Proceedings of the 27th Chinese Control Conference

July 16–18, 2008, Kunming, Yunnan, China

Decentralized Robust Adaptive Control of Uncertain Nonlinear

Systems Using Dual High Gain*
P. Krishnamurthy1, F. Khorrami1,2
1. IntelliTech Microsystems, Inc., 4931 Telsa Drive, Suite B., Bowie MD 20715
2. Control/Robotics Research Laboratory (CRRL), Dept. of Elec. and Comp. Engg., Polytechnic University, Brooklyn, NY
E-mail: [email protected]

Abstract: We propose a decentralized output-feedback controller for a general class of interconnected large-scale systems. Each
subsystem in the large-scale system is comprised of a triangular (strict-feedback-like) system coupled with an appended Input-
to-State Stable (ISS) dynamics. The design is based on our recent results on a new adaptive high-gain control design approach
utilizing a dual high-gain observer and controller architecture with a dynamic scaling. The technique provides strong robustness
properties and allows the system class to contain unknown functions dependent on all states and involving unknown parameters
(with no magnitude bounds required). Both the uncertain parameters and the appended dynamics are allowed to be coupled with
the entire system state.
Key Words: Decentralized, Large-scale systems, High gain, Output-feedback, Adaptive control, Nonlinear systems

1 INTRODUCTION centralized controllers were designed in [5, 6] for systems

of output-feedback canonical form including uncertain pa-
We consider a class of interconnected large-scale systems
rameters and polynomially bounded uncertainties. Decen-
with each subsystem being of the form
tralized output-feedback robust disturbance attenuation for
ż(i,m) = q(i,m) (z, x, u, t) , i = 1, · · · , sm + 1 systems in output-feedback canonical form with appended
stable linear dynamics with interconnections bounded by
ẋ(i,m) = φ(i,m) (z, x, u, t) + φ(i,i+1,m) (x(1,m) )x(i+1,m) ,
nonlinear functions of the outputs was addressed in [7].
i = 1, · · · , sm − 1 Using the Cascading Upper Diagonal Dominance (CUDD)
ẋ(i,m) = φ(i,m) (z, x, u, t) + φ(i,i+1,m) (x(1,m) )x(i+1,m) based technique in [8], a decentralized output-feedback dis-
+μ(i−sm ,m) (x(1,m) )um , i = sm , · · · , nm (1) turbance attenuation scheme was proposed in [9] for in-
terconnected large-scale systems with each subsystem be-
where1 xm = [x(1,m) , · · · , x(n,m) ]T ∈ Rnm is the state, ing in the generalized output-feedback canonical form [8]
ym = x(1,m) ∈ R the output, um ∈ R the input, and and with nonlinear appended dynamics. A decentralized
T T n +···+nzsm +1,m
zm = [z(1,m) , · · · , z(sm +1,m)
]T ∈ R z(1,m) controller based on the dual high-gain scaling based tech-
the state of the appended dynamics of the mth subsystem. nique was proposed in [10]. In this paper, we design a de-
M is the number of subsystems. Define x = [xT · · · , xT T centralized controller based on our recent results on a new
1 , M] ,
M dynamic high-gain based adaptive control design technique
u = [u1 , · · · , uM ]T , z = [z1T , · · · , zM
T T
] , nx = m=1 nm , [11]
sm +1 M which is applicable to a wide class of nonlinear trian-
nzm = i=1 nz(i,m) , and nz = m=1 nzm . φ(i,i+1,m) : gular systems. The bounds assumed on φi , i = 1, · · · , n,
R → R, i = 1, · · · , nm − 1, and μ(i,m) : R → R, i = (Assumption A2) allow cross-products of unknown param-
0, · · · , nm − sm are known continuous functions of their eters and unmeasured states with no magnitude bound or
n
arguments. q(i,m) : Rnz +nx +M+1 → R z(i,m) , i = sign information on the unknown parameters being required.
nz +nx +M+1
1, · · · , sm + 1 and φ(i,m) : R → R, i = While earlier control design techniques such as the clas-
1, · · · , nm , are uncertain continuous functions which can sic high-gain designs[12−15] and backstepping [16] cannot
contain both functional and parametric uncertainties. sm is handle cross-products of unknown parameters and unmea-
the relative degree of the mth subsystem. sured states, the dynamic scaling-based dual high-gain ob-
In this paper, we address the output-feedback (i.e., only ym server/controller design approach developed in our recent
measurement available to the controller for the mth sub- papers [17,18,11] provides a flexible framework which can ac-
system) decentralized control of the system (1). The de- comodate such cross-products.
centralized controller proposed in this paper addresses a
much wider class of systems compared to previous results 2 ASSUMPTIONS AND STATEMENT OF
in the literature. Early results in decentralized control fo- MAIN RESULT
cused on linear systems [1] and linearly bounded intercon-
nections [2,3] . In [4], higher order (i.e., polynomial type) The design is carried out under the Assumptions A1-A5,
interconnections were considered for large-scale systems as- each of which is required to hold for all m ∈ {1, · · · , M }.
suming matching conditions. Backstepping-based robust de- Assumption A1 A constant σm > 0 exists such that for all
x(1,m) ∈ R,
* This
work is supported in part by the NSF under grant ECS-0501539.
numbers (−∞, ∞), the set of nonnegative real numbers
1 The set of real
|φ(i,i+1,m) (x(1,m) )|  σm , 1  i  nm − 1 (2)
[0, ∞), and the set of real k-dimensional column vectors are denoted by R,
R+ , and Rk , respectively. x(nm +1,m) ≡ 0 is a dummy variable. |μ(0,m) (x(1,m) )|  σm > 0 (3)

