2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control

Mexico City, Mexico. September 26-28, 2012

On the Practical Stability for a class of Switched

System
C. Perez1a , A. Poznyak1 , V. Azhmyakov1
1
Department of Automatic Control. CINVESTAV
Av. Instituto Politécnico Nacional 2508. 07300 México D.F., México

condition. We are interested in creating of effective al-

Abstract— This paper addresses a problem of robust gorithms that constitute an adequate numeric approach
control for a class of switched systems with time- to robust stabilization based on control schemes of non-
delays. We deal with controllable nonlinear models that
satisfy a quasi-Lipschitz condition and include bounded switched systems from [17]. The presence of perturbations
perturbations. The stabilization issue is carried out in the state space motivates the application of an extended
by two linear feedback controls: the static and the version of invariant ellipsoid method that establishes the
switching one. We also apply the extended version robust stabilization issue as an optimization problem.
of the invariant ellipsoid method in combination with We restrict our attention here to a class of linear feed-
the Lyapunov-Krasovskii stability approach. Sufficient
conditions of practical stability are derived in terms back control laws. Control strategies for switched systems
of Linear Matrix Inequalities (LMIs). The result is can be conceptually addressed in different forms. One
illustrated by a numerical example. control strategy considers the switching signal as a design
variable. Meanwhile, another control problem is concerned
Keywords: LKF, Ellipsoid Method, Descriptor Method
to design a control law to stabilize the system for a class
of switching signals. In this case, the switching signal is a
I. Introduction prescribed function, so then we just have to deal with the
In many modern engineering applications such as in- design of a conventional feedback control law. In this paper
dustrial electronics, aircraft flight control, manufacturing we consider a static and a switching state-feedback control
systems, automotive control and chemical process are laws. Our motivation for these two control strategies is to
considered complex systems that exhibit various dynamic design feedback stable systems whether the information of
behaviours. Let us mention for example, logic controller the switching times is available or not. Rate of convergence
and human-machine interactions. Such systems are well of the trajectories achieved by feedback controls are of
represented in real-world engineering applications and are exponential nature.
called switched systems. By a switched system, we mean The outline of the paper is as follows. Section 2 contains
a dynamical system that contains a set of continuous- the problem formulation and some concepts related to
time subsystems and a rule that determine the switchings set-stability theory and ellipsoid method. In section 3,
between them. ([14], [15], [16]) our main contribution is presented. Section 4 contains
Despite of increasing interest in switched system theory an illustrative example where the outcomes are applied.
and the rapidly developing of a large amounts of results Section 5 summarizes the paper. The paper also contains
is considered yet an emergent research area (see [5], [14], an Appendix.
[15] and the reference therein). In this paper we study
II. Problem formulation and some basic
the problem of robust stabilization for a class of switched
concepts
systems.
We focus our attention to the problem of rejection of The class of switched system we study is given by the
bounded disturbances. One of the main tools for generat- following nonlinear delayed model
ing practically stable trajectories is the invariant ellipsoid

ẋ(t) = fq(t) x(t), x(t − hq(t) ) + Bq(t) u(t) + ν(t), ∀t ≥ 0,
method [9], [18]. We also refer to [12], [17] for some basic x(t) = φ(t), ∀t ∈ [−hmax , 0], (1)
theoretical details and further references.
Our paper deals with controllable nonlinear dynamic where x(t) ∈ Rn , t ∈ R+ , is the state vector and u(t) ∈ Rm
models with time-delays and bounded uncertainties. Some is the control input. Denote P = {1, 2, ..., M } a index
results about practical stability for time-delay systems are set for a finite M ∈ N. Let q : R+ 7→ P be a piecewise
derived in [21], but unlike this previous work, now we con- constant function. Hereafter, q is called the switching-
sider non-linearities that satisfy so called quasi-Lipschitz signal. fi : Rn × Rn 7→ Rn are non-linear functions,
Bi ∈ Rn×m are given system matrices, hq(t) ∈ {h1 , ..., hM }
a Correspondence to: [email protected]. are positive constant time-delays and hmax := maxi hi .

