Академический Документы
Профессиональный Документы
Культура Документы
. ~-
.*-. -
120 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 2, NO. 2, MAY 1994
in this simulation is to design a fuzzy controller such that the The free system of (1) is defined as
control system is asymptotically stable in the large, that is,
such that backing up control can be perfectly achieved from
all initial positions.
Recently, Tanaka, and Sugeno [3], [9] have derived some
conditions for ensuring stability of fuzzy dynamic systems.
In Section 11, we derive a stability condition by weakening
i=l
the basic stability condition (BSC) proposed by Tanaka and
Sugeno. This condition is named weak stability condition where it is assume that
(WSC). Section 111 shows two design problems: nonrobust
controller design and robust controller design. The former is a
design problem for fuzzy systems with no PPU. The latter is
a design problem for fuzzy systems with PPU. To realize two
design problems, we derive four stability conditions: nonrobust
stability condition (NSC), weak nonrobust stability condition
(WNSC), robust stability condition (RSC), and weak robust for all k . Each linear consequent equation represented by
stability condition (WRSC). In particular, WRSC is the most Aix(k) is called “subsystem.”
important condition. Because, we can introduce concept of The basic stability condition (BSC), proposed by Tanaka
robust stability for fuzzy control systems with PPU by using and Sugeno, for ensuring stability of ( 2 ) is given as follows.
the WRSC. To introduce robust stability, admissible region Theorem 2.1 131: Basic stability condition (BSC): The equi-
and variation region, which correspond to stability margin in librium of a fuzzy system described by (2) is asymptotically
the ordinary control theory, are defined. In Section IV, we stable in the large if there exists a common positive definite
develop a control system for backing up a computer simulated matrix P such that
truck-trailer model. We show that the designed fuzzy controller
guarantees stability of the control system under a condition.
11. BASICSTABILITYCONDITION for i = 1,2, . . . , T , that is, for all the subsystems.
AND WEAK STABILITY CONDITlON The proof of this theorem will be given in Appendix A.
This theorem is reduced to the Lyapunov stability theorem for
The fuzzy system, proposed by Takagi and Sugeno [lo],
linear discrete systems when T = 1. Theorem 2.1 gives, of
is described by fuzzy IF-THEN rules which locally represent
course, a sufficient condition for ensuring stability of (2). We
linear input-output relations of a system. This fuzzy system
may intuitively guess that a fuzzy system is asymptotically
is of the following form:
stable in the large if all its subsystems are stable, that is, if all
+
Rule: IF z ( k ) is dli and . . . and z ( k - n 1) is d,i
its Ai’s are stable matrices. However, this is not the case in
+
THEN xi(k 1) = Aix(k) B i u ( k ) , + general: we will discuss it in Example 2.1.
where
We should notice that the BSC of (4) depends only on A;.
XT(k) = [ s ( k ) ,s ( k - l ) ,. . . , z ( k - n + l)], In other words, the BSC does not depend on wi(k).This fact
becomes a key point when weakening the BSC.
UT(k) = [ u ( k ) , u ( k - l ) , - , u ( k - m + l)],
[Example 2.11 Let us consider the following fuzzy free
system;
i = 1 , 2 , .. . ,T and T is the number of IF-THEN rules. x;(k+l)
is the output from the i-th IF-THEN rule, and d;j is a fuzzy
set. Given a pair of (x(k),u(k)),the final output of the fuzzy Rulel: I F z ( k - l ) i s A l T H E N x l ( k + l ) =Alx(k),
system is inferred as follows. Rule2: I F z ( k - l ) i s d a T H E N x 2 ( k + l ) = A2x(k),
where
X(k)T = [ z ( k )
z(k - l)],
i=l
1 0 -05
A1 = [1:0
-1.0
0 1 4 2 = 1.0 [ -0.5
0 ]
where Fig. 2 shows membership functions of dl and d 2 . Fig. 3(a)
and (b) illustrates the behavior of the following linear systems
for the inSal.‘%ondition z(0) = -0.70 and z(1) = 0.90,
+ +
respectively: x(k 1) = Alx(k) and x(k 1) = A2x(k).
