TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6153 ~ 6163
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.564
6
6153
Re
cei
v
ed
Jan
uary 18, 201
4
;
Revi
sed Ma
rch 2
6
, 2014;
Acce
pted April 15, 2014
Dynamic Modeling Process of Neuro Fuzzy System to
Control the Inverted Pendulum System
Thar
w
a
t
O. S. Hanafy
1
, M.
K. Met
w
all
y
2
1
Computers a
n
d
S
y
stems En
g
i
ne
erin
g De
par
tment, Al
_Azha
r
Universit
y
, F
a
cult
y
of Engin
e
e
rin
g
,
Cairo, Eg
ypt
2
Departem
ent of Electrical En
gin
eeri
ng, Men
oufi
y
a U
n
ivers
i
t
y
, F
a
cult
y
of Engi
neer
in
g,
Meno
ufi
y
a, Eg
ypt
Corresp
on
din
g
author, e-mai
l
: thar
w
a
t.han
af
y@
ya
h
oo.com,
mohkame
l
20
0
7
@
y
a
hoo.com
Ab
stra
ct
T
he ana
lysis a
nd contro
l of compl
e
x pla
n
ts often
requ
ires the
princi
pl
es
of qual
itative p
r
ocess
m
o
dels sinc
e quantitat
iv
e, nam
e
ly analytic
al process
models are not avail
able. Qualitativ
e modeling is
one
prom
ising approach to the s
o
lution
of difficult tasks
aut
omation if
q
ualitative proc
es
s models
are
not
avail
a
b
l
e. T
h
is
contributi
on p
r
es
ents a new
concept of q
ualit
ative
dyn
a
m
ic pr
ocess
mo
de
lin
g usin
g so
calle
d Dyn
a
m
i
c
Adaptive N
e
uro fu
z
z
y Syst
ems. This
yi
el
ds the framew
ork of a new
systems the
o
ry the
essenti
a
ls of w
h
ich ar
e giv
en
in fu
rther secti
on of the p
a
p
e
r
. F
i
rst, an ide
n
tificatio
n
meth
od is pr
esent
e
d
,
usin
g a co
mbi
n
ation
of lin
gu
istic know
le
dge.
Next, a
stabi
lity
defin
ition f
o
r d
y
na
mic
neur
o fu
z
z
y
syste
m
s
as
w
e
ll as metho
d
s for stabil
i
ty ana
lysis is g
i
v
en. F
i
nal
ly, a n
euro fu
zz
y
mo
del-
base
d
ne
ur
o fu
zz
y
c
ontrol
l
er
desi
gn
meth
od
is deve
l
o
ped.
T
he ide
n
tifica
tion of re
al pr
obl
e
m
s an
d n
euro fu
zz
y
c
o
ntroll
er des
ign
for
inverted pendulum
system
dem
onstrate the
signific
anc
e of the new system
s theory.
Ke
y
w
ords
:
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 contra
st to commo
n ap
proa
ch
es of
A
daptive Neu
r
o Fu
zzy mo
deling [1], the dynamic
system i
s
co
mpletely de
scrib
ed i
n
the
neuro
fu
zzy
domain: th
e
neuro fu
zzy i
n
formatio
n a
bout
the previous
state is
directly
applied to
com
pute the
system’
s
cu
rre
nt state, i.
e. the delaye
d
neuro fuzzy output is fee
dba
ck to the
input wit
hou
t defuzzifi
cati
on. Knowle
d
ge pro
c
e
s
sin
g
in
su
ch dyna
mic neu
ro fu
zzy systems
requires a
ne
w inferen
c
e
method, the
inferen
c
e
with
interpol
ating rules.
The analy
s
is
and control o
f
complex pla
n
ts o
ften req
u
ire
s
the introdu
ction of qualitative
pro
c
e
ss m
o
d
e
ls si
nce qu
antitative, namely analyt
ical pro
c
e
ss m
odel
s are
no
t available. An
examination
of the qua
ntitative
and q
u
a
litative paradi
gms
will hel
p
to
identify their strength
s
a
n
d
wea
k
n
e
sse
s
and ho
w their divergent ap
proa
ch
es
can
comple
ment
each other. Howeve
r, hum
an
experts as operators
u
s
ually
are
capabl
e of accompli
shin
g
control tasks, taki
ng i
n
to
con
s
id
eratio
n
only impreci
s
e kno
w
led
ge
about the
pro
c
e
ss
which m
a
y describ
e by a set of rules
like
IF
valve is “op
en wi
de
”
THEN
liqui
d level is “ri
sing fast”.
Thus,
the be
havior of
an
operator
an
al
yzing
or controlling
a p
r
o
c
e
s
s stim
ulate
s
the ne
w
approa
ch
of neuro
fu
zzy modelin
g, system
s
a
nalysi
s, an
d
co
ntroller de
sign
pursue
d
in
t
h
is
contri
bution.
The n
e
w con
c
ept
allo
ws i
n
tegratin
g q
u
a
litative pro
c
ess
kno
w
le
d
ge into
mod
e
l
s of
these p
r
o
c
e
s
se
s like they
are found e.
g. in proc
e
ss or manufa
c
turing in
du
stri
es a
s
well as in
automotive systems [2].
