
22
2(
0
.
5
)
(
(
)
)
00
()
(
)
(0
)
4(
0
.
5
)
4(
0
.
5
)
t
kb
d
mm
ss
bb
VV
kk
e
(25
)
Since
0
()
0
bb
and
0
2
2(
0
.
5
)
0
()
0
4(
0.5
)
kb
t
m
b
k
e
,
it follows
that:
0
2
2(
0
.
5
)
0
()
(0)
4(
0
.
5
)
kb
t
m
ss
b
VV
k
e
(26)
Therefore,
0
2
2(
0
.
5
)
22
0
()
(0)
2(
0
.
5
)
kb
t
m
b
ss
k
e
(27)
From the a
b
o
v
e equation,
s
is bou
nde
d an
d it implies that
x
is bound
e
d
. Followin
g
[14],
the state vect
or
x
will remain in
x
for all
tT
.This
c
o
mpletes
the proof.
5. Simulation Stud
y
In this part, the followin
g
non-affine n
online
a
r sy
stem is sim
u
la
ted to illustrate the
effectivene
ss of the propo
sed a
daptive
neural network output feed
back tra
c
king
controller. T
h
e
non-affine no
nlinea
r syste
m
is de
scribe
d as follo
ws:
12
23
2
21
2
1
0.
15
0
.
1
1
si
n(
0.
1
)
xx
x
xu
x
u
u
yx
(28)
The tra
cki
ng
obje
c
tive is to make the
system outp
u
t
y
follow the
desi
r
ed traje
c
tor
y
0.5
s
in
0.5
d
yt
(
)
.
In (28),
0
f
u
whi
c
h
sati
sfy the
a
s
sumption.
T
he
simul
a
tion
pa
ramete
rs
are
sele
cted
as
follows:
20.
0
,
2.0
,
22.
0
k
.
T
he ad
aptive
gain
95
.
0
,
0.
02
. A
c
cording to
the
desi
gn proce
ss, we ca
n ge
t controlle
r an
d weig
hts up
date law a
s
follows:
20
20
p
uu
u
2
ˆ
22
2
p
ue
W
ˆˆ
95
(
0
.0
2
)
We
e
W
The syste
m
initial conditio
n
s are
12
(0)
0
,
(
0
)
0
xx
. The simulation re
su
lt using MATL
AB
is sh
own in Figure 1
-
4.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Neu
r
al Net
w
o
r
k A
ppro
a
ch for a
Class
of Un
certain Non
-
affine No
nlinea
r… (Hui
HU)
5291
Figure 1. Plots of Output Tr
acking of System
Figure 2.
Plots of the Weig
hts No
rm
Figure 3. Plots of Cont
rol Input
Figure 4. Plots of Output Erro
r
Figure 1 an
d
Figure
4 sho
w
s th
e re
sult
s of out
put tracking. It can
be se
en that
the actual
trajecto
ry con
v
erge
s rapidl
y to the de
sired on
e.
The
weig
hts n
o
rm
is
sho
w
n i
n
Figure 2
and
the
boun
ded
cont
rol input is in
dicate
d in Fig
u
re 3. The
s
e
simulatio
n
re
sults d
e
mon
s
trate the tracking
capability of the proposed
controlle
d and its
effectivenes
s for
cont
rol tracking of
uncertai
n non-
affine nonlin
e
a
r sy
stems.
6. Conclusio
n
This p
ape
r propo
se
s a ne
w output fee
d
back
ad
aptive neu
ral net
work
ada
ptive controlle
r
for a cla
ss of uncertain n
o
n
-
affine nonli
n
ear sy
st
em
s. The distin
gui
she
d
asp
e
ct
of the propo
sed
control alg
o
rit
h
m is th
at no
state ob
se
rve
r
is
em
ploye
d
.
Only the out
put er
ro
r is
used to ge
ne
rat
e
control i
nput
and
upd
ate l
a
ws. Th
e
sta
b
ility analysi
s
dep
end
s
he
avily on the
universal fu
n
c
tion
approximatio
n prop
erty, only one RBF
N
is em
pl
oyed to approximate the lumped un
ce
rt
ain
nonlin
ear fun
c
tion
(1
6). T
here
a
r
e
no
rest
rictiv
e
co
ndition
s o
n
t
he d
e
si
gn
co
nstant
s. So t
h
e
sy
st
em
con
s
t
r
uct
i
s
v
e
ry
si
mple.
Out
put
s t
r
a
c
king
e
r
ror an
d all
sta
t
es in the
clo
s
ed
-loo
p sy
stem
are
gu
ara
n
te
ed to
be
bo
unde
d by
Lyapun
ov ap
pr
oach. Simula
tion results
h
a
ve verifie
d
t
he
effectivene
ss
of the propo
sed app
roa
c
h.
Ackn
o
w
l
e
dg
ement
It is a project suppo
rted by Provincial
Na
tural Sci
e
nce Fo
und
ation of Huna
n, China
(Grant No.
13
JJ902
2), the
Re
sea
r
ch Fo
undatio
n of Educ
ation Bureau of Huna
n
Province, Ch
ina
(Grant No.09
B
022), the Pl
anne
d Scie
n
c
e a
nd T
e
ch
nology Proje
c
t of Hu
nan
Province, Chi
na
(Grant No.2
0
11FJ312
6). S
uppo
rt
ed by the Co
nst
r
u
c
t Program of the Key Disci
p
line in Huna
n
Province: Co
ntrol Sci
e
n
c
e
and En
ginee
ring Sci
e
n
c
e a
nd Te
ch
nolo
g
y Innovation
Team of
Hu
n
an
Prov
inc
e
: Co
mplex
Netw
or
k Co
ntrol.
