Anda di halaman 1dari 8

386

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

REFERENCES
[1] E. J. Davison, The robust control of a servomechanism problem
for linear time-invariant multivariable systems, IEEE Trans. Automat.
Contr., vol. 35, no. 2, pp. 131140, 1990.
[2] C. A. Desoer and Y. T. Wang, Linear time-invariant robust servomechanism problem: A self-contained exposition, Contr. Dynamic Syst., vol.
16, pp. 81129, 1980.
[3] F. D. Priscoli, Robust tracking for polynomial plants, in Proc. European Contr. Conf., June 1993, pp. 369373.
[4] B. A. Francis, The linear multivariable regulator problem, SIAM J.
Contr. Optimization, vol. 15, pp. 486505, 1977.
[5] B. A. Francis and W. Murray Wonham, The internal model principle
of control theory, Automatica, vol. 12, pp. 457465, 1976.
[6] J. Huang and C.-F. Lin, On a robust nonlinear servomechanism
problem, IEEE Trans. Automat. Contr., pp. 15101513, 1994; also
in Proc. 30th Conf. Decision Contr., Briton, England, Dec. 1991, pp.
25292530.
, Internal model principle and robust control of nonlinear sys[7]
tems, in Proc. 32nd Conf. Decision Contr., Dec. 1993, pp. 15011506.
[8] J. Huang, On the minimal robust servo-regulator for nonlinear systems, Syst. Contr. Lett., vol. 26, pp. 313320, 1995.
[9] J. Huang and W. J. Rugh, Stabilization on zero-error manifolds and
the nonlinear servomechanism problem, IEEE Trans. Automat. Contr.,
pp. 10091013, July 1992.
[10] W. Kang, P. K. De, and A. Isidori, Flight control in a windshear via
nonlinear h methods, in Proc. IEEE Contr. Decision Conf., 1992,
pp. 11351142.
[11] C. I. Byrnes, F. D. Priscoli, A. Isidori, and W. Kang, Structurally stable
output regulation of nonlinear systems, preprint.
[12] E. G. Albrekht, On the optimal stabilization of nonlinear systems, J.
Appl. Math. Mech., vol. 25, pp. 12541266, 1962.

New Results on Output-Feedback Variable


Structure Model-Reference Adaptive
Control: Design and Stability Analysis
Liu Hsu, Fernando Lizarralde, and Aldayr D. de Araujo

AbstractThe variable structure model-reference adaptive control (VSMRAC) redesign and stability analysis, presented earlier for the case
2, is generalized to the case of
of plants with relative degree n3
1. This paper presents a complete stability analysis in
arbitrary n3
the general case. The redesign is based on the explicit consideration of
input disturbances which may include the disturbance originated by the
uncertainty of the plant high-frequency gain. This leads to a considerably
simpler and less restrictive stability analysis as well as to a systematic
controller design for external disturbance rejection. The overall error
system is globally exponentially stable with respect to a small residual
set. Robustness to unmodeled dynamics is also discussed.

Index Terms Disturbance rejection, modal reference adaptive control, output feedback, robustness, stability, uncertain systems, variable
structure systems.

I. INTRODUCTION
In this paper, we consider the class of variable structure (VS) model
reference adaptive controllers proposed in [7], [8], and [10], which
only require input/output measurements to be implemented, referred
to as I/O VS-MRAC. Adaptation is achieved through signal synthesis,
instead of parameter adjustment. For simplicity we will omit the
acronym I/O from now on. The acronym MRAC will denote the
usual parameter adaptive model reference schemes.
The main interest in the VS-MRAC relies on its remarkable stability, disturbance rejection, and performance robustness properties, as
compared to parameter adaptive and robust linear controllers [7], [8],
[10]. Recent related works have also considered robustness issues
and the extension to multivariable plants [1], [2], [18]. Successful
full-scale experimental implementations of the VS-MRAC include
underwater vehicle control and robot manipulator control [3], [12].
The stability analysis for the case of relative degree n3 = 1 is rather
simple [10]; however, it becomes much more involved for n3  2
(see [8] for n3 = 2).
Based on the key observation that a plant with known highfrequency gain (HFG) is much simpler to analyze, a redesign of the
VS-MRAC was proposed in [11]. The resulting stability analysis is
less complicated than in previous works. Also, some restrictive assumptions made in earlier works [8] were removed. Details, however,
were presented only for the case n3 = 2.
This paper presents a complete stability analysis in the general
case including external disturbances. The extension to plants with
unmodeled dynamics is also discussed.
II. PRELIMINARIES
Consider an unknown single-input/single-output linear time inManuscript received January 21, 1994; revised November 22, 1994,
September 30, 1995, and May 30, 1996. This work was supported in part by
the Brazilian Research Council (CNPq).
L. Hsu and F. Lizarralde are with the Department of Electrical Engineering,
COPPE/Federal University of Rio de Janeiro, CP 68504, 21945/970 Rio de
Janeiro, Brazil.
A. D. de Araujo is with the Department of Electrical Engineering, UFRN,
59072/970 Natal, Brazil.
Publisher Item Identifier S 0018-9286(97)02037-0.