Modelin
g is
achi
eved u
s
i
ng a parti
cul
a
r cla
s
s of dynamic ne
uro
f
u
zzy
sy
st
e
m
s wh
er
e
the no
nlinea
r
static
ch
ara
c
t
e
risti
c
s of the
pro
c
e
s
s a
n
d
-
in contra
st to
comm
on
app
roa
c
he
s [1]
-
a
s
well its dyn
a
m
ics
are re
p
r
esented i
n
the ne
uro fu
zzy domai
n. T
o
be mo
re
specifi
c
, Figu
re 1
shows an
aut
onomous fi
rst
order dynam
i
c neuro fu
zzy system. The rule
base
may consi
s
t of
rules like
IF
y
k
-1 i
s
“
s
mall
”
Then
yk
is
“big”
.
Lingui
stic te
rms like “sm
a
ll” are mo
deled by ne
uro fu
zzy sets. The
kn
owle
dge
prop
agatio
n is ca
rri
ed out
by a neuro fuzzy infe
re
n
c
e
method. Since the ne
uro fuzzy outp
u
t i
s
feed ba
ck without a p
r
io
r defu
z
zificat
i
on, the
ling
u
istic i
n
form
ation ab
out
the syste
m
i
s
compl
e
tely m
odele
d
in the
neu
ro fu
zzy
domain. A
s
a
con
s
e
que
nce, a ne
w infe
ren
c
e
sche
m
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: 615
3 –
6163
6154
has to be d
e
rived for the followin
g
re
asons: An
inference method
is expecte
d to evaluate a
se
t
of neuro fu
zzy rules corre
s
pondi
ng to the human way of approxima
t
e reasonin
g
. Huma
n bein
g
s
are a
b
le to p
r
ocess o
n
ly su
ch n
euro fuzzy se
t
s
th
at might be
prop
erly adj
o
i
ned to ling
u
i
s
tic
values.
The
r
efore,
only th
ese
ki
nd
s of
interp
ret
abl
e
neuro fu
zzy
sets a
r
e
app
ropriate
inp
u
ts of
neuro fu
zzy
systems. Sin
c
e the
neu
ro f
u
zzy outp
u
t
o
f
a dynami
c
n
euro
fuzzy
system h
a
s to
be
pro
c
e
s
sed by
the inferen
c
e in su
bsequ
ent step
s,
it has to b
e
gu
arante
ed that
the inferen
c
e
maps inte
rp
re
table neu
ro fu
zzy inp
u
ts ont
o an interp
ret
able ne
uro fu
zzy outp
u
t.
Figure 1. Autonomo
u
s Fi
rst Order
Dyna
mic Ne
uro Fu
zzy System
In the
seq
uel,
neu
ro fu
zzy
numbe
rs
with
triang
ula
r
sh
aped
mem
b
e
r
shi
p
fun
c
tion
s, which
are often
use
d
to ch
ara
c
te
rize li
ngui
stic
values li
ke
“small” o
r
“bi
g
”,
will be u
s
e
d
as inte
rp
reta
ble
neuro fuzzy sets.
Conve
n
tional
rea
s
o
n
ing
method
s li
ke
“max
-mi
n
- i
n
feren
c
e
”
[3]
do not
gen
erate
an
interp
retabl
e
neuro fu
zzy
o
u
tput. The
r
ef
ore, a
ne
w n
e
uro fu
zzy infe
ren
c
e
method
, the “infe
r
en
ce
with interpola
t
ing rule
s” wa
s devel
oped
whi
c
h i
s
out
li
ned in the
se
con
d
sectio
n. This m
e
thod
i
s
the central el
ement of a new sy
stem theory cove
ring
processe
s repre
s
e
n
ted b
y
a set of neuro
fuzzy rul
e
s.
Within the scope of this sy
stem
s
theory
an identificati
on pro
c
e
dure
is develope
d
in
the se
con
d
section. Me
asurem
ents a
s
well a
s
heu
ristic
kn
owl
e
dge a
r
e u
s
e
d
to determi
ne
a
lingui
stic re
prese
n
tation of
the
process
dynamics. After that, t
he stability definition for dynam
ic
Neu
r
o fuzzy Systems is gi
ven and ap
proach fo
r stabi
lity analysis is briefly outlined.
The thi
r
d
se
ct
ion fo
cu
se
s o
n
a
ne
w d
e
si
gn
strategy
fo
r ne
uro fu
zzy
controlle
rs. T
h
is
ne
w
approa
ch e
n
able
s
the
synthesi
s
of n
euro
fuzz
y controlle
rs ex
clu
s
ively ba
sed on
qualit
ative
pro
c
e
ss
kno
w
led
ge. Final
ly, in the fourth sectio
n the main ch
ara
c
teri
st
ics of the ne
w Systems
theory a
r
e
de
monst
r
ated
fo
r an
inve
rted
pend
ulum
sy
stem. Fi
rst, th
e process i
s
identified. Th
e
n
,
the resulting
neuro fuzzy system model
is appli
ed to neuro fuzzy controlle
r.
