zyxwvutsrqponmlkjih
zyx
zyxwvutsrqponmlkj
zyxwvutsrqp
zyxwvut
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 41, NO. I 1, NOVEMBER 1996
+
CL have equally many zeros in the right-half
plane when p > 0.
3) The system ? c ( s ) / d ( s ) is stabilizable by a srdble controller.
Proof ( 1
2 ) : By the extended version of the zero exclusion
principle given above, and by using the fact that p o and p I are stable,
we deduce that the polynomials X p o (s) (1- X)pl ( s j are stable for
0 5 X 5 1 iff
2) p o ( s ) p l ( - s )
+
+
1691
Remarks: The equivalence between statements 1) and 2) is also
shown in [ 6 ] . Systems n ( s ) / d ( s )resulting from 110 and p l may be
nonproper. It may also be identically equal to infinity when d ( s ) 0.
This special case happens only when PO(.) = l i p ~ ( s for
) some real
X:, which is a trivial case. (We thank one of the reviewers for this
remark.)
The existence of a stable stabilizing controller for a system can be
checked by the techniques described in [ l ] and [2].These conditions
can thus be used as alternatives to the conditions given in [4] for
testing the stability of convex combinations of polynomials.
zyxwv
+ (1 - X ) p l ( j ~#) 0 0 < X < 1:
Xp~(s+
) (1 X)p~(s)has invariant
Xpo(,j-i’)
U,
E R.
(The family
degree when
0 < X < 1 since P O . P I are stabie and have same sign.) This last
condition can be expressed equivalently by
po(j&)
zyxwvutsrq
zyxwvutsr
+ ppl(,jd) i.0
__-
Multiplying both sides by
equivalent condition
111(
0
< p 5w
E
R.
j ~ =) p ~ ( - j w ) , we arrive at the
+
p o ( j ~ ) (p -~j ~ ) /1 ’# 0 0 < p , zu t R.
A second application of the extended zero exclusion principle leads
us to the conclusion.
(2 + 3): From the extended zero exclusion principle, the condition that po(s)pl ( - s ) p have equally many zeros in the right-half
plane when p > 0 is equivalent to
+
+
U O ( . ~ U ) ~ I ( - ~ L O/L
)
# 0 0 < p.
w
+
REFERENCES
11 B. D. 0. Anderson and E. I. Jury, “A note on the Youla-Bongiorno-Lu
condition,” Automutica, vol. 2, pp. 387-388, 1976.
121 V. Blondel and C. Lundvall, “A rational test for strong stabilization,”
Automutica, vol. 31, pp. 1197-1198, 1995.
131 B. R. Barmish, New Toolsfor Robustness of Linear Systems. New York
McMillan, 1994.
141 M. Fu and R. Barmish, “Maximal unidirectional perturbation bounds
for stability of polynomials and matrices,” Sq’st. Contr. Lett., vol. 11,
pp. 173-179, 1988.
PI D. C. Youla, J. J. Bongiorno, and C. N. Lu, “Single-loop feedback
stabilization of linear multivariable plants,” Automatica, vol. 10, pp,
159-173, 1974.
161 E. Zeheb, “Necessary and sufficient conditions for root clustering of
polytopcs of polynomials in a simply connected domain,” IEEE Trans.
Automat. Contr., vol. 34, pp. 986-990, 1989.
zyxwvutsrq
E R.
The decompositions po(.s) = p o o ( - s 2 )
s p o l ( - s 2 ) and til (s) =
p 1 ” ( - s 2 ) s p l l ( - s 2 ) lead to p”(,ju)
= POO(U,’)+ j ~ , p ~ ~
and( u ~ )
pi ( - . i ~=) p10(d~)-juip11(~~).Therefore,
anequivalent condition
is given by
+
[PO0(LO2)PlO(U2)
+
+
d2VOI (W”L)p11bJ2)
On Variable Structure Output Feedback Controllers
{L]
+.iubol(L,‘2)Pio(U,2)
- Poo(w’2)pll(-i’2j]# 0,
0 < p > 9 E R.