These linear systems are stable since A1 and A2 are stable
dij(z(k-j+l))is the grade of membership of s ( ( k - j + l ) ) matrices. However, the fuzzy system, which consists of these
in d;j. stable linear systems, is unstable as shown in Fig. 3(c).
TANAKA AND SANO: FUZZY CONTROL SYSTEMS 121 - __ .
2 *25 T I
Cases 1 and 2 are related to nonrobust controller design and
robust controller design, respectively. As shown in Section 11,
the final output of a fuzzy system is calculated by (1) if it has
1.5 no PPU, that is, (Case 1). Conversely, if it has PPU, that is,
(Case 2), the final output is calculated by
1
0.5
- 0
Y
1
-0.5
i=l
-1
-1.5 where Aw;(lc) denotes unknown PPU. A w ; ( k )is a value such
that
- 2
Next, we weaken the BSC of (4). If it is possible, we may for all k . Equation (6) is reduced to (1) when Aw;(k) = 0
be able to find Lyapunov functions more easily. Theorem 2.2 for all k and i.
gives a weakened condition for the BSC, that is, a WSC. We stabilize the fuzzy system, (1) or (6), by the following
Theorem 2.2: Weak stability condition (WSC): The equilib- controller.
rium of a fuzzy system described by (2) is asymptotically Rule: IF z ( k ) is dli and . . . and z ( k - n 1) is Ani +
stable in the large if there exists a positive definite matrix P THEN ui(k)= F i x ( k ) ,
such that where i = 1 , 2 , . . . ,T . The final output of this fuzzy controller
is calculated by
r
s= w i ( k ) w i ( k ) { ~ T ~-~ iP}
i=l
i<j (7)
- (Ai - Aj)TP(Ai- Aj)} < 0. (5)
1
1
-_
122 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 2, NO. 2. MAY 1%
I' A. Nonrobust Stability Conditions (Case 1 ) Proofi It follows directly from Theorem 2.1.
Assume that Aw;(k) = 0 for all k and i. By substituting The nonrobust design problem for Theorem 3.1 is to select
(7) into (l), we obtain Fj ( j = 1,2, . . . , T - ) which satisfies the NSC of (1 1) for a
common positive definite matrix P when Aj and Bj are given.
r r
Next, we weaken the NSC of Theorem 3.1 by applying
Theorem 2.2 to (10). Theorem 3.2 gives a weakened NSC,
this is, a WNSC.
Theorem 3.2: Weak nonrobust stability condition (WNSC):
The equilibrium of a fuzzy control system described by (10)
is asymptotically stable in the large if there exists a positive
definite matrix P such that
From (8),
r(T+l)/z
s= w ~ ( ~ ) ~ ~ ( ~ ) { H-TP}P H ~
i=l
+ C u i ( k ) W j ( k ) { ~-TP~+HTPH~
~, -P
i<j
totically stable in the large if there exists a common positive It is found that there does not exist a common positive definite
definite matrix P such that matrix P s u e t h t
H'PH~ -P <o
HTPH; -P <o i = 1,2,3
(1 1)
since HZ is an unstable matrix. Therefore, the NSC of (1 1 )
for i = 1 , 2 , . . . , T ( T + 1)/2. is not satisfied.
TANAKA AND SANO: FUZZY CONTROL SYSTEMS
-0.s
where -1
-1 -0.5 0 0.5 1 1.5 2
w m
Fig. 4. Positive region of fi(w~(E),
ulz(k)).
where
G;j = Ai + BiF,
Without loss of generality, (13) can be rewritten as follows.
T X T
CVi(k)HiX(k)
i=l
x(k+1) = TxT (14)
CWi(k)
i=l
into the above equation, we obtain where
H(;-l)XT+'j = Gi'j
fl(Wl(k),WZ(k)) =
+
1.06w?(k) 6.75w:(k)w2(k) v(i-l)XT+j(k) = (wi(k) + AwZ(k))wj(k)
+ 9.62w7(k)wg(k) for all i and j. By appling Theorem 2.1 to (14), we derive a
robust stability condition (RSC) for the fuzzy control system
- 5.29w:(k)w;(k)
(14) with PPU. Theorem 3.3 gives a RSC for (14).