2. Identifica
tion of D
y
namic Neuro fuzz
y
S
y
stems
Identification
of dynamic n
euro fu
zzy systems
requi
res the tra
n
sfer of crisp p
r
ocess
measurement
s into the domain of
neuro
fuzzy modeli
ng. An identification p
r
o
c
ed
ure for dyna
mic
neuro fuzzy systems
can
b
e
develop
ed
based on th
e
inferen
c
e
wit
h
interp
olatin
g rule
s Fig
u
re 1
illustrate
s the
identification
con
c
e
p
t. The
delayed
in
pu
ts and
output
s of the p
r
o
c
ess are u
s
ed
as
inputs of the
neuro fuzzy inferen
c
e
(seri
a
l-pa
ralle
l
structure). The
neuro fuzzy erro
r is calcula
t
ed
followin
g
Zad
eh’s
extensi
o
n pri
n
ci
ple a
s
the
differen
c
e b
e
twe
en t
he cri
s
p p
r
o
c
ess outp
u
t a
n
d
the neuro fuzzy model out
put:
Minimizi
ng bo
th, the mean squ
a
re
d ce
nter of the error.
(
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dynam
ic Mod
e
ling Pro
c
e
ss of Neuro F
u
zzy S
ystem
to Control the…
(Tha
rwat O.
S. Hanafy)
6155
And the fuzziness of the error.
(
2
)
Yields the proce
s
s model
. In Equation (2),
the inte
gral ove
r
the erro
r memb
ershi
p
function
defin
es
a m
e
a
s
ure of its fu
zzi
ness [1
1]. Th
e identific
a
tion is carr
ie
d ou
t in
two
s
t
ep
s
.
First, the
sig
n
i
ficant del
ays of the inp
u
t
and the
outp
u
t of the p
r
o
c
ess a
r
e d
e
termined to fix t
h
e
stru
cture of the neu
ro fuzzy model. Secon
d
, the
rul
e
base of the
neuro fu
zzy pro
c
e
ss mo
d
e
l is
identified mini
mizing Eq
uati
on (1
) and Eq
uation (2
).
2.1. Dete
rmining the Str
u
ctur
e of Dy
namic Neur
o
fuzz
y
S
y
stems
The
signifi
ca
nt delay
s of
the n
euro
fuzzy m
odel
ca
n be
det
ermin
ed a
p
p
l
ying a
pro
c
ed
ure si
milar to nonli
near
system i
dentificat
io
n algorith
m
s re
pre
s
ente
d
by neural nets [5
; 7]:
Figure 2. Neu
r
o Fu
zzy Mo
d
e
l in Parallel
-
seri
al Structu
r
e
Tange
nt plan
es of the sy
stem
s nonlin
earity
are e
s
timated on d
i
fferent point
s of the
operating do
main. To cal
c
ulate the ta
ngent pla
n
e
s
, matrice
s
a
r
e
built up from
measure
m
e
n
ts.
Usi
ng data
o
f
the output with a del
ay excee
d
ing
th
e pro
c
e
s
s’ order
re
sults in
ran
k
defi
c
ie
nt
matrices.
He
nce, th
e maxi
mum requi
re
d delay
of
the
output
s of lo
wer than
max
i
mum o
r
de
r
can
be dete
r
min
e
d
from the
tange
nt plan
e
s
. The
s
e
are
parallel to a
x
is sp
ann
ed
by insig
n
ifica
n
t
delayed o
u
tp
uts [7].
2.2. Identif
y
ing the Rule base
To illustrate the ba
sic id
ea
s of the ident
ificat
ion p
r
o
c
edure of the
rule b
a
se, it sufficie
n
t
to con
s
id
er th
e static
Neu
r
o fuzzy System depi
ct
ed in
Figure 3. It can be
sho
w
n
that the ce
nte
r
of the neu
ro fuzzy output o
f
the inferen
c
e only dep
en
ds o
n
the cen
t
ers of the
ne
uro fu
zzy in
p
u
ts
[5;5]. This rel
a
tionship is e
x
presse
d by the ce
nter eq
u
a
tion
.
^
In the first ste
p
, the cente
r
equatio
n is:
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: 615
3 –
6163
6156
For
Ne
uro fu
zzy Sy
stem
s with
multipl
e
inp
u
ts the
cente
r
equ
ation i
s
th
e pi
e
c
e
w
ise
multilinea
r interpol
ation fu
nction
span
n
ed by t
he ce
nters of the
neuro fuzzy premi
s
e
s
and
the
neuro fu
zzy
con
c
lu
sio
n
[5
]. Thus, in
case
of a
sin
g
le inp
u
t sy
stem the
cent
er e
quatio
n i
s
a
piecewi
s
e lin
ear inte
rpol
ation functio
n
. Figure 3
sho
w
s
an optimi
z
ation
re
sult. Obviou
sly, fo
u
r
rule
s had to b
e
:
Figure 3. SISO Static Neu
r
o Fuzzy System
Figure 4. Optimized
Cente
r
Equation
Identified. Therefo
r
e, the cent
ers of four premi
s
es
c(P1),
…., c(P5
) and four
co
nclu
sio
n
c(
C1
), …., c(
C5)
we
re fou
nd. T
he cent
er equ
ation i
s
the linea
r interpol
ation functio
n
f(c(E
))
spa
nne
d by c(P1), …., c(P5) an
d c(C1),
…., c(C5
). Be
cau
s
e of the
unste
adin
e
ss of the gradi
e
n
t
of J1 g
r
adi
e
n
t-ba
sed
se
a
r
ch
strategie
s
may
not a
pplicable. Fo
r sy
stems
of highe
r o
r
de
r,
evolutiona
ry algorith
m
s h
a
v
e been successfully appli
ed [5].