00189286/97$10.00 1997 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

THE ALGORITHMS

387

TABLE I
MRAC

FOR

AND

VS-MRAC

variant plant with a strictly proper rational transfer function given by1
W (s) = Kp Np (s)=Dp (s), with input u and output y, and a model
reference having input r and output ym characterized by the rational
transfer function M (s) = Km Nm (s)=Dm (s).
The objective of model following is to find a control law u(t)
such that the output error e0 := y 0 ym tends asymptotically to
some small residual interval for arbitrary initial conditions and an
arbitrary piecewise continuous uniformly bounded reference signal
r(t). The usual assumptions for MRAC are made [15, p. 183]. An
input disturbance d(t) is also included. Such a disturbance accounts
for the uncertainty of the HFG and for uniformly bounded external
disturbances.
A. Notation

1) Given a vector x 2 Rn , its elements are denoted


x1 ; x2 ; 1 1 1 ; xn . We denote by jxj, either the absolute
value of a scalar x, or the vector [jx1 j; 1 1 1 ; jxn j]T , if x is
a vector (x 2 Rn ).
2) Operators and convolution operators: For a precise meaning
of mixed time domain (state-space) and Laplace transform
domain (operator) representations, the following notations and
concepts are adopted. The output y of a linear time invariant
system with transfer function H (s) and input u is denoted
H (s)u. For (strictly) stable H (s), consider an arbitrary, possibly nonminimal, stable realization x_ = Ax + Bu; y =
Cx + Du. Let h(t) be the impulse response and y0 be the
transient term due to the initial state x0 (0) of the homogeneous
system x_ 0 = Ax0 and y 0 = Cx0 . The state x0 is called the
transient state. The superscript (1)0 will henceforth be used
to identify the transient terms. Then, the following notation is
adopted:

y(t) = H (s)u(t) = h(t) 3 u(t) + y0 (t):


Pure convolution operations will be denoted

h(t) 3 u(t)), thus, Hu = H 3 u + y0 .

H (s) 3 u

(=

1 The symbol s represents either the Laplace variable or the differential


operator (d=dt), according to the context.

Fig. 1. Block diagram of the VS-MRAC.

Fig. 2. VS implementation of the operator

L.

B. The Controller Structure


The block-diagram of the VS-MRAC is shown in Figs. 1 and 2.
The corresponding algorithm is presented in Table I alongside the
MRAC algorithm. One can verify that the computational complexity
is comparable.
The operator L(s) = L1 (s) 1 1 1 LN (s), where Li (s) = s + i ,
i > 0, and N := n3 0 1, is chosen such that ML is strictly
positive real (SPR). The operator L is an approximation of L. In
the VS-MRAC, the VS implementation of Fig. 2 is adopted. If
n3 = 1; L = L = 1, then the prediction error becomes null,
i.e., e^0  0, and one can remove the internal loops passing by
knom ML. The averaging filters Fi01 (i s) in Fig. 2 are low-pass
filters where Fi is a Hurwitz polynomial in i s and F (0) = 1,

388

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

( )=

+1

e.g., Fi i s
i s
. With sufficiently small i , they give an
approximation of the equivalent control signals Ui eq [17].
In the MRAC design, the control is parameterized as u !T !
1 ; 1 1 1 ; 2n01 (see Table I); !3 and 23n are the
2n r, where !T
matching parameters, i.e., with u u3 !3T ! 23n r, the closedloop transfer function matches M s exactly, i.e., y M s r. Note
=23n Kp =Km .
that 23n Km =Kp . We also define k3
To account for the available a priori knowledge we introduce
nom
U nom !nomT ! unom 2nom
n r U

( )

=[

]
= =
()
=1

+
= ()
=

where ! ;  n ; and k
are nominal values for !3 ; 23n ; and
3
k ; respectively. Then, the uncertainty upper bounds j and  are
; 1 1 1 ; n)
defined as (j
nom

j

nom
2

=1

nom

 j3 0 j

nom

;   j(k3 0 k )=k j:
nom

nom

C. A Formulation for the Uncertainty on Kp


nom
Following [11], if Kp
Kpnom
is some
k , where Kp
nominal value of Kp and k is the uncertainty on Kp , one can write

+1

N (s)
y = Kpnom p [u + du ]
Dp (s)

du =

=
^

1k u =  u:
K

(2)

knom = Kpnom =Km ; k3 = Kp =Km


 = (k3 0 knom )=knom ;  = 1 +  = k3 =knom :

()

()

()

nom

nom

where

U

()

()

( ) = ( ) ( )= ( ) ( )

nom

nom

nom

d(t) = du (t) + de (t) = u(t) + de (t)

gives the perfect model following control law, it is reasonable to


restrict the class of admissible control laws to that defined by

8t  0 : sup
ju(t)j  K! sup
k!(t)k + K
t
t

( )

()

(6)

where K! and Kred are arbitrary positive constants. This condition


is consistent with the modulation function fN to be determined
(Section IV) and assures that all signals of the system belong to L1e .
Condition (6) will be relaxed in Section IV to be satisfied modulo
some exponentially decaying term, due to initial conditions.
A. Error Equations
From Figs. 1 and 2, the auxiliary error equations can be expressed
by (i
; 1 1 1 ; N)

"0 = knom ML[0U0 0 L01 (U )]


"i = Fi01 [Ui01 ] 0 Li01 [Ui ]:

The dynamic system governing "0 can be represented as

x_ e

= Ac xe + k

nom

bc0 [0U0 0 L01 (U )];

"0 = hcT xe :

(7)

In order to obtain less conservative (smaller) modulation functions,


the following identity [15, p. 320] is used:

Wd (s) = 1 0 G^ 1 (s)

(8)

^ = [^ ^ ]
^ ( ) = ^ ( + 3)
^
= +
^
=
=
= (^ + ^ )

U = Ud + UN
(9)
1
3
Ud = (U 0 U ) +
0  n r 0 Wd de : (10)
k

01gf ; (uT 1 1 1 1 n01 ; and gf from


where G1 s
uT sI
Table I), i.e., G1 u is the contribution of the input filter to the ideal
du de
matching control u. Then, taking into account that d
(du
u) and noting that U 3
!3T !
U G1 u = (see
Appendix A), the signal U of (5) can be rearranged as
nom
2

After some algebraic manipulations, taking into account (10), the


;111;N 0 )
auxiliary errors can be rewritten as (i

=1

"0 = k ML[0U0 0 L01 (U )]


1
"i = Li01 0 Ui 0 F10;i1 Li0+1
;N (U ) + ei + 0i
"N = L01 0 (UN + F 01 Ud ) +  uN + eN + 0N
nom

where du is given by (2) and de is uniformly bounded by hypothesis.