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: 5284 – 52
92
5292
Referen
ces
[1]
F
i
trian Imad
ud
din, K
h
isb
u
ll
ah
Hud
ha, J
anat
ul Isla
h Mo
ha
mmad, His
ham
udd
in J
a
mal
u
d
d
in. Sim
u
l
a
tio
n
and
e
x
perim
en
tal i
n
vestig
atio
n o
n
ada
ptive
multi-ord
e
r pro
portio
nal-
i
nte
g
r
a
l c
ontrol
for
p
neum
atical
l
y
actuated
activ
e
susp
ens
ion
s
y
st
em usin
g kno
w
l
e
d
ge-
bas
ed
fuzz
y.
Int.
J. of Mode
lli
ng
, Identificati
o
n
and C
ontrol
. 2
011; 14(
1/2): 7
3
– 92.
[2]
Mrinal
Ka
nti S
a
rkar, Su
brata
Ban
e
rje
e
, Sa
kti Pr
asa
d
Gh
osha
l, T
apas
Kumar S
aha.
Adaptiv
e fuzz
y
param
eter scheduling
schem
e for GSA based optim
al pr
oportiona
l integral derivativ
e
and lag-lead
control
of a D
C
attraction t
y
pe l
e
vitatio
n
s
ystem.
Int. J. of
Automation and Control
. 2
0
12; 6(2): 1
74
–
192.
[3]
LIU Guo-ro
ng,
W
A
N Bai-
w
u
.
Direct a
dapti
v
e fuz
y
rob
u
st control for
a
class of n
o
n
l
ine
a
r MIMO
s
y
stems
. Co
ntrol T
heory&A
p
p
licatio
ns
. 20
02;
19: 693-6
98.
[4]
LIU Guo-ro
ng,
W
A
N Bai-
w
u
.
Indirect a
dapt
ive fuz
y
rob
u
s
t
control for a
class of no
nli
near MIMO
s
y
stems.
Co
ntrol an
d Decis
i
o
n
.
2002; 1
7
: 67
6-68
0.
[5]
T
O
NG Shao-chen
g, QU Lia
n
-jia
ng. F
u
zz
y adaptiv
e outp
u
t feedb
ack control for a cl
ass of MIMO
non
lin
ear s
y
ste
m
s.
Control an
d Decisi
on.
20
05; 20: 78
1-79
3.
[6]
T
O
NG Shao-c
hen
g, CHAI T
i
an-
yo
u. Ada
p
ti
ve fuzz
y o
u
tp
ut feed
back c
ontrol for
a cl
ass of MIMO
non
lin
ear s
y
ste
m
s.
Acta Electronic
a
Sinic
a
. 2
005; 23: 9
87-9
90.
[7]
R Boukezz
o
u
l
a, S Galichet,
L F
oull
o
r
. Fuzz
y
a
daptiv
e c
ontrol for n
on-
affine syste
m
s
. Proc. IEEE
Internat. Conf. F
u
zz
y
S
y
stems
.
2003; 23: 54
3
-
548.
[8]
J W
ang, SS Ge,
T
H
Lee.
Adaptiv
e fu
z
z
y
slidi
ng
mo
de
contro
l of a
class of no
nli
near syste
m
s.
Proc.T
hird Asian Co
ntrol Co
n
f. 2000; 170: 5
99-6
04.
[9]
Yih-Gua
ng L
e
u
, W
e
i-Yen W
ang, T
s
u-T
i
an
Lee. Obse
rv
er-
base
d
dir
e
ct a
daptiv
e fuzz
y-
neur
al co
ntrol
for non-affin
e
n
onli
n
e
a
r s
y
ste
m
s.
IEEE
Transactions on Neural Networks
.
200
5; 16: 853-
861.
[10]
H Du, SS Ge, JK Liu. Adap
tive neur
al net
w
o
rk
outp
u
t feedb
ack control
for a class of non-affin
e
n
o
n
l
i
n
e
a
r
sy
stems
w
i
th
un
mode
l
l
e
d
dy
na
mi
cs.
IET Control T
heory & App
lic
ation.
20
10; 5: 465-
477.
[11]
Hui Hu, Pe
ng Guo. Output F
eed
back Ne
ur
al Net
w
ork Ad
aptive T
r
ackin
g
Contro
l for Pure- F
eed
bac
k
Nonl
in
ear S
y
stems.
IJACT
. 2
012; 4(1
8
): 655
-663.
[12]
Hui
Hu, Gu
o-r
ong
Li
u, Pe
ng
Guo, F
e
n
g
H
uan
g.
Ne
ura
l
net
w
o
rk
ad
apti
v
e tracki
ng c
o
ntrol for
strict-
feedb
ack no
nl
i
near s
y
stems
w
i
t
h
o
u
t backst
epp
ing.
Jo
urn
a
l
of Centra
l So
uth Univ
ersity
(Scienc
e an
d
T
e
chno
logy)
. 2
012; 42(
10): 39
00-3
905.
[13]
SWKKB Park,
JJ Lee. Sm
oot
h variable struct
ure controller
for
robot manipulat
or us
ing v
i
rtual
plant.
Electron
ics Let
ters. 1995; 31(
23): 213
4-2
136
.
[14]
PA Ioann
ou, J Sun. Rob
u
st
a
daptiv
e contro
l. Engle
w
o
o
k C
l
i
ffs, NJ: Prentice-Hal
l
. 199
6.
Evaluation Warning : The document was created with Spire.PDF for Python.