Figure 5. Det
e
rmini
ng the
Neu
r
o Fu
zzy Output for a
Cri
s
p Input
Having
dete
r
mined th
e
ce
nters of th
e
premi
s
e
s
an
d con
c
lu
sion
s, the
sh
ape
s of th
e
membe
r
ship
function
s
are
dete
r
mine
d i
n
the
next
step
with
a m
e
thodol
ogy d
e
velope
d in
[5].
This ap
pro
a
ch guarantee
s a minimum fuzzine
s
s of
the error by minimizi
ng Equation (2) u
nde
r
con
s
id
eratio
n
of the
strag
g
ling
of
the
measurement
s. Fig
u
re
5
s
hows i
n
exte
nsio
n of Fi
gu
re 4
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dynam
ic Mod
e
ling Pro
c
e
ss of Neuro F
u
zzy S
ystem
to Control the…
(Tha
rwat O.
S. Hanafy)
6157
the neuro fuzzy output
Ŷ
computation fo
r a given cri
s
p input
e0 usi
ng the identified neu
ro fuzzy
model. Du
e to the crisp input, the ne
uro fuzzy
ou
tput of the
model is e
q
u
ivalent to the
interpol
ating
con
c
lu
sio
n
. The left and ri
ght foot
of the interp
olatin
g con
c
lu
sio
n
are calculat
ed
using L(IC)=gl(e0) and r(IC)=gr
(e0) respectively. gl(e) and
gr(e) are pi
ecew
ise multilinear
interpol
ation
function
s
sp
anne
d by th
e left an
d
ri
ght feet of t
he
con
c
lu
sio
n
mem
bersh
i
p
func
tions
,
L(c1), ….., L(c
5
) and r(c1),
….., r(c5)
,
res
p
ec
tively, and the centers
of the premis
e
membe
r
ship f
unctio
n
s
c(p1
), …., c(p5
). IF the
cri
s
p i
n
put e0
belo
n
g
s to th
e m
e
asu
r
em
ents, i
.
e.
e0 =ei, the membe
r
ship va
lues of the
co
rre
sp
ondi
ng
measured cri
s
p o
u
tput yi is al
ways g
r
e
a
ter
than ze
ro:
e0= ei
→μŶ
(y
i)>0.
Thus, the neuro fuzzy model output might be in
terpreted as a possibility distribut
ion [8].
Finally, it has to be e
m
p
hasi
z
e
d
the
in gene
ral li
ngui
stic
kno
w
led
ge is
a
pplied in
combi
nation
with the mea
s
ureme
n
ts. O
n
t
he one ha
nd, linguisti
c
kno
w
le
dge m
a
y be used f
o
r
situation
s
wh
ere no me
asurem
ents a
r
e
availabl
e. On the other
hand, rul
e
s
given by human
experts can be
take
n
as starting con
d
i
t
ions
for
the optimizatio
n
pro
c
ed
ur
e. For example, the
starting
condi
tions for the
optimiz
ation whose result
s
ar
e illustrat
ed in Figure
4 are the centers
of the four premise
s
an
d concl
u
si
on
s of the respe
c
tive rule
s.
3. Stabilit
y
Anal
y
s
is of Dy
nam
i
c Neuro Fuzz
y
S
y
stems
To sho
w
the
typical beh
avior of Dyn
a
mi
c Neuro fuzzy Syste
m
s and to o
b
tain an
approp
riate stability definition, it is sufficient to
con
s
id
er two a
sim
p
le auton
omo
u
s Neu
r
o fuzzy
System rep
r
e
s
ente
d
by the following two
rules:
IF
yk-1 is “ne
gative”
Then
yk is “positive”
IF
yk-1 is “positive”
Then
y
k
is “negative
”
The mem
b
e
r
ship fu
nction
s defin
ed o
n
the input
domain
are
sho
w
n in
Fi
gure
6.
Dep
endin
g
o
n
the output
membe
r
ship f
unctio
n
s, t
he
system exhi
bi
ts different dy
namic
beh
avior.
Given the ou
tput membe
r
ship fun
c
tion
s of Figu
re
7,
we obtai
n sy
stem 1
which
is stabl
e sin
c
e
the outp
u
t co
nverge
s to
th
e ne
uro
fuzzy
numb
e
r with
the
cente
r
0, the left foot
–2 a
nd th
e
right
foot +2.
Figu
re
8 d
epi
cts
the ne
uro
fu
zzy outp
u
t
resulting from a crisp initial
state y0=2. T
h
e
output me
m
bership
fun
c
tions of
syst
em 2
sh
own
in Fig
u
re 9
ca
use a
n
unsta
ble
system
behavio
r. Althoug
h the
ce
nter of th
e o
u
t
put co
nverg
e
s to 0
for
any
initial state, i
t
s left and
ri
g
h
t
foot moves to infinity (Figure 1
0
). Since t
he outp
u
t becom
es
fuzzi
er with
every step, the
spe
c
ificity of the output van
i
she
s
for
k
→∞
.