735
 M

Assumption A2 A continuous function Γm : R → R+ is
known such that +θm β(Υm ,k) (x(1,k) )x2(1,k) (9)
k=1

i
|φ(i,m) |  θm Γm (x(1,m) ) |x(j,m) | where θm is an unknown non-negative constant. αΥm is a
j=2 known class K∞ function. βΥm , Λ̃(i,m) , i = 2, · · · , sm ,
i
 M
  Λ̃(1,m,k) , and β(Υm ,k) , k = 1, · · · , M are known contin-
+ Λ(j,m) (|z(j,m) |)+ Λ(1,m,k) (|z(1,k) |) uous non-negative functions. Non-negative constants κΥm
j=2 k=1 and κ̃Υm exist such that |Υm |2  κΥm αΥm (|Υm |) and
M
 |Υm |2  κ̃Υm VΥm for all Υm ∈ Rnm −sm .
+θm Γ(m,k) (x(1,k) )|x(1,k) | Assumption A5 The z(i,m) , i = 1, · · · , sm + 1, subsystems
k=1 are ISS with ISS Lyapunov functions Vz(i,m) satisfying
for 1  i  sm (4)

nm V̇z(1,m)  −αz(1,m) (|z(1,m) |)
|φ(i,m) |  θm Γm (x(1,m) ) |x(j,m) | M

j=2 +θm β(z(1,m) ,k) (|x(1,k) |) (10)
s
m +1 M
 k=1
+ Λ(j,m) (|z(j,m) |) + Λ(1,m,k) (|z(1,k) |) V̇z(i,m)  −αz(i,m) (|z(i,m) |)
j=2 k=1
 M
 M
 +θm β(z(i,m) ,k) (|x(1,k) |)x2(1,k)
+um + θm Γ(m,k) (x(1,k) )|x(1,k) |
k=1
k=1
i
 
for sm + 1  i  nm (5) +βz(i,m) (|x(1,m) |) x2(j,m)
j=2
for all t ∈ R, xm ∈ Rnm , m = 1, · · · , M , z(i,m) ∈
n
R z(i,m) , i = 1, · · · , sm + 1, m = 1, · · · , M , and u ∈ for i = 2, · · · , sm (11)
RM , with θm  0 being an unknown constant and Γm , V̇z(sm +1,m)  −αz(sm +1,m) (|z(sm +1,m) |)
Λ(j,m) , j = 2, · · · , sm + 1, Λ(1,m,k) , k = 1, · · · , M , M
Γ(m,k) , k = 1, · · · , M , and Γ(m,) being known continuous +θm β(z(sm +1,m) ,k) (|x(1,k) |)x2(1,k)
non-negative functions. k=1
Assumption A3 Positive constants ρ(i,m) and ρ(i,m) exist nm