- 18.4wf(k)w;(k) - 5.29w:(k)wE(k) Theorem 3.3: Robust stability condition (RSC): The equi-
+ 9.62w;(k)w:(k) librium of .,a, fuzzy control system described by (14) is
+ 6 . 7 5 ~ 1 ( k ) d ( k+) 1.06w;(k) asymptogally stable in the large if there exists a common
positive definite matrix P such that
where ; P<o
H ~ P H- (15)
for i = 1,2;..,7- x T.
124 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 2, NO. 2, MAY 19%
--
V R G AR (17)
C . Admissible Region and Variation Region It is found that there does not exist a common positive definite
To introduce robust stability, we define admissible region matrix P such that
and variation region, which correspond to stability margin in
the ordinary control theory, from the WRSC of Theorem 3.4.
HTPH~ - P < o i = i,2,3,4
When a positive definite matrix P is given, by substituting H; since H2 and H3 are unstable matrices. Therefore, the RSC
and q ( k ) into (16) and by solving (16) for Aw;(k), we can of (15) is rst .tatisfied.
find an admissible region (AR) of Aw;(k)’s which guarantee Next, let us find an admissible region of Aw,(k)’s which
stability of the fuzzy system, (14), with PPU. On the other satisfies the WRSC of (16). Assume that
hand, we can find a variation region (VR) of Awi(k)’s, which
shows model error between a real system and an approximated
fuzzy model, if a mathematical model of the real system is
P= [:: 102].
TANAKA AND SANO: FUZZY CONTROL SYSTEMS
'tx
1
1
0.5
0 x(k)
0 5
-0.5
Fig. 6. Membership functions.
s=[I; I: 4
L
r - - - - - - - - - - - - - . - - -- .- -- .-
where
0.5
4 F - - - - - - - - - - - ---------_---
-1 -
0 0.2 0.4 0.6 0.8
Wl(k)
Fig. 1 1 .
0 0.2 0.4 0.6 0.8
Wl(k)
we obtain
0 . 5- I\l
I
Awl(k) = 1 - ~ l ( k ) ,
+
Awl(k) = - i ~ ( k ) 59 - wl(k),
x(k)< 4
4 <x(k)5 9
Awl(k)oPm
Admissble region ] Awl(k) = - ~ 1 ( k ) . 9 <x(k)
By substituting
Ai(x(k)) = w i ( k ) = 1, x(k)< o
-0.5 A l ( ~ ( k ) =) w l ( k ) = - ; z ( k ) + 1, 0 <x(k)2 5
U -1
0 0 . 2 0.4 0.6 0 . 8 1
Ai(x(k)) = w i ( k ) = 0,
into the above equations and by eliminating x ( k ) , we can
derive
5 <x(k)
0.5
I Admissbleregion ]
-
x3
Awi(k) 0
x3
-0.5
-1
0 0.2 0.4 0.6 0.8 1 ..........._..
V-q
‘ L .!I
w m
Fig. 12. Variation region (VR) for case c). Fig. 14. Truck-trailer model and its coordinate system.
1 .5
A . Models of a Truck-Trailer
We use the truck-trailer model formulated by Ichihashi [71
1
in this simulatiofl. After simplifying the truck trailer-model, we
0.5 approximate the simplified model by a fuzzy model with PPU.
0 Ichihashi used the following truck-trailer model.
-0.5 +
zo(k + 1) = z 0 ( k ) U . t/Z . tan [ ~ ( k ) ] (18)
-1 Zl(k) = z o ( k ) - d k ) (19)
-1.5 + +
z2(k 1) = z 2 ( k ) v . t / L . sin [ z l ( k ) ] (20)
5 3 ( k + 1) = z 3 ( k ) + U . t . cos [Zl(k)]
-=- case (a)
-2.5 case (b)
+
. sin [ { z z ( k+ 1) 22(k)}/2] (21)
5 4 ( k + 1) = Q ( k ) + U . t . cos [ X l ( k ) ]
-3 -+- case (c)
+ +
. cos [ { z 2 ( k 1) z2(k)}/2] (22)
Fig. 13. Control results.
where
O 5 sin(z) 5 X , -T 5x5n
x2(k) + t/{2L} . z l ( k ) O[deg],
the sine function can be exactly described by the state equation of the simplified model is approximated by
where y1 = 0, y2 = x and
5i=l i=li
GiYi/&W
w1= -
Y l - Y2 -X X
- sin(s)
[3
w2 =-.
