Copyright © IFAC Nonlinear Control Systems

Design, Tahoe City, California, USA, 1995

ADAPTIVE CONTROL
OF NONLINEAR SYSTEMS:
BASIC RESULTS AND APPLICATIONS *
Riccardo Marino

Dipartimento di Ingegneria Elettronica,

Universita di Roma 'Tor Vergata',
Via della Ricerca Scientifica, 00133 Roma, ITALY.

Abstract. Adaptive controls for nonlinear systems act cancellation of possibly nonlinear terms which
are defined as nonlinear dynamic (state or output may contain unknown or uncertain parameters. In
feedback) compensators capable of guaranteeing the case of linear systems once the pole placement
asymptotic tracking of an output reference signal theorem was established, the problem of adap-
for any unknown parameter vector and any initial tive pole placement was posed as a Model Ref-
condition for the closed loop system when time erence Adaptive Control (MRAC) : design a state
varying disturbances are not present. The purpose feedback control for controllable systems with un-
of this paper is to collect and illustrate the main known parameters so that the closed loop system
results available for single input single output non- asymptotically follows a linear system with preas-
linear systems in the design of adaptive state feed- signed eigenvalues. In the case of nonlinear sys-
back controls, adaptive observers, adaptive output tems the same problem may be posed. Since 1987
feedback controls and to briefly outline their de- the problem of designing controls for feedback lin-
velopments since 1987. The physical impact of earizable systems with uncertain parameters was
these results is illustrated by means of two clas- investigated in [16, 20 , 19]. In particular in [16, 19]
sical single input single output nonlinear control adaptive versions of feedback linearizing controls
problems containing uncertain parameters: tran- were designed by imposing Lipschitz conditions on
sient stabilization of a synchronous generator con- the nonlinearities while in [20] adaptive feedback
nected to an infinite bus subject to turbine failures linearizing controls were obtained under matching
and short circuits in transmission line; tracking conditions without imposing growth conditions.
control of a single link with unknown mass, iner- The interest of those results is easily illustrated
tia and length rotating on a vertical plane from by the simple example
angular measurements only. This paper is largely
based on examples, applications and results re-
ported in [15] .
Keywords. Nonlinear adaptive control, adaptive
observers, uncertain nonlinear systems. (1)
in which () is an unknown constant parameter. Let
INTRODUCTION B be the estimate of () in the linearizing control
(kl > 0, k2 > 0)
At the beginning of the eigthies two important
classes of non linear systems were defined for which (2)
constructive procedures to design nonlinear state
feedback control were determined: feedback lin- so that the corresponding closed loop system is
earizable systems and input-output feedback lin-
earizable systems. In the first case the state space X2
3
equations can be made linear (in certain state co-
A

(() - (})X2 - k1Xl - k 2 X 2

ordinates) by state feedback while in the second
case the input-output dynamics are made linear If () - B < 0 the origin is globally asymptotically
by a state feedback control which may possibly
render some dynamics unobservable from the out- stable. In the case in which () - B> 0 a limit cy-
put . Feedback linearization is a generalization cle exists, which is the boundary of the domain of
of pole placement for linear systems while input- attraction of the origin, and global asymptotic sta-
output linearization generalizes the zero-pole can- bility is not achieved. These drawbacks are over-
cellation technique. Both algorithms rely on ex- come by nonlinear adaptive techniques. In fact ,
the adaptive feedback linearizing control
·This work was supported by Ministero dell ' Universita.
e della Ricerca Scientifica e Tecnologica

869
with for a class of nonlinear systems under persistency
of excitation conditions following essentially the
(4) techniques developed in [8] for linear systems. In
view of the use of adaptive observers for out-
and P the symmetric positive definite solution of put feedback design, persistency of excitation is a
the Lyapunov matrix equation strong assumption. For this reason a more restric-
tive class of nonlinear systems for which adaptive
1
-k2 ]
T p+p [ 0
-kl -t ]=-1
observers may be designed without persistency of
excitation condition was obtained in r9]. The de-
sign procedure is illustrated by the fohowing sim-
ple example
assures that limt ..... oo IIx(t)11 = 0 for any initial con-
dition x(O), 8(0) and for any unknown parameter X2 + Bx~ ]
B, positive or negative. This conclusion may be [ Bx~ +u
drawn using the Lyapunov function
= Ac [ ~~ ] + bBxi + gu
6
whose time derivative is y Xl = Cex . (5)
V = _xT x. The observer is