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 49
978-1-4673-2169-3/12/$31.00 ©2012 IEEE
2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control
Mexico City, Mexico. September 26-28, 2012

Moreover, q(t) = i means that subsystem (fi , Bi ) is active. Let us now introduce some concepts of the set-stability
By ν(t) ∈ Rn we denote a bounded perturbation vector related with attractive sets for time-delay systems [4], [3],
associated with x(t), and φ(t) is a continuous function with [6], [10].
kφ(·)k := supθ∈[−hmax ,0] kφ(θ)k. Transitions between the Definition 2.1: A set D ⊂ Rn is called a positive
given subsystems occurs at “switching times” τr , r ∈ N, invariant set of system (6), if for any initial function
where function q(·) has discontinuities, unlike the time- φ(t) ∈ D for t ∈ [−hmax , 0] implies x(t) ∈ D for each
intervals [τr−1 , τr ) where it remains constant. Actually, in t ≥ 0.
this paper we deal with a specific case of two subsystems,
i.e. P = {1, 2}. Results derived are valid only for this Definition 2.2: [11] An attractive ellipsoid E (0, Q) for
specific case, but note that the proposed method can be (6) is called the set
easily extended for the case M > 2. In this situation the {x ∈ Rn |x⊺ Qx ≤ 1, Q = Q⊺ > 0} (8)
switching function is defined as
where Q is referred as the ellipsoid matrix, if this set is
q(t) = q(0)χI (t) + q̄ (1 − χI (t)) (2) asymptomatically Lyapunov stable.
where q̄ = P \ {q(0)} for a given q(0). Here χI (·) is the A system of the type (1) or (6) is associated with a
characteristic function of the time interval I = [τ2r , τ2r+1 ): set U of feasible control functions u(·) that stabilize the
system in a practical sense [13]. Admissible feedback con-

1 t ∈ [τ2r , τ2r+1 ), r = 0, 1, 2.... trol strategies are chosen from the class of linear retarded
χI (t) := (3) dynamic controllers
0 otherwise
From (2) it follows that the initial conditions are in fact w1 (x(t), hq(t) ) = K1 x(t) + K2 x(t − hq(t) ), (9)
composed by a pair (φ(t), q(0)). Consider τ0 = 0. w2 (x(t), q(t), hq(t) ) = K1,q(t) x(t)+K2,q(t) x(t−hq(t) ). (10)
We also introduce the following technical assumptions:
2
• kν(t)kK = ν (t)Kν ν(t) ≤ 1;
⊺ The control design we apply to (1), implemented as
ν
• the state x(t) and x(t−hi ) are completely measurable; u(t) = w1 (x(t), hq(t) ) or u(t) = w2 (x(t), q(t), hq(t) ), is
• fi (t, x, y) for i ∈ P, satisfies a quasi-Lipschitz condi-
determined by a selection of the suitable matrices
tion Θ1 := K1 K2 Rm×2n

(11)
fi (x, y) − Ai x − Ai y 2 ≤ δ0i

K11 K21 2m×2n
Qfi
(4) Θ2 := R (12)
2 2 K12 K22
+ δ1i kxkQδ + δ2i kykQδ
1i 2i
Hence Θ1 and Θ2 are referred as the dynamic controller
• the pairs (Ai , Bi ) and (Ai , Bi ) are controllable; matrices related to (9) and (10), respectively. The corre-
• min|[τr , τr+1 )| > hmax sponding closed-loop realizations Σk ,k = {1, 2}, of (6)
r
Here k · kKν , k · kQfi , k · kQδ1i and k · kQδ2i also are under (9) and (10) are defined as
called as weighted Euclidean norms, where Kν , Qfi , Qδ1i ẋ(t) =[G11 x(t) + G21 x(t − h1 )]χI (t)


and Qδ2i are given n × n-dimensional positive definite

+ [G12 x(t) + G22 x(t − h2 )](1 − χI (t))


symmetric matrices. δ0i > 0, δ1i ≥ 0 and δ2i ≥ 0 are Σk (13)
some given scalars. Also, we exclude the Zeno effect in