such that for all x(1,m) ∈ R +βz(sm +1,m) (|x(1,m) |)[ x2(j,m)
j=2

|φ(i,i+1,m) (x(1,m) )|  ρ(i,m) |φ(i−1,i,m) (x(1,m) )|,
+θm u2m ] (12)
for i = 2, · · · , sm − 1 (6)
|φ(i,i+1,m) (x(1,m) )|  ρ(i,m) |φ(i−1,i,m) (x(1,m) )|,
where θm is an unknown non-negative constant,
for i = 2, · · · , nm − 1. (7) αz(i,m) , i = 1, · · · , sm + 1, are known class K∞
functions, and βz(i,m) , i = 2, · · · , sm + 1 and
Assumption A4 The inverse dynamics of the mth subsys- β(z(i,m) ,k) , i = 1, · · · , sm + 1, k = 1, · · · , M are
tem of the large-scale system (1), i.e., known continuous non-negative functions. Furthermore,
non-negative constants κz(i,m) and κ̃z(i,m) exist such that2
Υ̇(i,m) = φ(sm +i,m) (z, x, v − ũ, t)
Λ2(i,m) (|z(i,m) |)+ Λ̃2(i,m) (|z(i,m) |)  κz(i,m) αz(i,m) (|z(i,m) |)
+φ(sm +i,sm +i+1,m) (x(1,m) )Υ(i+1,m)
and Λ2(i,m) (|z(i,m) |) + Λ̃2(i,m) (|z(i,m) |)  κ̃z(i,m) Vz(i,m) for
+μ(i,m) (x(1,m) )(vm − ũm ) n
all z(i,m) ∈ R z(i,m) , i = 2, · · · , sm +1. The following local
1  i  nm −sm (8) order hold as π → 0+ :
estimates
M
are Input-to-State Stable (ISS) with state Υm = (a) k=1 [Λ(1,k,m) (π) + Λ̃2(1,k,m) (π)] = O[αz(1,m) (π)]
2

[Υ(1,m) , · · · , Υ(nm −sm ,m) ]T = [x(sm +1,m) , · · · , x(nm ,m) ]T (b) M k=1 [Λ(1,k,m) (π) + Λ̃(1,k,m) (π)] = O[π]
M
and input [z, x̃m , v, t]T where Υ(nm −sm +1,m) ≡ 0 is a (c) k=1 β(z(1,k) ,m) (π) = O[π 2 ].
dummy variable, x̃m is the vector comprised of the entire Remark 1 Note that Assumptions A2, A4, and A5 allow
state x except (x(sm +1,m) , · · · , x(nm ,m) ), vm = um + ũm , cross-products between uncertain parameters and unmea-
φ(sm ,sm +1,m) x(1,m)
ũm = μ(0,m) (x(1,m) ) Υ(1,m) , v = [v1 , · · · , vM ]T , sured states. While earlier control design techniques such
as backstepping based designs would require knowledge of
and ũ = [ũ1 , · · · , ũM ]T . Furthermore, an ISS Lyapunov
a magnitude bound on the uncertain parameters in such a
function VΥm (Υm ) exists which satisfies
case, the dynamic dual high-gain scaling based design al-
 sm
 lows handling of such cross-products without any knowledge
2
V̇Υm  −αΥm (|Υm |) + θm βΥm (|x(1,m) |) vm + x2(i,m) of magnitude bounds on the uncertain parameters as shown
i=2 in [11]. For simplicity, it is assumed that the same func-
sm
 M
  tion Γm serves in the bounding of all uncertain functions in
+ Λ̃2(i,m) (|z(i,m) |) + Λ̃2(1,m,k) (|z(1,k) |)
2Λ for i = sm + 1 is a dummy variable identically equal to zero.
i=2 k=1 (i,m)