C. M. Kwan
By looking at the real and imaginary parts of this expression, we
deduce the equivalent condition that
Abstract-In a recent work by Zak and Hui [I], a sliding-mode
controller for multi-inputlmulti-output(MIMO) systems using static output feedback was proposed. Very nice geometric conditions for how to
Po”(LJ2)Plo(U2)
U,2Pol(u2)pll(U,2)2 0
design sliding surfaces were given. However, there are two restrictive
assumptions in it. One is that the uncertainties in the system must be
whenever
bounded by a known function of outputs which excludes some possible
U1 E R.
uncertainties in the A matrix if the system is described by the triple
L d ~ 0 1 ( U , ” ) p 1 0 ( L J 2 ) -p”o(Ll:2)plI (U”] = 0,
(A.4, C ) . The other one requires a matrix equality [l,(4.3)] to be held
In other words, the polynomial n ( s ) = y o o ( , ~ ) p l o ( s ) + . ~ ~ , o l ( s j l ~ l l (which
s ) may also be very difficult to sati3fy in many systems. In this paper,
must take positive values whenever d ( s ) = s ~ i ~ l ( s ) p l o (-s j we propose a modification of the sliding mode controller for a class of
poo(.s)pl~(s)]is equal to zero on the positive real axis. Since single-input/single-output(SISO) systems which can eliminate the abovementioned limitations and, under certain conditions, guarantee global
p o o ( O ) p 1 0 ( 0 ) = p o ( O ) p l ( O ) > 0, this condition is satisfied if and closed-loop stability. Hence the range of applicability of the method in
only if d ( s ) has an even number of zeros between each pair of positive 111 can be greatly broadened.
+
real zeros of n ( s ) .By the parity interlacing condition (see [SI), this
is equivalent to the requirement that IrL(s)/d(s) is stabilizable by a
stable controller.
U
Example: Let p o ( s ) = 10s“ s2
Gs
0.57 and p l ( s ) =
Ius3 2 2 8s 1.57. It is shown in [3] that although both
polynomials are stable, not all convex combinations of p o and p i are
stable. We verify this result by direct application of Theorem 1. From
P O (s) = - ( - s’ ) 0.5 7 s [ 10( - s’ ) 61 and pl ( s ) = - 2( -s 2 )
1.57+.s[-10(-s2)+8] we constnict n ( s ) = 10.s’-14s2+s4.86s and
d ( s ) = 1 0 0 -~138.~’+45.20~+0.8949.
~
The polynomial d ( s ) has a
single zero (at 0.5991) between the pair (0, 0.6368) of zeros of n ( s ) ,
and the system r r / d is therefore not stabilizable by a stable controller.
The convex combinations of 710 and 111 are thus not all stable. This
can be verified by checking that 2/3p0 1 / 3 p l is unstable.
+
+ + +
+ +
+
+
~
+
+
+
1. INTRODUCTION
Sliding mode control (or variable structure systems control) is
a popular robust control method among control engineers. It is
simple to design and completely robust to “matched” uncertainties.
Its importance and applications can be seen from two recent special
issues in the Intevnational Journal of’ Control 121 and the IEEE
TRANSACTIONS ON INDUSTRIAL
ELECTRONICS
[ 3 ] .One major drawback Of
sliding control is that states have to be available. Since in many
Manuscript reccived May 2, 1994; revised March 7, 1995 and July 1, 1995.
The author is with Intelligent AutomaLion Inc., Rockville, MD 20850 USA
(e-mail: ckwan@i-a-i.com).
Publisher ltem Identifier S 0018-9286(96)06767-0.
001 8-9286/96$05.00 0 1996 IEEE
zyxwvutsrqp
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL 41, NO 11, NOVEMBER 1996
1692
practical applications, only system outputs or part of system states
are available, this necessitates the development of sliding mode
control methods using output feedback. Emelyanov et al. [4] used
observers to reconstruct states. Zak and Hui [ I ] proposed a new static
output feedback scheme without using observers. Very nice geometric
conditions were given on the switching surface design. In this note,
we concentrate on a single-inputlsingle-output (SISO) linear system
of the form
zyxwvutsrqpo
zyxwvutsrqponm
zyxwvutsrq
zyxwvutsr
zyxwvutsrq
All
:y = c x
The second assumption is that to guarantee the sliding condition
m E 0. Zak and Hui [ l ] required that
(1)
for some matrix 124. Even for SISO systems, (8) is difficult to achieve.
If (8) cannot be satisfied, the authors in [ l , (4.6)] resorted to a
bounded controller which can only guarantee local stability.
In the next section, we shall propose a modification of the sliding
mode controller which will eliminate the two limitations and, at the
same time, guarantee the global closed-loop stability. It should be
emphasized here that our controller is no longer static. Our simple
dynamic controller can guarantee global stability under the following
conditions.
1) Solvability of equation S = F C , i.e., Zak and Hui [ l , Th.
with .rl E Rn-‘ , . r 2 E R..c = [x: X 2 I T , u E R , and :y t R. -4,,
are matrices with appropriate dimensions. Assume the uncertainties
A&, A&, Ah, d ( t ) are all bounded and satisfy the so-called
“matching” conditions. i.e.,
4.31, is needed.
2) Existence of S is needed [ I , Th. 4.11.