X
Kawamoto [13] has reported that nonlinear models can be
exactly described by fuzzy models using this idea. A disad- + .(IC).
[I: 3
Q ( k + 1) =
0
d . ~ . t1
0
0
because
v.t
Ix2(k) +p ( k )
(34) for
179.997' < zz(k)+ 'U. t / { 2 L } . ~ < -179.997'
l ( k )
where d = 1Op2/7r. Of course, this system is theoretically This means that the dynamics of the simplified model can
controllable. be interpolated by two linear systems, A l x ( k ) b l u ( k ) and +
From (31) and (34), a fuzzy model, which approximately
represents the dynamics of the truck-trailer, can be derived as
+
A2z(k) b 2 u ( k ) , under
follows. 179.997" < ~ 2 ( k+) TJ . t / { 2 L } . ~ l ( k<) -179.997".
Rule : If z 2 ( k ) + 1 1 . t / { 2 L } . zl(k) isabout 0 (rad). That is, for j w i ( k ) E [0,1] and all k, (36) exactly describes
+ +
thenx(k 1) = A l x ( k ) b l u ( k ) the dynamics of the simplified model.
+
Rule : If x 2 ( k ) U ' t / { 2 L } . x 1 ( k ) isabout 7r 2
i=l
where w , ( k ) is membership value of the fuzzy set in Rule where A , w z ( k )is a value such that
i. Fig. 15 shows the fuzzy sets of "about 0 (rad)" and
"about 7r (rad) or -7r (rad)." We define the fuzzy sets as -1 5 AW,(k) I 1.
for all i
simple triangles. The fuzzy sets are unimodal and normal,
f.
+
u i z ( k ) Awt(k) >_ 0.
for all i
however, the dynamics of the approximated fuzzy model does
not exactly describe that of the simplified model. Let us show CiWdk) + Awx(k))> 0
that an exact approximation can be realized by using a fuzzy 1=l
model with PPU instead of (35). for all k . Equation (37), a fuzzy system with PPU, exactly
The simplified model can be generally represented by describes the dynamics of the simplified model under
x(k + 1) = f ( X ( k ) , ?L(k)) 179.997' < z Z ( k ) + TJ . t / { 2 L } . ~ i ( k )< -179.997'.
where f is a nonlinear function. Then, the condition, B. Controller Design and Robust Stability Analysis
Aix(k) + biu(k) I f ( x ( k - ) . u ( k ) )I A2x(k) + b n u ( k ) We design a fuzzy controller for backing up the truck-trailer
described by (37). The main idea of the controller design is
is satisfied under
to derive each control rule so as to compensate each rule of a
179.997' < ~ 2 ( k+) T Jt./ { 2 L } . x l ( k ) < -179.997' fuzzy system. Fig. 16 shows concept of the controller design.
130 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 2, NO. 2, MAY 1994 - -. -
Rule r t
Rule: +
If22(k) w-t/{2L} .z1(k)isaboutO(rad), It is sufficient to find a V R for
thenu(k) = flx(k), 0" 5 p ( k ) 5 180"
Rule : If 2 2 ( k ) +
t/{2L} . z l ( k ) isabout 7r (rad) because each membership function is symmetric with respect
or -7r (rad), thenu(k) = f2x(k), to the ordinate ( p ( k ) = 0) as shown in Fig. 15. Moreover,
where f1 and f 2 are feedback gains. We use the exact same sin ( p ( k ) ) , is also symmetric with respect to the ordinate
fuzzy sets in the premise part of the fuzzy controller. The ( p ( k ) = 0). In other words, the same V R is obtained also
purpose of controller design is to determine feedback gains for p ( k ) < 0. The membership function of the fuzzy set A1
of fi and f 2 . The following feedback gains are used in the defined in Fig. 15 is represented by
simulation:
f i l = 1.2837, f12 = -0.4139, f13 = 0.0201
when
Riccati equation for linear discrete systems was used to de- 0" 5 p ( k ) 5 180". (40)
termine these feedback gains because each consequent part is
From (30), we can derive the following equation.