736
(5). To remove possible conservativeness, different bound-
ing functions can be introduced; while this would increase ⎫
Pom Aom (x(1,m) ) + ATom (x(1,m) )Pom ⎪
⎪
the algebraic tediousness in deriving the final controller ex- ⎪
⎪
pressions, the design would follow along the same lines as  −νom |φ(1,2,m) (x(1,m) )|Inm ⎬
ν om Inm (13)
here. ⎪
⎪
 Pom (Dom − 12 Inm ) + (Dom − 12 Inm )Pom ⎪
⎪
Remark 2 Assumption A3 requires ratios of the “upper- ⎭
 ν om Inm
diagonal” terms φ(i,i+1,m) to be bounded. The condition (6) ⎫
requires the upper-diagonal terms closer to the input to be Pcm Acm (x(1,m) ) + AT
cm (x(1,m) )Pcm ⎪
⎪
⎪
⎪
larger while condition (7) requires the upper-diagonal terms  −νcm |φ(2,3,m) (x(1,m) )|Inm−1 ⎬
closer to the output to be larger. The conditions (6) and (7) ν cm Inm−1 (14)
⎪
⎪
constitute the cascading dominance assumptions [19] in the  Pcm (Dcm − 12 Inm−1 )+(Dcm − 12 Inm−1 )Pcm ⎪
⎪
⎭
controller context and observer context, respectively, and are  ν cm Inm−1
related to uniform solvability of coupled Lyapunov inequali-
where
ties [17,20] which are instrumental in the design of controller
and observer gains in a dual dynamic high-gain design. Aom = −[g(1,m) , · · · , g(nm ,m) ]T Cm
Theorem 1 Under Assumptions A1-A5, a decentralized +upperdiag(φ(1,2,m) , · · · , φ(nm −1,nm ,m) )(15)
output-feedback control law (i.e., with um depending only
on ym ) can be designed such that for all initial conditions of Acm = −Bm [k(2,m) , · · · , k(sm ,m) ]T
the plant state (x, z) and the controller state (which, as in the +upperdiag(φ(2,3,m) , · · · , φ(sm −1,sm ,m) ) (16)
controller constructed in Section 3, includes an adaptation Cm = [1, 0, · · · , 0] ∈ Rnm (17)
parameter and a high-gain scaling parameter for each sub-
Bm = [0, · · · , 0, 1] ∈ Rsm −1 (18)
system), the following properties are satisfied by all solution
trajectories of the resulting closed-loop system: Dom = diag(1, 1, 2, 3, · · · , nm − 1) (19)
Dcm = diag(1, 2, 3, · · · , nm − 1) (20)
• Solution trajectories exist on the time interval [0, ∞) Ik denotes an identity matrix of dimension k × k,
and all closed-loop signals are bounded on [0, ∞). diag(a1 , · · · , ak ) denotes the k × k diagonal matrix with the
ith diagonal element being ai , and upperdiag(a1 , · · · , ak )
denotes the (k + 1) × (k + 1) matrix which has its upper
• The overall state of the large-scale system, i.e.,
diagonal elements (i.e., elements at positions (i, i + 1), i =
[xT , z T ]T asymptotically converges to zero as t → ∞.
1, · · · , k) equal to ai , i = 1, · · · , k, and zeros everywhere
Also, the control inputs um , m = 1, · · · , M asymptoti-
else.
cally converge to zero as t → ∞.
Furthermore, by Theorem A1 in [17],

3 PROOF OF THEOREM 1 g(1,m) (x1 ), · · · , g(n,m) (x1 )

A decentralized output-feedback dynamic controller satisfy- can be picked to be linear constant-coefficient combinations
ing the assertions in Theorem 1 is designed in this section of φ(1,2,m) (x(1,m) ), · · · , φ(nm −1,nm ,m) (x(1,m) ). Hence, us-
based on the technique introduced in [11] where a dynamic ing