11. MAINRESULTS
with
D,’s
unknown but bounded quantities. Hence (1) can be written
From (7), we can express
~ ‘ in
2
terms of
TI
and
0
as
as
Note that SS is a scalar in SISO systems. Using (9), (2) can be
recasted as
zyxwvutsrqpon
zyxwvutsr
I.,+
zyxwvutsrqpo
zyxwvutsrqp
There are two major assumptions in the paper by Zak and Hui [l].
The first assumption of [ l ] is that
The eigenvalues of (All
AltS;’S1) in (loa) are {-XI. - X 2 .
. . . , -XVL--1} which can be done by suitably choosing SI.
S2 such
that [l, Th. 4.11 is satisfied.
To guarantee the sliding condition (T E 0. we differentiate the
sliding variable in ( 5 ) and use (IO) to get
~
with p a known function of outputs. This is quite restrictive because
it can be seen easily that (3) does not satisfy (4). In other words,
AArl, AA22 may not be allowed to exist in the system. For example,
let us consider
x = [ a21
0
y=[l
CI.22
+ Aa2t
[:‘]U
O]C.
The uncertainty is E = A a 2 2 x 2 . It is not possible, in this case, to
Ax! r” 3s 2 fttftctior? 9f I .
The sliding variable is defined as [I]
where
h!f
with F a constant matrix. The choice of F is related to the design
of state switching surface
0
=SX
(6)
or, with no loss of generality
The matrix F should be selected to satisfy
S = FC.
Geometric conditions were given in [ 11 for how to choose F and S so
that ( 1 2 - 1)prescribed distinct, nonzero, and real negative eigenvalues
{-A1,-A2;...
.-An-l}
with A,’s>O can be assigned.
N
=si(411 AA12S,1S1) + s2(-421 A ~ ~ S ~ ’ S I )
+ Szb(D1 - D2s,1sl)
= S,.A,,S,’
+ S a ( A a a +bDZ)SF1.
-
-
Since (11) contains 2 1 , Zak and Hui used (8) to avoid the measurement of states. If (8) cannot be satisfied, a bounded controller
is used which can only guarantee local stability. As we shall see in
a moment, we can eliminate these limitations while, under certain
conditions, guaranteeing global stability by a slight modification of
the controller. First we need the following lemma.
Lemma: Consider (loa). Let A,,
be the minimum value
of { A I , A ~ , . . . , A ~ - ~ } where {-A1,-A2:...
-Xn-l}
are the
eigenvalues of ( A l l - A12ST1Sl) with A’s all positive real values.
Then we have
i) j ( e x p ( A l l - 2412S21Sl)t((
5 k e x p ( - ~ , t ) for some k > ~ .
ii) Ilrl(t)ll is bounded by ~ ( f for
) all time, where u l ( t ) is the
zyxwvutsrqpo
zyxwvutsr
zyxwvutsrqponml
zyxwvutsrq
1693
IbEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 41, NO. 11, NOVEMBER 1996
REFERENCES
solution of
+ k11412s;-11( IoI,
i ( t ) = -X,,u,(t)
.i(O)
5
k ~ ~ . E L ( O j ~ ~ . [ I ] S. H. Zak and S. Hui, “On variable structure output feedback controllers
for uncertain dynamic systems,” IEEE Trans. Automat. Contr., vol. 38,
(12)
no. IO, pp. 1509-1512, 1993.
[21 A. Sabanovic, N. Sabanovic, and K. Ohnishi, “Sliding modes in power
Proof: It is obvious that i) hoilds for some positive constant k .
converters and motion control systems,” Special Issue on Sliding Mode
To see ii), we solve (loa) to yield
Control, Int. J, Contr., vol. 57, no. 5 , pp. 1237-1259, 1993.
[3] J. Y. Hung, W. Gao, and J. C. Hung, “Variable structure control:
(t)l)= I I C ( A 1 1 - A l a S 2 1 . w t I1 IIx1(0)lI
A survey,” Special Issue on Sliding Mode Control, IEEE Trans. Ind.
Electronics, vol. 40, no. 1, pp. 2-22,, 1993.
+
1 1 e( A I I - A I z S; ’ -71) ( t- 7 )II
141 S. V. Emelyanov, S. K. Korovin, A. L. Nersisian, and Y. E. Nisenzon,
“Output feedback stabilization of uncertain plants: A variable structure
’ (IA12S,11(
IffIdT
approach,” Int. J. Contr., vol. 55, no. 1, pp. 61-81, 1992.
[ 5 ] E. Bailey and A. Arapostathis, “Simple sliding mode control scheme
< kc-x“tllSL(o)II
k*c-x-(‘-’)
applied to robot manipulators,” Int. J. Contr., pp. 1197-1209, 1987.