These simpl
e
exampl
es suggest
the followi
ng stability definition for
Dynami
c
Neuro
fuzzy System
s: An equilibri
um poi
nt of a Dynami
c
Neuro Fuzzy System marked by a crisp value
R0 is
stable if
:
a)
R0 is an asy
m
ptotically st
able equilibri
um point for the center of the output c(Y
k
)
b)
The feet of the neuro fuzzy output
stay in a boun
ded
environ
ment
of R0.
In the examples above R0=0
marks the
equilibrium point.
System 1 has a stable eq
uilibriu
m
poin
t, wher
ea
s the equilibri
um
point of system 2 is
unsta
ble. Sin
c
e it i
s
sufficient to exami
ne the m
appi
ng of the
cri
s
p p
a
ramete
rs of the
neu
ro
fuzzy in
put o
n
to the crisp
para
m
eters o
f
the
neuro f
u
zzy output,
conve
n
tional
method
s for t
he
stability analysis of nonli
n
e
a
r syste
m
s can be appli
e
d
.
If all interpolating premi
s
e
s
defined o
n
yk-
1,…., yk-n are fuzzi
er tha
n
the interpol
ating co
ncl
u
si
o
n
with the sa
me ce
nter de
fined on yk, it is
only nece
s
sa
ry to analyze
the mapping
of the c
enters of the neuro fuzzy
input
onto the neu
ro
fuz
z
y
output [5].
With a
co
nst
ant neu
ro
fuzzy Uk
re
sults a di
screte
n
online
a
r
syst
em de
scrib
e
d
by the
cente
r
e
quati
on
c(Y
k
)=f(c(Yk-1
),……
…, c(Y
k
-n)).
Wit
h
the
cente
r
s c(Y
k
),
c(Yk-1),……
…, c(Yk-
n)
of the
neu
ro fu
zzy outp
u
t Yk
and
its delay
s
Y
k
-1,
………,
Y
k
-n.
To analy
z
e su
ch a syste
m
,
method
s
b
a
s
ed
on co
m
m
on stability
analysi
s
a
ppro
a
che
s
may
be use
d
.
The “Co
n
ve
x
De
comp
ositio
n” [9, 10] as an efficient
nume
r
ical st
ability analysis metho
d
a
nd an ap
pro
a
ch
based “integ
ral
Ljap
unov F
unctio
n
”
[1
1] have
be
en su
ccessfully
a
p
p
lied
t Dyna
mic Neu
r
o
fu
zzy
S
y
st
ems.
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: 615
3 –
6163
6158
Con
s
id
erin
g first o
r
de
r Dy
namic
Ne
uro
fuzzy
Syste
m
s, the regio
n
of attractio
n
of a
n
equilib
rium p
o
int can eve
n
be analyticall
y
determine
d [5, 12].
Figure 6. Membershi
p
Fun
c
tion
s Defin
e
d
for
y
k-1
Figure 7. Output Membe
r
ship Fun
c
tion
of
System 1
Figure 8. Dyn
a
mic Beh
a
vio
r
of System 1
Fi
gure 9. Output Membe
r
ship Fun
c
tion
s of
System 2
Figure 10. Dy
namic Be
havi
o
r of System 2
4. Neuro Fu
zz
y
Model Based Ne
uro F
u
zz
y
Controller Design
This sectio
n
outline
s
a
new ne
uro
f
u
zz
y controller
synthe
si
s app
roa
c
h
u
s
ing
a
qualitative (n
euro fu
zzy) p
r
ocess mo
del
. In Figure
1
1
,
the stru
cture of t
he co
ntrolled ne
uro fuzzy
system i
s
de
picted. T
he
p
l
ant is mod
e
l
ed by
th
e se
con
d
o
r
de
r Dynamic Ne
uro
fuzzy
Syst
em.
The
cont
rolle
r dete
r
min
e
s
the control
si
gnal
Uk
fro
m
the n
euro fu
zzy
model
ou
tput Yk
and t
h
e
comm
and variables Wk. As mentioned above,
the center of the neuro fuzzy model output
exclu
s
ively d
epen
ds
on th
e ce
nters of t
he inferen
c
e
inputs. Th
ere
f
ore, the
cent
er of the o
u
tp
u
t
can
only m
a
n
i
pulated
by th
e center of
control
sig
nal
s. Thu
s
, given
the center of
the ne
uro
fuzzy
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dynam
ic Mod
e
ling Pro
c
e
ss of Neuro F
u
zzy S
ystem
to Control the…
(Tha
rwat O.
S. Hanafy)
6159
model o
u
tput
and the
ce
nter of the
comman
d
an approp
riate
crisp co
ntrol sign
al
ca
n
b
e
determi
ned. Con
s
e
quently
, the center e
quation of
the neuro fuzzy controll
er is
determi
ned from
the center eq
uation
of the
neu
ro fu
zzy
pro
c
e
s
s
mo
d
e
l. For a
dyn
a
mic
neu
ro f
u
zzy sy
stem
of
orde
r n, th
e
cente
r
e
quati
on is given
by c(Y
k
)=f(c(Yk-1
),
……, c(Yk-n), c(Uk-
δ
),…. c
(
U
k
-
m
)).
c(Y
k
),
c(Y
k
-n
),
c(
U
δ
),
…. c(Uk-m)
re
pre
s
ent
the cente
r
s of the
neu
ro fuzzy inp
u
t
s and
output
s
and their del
ay.