 Assumption A3, a positive constant Gm exists such that
nm 2
i=1 g(i,m) (x(1,m) )  Gm |φ(1,2,m) (x(1,m) )|.
dual high-gain design was constructed to address uncertain
nonlinear system wherein unknown parameters can appear
coupled with unmeasured states (with no a priori bounds on 3.2 Observer Design
unknown parameters, i.e., global in unknown parameters). A full-order observer for the mth subsystem is given by
The design utilizes a high-gain observer and a high-gain con-
trol law for each subsystem as described below. In addition, x̂˙ (1,m) = φ(1,2,m) (x(1,m) )x̂(2,m)
an adaptation parameter and a dynamic scaling parameter are ṙm
introduced and their dynamics are designed. − (x̂(1,m) − x(1,m) )
rm
−rm g(1,m) (x(1,m) )[x̂(1,m) − x(1,m) ]
3.1 Coupled Lyapunov Inequalities
˙x̂(i,m) = φ(i,i+1,m) (x(1,m) )x̂(i+1,m)
As discussed in Remark 2, Assumption A3 is central
i
to the dual dynamic high-gain scaling based control de- −rm g(i,m) (x(1,m) )[x̂(1,m) − x(1,m) ]
sign technique and ensures solvability of the key coupled for i = 2, · · · , sm − 1 (21)
Lyapunov inequalities that arise in the design. Specifi- x̂˙ (nm ,m) = φ(i,i+1,m) (x(1,m) )x̂(i+1,m) +
cally, using Theorems A1 and A2 in [17], Assumptions
μ(nm −sm ,m) (x(1,m) )um
A1 and A3 are necessary and sufficient for the exis-
i
tence of functions g(1,m) (x(1,m) ), · · · , g(nm ,m) (x(1,m) ) and −rm g(i,m) (x(1,m) )[x̂(1,m) − x(1,m) ]
k(2,m) (x(1,m) ), · · · , k(sm ,m) (x(1,m) ), symmetric positive- for i = sm , · · · , nm (22)
definite matrices Pom and Pcm , and positive constants νom ,
ν om , ν om , νcm , ν cm , and ν cm to satisfy (13) and (14) for all where rm is a dynamic high-gain scaling parameter (whose
x(1,m) ∈ R dynamics will be designed later) and g(i,m) : R → R, i =

737
1, · · · , nm are continuous functions chosen to ensure solv- 3.4 Dynamics of the High-Gain Scaling Parameter and
ability of the coupled Lyapunov inequalities (13). The ob- Adaptation Parameter
server errors e(i,m) and scaled observer errors (i,m) are de- As in [17], the dynamics of the high-gain scaling parameters
e
fined as e(i,m) = x̂(i,m) − x(i,m) and (i,m) = r(i,m) i−1 for rm , m = 1, · · · , M are designed as
m
1  i  nm . The dynamics of the scaled observer error  
˙
vector m = [(1,m) , · · · , (n,m) ]T are given by ṙm = λm Rm (x(1,m) , θ̂m , θ̂m ) − rm
ṙm ˙
˙m = rm Aom m − Do m − Φm (23) ×Ωm (rm , x(1,m) , θ̂m , θ̂m ) (31)
rm m
φ(i,m) with initial value rm (0)  1 and with λm being any contin-
Φm = [Φ(1,m) , · · · , Φ(nm ,m) ]T , Φ(i,m) = i−1 (24) uous non-negative function such that λm (π) = 1 for π  0
rm
and λm (π) = 0 for π  −rm with rm being any positive
where Aom and Dom are defined in (15) and (19), respec- constant. The functions Rm and Ωm will be constructed in
tively. The term − ṙrm
m
e1m introduced in the dynamics of the stability analysis later in this section. The dynamics of