]Iz,
1‘
zy
zyxwvutsrqp
zyxwvutsrqpo
+
-
. IIAlzs,’ 1 1 101
5 e-X”i,a,(0) +.
‘
IIAlaS;l
if ~ (
l
(LT
I’
&-Am“(-‘)
1 1 Iff1 d r = w ( t ) ,
0 5) kll.cl(0)ll > 0
Discrete-Time Neural Net Controller for a
Class of Nonlinear Dynamical Systems
where ui ( t )satisfies (1 2). Thus, by comparing both sides of the above
inequality, one can see that w(t) 5 I I z l ( t ) l I for all time if ~ ( 0 is)
sufficiently large.
S. Jagannathan and F. L. Lewis
(-1.E.D.
Abstract-A family of two-layer discrete-time neural net (NN) controllers
is presented for the control of a class of mnth-order multi-input-multioutput (MIMO) dynamical system. No initial learning phase is needed
so that control action is immediate; in other words, the neural network
(NN) controller exhibits a learning-while-functioning-feature instead of
a learning-then-control feature. A two-layer NN is used which is linear
in the tunable weights. However, this is a far milder assumption than
the adaptive control requirement of linearity in the parameters, since
the universal approximation property of the NN means that any smooth
U = (Sab)-’(l
- D;”””)-’[-Hw(t)
- Glal - 51 sgri ( a ) (13) nonlinear function can be reconstructed. The structure of the neural net
controller is derived using a filtered error approach. It is indicated that
where
delta-rule-based tuning, when employed for closed-loop control, can yield
unbounded N N weights if: 1) the net cannot exactly reconstruct a certain
(
1
4
4
H L
required function, or 2) there are bounded unknown disturbances acting
(14b) on the dynamical system. Certainty equivalence is not used, overcoming
G 21
1
N
1
a major problem in discrete-time adalptive control. In this paper, new
b 2 s11p[ISzhD,,(tjI]+ q ,
‘t)>O.
( 1 4 ~ ) on-line tuning algorithms for discrete-time systems are derived which are
Note that o = F g , and hence (13) does not require states. Substituting similar to f-modification for the case of continuous-time systems that
include a modification to the learning rate parameter and a correction
(13) into ( l l ) , one gets
term to the standard delta rule. These improved weight-tuning algorithms
guarantee tracking as well as bounded NN weights in nonideal situations
so that persistency of excitation (PE) is not needed.
The hound on Ilxl (0) 11 can be estimated by following the definition
of each state. From their physical meanings, we can easily get a crude
estimation of I(z1(0)11.
Now let us go back to (1 1). Assume 0 < D F n < I D3 I < Dinax< 1
which is a quite reasonable assumption. The sliding controller is
chosen as
zyxw
zyxwvuts
Il-~~lI
zyxwvutsr
zyxwvutsr
+
Noting w ( t ) 2 I l z ~ ( f ) l from
l
the Lemma, the fact that 1 DJ
1 - Dinax,and using the inequalities in (14), (IS) then becomes
a& 5 - r / l n l < 0.
I. INTRODUCTION
2
(16)
Hence o tends to zero in finite time [SI.The time to reach sliding
mode is determined by q. The larger the 1 1 , the shorter the time will
be to reach sliding.
111. CONCLUSIONS
We proposed an output feedback sliding controller for a class
of SISO systems. The proposed controller is a modification of the
controller in [l]. Two major limitatlions of the scheme in [ I ] have been
eliminated by using a simple dynamic controller. This implies that the
applicability of the scheme by Zak and Hui can be greatly broadened.
Another merit is that we can guarantee global stability under certain
conditions [l, Ths. 4.1 and 4.31 by using output feedback only.
Considerable research has been conducted in system identification
or identification-based neural network (NN) control [SI, [IO], [12],
[131 and little in the use of direct closed-loop NN controllers that yield
guaranteed performance. On the other hand, some results presenting
the relations between NN and direct adaptive control [2],as well
as some notions on NN for robot control, are given in [4]. A direct
continuous-time neural net robot controller was proposed in [4] which
guarantees closed-loop tracking performance. However, little about
Manuscript received August 15, 1993; revised April 30,1994 and November
22, 199.5.This work was supported in part by the National Science Foundation
under Grant IRT-9216545.
S. Jagannathan is with the Automated Analysis Corporation, Peoria IL
61602 USA (e-mail: saranj @cujo.tc.cat.com).
F. L. Lewis is with the Automation and Robotics Research Institute, The
University of Tcxas at Arlington, Fort Worth, TX 76118 USA.
Publisher Item Identifier S 0018-9286(96)04373-5.
0018-9286/96$05,00 0 1996 IEEE