It is also assumed that an instantaneous upper bound de t of de t

nom

This section considers the problem of redesign and stability analysis of the VS-MRAC (Fig. 1), for a plant (2) with known HFG
Kpnom , under the action of a general input disturbance

(4)

nom
2

nom

III. STABILITY ANALYSIS

(3)

= U^ 0 U + (1=k ) 0  n r 0 Wd d
(5)
 d , and W d (s) = hcT (sI 0 Ac )0 b0
with Wd (s) = [k
M (s)]0 W
is the closed-loop transfer function from d to e with control u = u
^.
Since there exists U 3 = !3T ! such that u3 = U 3 + 3n r 0 Wd de

=1

When dealing with systems described by differential equations with


discontinuous right-hand sides, it is necessary to define precisely the
meaning of a solution for such systems. Here, Filippovs definition
is assumed [6]. Note that u or d are not necessarily functions of
t in the usual sense when sliding modes take place. In order to
avoid clutter, we will denote by u t and du t the locally integrable
functions which are equivalent to u and du , respectively, in the sense
of equivalent control, along any given Filippov solution z t of the
closed-loop system. It should be stressed that z t is, by definition,
absolutely continuous. Then, along any such solution, u or du can
be replaced by u t or du t , respectively, in the right-hand side of
the governing differential equations. Although the equivalent control
ut
ueq t is not directly available, filtering u with any strictly
G s ueq t .
proper filter G s gives G s u G s u t

()

^=^

D. Differential Equations with Discontinuous Right-Hand Side

()=

red

Thus, the uncertainty on the HFG can be formulated as an input


disturbance du . External input disturbance rejection can be achieved
by adequate redesign of the modulation functions fi . Here, the
difficulty is that no prior bound is available since du is state dependent
in the closed-loop system. Notwithstanding, we show in Section IV
that the modulation functions for known k3 can be appropriately
redesigned so as to solve the problem for uncertain k3 . The following
notation will also be adopted (note that  > ):

()

^= ^+

_ = Ac e + k bc(0UN 0 U )
e = hcT e
I/O form: e = k
M (s)[0UN 0 U ]

pnom

()

() ( )

State-Space form: e

(1)

The relay modulation functions fi are calculated from the regressor


vector ! , the filtered signals (Table I), and the uncertainty bounds (1).

=
1

( )

is known and satisfies de t  jde t j 8t . The term du allows


the case of unknown HFG to be embedded in the case of known
HFG. The redesign is centered on the derivation of conditions to be
satisfied by the modulation functions in order to guarantee stability
and disturbance rejection.
It is important to remark that the standard error equations (see
nom
U 2nom
[15]) have to be rewritten in terms of u
n r and k
3
3
3
3
instead of u
U
2n r and k , respectively, where U !T !
is such that u is the ideal control law that leads to perfect model
N (s)
following of the plant (2) (Kpnom D (s) ), with du  de  . The
error equations become

;N

(11)
(12)
(13)

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

where Li;j = jk=i Lk (Li;j = 1 if j < i), Fi;j is defined in a


similar way and (by convention, e1  0)

01[UN ]
uN = (F1;N 0 1)F10;N1 LN
i01
1
01
ei =
Lj;i01 Fj0+1
;i ["j ] = Li01 Fi [e;i01 + "i01 ]

(14)
(15)

j =1

0i = (knom MF1;i Li;N )01 ["0 ]:

(16)

Remark 1: In [7] the averaging filters Fi0 s were believed to be


necessarily of order 2 or higher. It is sufficient, however, to use firstorder filters. The essential tool that has led to the new result is Lemma
2, which gives a bound for the solution of error (34), where is an
output disturbance, while d(t) is an input disturbance.
Remark 2: Note that the and  terms in (12) and (13) are
absolutely continuous since the operator in (14) is strictly proper,
and those in (15) and (16) are proper.
B. Bounds for the Auxiliary Errors
The error system to be considered here is composed of (3), (7),
(12), and (13). Let x0F L denote the transient state corresponding to:
1
L01 in (7), Wd in (5), F10;i1 Li0+1
;N in (12), and all the remaining
operators associated with uN ; ei ; 0i in (14)(16). Since all these
operators are stable, there exist positive KF L and aF L such that

x0F L (t)

 KF L e0a

x0F L (0) :

In order to fully account for the initial conditions, the following


state vector z is used:
T

z = (z ) ; "N ; e
T
(z )T = xTe ; "1 ; "2 ; 1 1 1 ; "N 01 ; x0F L
:
0

3 (U )

and fN

 F 0;N 3 Ud
1

then the auxiliary errors "0 (N = 0) or "i (i 0 0; 1 1 1 ; N


tend to zero, at least exponentially. Moreover

The following stability theorem will be demonstrated for the full


error system given by (3), (7), (12), and (13). For N = 0(n3 = 1)
exponential stability follows immediatly from (3), (4), and Lemma 1.
Theorem 2 (N  1): Assume that (6) is satisfied and that the
modulation functions satisfy (18). Then, for sufficiently small  >
0( = maxi i ), the error system (3), (7), (12), and (13) with state
z as defined in (17) is globally exponentially stable with respect to
a residual set of order  , i.e., there exist positive constants K and a
such that 8z (0); 8t  0; kz (t)k  Ke0at kz (0)k + O( ).
Proof: See Appendix D.