If a linear model reference is defined the same pro-

cedure leads to the design of a MRAC which guar-
antees that the closed loop system follows a linear
[!~ ] = Ac [ ~~ ] + b8y2 + gu + [ ~~ ] (y - id
(6)
one even though the parameter estimation error
may not converge to zero. with 8 being the estimate of the unknown constant
Matching conditions, which are satisfied in the parameter B. We look for an update law
above example, were relaxed by extended match-
ing conditions in [2] and by pure feedback con- 8= w(y,8,xl,i 2,u)
ditions in [4] so that a larger class of feedback
linearizable uncertain systems with non Lipschitz such that the observer error x= X - i, whose
nonlinearities can be adaptively controlled. The dynamics are (8 = B - 8)
disadvantage in [4] of requiring several estimates
for a single unknown parameter was removed in f7]
where only one parameter estimate is needed for = [-klXl + X2 :+- !y2 ]
every unknown parameter: this allows a consider- -k2Xl + By
able reduction of the order of the resulting adap-
tive control. An important feature of these adap- (Ae - kc c) [ ~~ ] + b8y2
tive controls is to guarantee asymptotic following
of a linear reference model without persistency of tends to zero as t -+ 00 for any unknown value of
excitation conditions . At the present state of the B. We choose kl and k2 as follows P, > 0)
art however adaptive feedback linearizing controls
are not available for systems such as
[ ~~ ] = Aeb + >'b
Xl X2 + BX1X3
x2 = x3 so that, denoting A = Ae - kc e, we obtain
x3 U
-1 1
which does not satisfy the pure feedback condi- cc (sl - A) b = --, .
tions even though it is feedback linearizable for
s+"
any value of the unknown parameter B. This implies that there exist positive definite ma-
As it happened for linear systems, once adap- trices P and Q such that , according to Meyer-
tive state feedback problems are solved it is nat- Kalman-Yacubovich Lemma [17, 18]
ural to begin investigating adaptive output fed-
dback ones possibly on the basis of adaptive state
observers . Unlike the linear case, for non linear
systems output feedback control algorithms are
much less developed than state feedback ones . The update law
Moreover, as shown in [6], observers may be built
for a restrictive class of nonlinear systems with
no uncertainties: the design procedure involves (7)
cancellation of nonlinear terms and state space
change of coordinates. When uncertain param- gives together with (6), an adaptive observer. In
eters are present the resulting observer no longer fact, since B is constant,
rovides asymptotic estimates of the states. In
r1] it was shown how to design adaptive observers -
B = -y (y -
2
id

870
and the error dynamics are Let

x Xlr XZr - k 3 X l r - k4 xz r + U r
XZr = .-k3 X lr - k 4 xzr + U r
Yr = Xlr

Differentiating the function be the reference model which generates the ref-
erence output Yr' The gains kc = (k3, k 4 ) are
chosen so that the characteristic polynomial of
(Ac - bk c) is (s + l)(s + >'c), i.e. there is a zero-
with respect to time and recalling that Pb = c~ pole cancellation in the transfer function cc[sI -
we have (Ac - bkc)]-lb = l/(s + >'c) (>'c is a positive
real). In both zero-pole cancellations we make
use of the crucial property enjoyed by the triples
(Ac - kocc , b, cc) and (Ac -bk c, b, cc) of being mini-
mum phase and of relative degree one. The output
feedback control is designed as
If the output yet) is bounded , it follows that
(9)
lim
t-co
IIx(t)11 = 0
in which the state estimates are provided by the
observer. Both the control and the observer try
and (6), (7) define an adaptive observer. Also to cancel the nonlinearity 8y2 by using the same
in this case the parameter estimation error is not
guaranteed to converge to zero estimate 0 of the unknown parameter 8 which is
On the basis of [9], for the same class of sys- now adjusted according to the update law (ry suf-
tems considered in [1] it was shown in [13, 12] how ficiently large positive real)
to design adaptive observers for equivalent states
via filtered transformations without requiring per-
sistency of excitation . The equivalent states are
related to the original ones through unknown pa- which reacts both to the tracking error y - Yr and
rameters and signals generated by known auxil- to the observer error y - Xl. The stability of the
iary filters (filtered transformations) . This result closed loop system can be assessed by expressing
leads immediately to the construction of adaptive the dynamics of the tracking error e = x - X r ,
output feedback controls for a class of nonlinear observer error x = x - X and estimation error
systems with relative degree one as shown in [12]. 9=8-0
For relative degree one systems the design proce-
dure is not a complex one as shown by the follow- (Ac - bkc)e + by 29+ bkcx
ing example -
x (Ac - kocc)x + by 2-8
[~ ~] x + [ ~ ] (u + 8yz) -y2cc(e + ryx)
and using the function
Acx + b(u + 8y2)
t::.
y [ 1 o ]X = Cc X .

An adaptive observer is with Pc and Po positive definite solutions of

Xz + u + OyZ + kl (y - xd
(Ac - bkcf Pc + Pc(Ac - bkc)
• Z Pc b
u+8y +kz(y-xd
(Ac - kocc)T Po + Po(Ac - koc c)
= yZ(y _ xd (8)

with (>'0 positive real) with Qc and Qo suitable positive definite matrices,
according to Meyer-Kalman-Yacubovich Lemma
[17, 18]. The time derivative of V is in fact

dV _ T T - -T -
dt - -e Qc e + 2e Pcbkc x - ryx Qox .
so that the characteristic polynomial of(Ac-kocc)
is (s + 1)(s + >'0), i.e. there is a zero-pole cancel-
latIOn in the transfer function This allows us to conclude that both the traking
error and the observer error converge to zero while
1 the convergence to zero of the estimation error is
Cc [ sI - (Ac - koc c)]
-1
b =- -.
s + >'0
not guaranteed.