δ
is the difference ord
e
r
of the cente
r
equatio
n. To dedu
ce the
center eq
uat
ion
of the neuro
fuzzy co
ntro
ller, app
roa
c
hes fo
r co
ntroller synth
e
si
s of time-di
s
crete
nonlin
e
a
r
system
s ca
n be applie
d. In [5] the cen
t
er equatio
n is determine
by input/outp
u
t lineari
z
atio
n.
The pro
b
lem
of handling
a zero dyn
a
mics whi
c
h
may occu
r whe
n
usin
g this method
s is
discu
s
sed [5, 13].
The exam
ple
depi
cted in
Figure (1
1)
d
e
mon
s
tr
ate
s
the ba
sic i
d
e
a
s of the
neu
ro fuzzy
model ba
se
d controlle
r syn
t
hesi
s
. The u
nder lying
set
of rules i
s
:
IF
Yk
-1 =
A and Uk
-1=
X
Th
e
n
Yk=
AX
IF
Yk
-1 =
B and Uk
-1=
X
Th
e
n
Yk=
BX
Figure 11. Structure of
a Controlle
d Dyn
a
mic Neu
r
o F
u
zzy System
IF Yk-1 = C a
nd Uk
-1
=X Then Yk=CX
IF Yk-1 =
A and Uk-1=
Z
Then Yk=AZ
IF Yk-1 =
B and Uk-1=
Z
Then Yk=BZ
IF Yk-1 =
C and Uk-1=
Z
Then Yk=CZ
The p
r
emi
s
e
s
A, B and C are d
e
fined f
o
r the d
e
laye
d output yk-1
and the p
r
e
m
ise
s
X
and Z a
r
e d
e
fined for
Uk-1.
The con
c
lu
si
ons AX,……,
CZ defin
ed
on yk a
r
e a
ssumed n
o
t to be
fuzzi
er than
o
ne of th
e p
r
e
m
ise
s
A, B
or C. T
heref
ore
,
it is
sufficie
n
t to co
nsi
d
e
r
t
he ma
ppin
g
o
f
the cente
r
s c(Yk-1) a
nd
c(Uk-1
)
onto th
e cente
r
c(Y
k
). Usin
g the
inferen
c
e
wi
th interpol
ating
rule
s to evalu
a
te the neu
ro
fuzzy rule se
t, it c
an be sh
own [5, 12] th
at c(Y
k
)
=r(c(Yk-1
)
)+h
(
c(Y
k
-
1)). c
(
Uk
-1
)
Hold
s. Assu
ming h c((Y
k-1))
≠
0
c(Y
k
-1
),
the control law:
Ensures that the cent
er
of the output c (Yk) reaches
a desi
red equilibrium
point y
R
within a
sing
le step. With
out a bou
nd
ed cont
rol
si
gnal, the reg
i
on of attra
c
tion equ
als t
h
e
domain
of d
e
f
inition. Du
e t
o
the m
a
xim
u
m differen
c
e orde
r (
δ
=n
=1)
a ze
ro
dy
namics do
se no
t
occur [5, 13].
Ho
weve
r, in
pra
c
tical
ap
pl
ication
s
a bo
unde
d
contro
l sig
nal m
u
st
be
con
s
ide
r
e
d
.
No
w,
a re
gio
n
of attraction of the equilibri
um poi
nt might be determin
ed usi
n
g a Lyapuno
v
function, e.g.
V(c(Yk))=
(
c
(
Y
k
)-yR)
2
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: 615
3 –
6163
6160
Thus, the first step of the
controlle
r de
sign is
to form
ulate the cont
rol la
w Equati
on (3
).
Next, the regi
on of the attraction of the
desi
r
ed
e
quili
brium p
o
int is determin
ed
con
s
id
erin
g the
boun
ds
of th
e co
ntrol
sig
nal. Fro
m
Eq
uation
(3), a
n
ade
quate
manipul
ation variable
migh
t
be
cal
c
ulate
d
for each
cente
r
c(Y
k
) of the
p
r
eviou
s
dete
r
mined region
of attraction.
We sta
r
t with
the neuro fuzzy cont
rolle
r rule set:
IF Yk =
A
Then Uk=
u
A
IF Yk =
B
Then Uk=
u
B
IF Yk =
A
Then Uk=
u
C
The premise
s
A, B and C are
kn
own from
the rule set repre
s
entin
g the pro
c
e
s
s
behavio
r, wh
erea
s the crisp con
c
lu
sion
s uA, uB
and u
C
are
cal
c
ulat
ed usin
g Equ
a
tion (3). Th
u
s
,
the con
c
lu
sio
n
of the first rule is given b
y
:
Due to the
singleton
s u
s
e
d
as con
c
lu
si
ons, cri
s
p co
ntrolle
r input
s lead to a cri
s
p cont
rol
output. Only
for the
cri
s
p
inputs c(A),
c(B
)
, an
d
c(C) the
evalu
a
tion of
the
controlle
r
rul
e
set
usin
g the inf
e
ren
c
e
with i
n
terpol
ating
rules yi
el
ds t
he same
out
put as the
crisp
control
l
a
w
Equation
(3
).