D. Unmodeled Dynamics and Robustness


Unmodeled dynamics can also be assessed. Consider, as an exs
ample, y = s1(0
s01) u, where  > 0 is the unmodeled dynamics
parameter. The nominal plant ( = 0) has n3 = 2 and is minimum
phase, but the actual plant is not. The unmodeled dynamics can
be formally embedded in an input disturbance d = 0su. For
simplicity, we assume k3 known and u = 0U1 (U nom = 0;
3
2nom
n = 0). Then, Ud = u + Wd (su) and (F = F1 ; L = L1 )

"0 = k3 ML
"1 = L01

0U 0 L0 u3 + sL0 Wd U
0

0 U1 0 F 01 u3

(24)

O()

sWd
+
U + 01 :
FL 1

(25)

(17)

In what follows, all K s and as denote positive constants, operator


norms (k 1 k) are L1 induced norms [5], and EXP and EXP0
denote any term of the form K kz (0)ke0at and K kz 0 (0)ke0at ,
respectively, where K and a are (generic) positive constants.
Theorem 1: Consider the auxiliary errors (11)(13). If condition
(6) is satisfied and the relay modulation functions satisfy (i =
0; 1 1 1 ; N 0 1)
1
fi  F10;i1 Li0+1
;N

C. Error System Stability

O()

8t  0 :

389

(18)

0 1( 0))

j"i (t)j; kxe (t)k  EXP


j"N (t)j  2 KeN C (t) + EXP
0

(19)
(20)

Applying Lemma 1 to (24), with f0 slightly modified to f0 =


f0 +(k+G (s))f1 (the -term here dominates the -term in (24) with
 , according to Lemma 3), and Lemma 2 to (25),
appropriate filter G
one easily concludes that "0  EXP0 and "N  2KeN C (t) +
EXP as in Theorem 1. Noting that (3) is now given by e_ =
Ac e + k3 bc (0U1 0 u3 + Wd sU1 ), we rewrite (25) in a form similar
to (41) (Appendix D), "1 = L01 [0U1 0 u3 + Wd sU1 ]+ u1 + 01 ,
with u1 := (F 0 1)F 01 L01 [u3 0 Wd sU1 ]. Then, as before, we
define e^1 := "1 0 ( u1 + 01 ) and e := e 0 k3 bc e^1 . Now, referring to
Appendix D, (42) is then obtained (knom = k3 ) and all the remaining
inequalities (43)(51) hold with  replaced by  + . Thus one can
conclude that global exponential stability with respect to a residual set
of order  +  holds for sufficiently small  + . A slightly different
result holds if f0 remains unchanged. Details for more general cases
will be presented in a separate paper.

and

jei (t)j; j i (t)j  EXP ; i = 1; 1 1 1 ; N


j uN (t)j   K N C (t) + EXP
0

(21)

IV. THE RELAY MODULATION FUNCTIONS

(22)

Explicit formulas for the modulation functions are now derived.


01
nom
) 0 2nom
Noting that 23n 0 2nom
n =  ((1=k
n ), the expressions
(18) can be rewritten as ( =  0 1; i = 0; 1 1 1 ; N 0 1)

where

C1 (t) = sup k!(t)k; C (t) = M C1 (t) + Mred


t

(23)

with some positive constants M ; Mred and  := maxi i .


Proof: See Appendix C.
Remark 3: In [9] a similar analysis was carried out with "i =
F 01 
1

U + ei + 0i , instead of (12),
Li01 [0Ui 0 Li0+1
;N (U )] + F L
0
1
 j, instead of (18). The results, however,
and fi  jLi+1;N 3 U
required excessively small time constants of the averaging filters
F 01 (s), e.g., i =  2
and sufficiently small  ; whereas, here
this restriction is eliminated and we can even take i   .

1
~ nom 0 (u 0 U nom ) 0 Wd 3 de
fi  F10;i1 Li0+1
;N 3 U
0
1
nom
01
fN  F1;N 3 U~
0 23n 0 2nom
n r 0  Wd 3 de

(26)
(27)

~ nom := (U 3 0 U nom ). Defining i and i as in Table I [see


where U
also (8) and the notation below (13)] and
1
01
^
i = Li0+1
;N Wd de = Li+1;N (1 0 G1 )de
N = 01 Wd de = 01 (1 0 G^ 1 )de

(28)
(29)

390

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

one easily concludes that (26) and (27) are satisfied, modulo expo;111;N 0 )
nentially decaying terms if2 (i

=0

= F 0;i !T ji j +  i 0 ! T i + ji j


fN = F 0;N !T j! j +  n jrj + jN j
fi

1
1

nom

(30)
(31)



where <     ;   jj.


Note that Theorems 1 and 2 hold for these modulation functions,
which do not use convolution operators. Indeed, the additional
0
terms which
transient terms can be easily included in the
appear in the proofs of both Theorems.
sI
.
The poles of G1 s are the (known) roots of
Then, for  sufficiently small and  2 ;  , bounds for i and N
can be established via Lemma 3 and Corollary 1 (see Appendix B),
using signals filtered from de t . Similarly, ji j and ji 0 nomT i j
can be bounded using signals filtered from j uj
!T j! j and
nom
nom
u
u0U
. Note that one should not use ju 0 U nom j
in order to avoid too conservative results since u is a discontinuous
control law and should be smoothed with a filter prior to taking its
absolute value). Then, the following modulation functions can be
; 1 1 1 ; N 0 ):
derived (i

^()

( )

EXP
det( + 3) = 0

(0 ]

1 := 

:=

=0
1
fi = F^i0 Li0 ;N j1
uj + 
 Li0 ;N 1u + ci d^ei
fN = F^N0 j1
uj +  n jrj + cN d^eN (t) + 0 de
1