871
 As it happened for linear systems once the rel- Definition 2.2 Assume that p ~ n for system
ative degree one case had been solved, the solu- {11}. Let
tion of the relative degree two case readily fol-
lowed while the solution of the general case was M t = {x E Rn : hex) = Yr(t),
considerably harder and led to complex control
algorithms: the complete solution was given in . . . , Lj;lh(x) = y}P-1)(t)}
[14, 12] for a class of nonlinear systems with any
relative degree which contains,as a special case, be the time-varying (n - p )-dimensional integral
minimum phase linear systems with known rel- manifold called tracking manifold. The tracking
ative degree, known sign of the high frequency dynamics are defined as the dynamics constrained
gain, known upper bound of state space dimen- in Mt . 0
sions. In conclusion the basic adaptive result for
linear systems (see [17, 18]) has been generalized Definition 2.3 The disturbance characteristic
to a class of nonlinear systems with non Lipschitz index 11 is defined as the integer such that:
nonlinearities depending, in suitable coordinates , LqL~h(x) = 0, 0 ~ i ~ 11- 2, Tix E Rn , VB E 0 ,
on the output only. A different solution was pre- LqLj- 1h(x) #0, for some 0 E 0 , for some
sented in [5, 3] for a more restrictive class on non-
linear systems with any relative degree. Unlike x E Rn . We set 11 = 00 if LqL~h(x) = 0, i ~ 0,
the case of linear systems for which the solution Tix E Rn, VB E 0. 0
of the adaptive output feedback control problem
was obtained once non adaptive output feedback Definition 2.4 The adaptive tracking problem is
control problems were fully solved by Separation said to be globally solvable for system {11} if for
Theorem, for nonlinear systems the adaptive out- any smooth bounded reference trajectory Yr(t) with
put feedback control designed in [14] solved new bounded time derivatives yP) , . .. , y~p) a dynamic
problems (see [11]) in the non adaptive setting as state feedback control
well, that is for nonlinear systems with known pa-
rameters .
tU = J.l(X , w, Yr" "' Y~p» ) , wE R'
u u ( x , W, Yr , ... , y~p»)
ADAPTIVE STATE FEEDBACK
exists such that:
We consider nonlinear systems
(i) for any initial condition x(O) , w(O), IIx(t)lI,
p Ilw(t)1I are bounded Tit ~ 0,
x f(x) + g(x)u + L Oiqi(X) (ii) limt_oo(y(t) - Yr(t)) = O. 0
i=l
Theorem 2.1 Assume that for system {11}:
fe(x) + g(x)u
y hex) (11) (i) the global relative degree p is well defined and
v ~ p ~ n,
which are affected by an unknown constant pa-
rameter vector (ii) the tracking dynamics are bounded input
bounded state,
(iii) d(LqjL~h) E span{dh , . . . , d(L~h)} , 11-1 ~
i ~ p - 2, 1 ~ j ~ p , Tix E Rn ,
whose nominal value is zero so that the nominal
system is (iv) the vector fields
i: = f(x) + g(x)u
y hex) (12) f
where x E Rn is the state vector , u E R is the
control , h : c oo (Rn , R) is the output function , f ,
g , qi, 1 ~ i ~ p , are smooth vector fields in Rn ,
with g(x) # 0, for every x ERn.
are complete.
Definition 2.1 The global relative degree p is de- Then , the adaptive tracking problem is globally
fined as the integer such that for every 0 E n solvable for system {11}. 0

LgL~8h(x) 0, Tix ERn , 0 ~ i ~ p- 2 Consider the worst case in which v = 1. By virtue

of conditions (i) and (iv) there exists a global
LgLj;lh(x) # 0, Tix ERn . change of coordinates

If LgL~8 h(x) = 0, Tix E Rn , Tii ~ 0 we say that Zi L~-lh(x), 1 ~ i ~ p

p = 00 . 0 Z] · 4Jj(x) , p+l~j~n (13)

872
with tP;(x) , p+ 1:5 i:5 n, such that (dtP;, g) = 0, (v) th e vector fields
which transforms (11) into
Uh
J
P j f - gL L P- 1 h
Zj = Zj+1 + LB;Lq.L~-1h(x) 9 J
;=1 1
Co 9 =
Zj+1+ WJ(z)B, 1:5 i:5 p-l
Zp 1
Ljh(x) + L gLj- h(x)u are complete.
P
Then , the adaptive tracking problem zs globally
+ LB;Lq.Lr 1h (x) solvable for system (11). 0
;=1

Co Ljh(x) + LgLr1h(x)u + W;(z)B

P
LJtPp+j(z) +L B;Lq.tPp+j(z) ADAPTIVE OBSERVERS
;=1
Co In this section we shall consider single output non-
tPj( Z, (J) , l:5i:5n-p linear systems with unknown constant parameters
Y Z1 (14) (Ji

in which LgLj- 1h(x) :I 0, Vx E Rn , by definition

P

of global relative degree p . Assumption (iii) guar-

x !(x)+g(x,u)+ L(Jiq;(X,U)
;=1
antees when v = 1 that the matrices W; in (14)
only depend on Z l, . . . , Z; for 1 :5 i :5 p - 1. We Y hex) (17)
introduce a new control variable v , defined as
in which x E Rn, U E Rm, (J = [(Jl, ... , (Jp]T E RP ,
yE R , f(O) 0, h(O) = 0, g(x ,O) =
0, Vx ERn . =
We will restrict ourselves to those systems (17)
which are transformable by (J-independent diffeo-
which substituted in (14) gives morphisms into linear observable systems with
output injection terms depending on B. They are
Zj+1 + WJ(Z1 ," " zj)B , l:5i:5p-l identified by the following theorem , which is based
on the results obtained in [6] .
v+ W;( Z1, ... , Zn)(J
~(Zt ,Z l , . . . , zp, B) Theorem 3.3 There exists a global diffeomor-
phism
Y Zl (16)
Z = T(x) , T(O) = 0,
where Zt [Zp+1, "" zn]T
and ~ = [tPl , ' . . , tPn-p]T. Assumption (ii) allows transforming system (17) into
to prove boundedness of the variables Zt in the
closed loop system whose adaptive control is de-
signed according to the algorithm given in [4] and
modified in [7] . Actually in [4] the class of systems
01
.
00
.
.... ..
.
00
.
00
.
1z + r 11 (y ,, u)
12(Y
.
u) 1
identified by the following theorem are considered . . . . . .
ro
# • • • • •

.
They coincide with the systems identified in The- 0 . .. 1 0 In(Y , u)
orem 2.1 , which were introduced in [10], when the
relative degree is equal to one or equal to n .
~.
l/Ji1(Y, u)
l/Jd y, u)
1
Theorem 2.2 Assume that for system (11):
+ ~(J, .