If c(Y
k
) is
somewhe
r
e
b
e
twee
n the
s
e
pa
rticula
r
va
lues, th
e
con
t
roller outp
u
t is
determi
ned b
y
interpolatio
n. It might be
nece
s
sary
to
add more rul
e
s if the cha
r
acteri
stic of t
he
neuro fu
zzy
controlle
r diff
ers too
mu
ch
from th
e
no
nlinea
r
cha
r
a
c
teri
stics of t
he n
euro fu
zzy
control law E
quation (3).
With the infe
ren
c
e with
in
terpolatin
g ru
les, the neu
ro fuzzy rule
set
resulting
so
far may
be
use
d
to
de
si
gn a
ne
uro f
u
zzy controll
er; the
ne
uro fuzzy
rule
set
resulting
so f
a
r m
a
y b
e
u
s
ed to
de
sign
a ne
uro fu
zzy
co
ntroll
er.
Due to
the
crisp con
c
lu
sion
s, a
defuzzificatio
n
is n
o
t req
u
ired
for
cri
s
p inp
u
ts of t
he
cont
rolle
r. Becau
s
e
of the
pie
c
e
w
i
s
e
multilinea
r ce
nter eq
uation
,
such a co
n
t
roller h
a
s ch
ara
c
t
e
ri
st
ic
s con
s
i
s
t
i
ng
of
regio
n
s whe
r
e
multilinea
r fu
nction
s a
r
e
d
e
fined. Howe
ver, we
obtai
n a
cont
rolle
r with the
sam
e
characte
rist
ics
if the neuro f
u
zzy rule
set
is evaluate
d
with t
he conv
entional
sum
-
prod
-infe
r
en
ce com
b
ine
d
with
a cente
r
of singl
eton
s d
e
fuzzificatio
n. Only
the premise m
e
mb
ership fun
c
tions h
a
ve to be
manipul
ated i
n
the follo
win
g
way: the
ce
nters of a
ll p
r
emise
s
are
kept but the fe
et are
moved
to
the centers
o
f
the adj
acen
t premi
s
e
s
[5
]. The resu
lt
is a
ne
uro
fu
zzy
rule
set
with tria
ngul
a
r
membe
r
ship
function
s fo
r
the premi
s
e
s
and
sin
g
leto
ns fo
r the
co
nclu
sio
n
s. T
h
is
set of
rule
s i
s
use
d
for a
ne
uro fu
zzy
con
t
roller,
whi
c
h
can
be evalu
a
ted with
wel
l
-kn
o
wn meth
ods. Th
us, th
e
final tuning of the controlle
r in the closed
loop
with the real pro
c
e
s
s might be accomplished
with
comm
on software tools.
5. Identifica
tion and Co
ntrol of an Inv
e
rted Pe
ndul
um Sy
stem
The con
s
ide
r
ed inverte
d
p
endul
um sy
stem is
d
epi
cte
d
in Figu
re 1
2
. Input and
output of
the pro
c
e
ss a
r
e the force a
nd the angl
e of the pendul
um, respe
c
tively.
Figure 12. Inverted Pen
dul
um System
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dynam
ic Mod
e
ling Pro
c
e
ss of Neuro F
u
zzy S
ystem
to Control the…
(Tha
rwat O.
S. Hanafy)
6161
Figure 13. Structure of the
Neu
r
o Fu
zzy Process Mod
e
l
In the first
step, the st
ru
cture of th
e ne
uro fu
zzy pro
c
e
ss
model i
s
ide
n
tified Fi
gure
12
.
The ide
n
tifica
tion of the rul
e
ba
se was
carri
ed out
in t
he second
step an
d yielde
d 35 rule
s. From
the dynamic
neuro fuzzy model a ne
uro fuzzy c
ontroller with 5
5
rule
s wa
s de
sign
ed followi
n
g
the pro
c
ed
ure outlined in
se
ction 4.
Th
e resulting cl
ose
d
system i
s
given in Fig
u
re 14
(a, b, c).
Figure 14(a).
Inverted Pen
dulum Syste
m
Rep
r
e
s
ent
ed by Fuzzy Model
Figure 14(b).
Inverted Pen
dulum Syste
m
Cont
rolled
by a Neuro F
u
zzy Model B
a
se
d Ne
uro
Fuzz
y C
ont
r
o
ller
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: 615
3 –
6163
6162
Figure 14(c). System Re
sp
ons
e of Neu
r
o Fuzy Mod
e
l
6. Conclusio
n
This
co
ntribut
ion presente
d
the fram
ework
of a n
e
w qualitative systems th
eory base
d
on Dyn
a
mic
Neu
r
o fu
zzy Systems
whe
r
e
kno
w
led
g
e
about the
proce
s
s be
havi
o
r i
s
de
scrib
e
d
by
the a
set of rules. T
he dyn
a
mics of the
pro
c
e
s
s be
ha
vior is
model
ed by a
pprop
riate time
del
ay
and fe
ed
ba
ck of th
e n
euro fuzzy o
u
tpu
t
to the
sy
ste
m
’s in
put
without p
r
eviou
s
defu
zzifi
catio
n
.