represented by a linear state equation. The detailed derivation
of feedback gains f1 and f2 will be given in Appendix B. +
(wl(k) Awl(k)) . p ( k ) = sin p ( k ) . (41)
Next, we consider robust stability of the backing up control
system which consists of the simplified model and the fuzzy By solving the above equation for Awl(k), we obtain
controller. From AI, Az, bl, and b2 of the approximated
fuzzy model and f1 and f 2 of the fuzzy controller, we obtain
0.448 0.296 On the other hand, by solving (39) for p ( k ) ,
-0.364 1
0.364 -2 1
-0:14i7
0.666 0.051 -0.383 x 10-31 By substituting (43) into (42), we obtain
[-0.364
0.364
0.448
1
-2
0.296
O1
-0.014
1.
-0.364 1 V R is a region which satisfies (44) for 0 5 w1(k) < 1 and
0.116 x -.0.637 x low2 1 -1 5 Awl(k) 5 1. Fig. 26 shows the V R . The V R (the line)
H4= [ 0.666
-0.364
0.116 x
0.051
1
-0.637 x lov2
-0.383 x
0
1 1
perfectly belongs to the AR when
1
113.9 -92.61 2.540 Therefore,&t. cgntrol system at least satisfies the WRSC when
-179.29[deg] < x2(k) + z1. t/{2L}. x1(k) < 171.29[deg]
2.540 -3.038 0.5503
(46)
as a common positive definite matrix P. This matrix P was for all k. In other words, under the condition of (46), the
selected by the construction procedure of the literature [12]. designed fuzzy controller guarantees stability of the control
TANAKA AND S A N 0 FUZZY CONTROL SYSTEMS 131 . -_
Initiai, x3
posj t i o n \
!
20
lni!i!l
pas I :I \PO
\
0
-20
Fig. 17. Control result for Case 1 (simplified model).
Fig. 20. Control result for Case 4 (simplified model).
Initial VP
TABLE I
POSITIONS
INITIAL OF TRUCK-T~AILER
10
Case 4
1-20
Fig. 18. Control result for Case 2 (simplified model).
x3
t20 +-20
Fig. 21. Control result for Case 1 (Inchihashi’s model).
lnitiai, ~
Initial, x3
Lo
system which consists of the simplified model and the fuzzy
controller, that is, the backing up control can be perfectly
achieved from all initial positions.
The line of V R is close to the straight line of A w l ( k ) = 0
in the all range of 2u1 ( I C ) . This means that the dynamics of the
approximated fuzzy system agree well that of the simplified Fig. 22. Control result for Case 2 (Inchihashi’s model).
model, that is, that a fairly good approximation is realized.
model. The following points can be pointed out from the
simulation results.
C. Simulation Results 1) The designed fuzzy controller can effectively achieve
In this simulation, we use two kinds of the controlled backing up control even from difficult initial positions
objects: the Ichihashi’s model and the simplified model. Table such as Cases 2 and 4 in both of the Ichihashi’s model
I shows four cases of initial positions used in this simulation. and the simplified model.
Case 2 and Case 4 require to turn the truck-trailer in order to 2) The dynamics of the simplified model agree well with
realize a perfect backing up control. In particular, Case 4 is thakef‘rhe Ichihashi’s model.
a jackknife position. ~
Nobody can deny the first point, because the control system
Figs. 17-20 shows simulation results for four cases of initial is asymptotically stable in the large under the condition of (46).
positions in the simplified model. Figs. 21-24 show simulation The second point shows that a fairly good approximation is
results for four cases of initial positions in the Ichihashi’s realized.
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 2, NO. 2, MAY 1%
-_
132
?O
Initial ~
-20
Fig. 23. Control result for Case 3 (Inchihashi’s model).