(i) the global relative degree p is well defined and

Co
;=1
r l/Jin(Y, u) .

v :5 p :5 n , = A oz + I (y , u) + W(y , u)(J
(ii) the tracking dynamics are bounded input
y = [0 0 . . . 0 1] Z = Co Z (18)
bounded state, if, and only if, in Rn :
(iii) the distributions 9; = span {g , . .. , a~ g} are (i) rank {dh , . . . , d(Lj-lh)} =n,
involutive and of constant rank i + 1 in Rn
for 0 :5 i :5 p - 1,
(ii) [ad~r, adjr] = 0, 0:5 i , j :5 n - 1,
(iv) ad q .0 C 9j, 1 :5 i :5 p , 0 :5 j :5 p - 2,
Vx E f(n , (iii) [g , adjr] =0, 0:5 j:5 n - 2, Vu E R m ,

873
 (iv) [q; , ad~r] = 0, 0 ~ j ~ n-2 , 'Vu E R m , Proof. The error equation is
1 ~ i ~ p,
z Ai + b,BT (y, u, t)O
(v) the vector fields a~r, 0 ~ i ~ n - 1, are
complete,
8 -f,B(y,u,t) coi (22)

where r is the vector field satisfying where A =

Ao + kc o , i = z - i and 0 B- =
B. According to Definition 3.6 bn > 0 so that by
(dh,r) 1 [01 virtue of (21) we can write

[ (d{L/~1h), r) = i bn{sn + k n _ 1s n - 1 + ... + k 1) =

{s + A)(bn s n - 1 + .. . + bt}
Definition 3.5 A global adaptive observer for which implies
system (17) is a finite dimensional system
-1 bn
tU= a1(w,y(t), u(t)) , w(O) = Wo, wE R r co{s! - A) b= --, . (23)
s+"
x = a2(w, yet), u(t)) , X ERn
There are n - 1 pole-zero cancellations in (23) .
driven by the inputs u(t), yet), such that for every Matrix A is Hurwitz since A > 0 and by as-
x(O) E Rn, Wo E Rr, for any value of the unknown sumption the polynomial bn s n - 1 + ... + bI is Hur-
parameter B and for any bounded IIx(t)l!' lIu(t)IL witz. Therefore, the triple (A, b, co) satisfies the
'Vt ~ 0: strict positive real condition and Meyer-Kalman-
Yacubovich Lemma [17] applies. At this point the
(i) Ilw(t)1I and IIx(t)-x(t)11 are bounded, 'Vt ~ 0; conclusion follows by simple arguments. 0
(ii) lim Ilx(t) - x{t)11 =
t-oo
o. 0 Definition 3.7 An input-output filtered transfor-
mation is defined as a global state space transfor-
Definition 3.6 The system mation, which may depend on the unknown B,

i = Aoz + r(y, u) + b,BT (y, u, t)B z = T(x , ~, B)

y = CoZ (19) with z E Rn , X E Rn , ~ E R' , B E RP, T a Coo
mapping from Rn X R' x RP _ Rn and ~(t) a
with Z E Rn, y E R, u E R m , B E RP , signal generated by the auxiliary linear, asymptot-
r a smooth function mapping R x ~ _ Rn , ically stable filter (A is a Hurwitz matrix) driven
b = [b 1 , ... , bnf E Rn a vector with bn > 0 by the known signals u and y
such that the polynomial bn s n - 1 + ... + bl is Hur-
witz, i. e. all its zeros have negative real part, ~=A~+a(y,u), ~(O) = 0
,B: R x R X R+ - RP with continuous compo- such that there exists a Coo function S( z, ~, B) sat-
nents, (Ao, co) the Brunovsky observer canonical isfying
form is said to be in adaptive observer form. 0
z=T(S(z,~ , B), ~,B), 'V~ER', 'VBERP.
Systems in adaptive observer form (19) belong
to a restricted subset of systems (18), since the o
matrix \}iCy, u) can not be expressed in general as
Theorem 3.5 Given the system
b,BT (y , u) . We will first show how to design adap-
tive observers for systems in adaptive observer i Aoz + r(Y, u) + \}iCy, u)B
form with ,B bounded . We will then show how P
to transform the more general system (18) into
adaptive observer form.
Aoz + r(y, u) + 2: 'I/J;(y, u)B;
;=1
Theorem 3.4 Given a system in adaptive ob- (24)
server form (19) if ,B(y, u, t) is uniformly bounded
for every (y, u) bounded, then the system and a vector b = [b l , . .. , bn - 1, If such that

z (Ao + kco)z + r(y, u) + b,BT (y, u, t)O - ky sn-1 + bn _ I s n - 2 + ... + b2 s + bl =

n-1
f,B(y, u, t)(y - coz) (20) IT (s + Ai)
;=1
is a global adaptive observer for (19), where z E
Rn, 0 E RP, f is any symmetric positive definite is a Hurwitz polynomial (A;, 1 ~ i ~ n , are arbi-
matrix and k = [k1' . .. ,knf is given by trary positive reals) , the filtered transformation
P
(21) Zj - 2:~i[i]B; , l~j~n-l
;=1
with A an arbitrary positive real. TIn (25)