Hen
c
e, an im
portant featu
r
e of this theory is t
he particular procedu
re for rule pro
pagatio
n, whi
c
h
wa
s develo
p
ed for this
cl
ass of syste
m
s and i
s
ca
lled infere
nce
with interpol
ating rule
s. T
h
e
essential
s
of this system
s t
heory
we
re
outlined: In
a
ddition to
rul
e
ba
se
d mo
d
e
ling
by hum
an
experts an identification me
thod allows to obtain a Dynamic Neuro fuzzy System fr
om
measurement
s. A new
sta
b
ility definition and diffe
re
nt approa
che
s
for a
nalytical and n
u
me
rical
stability anal
ysis
were b
r
iefly describ
ed. Mo
reove
r
, a neu
ro f
u
zzy-mo
del
based
contro
ller
synthe
sis me
thod wa
s giv
en. Fi
nally, as practi
cal d
e
mon
s
tratio
n
a inverted p
endul
um syst
em
wa
s ide
n
tified from
mea
s
ureme
n
ts,
a ne
uro
fuzzy controll
er wa
s d
e
si
g
ned
usin
g t
h
e
identification neuro fuzzy process
model
and
t
he closed
l
oop behav
ior was
presented.
Con
c
lu
ding, t
he ne
w
syst
ems th
eory
e
nable
s
q
ualit
ative modeli
n
g and
sim
u
la
tion as well
as
system
s an
al
ysis an
d con
t
roller d
e
si
gn
of co
mplex
dynamic
pro
c
esse
s. Since
the qualitati
v
e
approa
ch i
s
often the o
n
l
y
way to o
b
t
ain an
app
ropriate
proce
s
s re
present
ation, the n
e
w
con
c
e
p
t of the qualitative
systems the
o
ry offers
a
con
s
ide
r
abl
e pot
ential toward
s the a
u
tomat
i
on
of
t
h
is sy
st
em
clas
s.
Referen
ces
[1]
HR, BT. Dy
namic Fuzzy
Sy
stems for
Qualita
t
ive Process M
ode
lin
g. 199
9.
[2]
Abraham A, Nath B.
Evolutio
nary Des
i
gn
of Neuro-fu
zz
y
s
ystems-A Gen
e
ric F
r
amew
ork
. Proceedi
n
g
of the 4th Japa
n-Austral
i
a jo
in
t W
o
rkshop on
In
telli
gent a
nd
Evoluti
onar
y
s
ystems, Japan. 200
0.
[3]
Abraham A, Nath B.
Designi
ng Optimal Ne
uro-fu
zz
y Syst
ems for Intell
ig
ent Contro
l.
Procee
din
g
s of
the si
xth Intern
ation
a
l C
onfer
ence
on C
ontr
o
l Auto
m
a
tion
Rob
o
tics Com
puter Visi
on (I
CARCV 2
0
0
0
).
Sing
apor
e. 200
0.
[4]
Jang R, Ne
uro
-
fuzz
y
Mod
e
li
n
g
: Architecture
s, Anal
y
s
es an
d Appl
icatio
ns, PhD thesis, Univers
i
t
y
o
f
Califor
ni
a, Berkele
y
. 19
92.
[5
]
Ka
sa
bo
v N
,
Qu
n
So
ng
. Dy
nami
c
Evo
l
vi
ng
F
u
zzy
Ne
u
r
al
Ne
tw
orks
w
i
th
‘mo
u
t
- o
f
-n’
Acti
va
ti
on
No
de
s
for on-lin
e Ad
aptive S
y
stem
s.
T
e
chnical
Rep
o
rt
T
R
99-04, Dep
a
rtme
nt of information scie
n
ce,
Univers
i
t
y
of Otago. 19
99.
[6]
Juan
g Chi
a
F
e
ng, Lin C
h
in T
eng. an
onl
ine
self c
onstructin
g
Neur
al F
u
zz
y Infere
nce Ne
t
w
ork a
nd its
Applications. IEEE
T
r
ansactions on Fu
zz
y
s
y
stems. 1998; 16(1): 12-
32.
[7]
Sulzb
e
rger SM
,
T
schicholg-G
u
rman N
N
, Ve
stli SJ.
FUN:
Optimi
z
a
t
i
o
n
of
fu
zz
y
Ru
le Ba
sed Syste
m
s
Using
Ne
ural
Netw
orks.
Proceed
ings
of IE
EE Conf
erenc
e on
Ne
ural
N
e
t
w
o
r
ks, San
F
r
ancisc
o
. 19
93
;
312-
316.
[8]
T
ano S, Oy
am
a T
,
Arnould
T
.
Deep comb
inati
on of fuzz
y Infere
nce a
n
d
Neur
al N
e
t
w
ork in F
u
zz
y
Inference. F
u
z
z
y
Sets an
d S
y
stems. 1996; 8
2
(2): 151-
16
0.
[9]
Nauck D, Kla
w
onn F
,
Kruse R
.
F
oundatio
ns
of
Neuro-fuzz
y
S
y
stems, Joh
n
W
ile
y
a
nd So
n
s
. 1997.
[10]
Lin C T
,
Lee CSG. Neural
Net
w
ork b
a
s
ed F
u
zz
y
Lo
g
i
c Contro
l an
d Decis
i
on S
y
stem.
IEEE
T
r
ansactio
n
s o
n
Co
mp
ut.
40(12): 132
0-1
991
.
Evaluation Warning : The document was created with Spire.PDF for Python.