1
+1

1
+1

^ =1 ^

nom

(32)

(33)

k
where F0
; Fi01
for i
; 1 1 1 ; N ; the posis+
tive constants kF i ; F i can be precalculated from F10;i1 ci=0;111;N
can be calculated from (28) and (29) and the uncertainty bounds;

dei
pi;k j, with pi;k being the poles of
kj
s+ de i
0
1
Li+1;N 0 G1 . Note that either (31) or (33) implies that (6) is
satisfied.
1
1
r. ChoosExample: Consider y
s(s01) u and ym
(s+1)(s+2)
0
4
gf
we have G1 s
.
Choosing
L
s
, it is
ing
s+1
and 1
: are possible choices since
easy to verify that c1
c1 1 e0 t  L01 fL01 0 G1 g or L01 fG1 g (L01 denotes the
inverse Laplace transform). Of course, in general, G1 is not known,
then some uncertainty bound for u [see below (8)] has to be used
to determine ci and i .

= 1

 ; = min Re( )
(1 ^ )
=
=
^ ( )=
3= =1
=5
= 09
^
(1 ^ )

= +1

V. SIMULATION RESULTS
This section presents an example which highlights the performance
of the VS-MRAC. The actual system is

[0 0 2]
=0
()

0s

= se 0 1 [u + d]

where  2 ; : and d is a bounded input disturbance. A similar


example was utilized in [4]. The nominal plant was assumed without
1
). The reference model is chosen to be ym
delay (
s+1 r .
sqw : t ; r sqw : t ,
For the simulations presented here d
where sqw 1 denotes a unit square wave. In order to avoid chattering
in the VS-MRAC [8], the signal U0 was filtered by s1+1 inserted
between points 1 and 2 of Fig. 1, before being fed to the plant and
here. From (31)
the operator L01 (see Fig. 1). Note that L
jyj and
or (33), the modulation function was chosen as f0

: . As shown in Fig. 3, the MRAC [13], [14] fails completely
under the action of the external disturbance, while the VS-MRAC
and the modified MRAC of [4] maintain remarkable model following
performance. Fig. 4(a) illustrates the progressive degradation of the
VS-MRAC performance in the presence of unmodeled dynamics. For

=3

(0 3 ) =
=1

= 0 05

2 See

Section II-A for notations.

(b)

^ =

(a)

(0 1 )

= 3 +4

Fig. 3.  = 0; (a) VS-MRAC: (dashed line) ym , (solid line) y for d = 0,


(dotted line) y for d = 3sqw (0:3t); and (b) MRAC: (dashed line) ym , (solid
line) MRAC [14] y for = 10; d = 3sqw (0:3t), (dotted line) MRAC [4]
y for = 1; d = 3sqw(0:3t).

0 06

  : ideal sliding is preserved and no oscillation appears at


the output. An oscillation grows continuously when  is increased
beyond  = 0:06 (in Fig. 4(a),  = 0:073). On the other hand,
Fig. 4(b) shows how the MRAC, based on  -modification, projection,
and the swapping lemma proposed in [4], is explosively unstable for
 = 0:065, similarly to what occurs with high-gain linear robust
control [8]. To be fair,  = 0:05 was also used as the small time
constant that appears in [4], and the corresponding controller was
carefully tuned to achieve its best performance.
The VS-MRAC was shown to outperform standard MRAC algorithms in several previous works (see Section I). The latter example
shows that it also outperforms more recent versions of parameter
adaptive schemes [4], [14].
VI. CONCLUSION
In this paper, earlier results about design and stability analysis
of the VS-MRAC, taking into account the effect of the averaging
filters, were generalized to the case of arbitrary relative degree.
Global exponential stability of some small residual set in the error
space was obtained. The present analysis reveals that the averaging
filters can be chosen of order one, contrary to what was believed
originally. Moreover, the new design can also be used for more

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

391

Lemma 1: Consider the I/O relationship

"0 (t) = M (s)[u + d(t) + (t)]


 (s) is SPR and d(t) and  (t) are LI. Let xe be the state of a
where M
 . If u = 0f (t)sgn("0 ),
stable, possibly nonminimal, realization of M
where f is LI and f (t)  jd(t)j 8t, then the inequality

j" (t)j and kxe (t)k  [c kxe (0)k + c R]e0 t

(a)

holds 8t  0 and for some positive constants c1 ; c2 ; 1 . Moreover,


if f (t) satisfies f (t)  jd(t)j +  (8t), with arbitrarily small positive
constant , then "0 (t) becomes identically zero after some finite time
tr  0 (for a proof see [10] and [11]).
Lemma 2: Consider the I/O relationship

"i (t) = Li01 (s)[u + d(t)] + (t) + (t)

(34)

where Li (s) = s + i ( i > 0); d(t) is LI, and (t) and  (t) are
absolutely continuous (8t). If u = 0f (t)sgn("i ), where f is LI and
f (t)  jd(t)j 8t, then the signals "i (t) and e^i (t) := "i (t) 0 (t)
are bounded by

(b)

Fig. 4. d = 0; (a) VS-MRAC: (dashed line) ym ,


 = 0:073; and (b) MRAC [4]: y for  = 0:065.