874
with ~[i] E Rn-l satisfying for 1 ~ i ~ p Theorem 3.5 applies to a realization in observer
form of (28) and guarantees the existence of a

[! 1~[i]
0 0 filtered transformation into an adaptive observer
0 0 -b
-h,2 form (19). Theorem 3.4 applies and leads to the
i[i] 1 0
-~3 construction of an adaptive observer. 0

0 1 -b n- 1
0
1
0
0 -b
-h,2 1 ADAPTIVE OUTPUT FEEDBACK

+ [l 0 1
:
-bn- 1
"pi(Y, u) In this section we consider nonlinear uncertain
systems
(26)
transforms (24) into the adaptive observer form
x = f(x) + [90(X) + tB;g;(X)] u + tB;q;(X)
6.
P f(x) + g(x, B)u + Q(x)B
iJ = Ao77 + ,(y, u) + b l)~n-l[i] + "pin(Y, u))9i Y h(x) (29)
i=1
6.
where x E Rn is the state vector, u E R is the
Ao77 + ,(y, u) + b{3(y, u, ~)9 control, £I = [£1 1 , . .. ,Bpf is the constant uncertain
y = Co 77 (27) parameter vector belonging to f2, a closed subset
in RP, h : Coo (R" , R) is the output function, f, go,
with {3 = L:f-l (~n-l [i] + "pin) bounded for every gi, q;, 1 ~ i ~ p, are smooth vector fields in Rn,
bounded u and y. with go(x) :I 0, for every x E Rn. We assume that
h(O) = 0 and that £I = 0 is the nominal parameter.
Proof. Differentiatin?; (25) with respect to time We will address the design of adaptive tracking
and taking (24) and (26) into account, we obtain controls defined as follows.
P Definition 4.8 A global adaptive output feedback
tracking control for system (29) is a finite dimen-
771 = 'I+Lbl(~n-di]+"pin)9i sional system
;=1
77j Zj-l +,j w
= p( w, y(t) , t), w(O) = Wo, w E R'
P u = u(w, y(t), t), uE R
+ L( -~j-di] + bj(~n_l[i] + "pin))9i such th at for every x(O) E Rn, Wo E R' and for
i=1 every bounded smooth output reference Yr(t) with
P bounded time derivatives, Ilx(t)II, IIw(t)1I and u(t)
77j-l +,j + L bj(~n-di] + "pin)9;, are bounded for every t ~ 0, and
;=1
lim [y(t) - Yr(t)] =0 .
2~j~n t-oo
o
which may be rewritten in matrix form as (27).
For every y and u bounded, ~n-di], 1 ~ i ~ p, Theorem 4.6 Let system (29) be of relative de-
is also bounded and therefore the matrix {3 has gree p for every £I E f2. Then , system (29) is
bounded entries . 0 transformable by a global state space diffeomor-
Theorem 3.5 shows that nonlinear systems (17) phism (independent of B)
satisfying the conditions of Theorem 3.3 are trans- (=T(x), T(O)=O, (ERn
formable into the system (27) in adaptive observer
form whose state 7J is related to the state Z (which into
is in turn related to the original state x by a £1- P
independent transformation) by a £I-dependent fil- ( Ac( + b(B)u(y)u + "po(y) +L "p;(y)B;
tered transformation (25) , (26). If Theorem 3.4
i=l
applies to the transformed system, it provides an
adaptive observer for a state of the system. Y Cc ( (30)
In the case of linear systems with unknown pa- with u: R -> R, "pi : R -> Rn, and (Ac,b,cc) in
rameters we have the following result as a special observer canonical form
case.
0

[~
1 0

J
Remark 3.1 Consider a single input, single out-
put linear system represented by the transfer func- 0 1
tion b(B) = bp(B)
0 0
0 0
bn(B)
[ 1 0 0 o]

875
if and only if: the filters (1 ~ j ~ n - 1, 0 ~ i ~ p)
(i) rank{dh, d(Lfh), ... , d(L,-lh)} = n, 1 . .. 0
-,\2 ... 0
(ii) [ad~r, a~r] = 0, 0 ~ i,j ~ n - 1, itfj]
[T 0
...
-'\;-1
r ] pul
(iii) [q;,ad}r] = 0, 1 ~ i ~ p, 0 ~ k ~ n-2, 0 0 -,\;

~
(iv) there exist a smooth function u : R
---> Rand
n-p+1 reals depending on 0, bp(O), ... , bn(O),
such that +[ ] u(y).
n-p+1
g = (u h) L bn- H1 (0)acfc=J)r ... 0 o ]pfj]

!]
0 yfj] [ 1 0 (33)
i=l 1 0 .. .
[ -d,
-d3 0 1 .. .
with bp(fl)sn- p+ ... +bn_1(fl)s+bn(fl) a Hur-
~[i]
witz polynomial with bp(fl) of known and con-
stant sign for every fl E f.!, -d~_l 0 0 " .
,[I]
-dn 0 0 ...
(v) the vector fields a~r, 0 S i ~ n - 1, are