x̂1m in (22) contributes a value of 1 to the (1, 1) element of the adaptation parameter are designed as
Dom thus ensuring that the matrix (Dom − 12 Inm ) is positive-
˙
definite. This is crucial to the solvability of the second Lya- θ̂m = Θ(1,m) (x(1,m) )x2(1,m)
punov inequality in (13).
+Θ(2,m) (x(1,m) , rm )(x̂(1,m) − x(1,m) )2 (32)
3.3 Control Law Design
with θ̂m (0) > 0, and with Θ1 : R → R+ , Θ2 : R2 → R+ ,
The control law for the mth subsystem is given by
and γ : R4 → R+ being continuous functions.
 sm
rsm 
um = − m k(i,m) (x(1,m) )η(i,m) 3.5 Closed-Loop Stability Analysis
μ(0,m) i=2
 The closed-loop stability analysis is based on constructing an
x̂(s +1,m) overall Lyapunov function V (x, z, θ̂1 , · · · , θ̂M , r1 , · · · , rM )
+φ(sm ,sm +1,m) (x(1,m) ) msm (25)
rm which has the property that V̇ is negative when either the
x̂ +ζ (x ,θ̂ )
 high-gain scaling parameters r1 , · · · , rM or their derivatives
η(2,m) = (2,m) m
rm
(1,m) m
are large (as specified through appropriate functions of the
x̂(i,m) (26)
η(i,m) = rmi−1 , i = 3, · · · , sm available outputs and of the unknown parameters) and then
using the adaptive technique introduced in [11] to replace the
where the controller gain functions k(i,m) : R → R, i =
unknown parameters with on-line bound estimates. Based on
2, · · · , sm are chosen to ensure solvability of the coupled
the techniques introduced in [11, 17], a suitable
M Lyapunov
Lyapunov inequalities (14) and ζ : R2 → R is a continu-
function is constructed to be of form V = m=1 Vm with
ous function which can be picked to be of the form
β(1,m) 2
ζm (x(1,m) , θ̂m ) = (1 + θ̂m )x(1,m) ζ(1,m) (x(1,m) ) (27) Vm = cm rm T T
m Pom m + rm ηm Pcm ηm + x(1,m)
2
with ζ(1,m) being a continuously differentiable function VΥm
+(cm κΥm + 1) 2sm −1
and θ̂m a parameter estimator. The dynamics of ηm = rm
[η(2,m) , · · · , η(sm ,m) ]T are s
m +1
κz(i,m) Vz(i,m)
+(cm + 1)
ṙm 2i− 5
rm 2
η̇m = rm Acm ηm − Dc ηm − rm G(2,m) e(1,m) i=2
rm m
+(1 + G∗m θm
∗
)Ṽz(1,m) (33)
+Hm (η(2,m) − (2,m) ) + Ξm (28)
∗ 2
with Acm and Dcm defined in (16) and (20), respectively, where for m = 1, · · · , M , θm = 1 + θm + θm , cm , cm , and
∗
G(2,m) = [g(2,m) , · · · , g(n,m) ]T , and Gm are appropriate positive constants, β(1,m) is a positive
  constant free to be picked by the designer, , and Ṽz(1,m) is
 a Lyapunov function for the z(1,m) subsystem obtained by
Hm = (1 + θ̂m ) ζ(1,m) x(1,m) + ζ(1,m) φ(1,2,m) ,
T reshaping Vz(1,m) to meet local order restrictions based on
0, · · · , 0 (29) z(1,m) dependent terms which arise from the uncertain terms
 in the system. The design functions ζ(1,m) , Θ(1,m) , Θ(2,m) ,
1 ˙ Rm , and Ωm are constructed using techniques analogous to
Ξm = θ̂m ζ(1,m) x(1,m)
rm [11, 17]. Using Assumption A5 and algebraic computations
 as in [11, 17], it can be shown that the Lyapunov function V