(solid line)

for

effective external disturbance rejection. The robustness to unmodeled


dynamics is briefly discussed, showing that the methods presented
here can also be used to assess this important issue.

je^i (t)j and j"i (t)j  je^i (0)je0 t + 2

Re0 min( ;)t + sup j j :


t
(35)

Proof: Let e^i := "i 0 ( +  ). Then, e^_i = 0 e^i + (u("i ) + d).


 by: m
_ = 0 i m; m
 = m + , where , absolutely
Define m and m
continuous, is to be chosen later. One can write the comparison
d . Then one has d [m
 0 ) + dt
_ = 0 i (m
equation m
dt  0 e^i ] 
0 i [m 0 e^i ] + [f sgn("i ) 0 d] + dtd + i . Now, we require
d
)j +  (t) (8t). A satisfying the later
dt  0 i ; (t)  j 0(tmin(
;)t + sup j (t)j.
inequalities is (t) = Re
t
It is easy to verify that m(0) > 0 implies

m(t) > 0; m
 (t) > sup j (t)j;
t

8t:

(36)

APPENDEXES

Then, if m
 (t0 ) = e^i (t0 ) for any t0  0, one concludes that
"i (t0 ) > 0 since, from (36), m
 (t0 ) > supt0 j (t0 )j; 8t0  0.
d [m
 0 e^i ]  0 (t = t0 ). Similarly,
Therefore, under these conditions dt
d [m
 (t0 ) = e^i (t0 ) one obtains dt
 0 e^i ]  0 (t = t0 ).
if 0m

U^

je^i (t )j  m (t ) implies je^i (t)j  m (t); 8t  t . This in turn

Thus using the Comparison Lemma in [6, Th. 7], we conclude that

A. Relationship Between U 3 and

^ + G
^ 1 u) are
The key formulas needed to derive U 3 = (U
those given in [16, pp. 105 and 106] for the transfer functions
G31 (s); G32 (s); G^ 1 (s); and G^ 2 (s), giving U 3 = G31 (s)u + G32 (s)y
N (s)
^ = G
^ 1 (s)u +
(which matches W (s) = Kp D (s) to M (s)) and U
(
s
)
N
^ (s) = Kpnom
G^ 2 (s)y (which matches W
to M (s)).

D (s)

B. Auxiliary Lemmas

In what follows, we always consider t 2 R+ so that 8t means


t  0 (except otherwise stated). We denote locally integrable, in
the sense of Lebesgue, by LI (ueq (t) is LI, c.f. Section II-D) and
omit the term almost everywhere since its need is believed obvious
where necessary [5]. In the Lemmas to follow, all input and output
signals are scalars; the signal  (t) is exponentially decreasing, i.e.,
j(t)j  Re0t for some positive scalars ; R; and 8t; all transfer
functions are rational, stable, and strictly proper.

results in (35), where t0 = 0 without loss of generality.


Lemma 3: Consider the stable strictly proper I/O relationship
z = W (s)d. Let 0 be a positive constant satisfying 0 < 0 <
minj jRe(pj )j (pj are the poles of W (s)) and d(t) be an instantaneous upper bound of d(t), i.e., jd(t)j  d(t) 8t. Then, there exists
a positive constant c1 such that the impulse response w(t) satisfies
jw(t)j  c1 0 e0 t and the following inequalities hold:

jW 3 d(t)j  c s + 3 d(t)
jz(t) 0 z (t)j  c jd^(t) 0 d^(t) j; d^ = s + d
jz(t)j  c d^(t) + exp
1

1
1

(37)
(38)
(39)

where z 0 ; d^0 ; and exp depend on the initial conditions and decay
exponentially to zero with rate 0 (for a proof see [13]).

392

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

Corollary 1 (Lemma 3): Consider z = GF (s)GL (s)d =


GF (s) s+1 G L (s)d, where GF ; GL are rational, stable, strictly
L has positive impulse response, > 0. If  2 [0; ]
proper, and G
and  is sufficiently small, there exists k > 0 such that (38) and
(39) hold with

d^(t) = kGL d(t):


> 0, such that jz (t) 0
z0 (t)j  ( sk+ ) 3jGL 3jdjj  (( sk+ )( s+1 )G L ) 3jdj. The latter
inequality comes from the positive impulse response assumption.
However, L01 [( sk+ )( s+1 )] = ( k0 )(e0 t 0 e0 ) 
ke0 t , for sufficiently small  and some k > 0. Thus, jz 0 z 0 j 
kG L 3 e0 t 3 jdj = kGL 3 jdj.
Proof: By Lemma 3, there exist kF ; F

C. Proof of Theorem 1
Proof: By Lemma 1, "0 (11), as well as the state xe , converge
to zero, at least exponentially, if the signal f0 satisfies (18). Note
that the transient part of L01 is represented by  (t), which is thus
bounded by EXP0 . Consequently, one concludes j"0 (t)j  EXP0 .
From (16) one can write 0i = Hi (s)"0 = hi (t) 3 "0 (t) + 00i (t).
Since j00i j  EXP0 and hi (t) 3 "0 (t)  EXP0 , then j0i j  EXP0 .
Now, for i = 1, since e1  0, (12) results in j"1 (t)j  EXP0 from
Lemma 2. With similar argument, we recursively conclude from (12),
(15), and Lemma 2 that jei (t)j  EXP0 and j"i (t)j  EXP0 (i =
2; 1 1 1 ; N 0 1).
Now consider "N [see (13)]. Note that, from (6) and (23), M and
Mred can be chosen so that supt jUN (t)j  C (t). Since e;N 01 and
"N 01 are bounded by EXP0 , so is eN , i.e., jeN (t)j  EXP0 .
From (14) one has

0 (t)  F1;N 0 1 C (t) = K C (t)


sup uN (t) 0 uN
N
F1;N LN
t

(40)

O( )

0 (t) is bounded by EXP0 . Then,


where  := maxi i and uN
applying Lemma 2 to (13), which can be rewritten as "N =
0 ] + [ 0 + eN + 0N ], one
LN01 [0UN 0 U ] + [ uN 0 uN
uN
0 j  j uN 0 0 j 
readily concludes (20). Since supt j uN 0 uN
uN
j uN j 0 EXP0 , then (22) follows from (40).
D. Proof of Theorem 2
Proof: It is convenient to rewrite (13) as