~ ~,(y)
1 0 0
complete, [ -d,
-d3 0 1 0

where r is the vector field satisfying

+ : ]
-dn - 1 0 0 1
-dn
1 [ 1
0 0 0
(dh, r) 0 (34)
[ (d(L,~lh), r) = i with pfj] E Ri and ';[i] E Rn-I , transforming sys-
tem (31) into the adaptive tracking form
o
i A cz + deign(bl(~))U(Y)U
Definition 4.9 The system

z. = Acz + b [U(y)U
- - + oo(y, t) + f..
Lt fl;o;(y , t) 1 + ~6iy(j - 1] + ~ fl;(1/!i1(Y) + 6 [iD)
, ;=1
y = CcZ (35)

in which (Ac, cc) are in observer canonical form , withflo = 1 and(,,6 2 , •.. , 6n , 01 , ... , fl p ) unknown
b = [1 , b2 , • .. , bn]T is a Hurwitz vector, o;(y, t) , parameters. Ify(t) is bounded then y[i](t) , 1 ~ i ~
o ~ i ~ p , and u{y) are known smooth functions , n - 1 are also bounded.
, and 0; , 1 ~ i ~ p , are unknown constant param- Theorem 4.7 Consider system (29) with relative
eters with, > 0, is said to be in adaptive tracking degree p = 1. If assumptions (i)-(v) of Theorem
form . 0
4.6 are satisfied, then there exists a global adap-
tive output feedback tracking control for every un-
Lemma 4.1 Consider the system (30) with p =1 known value of the parameter vector fl for system
(0 0 = 1) (29).
p Proof By virtue of Theorem 4.6, system (29) is
( Ac( + b(fl)u(y)u + L fl; 1/!;(y) transformable into system (30) by a global state
;=0 space diffeomorphism . By virtue of Lemma 4.1 fil-
(31) tered transformations transform system (30) into
the adaptive tracking form (35). Since the triple
Given any constant Hurwitz vector d (Ac , d, cc) is observable , there exists a vector kc
such that for any positive real Ac
[1, d2 , .. . , dnf such that
n-l
det(sI - Ac - dk c) = (s + Ac)
sn-1+d 2 s n- 2 + . +dn _ 1s+dn = IT (s+'\;) (32) (sn-l + d 2 s n- 2 + ... + dn-1s + dn )
;=1 and
with A; , 1 ~ i ~ n - 1 arbitrary positive reals, -1 1
there exists a filtered transformation depending on Cc ( sI - A c - dk c ) d = --\-
s+ Ac

876
is strictly positive real. Define the reference model while the observer error dynamics are
ir = (Ac + dkc)zr + du r , Zr ERn
Yr = CcZr · (36) Z = (Ac+koCc)Z+d(bli7(y)u+t,bjY[j-1]
Note that with zr(O) = 0 we have
fir + >'cYr = U r (37) + tBi(tPil(Y) +~Ili])) .
which gives ur(t). Let e = Z - Zr be the tracking
error whose dynamics are, according to (35), (42)
1 n
e = (Ac + dkc)e + d [ -i7(y)u + L: bjy[j - 1] Since both triples (Ac + dk c , d, cc) and (Ac +
koc c, d, cc) are strictly positive real by Meyer-
, j=2
-Kalman-Yacubovich Lemma there exist two pos-

~Bi(tPil(Y)+6[i])-kcz-ur]
itive definite matrices Pc and Po, two vectors qc
+ (38) and qo, and two positive reals ic and io such that

The state estimate i is provided by the adaptive (Ac + dkcl Pc + Pc(Ac + dk c) -qcq'[ - ic l
observer Pc d cTc
-qoq~ - ioI
Z = Acz + d [8 1 i7(Y)U + ~ 8j y[j - 1] + tPOl(Y)
(Ao + koccf Po + Po(Ac + koc c)
Pod T
Cc •

+6[0]+ tOi(tPil(Y)+~di])] - ko(Y-Y)

Define the function (b
- - T
[B 1 , •.. , Bp] , TJ > 0)
= [b 1 , .. ., bnY, B=

y ccz (39)
in which 81 is an estimate of 01 = 1h and ko is
chosen so that
whose time derivative is
det(sI - Ac - koc c) = (s + >'0)
(sn-l + d2 s n - 2 + . . . + dn ) v < -iclleW - TJiolliW + 211ellllc'[ kcllllill
which implies that
+2 !: (60 + (y - Yr )i7(Y)(t, 8j y[j - 1]
cc(sI - Ac - k occ)-ld = ~
s+ p
+ L: Bi(tP;I(Y) + 6[i])
Ao

is strictly positive real for any positive real >'0.

The control u is defined as ;=1

u= __ 8
u(y)
0
[t 8jy[j-1]+t O;(tPil(Y)
j=2 ;=1 n

+28 1 (6 1 + TJ(Y - y)i7(y)u) + 2:L bj

+6[i]) - kcZ - U r + tPOl(Y) + 6[0]] (40) j=2

in which 80 is an estimate of 00 = ,. Substitut- (6i + ((y - Yr) + TJ(Y - y))y[j -1])

ing (40) in the tracking error dynamics (38) we
have(Z = Z - z)
+2 t B; (B;
;=1
+ ((y - Yr) + TJ(Y - y))
e = (Ac+dkc)e-dkci+d[tB;(tPil(Y)
(tPil (y) + ~Ili])) .
n

+6[i]) + L: bjy[j - 1] Defining the parameter estimates dynamics

j=2

p
+ I: B;(tP;I(Y) + 6[i])
;=1 ;=1

(41)