+(1+ θ̂m) ζ(1,m) x(1,m) +ζ(1,m) satisfies an inequality of the form

× φ(1,m) − (1 + θ̂m )ζ(1,m) x(1,m) φ(1,2,m) , M 
αΥm (|Υm |)
T V̇  − a∗m rm2
|m |2 + 2sm −1
m=1
2rm
0, · · · , 0 (30) 2
+rm |φ(2,3,m) (x(1,m) )||ηm |2 + x2(1,m) ζm
∗
(x(1,m) )
s
m +1

where ζ(1,m) (x(1,m) ) denotes the partial derivative evaluated κz(i,m) αz(i,m) (|z(i,m) |)
+α̃z(1,m) (|z(1,m) |) + 2i− 52
at x(1,m) of ζ(1,m) with respect to its argument. i=2 2rm

738
 ∗
+(θ̂m − θm )q(2,m) (x(1,m) )x2(1,m) ation. IEEE Trans. on Automatic Control, 2001, 46: 1623-
 1629.
+rm [|m |2 + |ηm |2 ] θm
∗
w(2,m) (x(1,m) , θ̂m ) − θ̂m [8] KRISHNAMURTHY P, KHORRAMI F, JIANG Z P. Global
output feedback tracking for nonlinear systems in generalized
1 output-feedback canonical form. IEEE Trans. on Automatic
− w2 (x(1,m) ,θ̂m )(1+x2(1,m) )
β(3,m) (2,m) Control, 2002, 47: 814-819.
 [9] KRISHNAMURTHY P, KHORRAMI F. Decentralized con-
−β(4,m) (34) trol and disturbance attenuation for large-scale nonlinear sys-
tems in generalized output-feedback canonical form. Auto-
where a∗m , β(3,m) , and β(4,m) are positive constants and ζm∗
, matica, 2003, 39: 1923-1933.
α̃z(1,m) , w(2,m) , and q(2,m) are continuous functions of the [10] KRISHNAMURTHY P, KHORRAMI F. Application of the
dual high-gain scaling technique to decentralized control and
indicated arguments. The stability properties of the closed-
disturbance attenuation. in Proc. of the IEEE Conference on
loop system can be inferred from (34) in a manner analogous Decision and Control, (San Diego, CA), Dec. 2006.
to [11, 17]. The central observation is that there are two pos- [11] KRISHNAMURTHY P, KHORRAMI F. Dual high-gain-
sible cases for each subsystem: based adaptive output-feedback control for a class of non-
Case A1 : supt∈[0,tf ) θ̂m (t)  θ m linear systems. International Journal of Adaptive Control and
Case A2 : supt∈[0,tf ) θ̂m (t) > θ m . Signal Processing, 2008, 22: 23-42.
[12] KHALIL H K, SABERI A. Adaptive stabilization of a class
Under Case A1 , it is proved from the designed dynamics of
of nonlinear systems using high-gain feedback. IEEE Trans.
the adaptation parameter and the Lyapunov analysis above on Automatic Control, 1987, 32: 1031-1035.
that all solution
 trajectories are bounded and that further- [13] TEEL A R, PRALY L. Global stabilizability and observ-
more Vm and Vm are bounded on the maximal interval of ability imply semi-global stabilizability by output feedback,.
existence of solutions (whence it is inferred that solutions Systems and Control Letters, 1994, 22: 313-325.
exist for all time). On the other hand, under Case A2 , it [14] KHALIL H K. Adaptive output feedback control of nonlinear
is proved that supt∈[0,tf ) θ̂m (t) is nevertheless bounded and systems represented by input-output models. IEEE Trans. on
that boundedness and convergence properties of the solution Automatic Control, 1996, 41: 177-188.
trajectories as summarized in Theorem 1 follow. Details are [15] ILCHMANN A. “High-gain adaptive control: an overview.”
omitted here for brevity. in IEE Colloquium on Adaptive Control (Digest No:
1996/139), (London, UK), 1996: 1-4.
REFERENCES [16] KRSTIĆ M, KANELLAKOPOULOS I, KOKOTOVIĆ P V.
[1] JAMSHIDI M. Large-Scale Systems: Modeling and Control. Nonlinear and Adaptive Control Design. New York: Wiley,
New York: North-Holland, 1983. 1995.
[2] ÖZGÜNER Ü. Near optimal control of composite systems: [17] KRISHNAMURTHY P, KHORRAMI F. Dynamic high-gain
the multi-time-scale approach. IEEE Trans. on Automatic scaling: state and output feedback with application to sys-
Control, 1979, 24: 652-655. tems with ISS appended dynamics driven by all states. IEEE
[3] KHALIL H K, SABERI A. Decentralized stabilization of Trans. on Automatic Control, 2004, 49: 2219-2239.
nonlinear interconnected systems using high-gain feedback. [18] KRISHNAMURTHY P, KHORRAMI F. On uniform solv-
IEEE Trans. on Automatic Control, 1982, 27: 265-268. ability of parameter-dependent lyapunov inequalities and ap-
[4] SHI L, SINGH S K. Decentralized adaptive controller design plications to various problems. SIAM Journal on Control and
for large-scale systems with higher order uncertainties. IEEE Optimization, 2006, 45: 1147-1164.
Trans. on Automatic Control, 1992, 37: 1106-1118. [19] KRISHNAMURTHY P, KHORRAMI F, JIANG Z P. Global
[5] JAIN S, KHORRAMI F. Decentralized adaptive control of a output feedback tracking for nonlinear systems in generalized
class of large-scale interconnected nonlinear systems. IEEE output-feedback canonical form. IEEE Trans. on Automatic
Trans. on Automatic Control, 1997, 42: 136-154. Control, 2002, 47: 814-819.
[6] JAIN S, KHORRAMI F. Decentralized adaptive output feed- [20] KRISHNAMURTHY P, KHORRAMI F, CHANDRA R S.
back design for large-scale nonlinear systems. IEEE Trans. Global high-gain-based observer and backstepping controller
on Automatic Control, 1997, 42: 729-735. for generalized output-feedback canonical form. IEEE Trans.
[7] JIANG Z P, REPPERGER D,HILL D J. Decentralized non- on Automatic Control, 2003, 48: 2277-2284.
linear output-feedback stabilization with disturbance attenu-

739