"N = LN01[0UN 0 U ] + uN + eN + 0N


(41)
1 L01 [U ] can be bounded by C (t)
where uN = (F1;N 0 1)F10;N
N
0 (t)j   K
 N C (t).
similarly as uN in (40), i.e., supt j uN (t) 0 uN
0
1
Note that, in the form (41), LN operates on the same input as
the one in (3). From (3) and (13), the model following error can
be rewritten as: e_ = Ac e + knom bc [e^_ N + N e^N ], where e^N :=
"N 0 ( uN + eN + 0N ). Then, in order to eliminate the derivative
term e^_ N , a variable transformation e := e 0 knom bc e^N is performed
yielding

e_ = Ac e + knom (Ac + I N )bc e^N :

(42)

The bound (20) (Theorem 1) and the exponential stability of Ac imply


 (t)+EXP with C (t) defined
that e(t) is bounded by ke(t)k   KC

in (23). Moreover, as explained below

ke(t)k  Ke C (t) + EXP; and je0 (t)j  K0 C (t) + EXP
C1 (t)  K1 C (t) + K2 kz (0)k + Km
0
C (t)  Kred + K4 kz (0)k :
1 0 K3

(43)
(44)
(45)

Indeed, (43) follows from e = e 0 knom bc e^N . From the relation


T = [v T ym v T ] is the regressor
where !m
m1
m2
vector corresponding to the reference model and
is a constant
matrix, it follows that k! k  Km + K
kek and, since C1 (t) =
supt k!(t)k, from (43) we obtain (44), where K2 kz(0)k comes
from the initial value of the term EXP appearing in the bound (43)
of kek. Now, from (23) and (44), C (t) can also be bounded by
0 , whereby, after a simple
C (t)  K3 C (t) + K4 kz (0)k + Kred
algebraic manipulation one obtains (45), which is valid for  < K301 .
Now, as explained below, we can also write

! = !m +
e,

kz0 (t)k  Kz e0a t kz0 (0)k


kze (t)k  K5 kze (0)k + kz0 (0)k + O( ) + EXP:

(46)
(47)

Indeed, the variables xe ; "i=0;111;N 01 ; x0FL were shown to be


bounded by EXP0 in Theorem 1. Therefore, with the partition
z T = [(z 0 )T ; zeT ], where zeT := ["N ; eT ] one gets (46), where
only the initial condition on z 0 appears. Now, from (20), (43),
and (45) follows (47), where O( ) is independent of the initial
conditions. Noting that the initial time is irrelevant in deriving the
above expressions, we can write, for arbitrary t  t0  0, some
T1 > 0 and  < 1

kze (t)k  K5 + K6 e0a(t0t ) kze (t0 )k + kz0 (t0 )k + O( )


kz0 (t)k  Kz e0a (t0t ) kz0 (t0 )k
kze (t0 + T1 )k   kze (t0 )k + kz0 (t0 )k + O( )
kz0 (t0 + T1 )k  kz0 (t0 )k:

(48)
(49)

(50)
(51)

Equations (50) and (51) are obtained from (48) and (49) as follows:
for  < K501 , there exists T1 > 0 such that  = max[ K5 +
K6 e0aT ; Kz e0a T ] < 1. Then, the simple linear recursive inequalities (50) and (51) hold and easily lead to the conclusion that, for 
small enough, the error system is globally exponentially stable with
respect to a residual set of order  .

REFERENCES
[1] C. J. Chien, A. C. Wu, L. C. Fu, and K. C. Sun, A robust model
reference using VS design: The general case for multivariable plants,
in Proc. Amer. Contr. Conf., Chicago, IL, 1992, pp. 27302734.
[2] R. R. Costa and L. Hsu, Robustness of VS-MRAC with respect to
unmodeled dynamics and external disturbance, Int. J. Adaptive Contr.
Signal Process., vol. 6, no. 1, pp. 1933, 1992.
[3] J. P. V. S. Cunha, R. R. Costa, and L. Hsu, Design of a new high
performance VS position control of ROVs, IEEE J. Oceanic Eng.,
vol. 20, no. 1, pp. 4255, 1995.
[4] A. Datta and P. Ioannou, Performance improvement versus robust
stability in MRAC, in Proc. IEEE Conf. Dec. Contr., Brighton, 1991,
pp. 748753.
[5] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output
Properties. New York: Academic, 1975.
[6] A. F. Filippov, Differential equations with discontinuous right-hand
side, Amer. Math. Soc. Translations, vol. 42, no. 2, pp. 199231, 1964.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 3, MARCH 1997

393

Index TermsLyapunov stability, robust control, uncertain systems.