877
 y)c5o (~c5jy(j - 1]
Proof. Let
-77(Y -
i Acz + bu - ay,
p
Y = cc z , yER
+L B;(~i1(Y) + {di])
;=1 be an observable realization of W( s) with n n*, =
b = [O, ... ,bp, . . . ,bno]T and a =
fa1, . .. , ano]T
-kc z - U r + ~01(Y) + {dO]) unknown vectors. This system satisfies conditions
(i)-(v) of Theorem 4.6, since it is minimum phase
((y - Yr) + 77(Y - y) }~i1(Y) + 6 [i]), and bp is of known sign. Therefore either Theorem
4.7 or Theorem 4.8 apply. 0
lSiSp
((y - Yr) + 77(Y - y))y(j -1], PHYSICAL APPLICATIONS
(44)
Example 5.1 Consider the reduced order model
with of a synchronous generator connected to an infi-
nite bus

77 > 6 = w - w.
W = ()1 - ()2w~1 sin 6 + ()3 sin 6 cos 6
we have ~I vI - ()4w~1 + ()s cos 6 .
(45) The parameters ()1, ()2, ()3 , ()4, ()s may undergo
an abrupt change due to short circuits in the line
connecting the generator to the infinite bus or to
for a suitable ( > O. This implies that e(t) , E(t) , turbine failure affecting ()1 . Those changes are
8i (t), 0 S i S n , B;(t) , 1 S i Spare bounded unknown and the task of the control algorithm is
for every t ? 0 and in particular that yet) - Yr(t) to prevent the generator to go out of step , i.e. to
is bounded which, since Yr(t) is bounded , implies keep 6 as close as possible to the stable operating
that yet) is bounded . Since yet) is bounded, from point 6e with w =
w. by acting on the control
(33)-(34) {i(t) , 0 S i S p and y[i](t) , 1 Si S n-1 input v J. We show how to solve this problem by
are bounded as well. From (41) and (42) it follows a state feedback adaptive control. For simplicity
that Ile(t)11 and lIi(t)11 are bounded . Taking (43), we do not introduce a reference model. Define
(45) into account we may apply Barbalat 's Lemma (A1 > 0)
[17] and conclude that for any initial condition

t~~ II[ E(t)

e( t)
] 11 =0 . with
o
Theorem 4 .8 Consider system (29) with relative so that the first two equations
degree p, 2 S p S n. If assumptions (i)-(v)
of Theorem 4.6 are satisfied, then there exists a 6 = w -wo
global adaptive output feedback tracking control for W = ()1 - ()2w ~1 sin 6 + ()3 sin 6 cos 6
every unknown value of the parameter vector () for
system (29) . may be rewritten as

Proof. See [14] . o -A18 +w

w ()1 - sin 6 + ()3 sin 6 cos 6
()2W ~ I
Theorem 4.9 Consider a linear system with +A1(W - w.) . (47)
transfer function
Introducing the new parameter (recall that ()2 >
(46)
0) ()6 = 1/()2 , the parameter estimates B; and er-
rors Oi = ()i - Bi , and defining (A2 > 0)
Assume that this system is minimum phase, has 1 · . .
constant and known relative degree p with high fre- - .- ()6(A2W + ()1 + ()3 sin 6 cos 6
wsm6
quency gain bp of known sign , and an upper bound
+A1(W - w.))
n* for n is known. Then, there exists a global
adaptive output feedback tracking control. ~I

878
equation (47) may be rewritten as Defining the adaptation laws

-A16 +w 01 WZ1 W + W31;fij

W -A2W - (}2w;fi j sin 6 + B1
+ B3 sin 6 cos 6 O2 W32;fij - w sin 6w;fij
+06(}2(A2W + 01 + 03 sin 6 cos 6 + A1(W - w.)) 03 = W23 W + W33;fij
- A2W - (}2w;fi j sin 6 04 = W341/;j

+ [W21 W23 W26] [ ~~

(}2(}6
1 Os
86
W35'1/;j
WZ6 W
we have
while
. -2
V=-A 1 6 -AzW -A3'1/;j+6w
- Z -z --
;fij = Vj+8 4 (-w'l/;j)+8scos6
which allows us to conclude that 11(6,w,;fij )11 tends
06 ~ ~ to zero. 0
--.-[A1Az(W -w.) + 81 + 83 sin6cos6
wsm6
Example 5.2 Consider a single-link rigid robot
+03(COS Z 6 - sin 2 6)(w - ws ) rotating on a vertical plane modeled as
-(A2W + 01 + 03 sin 6 cos 6
06 06(w-w.)cos6)]
1
I q·· + 2mg I cos q = U

+A1 (w - w »)( - - - -=..:.....--....;..:..--

• w sin 6 w sin 2 6 with unknown moment of inertia I and unknown
constant mgl. The model may be rewritten in
8 [ (AI + A2)06
+ 1
wsmu
. >: state space form as (Xl q, Xz q) = =
06 A A Xl X2
+ w 2 sin 6 (A2 W + 81 + 83 sin 6 cos 6 1
Xz Iu+8cosxl
+A1(W - w.))] + 82[06(A1 + A2)'I/;j
y Xl (48)
'l/;j 06
+ 81 + (}3 sin 6 cos 6
A A

---(A2W with () = -(mgl)j(2I). The system has relative

W
degree two. Theorem 4.8 applies and gives the
+A1(W - w.))] following design procedure. Introduce the filter
II [ (AI + A2)06 cos 6 r.p = -<p +U
+U3 - -'--'------'----
W