0 j 0.2 0.4
W1($.6
0.8 1
x3 0.0451
APPENDIX A
THE PROOF OF THEOREM
2.1
A lemma is necessary in order to prove the condition of
Theorem 2.1. The proof of the lemma is given in the literature
~31.
Lemma 131 If P is a positive definite matrix such
that
-
I
L - Wk) ATPA-P<O and BTPB-P<O
0.045 1
Fig. 25. Admissible region of the truck-trailer system.
then
V. CONCLUSION
A robust stabilization problem for fuzzy systems has been A~PBB~PA - 2~ < 0. +
discussed in accordance with the definition of stability in
the sense of Lyapunov. We have considered two design
The proof of Theorem 2.1 Consider the scale func-
problems: nonrobust controller design and robust controller
tion V(x(k)) such that
design. The former is a design problem for fuzzy systems
with no premise parameter uncertainty. The latter is a design
problem for fuzzy systems with premise parameter uncertainty. V(z(lc))= xT(k)Px(k)
To realize two design problems, we have derived four stability
conditions from a basic stability condition proposed by Tanaka
and Sugeno: nonrobust condition, weak nonrobust condition, where P is a positive definite matrix. This function satisfies
robust condition and weak robust condition. We have intro- the following properties:
duced concept of robust stability for fuzzy control systems
with premise parameter uncertainty from the weak robust
condition. To introduce robust-stability, admissible region and a) V ( 0 )= 0
variation region, which oorrespond to stability margin in the b) V ( x ( k ) )> 0 for x ( k ) # 0
ordinary control theory, have been defined. Furthermore, we
have developed a control system for backing up a computer c) V(x(lc)) approaches infinity as llx(k)(l-+ 00.
-
TANAKA AND SANO: FUZZY CONTROL SYSTEMS 133 - -. .
P
fi = - ( T + bTPlbl)-’bTP1Al
r ~ 10,qz
where f1 = [ f ~ ~ , f 1 2 r f 1 3 1 = ~ = 1,43 = 1, and
T = 1000, we obtain
REFERENCES
Q=
T
[“‘
>O.
0
0
42
0 :I
43
41 > 0742 > 0, and 43 >0
199-249, 1975.
[15] -, “The concept of a linguistic variable and its application to ap-
proximate reasoning part 2,” Information Sciences, vol. 8, pp. 301-357,
1975.
[16] -, “The concept of a linguistic variable and its application to
approximate reasoning part 3,” Inform. Sci., vol. 9, pp. 43-80, 1975.
134 IEEE TRANSACIIONS ON FUZZY SYSTEMS. VOL. 2. NO. 2, MAY 19%
Kazuo Tanaka (S’87-M’91) received the B.S. Manabu San0 (M’93) received the B.S. and MS.
and M.S. degrees in electrical engineering from degrees in mechanical engineering from Kanazawa
Hosei University, Tokyo, Japan, in 1985 and 1987, University, Kanazawa, Japan, in 1973 and 1975. and
and Ph.D. degree, in systems science from Tokyo Dr.Eng. degrees in control engineering from Tokyo
Institute of Technology, in 1990, respectively. Institute of Technology, in 1983, respectively.
He is currently an Associate Professor in the He is an Associate Professor in the Mechanical
Mechanical Systems Engineering Department at Systems Engineering Department at Kanazawa Uni-
Kanazawa University. He was a Visiting Scientist versity. His research interests include fuuy l+c
in Computer Science at the University of North control, pattern recognition, and computational neu-
Carolina at Chapel Hill. He received the Young ral networks.
c Engineer Award from the Japan Society for Fuzzy
Theory and Systems in 1990, the Theoretical Papers Award at the 1990
Annual NAFTPS Meeting in Toronto, Canada, in 1990, and an Award at the
Joint Hungarian-Japanese Symposium on Fuzzy Systems and Applications in
Budapest, Hungary, in 1991, and the Young Engineer Award from the Japan
Society of Mechanical Engineers, in 1994. His research interests include
principle, analysis and design of intelligent control systems such as fuzzy
control, near0 control, evolutionary control and so on.