and stabilizing robust controllers have been developed, for example,


the results in [1], [3], [10], [16][19], and [21]. The uncertainties
admissible in these results can be of either matched type or unmatched
type, and they are generally bounded by known nonlinear functions
of the state. However, input-related uncertainties analyzed in the
existing results are assumed to be of fixed and known signs; that is,
the direction of motion of the system is fixed and known under any
choice of control. The assumption of priori knowledge of the control
direction being known makes robust control design straightforward.
In [5], the problem of designing robust control was studied for
the class of systems with an unknown but constant and boundedaway-from-zero control direction. The objective of this paper is
to generalize the results in [5] so that systems with an unknown
and time-varying control direction can be treated. This extension is
nontrivial since the time-varying control direction is allowed to cross
zero an infinite number of times and since identification must now be
done fast and continuously to keep tracking the changing control
direction. As before, the Lyapunovs direct method is utilized to
proceed with the design and to ensure the stability of uniform ultimate
boundedness. Similar to the results introduced in [5], continuity of
the proposed robust control is achieved by designing the so-called
shifting law, which continuously and smoothly steers the sign of our
robust control.
In the area of adaptive control, many results dealing with systems
containing an unknown constant control direction, referred to therein
as the unknown high-frequency gain, have been obtained. For example, Nussbaum [15] proposed for linear time-invariant systems an
adaptive control algorithm whose sign changes an infinite number
of times as its argument tends to infinity. In [12], Mudgett and his
coworkers developed an adaptive control design scheme for general
linear systems without requiring the knowledge of high-frequency
gain by following the same principle as that in [15]. In the effort
of extending these results to nonlinear systems having unknown
control directions, two adaptive control schemes have been proposed
for simple first-order nonlinear systems [2], [6]. In [2] and [6], a
hysteresis method was used to construct an adaptive control that
jumps a finite number of times over the infinite time horizon. Along
the line of nonlinear robust control, a continuous robust control design
scheme was developed in [5] for uncertain nonlinear systems with an
unknown but constant control direction. Compared with the existing
result, the proposed result is the first attempt in addressing the control
design problem in the presence of an unknown but smooth control
direction, whose sign change is of unlimited number and may occur
anytime.
The paper is broken into the following sections. System description,
necessary assumptions, and the problem of designing robust control
are described in Section II. Robust control design is proceeded in
Section III by the Lyapunovs direct method. A simulation example
is presented in Section IV.

I. INTRODUCTION

II. PROBLEM FORMULATION

For more than one decade, robust control design of nonlinear


uncertain systems has been an active area of research. Several important classes of stabilizable uncertain systems have been identified,

We shall consider the nonlinear uncertain system described by

[7] L. Hsu, Variable structure model reference adaptive control using only
I/O measurement: General case, IEEE Trans. Automat. Contr., vol. 35,
no. 11, pp. 12381243, 1990.
[8] L. Hsu, A. D. Araujo, and R. R. Costa, Analysis and design of I/O based
variable structure adaptive control, IEEE Trans. Automat. Contr., vol.
39, no. 1, pp. 421, 1994.
[9] L. Hsu, A. D. Araujo, and F. Lizarralde, New results on VS-MRAC:
Design and stability analysis, in Proc. Amer. Contr. Conf., San Francisco, 1993, pp. 10911095.
[10] L. Hsu and R. R. Costa, Variable structure model reference adaptive
control using only input and output measurement: Part I, Int. J. Contr.,
vol. 49, no. 2, pp. 399416, 1989.
[11] L. Hsu and F. Lizarralde, Redesign and stability analysis of VS-MRAC
systems, in Proc. Amer. Contr. Conf., Chicago, 1992, pp. 27252729.
, Experimental results on variable structure adaptive robot control
[12]
without velocity measurement, in Proc. Amer. Contr. Conf., Seattle,
1995, pp. 23172321.
[13] P. Ioannou and K. Tsakalis, A robust direct adaptive controller, IEEE
Trans. Automat. Contr., vol. 31, no. 11, pp. 10331043, 1986.
[14] S. M. Naik, P. R. Kumar, and B. E. Ydstie, Robust continuous-time
adaptive control by parameter projection, IEEE Trans. Automat. Contr.,
vol. 37, no. 2, pp. 182197, 1992.
[15] K. Narendra and A. Annaswamy, Stable Adaptive Systems. Englewood
Cliffs, NJ: Prentice-Hall, 1989.
[16] S. S. Sastry and M. Bodson, Adaptive Control: Stability, Convergence
and Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989.
[17] V. I. Utkin, Sliding Modes and Their Application in Variable Structure
Systems. MIR, 1978.
[18] A. C. Wu, L. C. Fu, and C. F Hsu, Robust MRAC for plant with
arbitrary relative degree using variable structure design, in Proc. Amer.
Contr. Conf., Chicago, 1992, pp. 27352739.

Robust Control Design for Nonlinear Uncertain Systems


with an Unknown Time-Varying Control Direction
Joseph Kaloust and Z. Qu
AbstractIn this paper, a continuous robust control design approach is
proposed for first-order nonlinear systems whose dynamics contain both
nonlinear uncertainty and an unknown time-varying control direction.
The so-called control direction is the multiplier of the control term in the
dynamic equation, and it effectively represents the direction of motion
under any given control. A nonlinear robust control is designed to on-line
and continuously identify sign changes of the unknown control direction
and to guarantee stability of uniform ultimate boundedness. The proposed
robust control design requires only three conditions: the nominal system is
stable, the control direction is smooth, and the uncertainty in the system
is bounded by a known function. The necessity of these conditions is
established in this paper. Continuity of the proposed robust control is
achieved by using a so-called shifting law that changes smoothly the sign
of robust control and tracks the change of the unknown control direction.
Analysis and design is shown by using the Lyapunovs direct method.

Manuscript received March 14, 1995; revised September 8, 1995, April 20,
1996, and July 1, 1996.
J. Kaloust is with Loral Vought Systems, Dallas, TX 75265-0003 USA.
Z. Qu is with the Department of Electrical and Computer Engineering,
University of Central Florida, Orlando, FL 32816 USA.
Publisher Item Identifier S 0018-9286(97)02046-1.

_ = ( )+1 ( )+ ( ) ( )

f x; t

x; t

a t u x; t

(1)

( )

where x 2 < is the state, u 2 < is the control input, f x; t


and f 0 x; t are known and unknown dynamics of the system,
respectively. Symbol a t characterizes the time-varying direction
of motion with respect to a given control. The system is chosen to be
scalar in order to highlight the development of robust control design

1 ( )

00189286/97$10.00 1997 IEEE

()

Anda mungkin juga menyukai