06 cos 6 which along with a filtered transformation trans-

+
A A

2 (A2W + 81 + (}3 sin 6 cos 6 forms (48) into

W
+A1(W - w.))]
5
'-
Z = Ac z + I_ 1[1]
1 <p + [0]
cos 8 y
D. Vj + W30 + L W3i(}i . y = cc z . (49)
i=l
Introduce the filter
Define (A3 > 0)

5
i = -~ + [-1 1] [ co~ y ] = -~ + cos y
Vj -A3;fij + 02 WW sin 6 - W30 - L W3i Oi which along with a filtered transformation allows
i=l us to transform (49) into

so that
i = Acz + [ ~ ] (~<p + 8~)
5
Jj = -A3;fij +02wwsin6+ L W3i Bi . y = Cc Z (50)
;=1 which may be rewritten as
Consider the function <pO - j ([ Ac 1] [ ~~ ] + O~ - U r)

i Ac z + [ ~ ] (~<p. + ()~ + ~~)

879
According to the proof of Theorem 4.8 the result- [8] G. Luders and K.S . Narendra. An adaptive
ing adaptive tracking control (S is the estimate of observer and identifier for a linear system.
IEEE '['rans. Automatic Control, 18:496-499 ,
6 = J) is 1973.

Ac i + [1] ' . [-I-A]

1 (6<p+8~)- -Ao 0 (y-y)
[9] R. Marino. Adaptive observers for single out-
put nonlinear systems. IEEE Trans. Auto-
matic Control, 35:1054-1058, 1990.
y
[10] R . Marino, W. Respondek, A. J.
o [(y - Yr) + T}(Y - y)]~ van der Schaft, and P. Tomei. Nonlinear Hoo
~~ ] + O~ -
almost disturbance decoupling. Systems and
1 (y - Yr) ([ Ac 1] [ ur) Control Letters, 23 :159-168, 1994.
[11] R. Marino and P. Tomei. Dynamic output
-T}(Y - y)1 ([ Ac 1] [ ~~ ] + O~ - U r) feedback linearization and global stabiliza-
tion . Systems and Control Letters, 17:115-
[=~: ~1] ~
121, 1991.
Zr +[ ] Ur
[12] R. Marino and P. Tomei. Global adaptive ob-
Yr [1 0] Zr servers and output-feedback stabilization for
a<p- ~ a<p- ~ a<p- ~ a<p-.) a class of non linear systems. In P.V. Koko-
U
- ( of 1+ ai Z + oB 8 + a~ ~ tovic, editor, Foundations of Adaptive Con-
trol, pages 455-494, Springer-Verlag, Berlin,
-k1(<p-<p-)+<p . 1991.
[13] R. Marino and P. Tomei. Global adaptive
Note that u depends on f , i , B and ~ which are observers for nonlinear systems via filtered
known signals. 0 transformations . IEEE Trans. Automatic
Control, 37:1239-1245, 1992.
References [14] R. Marino and P. Tomei. Global adaptive
[1] G. Bastin and M. Gevers . Stable adaptive output-feedback control of non linear systems.
observers for nonlinear time varying systems . Part I: linear parameterization. IEEE Trans.
IEEE Trans. Automatic Control, 33:650-658 , Automatic Control, 38:17-32, 1993.
1988.
[15] R. Marino and P. Tomei. Nonlinear Con-
[2] I. Kanellakopoulos , P . V. Kokotovic, and R . trol Design: geometric, adaptive, and robust.
Marino. An extended direct scheme for ro- Prentice Hall, 1995.
bust adaptive nonlinear control. Automatica,
27:247-255, 1991. [16] K. Nam and A. Arapostathis. A model
reference adaptive control scheme for pure-
[3] I. Kanellakopoulos, P. V. Kokotovic , and feedback nonlinear systems . IEEE '['rans.
A. S. Morse. Adaptive output-feedback con- Automatic Control, 33:803-811, 1988.
trol of systems with output nonlinearities. In
P.V. Kokotovic, editor, Foundations of Adap- [17] K. S. N arendra and A. M. Annaswamy. Stable
tive Control, pages 495-525, Springer-Verlag, Adaptive Systems. Prentice Hall, Englewood
Berlin , 1991. Cliffs, 1989.
[18] S. S. Sastry and M. Bodson. Adaptive Con-
[4] I. Kanellakopoulos, P. V. Kokotovic, and trol: Stability, Convergence , and Robustness.
A. S. Morse. Systematic design of adaptive Pr entice Hall, 1989.
controllers for feedback linearizable systems .
IEEE Trans . Automatic Control, 36:1241- [19] S. S. Sastry and A. Isidori . Adaptive control
1253, 1991. of linearizable systems. IEEE Trans . Auto-
matic Control, 34:1123-1131 , 1989.
[5] I. Kanellakopoulos, P.V. Kokotovic , and
A. S. Morse. Adaptive output-feedback con- [20] D. Taylor, P. V. Kokotovic, R . Marino, and
trol of systems with output nonlinearities . I. Kanellakopoulos . Adaptive regulation of
IEEE Trans . Automatic Control, 37:1666- nonlinear systems with unmodeled dynamics.
1682, 1992. IEEE Trans . Automatic Control, 34:405-412 ,
1989.
[6] A. J. Krener and A. Isidori . Linearization
by output injection and nonlinear observers .
Systems and Control Letters, 3:47-52 , 1983 .

[7] M. Krstic, I. Kanellakopoulos, and P.V .

Kokotovic. Adaptive nonlinear control with-
out overparametrization. Systems and Con-
trol Letters, 19:177-185, 1992.

880