 + ξ(t), ∀t > 0

x(t) =φ(t), ∀t ∈ [−hmax , 0],

the switching signal, i.e., infinite switchings in finite time
are not considered. Switched systems with Zeno effect are where k = 1 corresponds to w1 ,G1i := Ai + Bi K1 , G2i :=
addressed in [1]. Ai + Bi K2 ; and k = 2 corresponds to w2 , G1i := Ai +
We define the auxiliary function Bi K1i , G2i := Ai + Bi K2i , i = {1, 2}.
Now consider E (0, Q) associated with the closed-loop

ξ(t) :=fq(t) x(t), x(t − hq(t) ) − Aq(t) x(t)
(5) system (13). Our aim is to generate control strategies
− Aq(t) x(t − hq(t) ) + υ(t)
of the type (9)-(10), and determine suitable dynamic
and rewrite the system (1) as follows: controller matrices Θk in order to guarantee that E (0, Q)
is an attractive ellipsoid for Σk . From the point of view of
ẋ(t) =Aq(t) x(t) + Aq(t) x(t − hq(t) ) + Bq(t) u(t)
control applications, we are interested in constructing an
+ ξ(t), ∀t ≥ 0 (6) ellipsoid of “minimal” size.
x(t) =φ(t), ∀t ∈ [−hmax , 0]. The minimality property can be formalized by the fol-
lowing problem
Using the quasi-Lipschitz condition we obtain
minimize tr(Q−1
k )
kξ(t)k2Qf ≤ δ̂i + δ1i kx(t)k2Qδ + δ2i kx(t − hi )k2Qδ , (7)
i 1i 2i subject to (14)
where δ̂i = δ0i + 1, i ∈ P = {1, 2}. Qk ⊺ = Qk > 0, Qk ∈ Γ1 (x(·)), Θk ∈ Γ2 (x(·))

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 50
978-1-4673-2169-3/12/$31.00 ©2012 IEEE
2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control
Mexico City, Mexico. September 26-28, 2012

where Γ1 (x(·)) is a set of symmetric and positive definite Proof: The derivative of the first term can be com-
n-matrices that guarantee the invariance and attractivity puted easily as 2x⊺ (t)P ẋ(t). The derivative of the integral
property of the correspond ellipsoid E (0, Q). Γ2 (x(·)) is a terms can be computed using the well-known differenti-
subset of the space of (km × 2n)−dimensional matrices. ating formula for integral with variable limits (see [20]).
This subset describes the admissible dynamics controller
matrices.
n o We also assume that (14) has an optimal solution Our aim now is to verify the conditions from Lemma
Q̂, Θ̂k . The main problem is now to give a constructive 3.1 for the combined function V(t) = V (x(t), ẋ(t)). We
characterization of the above set restrictions. are now ready to formulate our main result.
III. The Attractive Ellipsoid Method
Theorem 3.3: Let a ∈ R+ and
The conventional invariant ellipsoid method use a
Lyapunov-based analysis of attractive sets ([18]). The
extended ellipsoid method we propose is based on the η(t) = x(t), ẋ(t), x(t − h1 ), x(t − h2 ),
following abstract result (see [19]). #⊺
Z t Z t−h1
Lemma 3.1: Let a continuously differentiable function ẋ(s)ds, ẋ(s)ds, ξ(t) ,
V : R+ 7→ R+ satisfies the following differential inequality t−h1 t−h2

V̇(t) ≤ −αV + β, (15)

an extended state vector. Then, for the derivative of
where α > 0 and β > 0. Then V (x(t), ẋ(t)) the following estimation:

β β
V(t) ≤ + V(0) − exp (−αt) , ∀t ≥ 0, (16) dV
α α (x(t), ẋ(t)) ≤ η ⊺ (t) [χ(t)W1 + (1 − χ(t))W2 ] η(t)
dt (19)
β
and moreover lim V(t) ≤ α. − aV (x(t), ẋ(t)) + δ̂1 χ(t) + (1 − χ(t))δ̂2
t→∞
This lemma can be proved using an extension of well-
known Bellman-Gronwall inequality [2]. We call function holds.
V(·) from Lemma 3.1 a “storage function”. Here δ̂i are given, W1 (P, S1 , S2 , R1 , R2 , Θ1 , F ) and
Consider the following Lyapunov-Krasovskii functional W2 (P, S1 , S2 , R1 , R2 , Θ1 , F ) are symmetric matrix func-
(LKF) [8] : tions and F = diag(F1 , F2 , F3 , F3 , F4 , F4 , F5 ) where Fj ,
t−hi−1 j = 1, ..., 5 are some matrices of suitable dimensions.
2
Proof: The second integral term in (18) can be
X Z
V (x(t), ẋ(t)) = x⊺ (t)P x(t) + ea(s−t) x⊺ (s)Si x(s) ds
estimated using de Jensen Inequality as:
i=1
t−hi

2
X
−hi−1
Z Zt t−hi−1
Z t−hi−1
Z
+ (hi − hi−1 ) ea(s−t) ẋ⊺ (s)Ri ẋ(s) dsdθ (17) e a(s−t) ⊺
ẋ (s)Ri ẋ(s)ds ≥ e −ahi
ẋ⊺ (s)Ri ẋ(s)ds
i=1 −hi t+θ
t−hi t−hi
where P , Si , Ri are positive definite symmetric n-
   
t−hi−1 t−hi−1
Z Z
matrices, h2 > h1 > h0 = 0 are the constant delays and ≥ e−ahi 

ẋ⊺ (s)ds Ri 
 
ẋ(s)ds .

a > 0. This LKF is associated with the trajectory x(t) and t−hi t−hi
its derivative ẋ(t). Let us now compute the derivative of
this LKF.
Lemma 3.2: The derivative dV (x(t), ẋ(t))/dt along the Adding and subtracting the term ax⊺ (t)P x(t) we can
trajectories (x(·), ˙·(t)) is given by estimate the derivative of the LKF as follows
2
dV
⊺
X
2x (t)P ẋ(t) + e−ahi−1 x⊺ (t − hi−1 )Si x(t − hi−1 ) (x(t), ẋ(t)) ≤ −aV (x(t), ẋ(t)) + η ⊺ (t)Wη(t)
i=1 dt
− e−ahi x⊺ (t − hi )Si x(t − hi )
Z t−hi−1 where
−a ea(s−t) x⊺ (s)Si x(s)ds
t−hi
2
X (18) W = diag W1 , e−ah1 (S2 − S1 ), e−ah2 S2 , −h1 e−ah1 R1 ,
+ (hi − hi−1 ) (hi − hi−1 ) ẋ⊺ (t)Ri ẋ(t)
i=1
−(h2 − h1 )e−ah2 R2 , 0̄ ,
Z t−hi−1
aP + S1 P
− ea(s−t) ẋ⊺ (s)Ri ẋ(s)ds W1 = ,
t−hi
P h21 R1 + (h2 − h1 )2 R2
Z −hi−1 Z t
−a ea(s−t) ẋ⊺ (s)Ri ẋ(s) dsdθ
−hi t+θ and 0̄ is a zero n-dimensional matrix.

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 51
978-1-4673-2169-3/12/$31.00 ©2012 IEEE
2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control
Mexico City, Mexico. September 26-28, 2012

Following the “descriptor method” for time-delay sys- has optimal solution Υ̂ := Ŷ , Ψ̂, Q̂r , Ĥ1i , Ĥ2i , X̂1 , X̂2 ,
tems ([7]) we add the term D(t) ρˆi , â . Then for α = â and β = max(δ̂1 , δ̂2 ), the ellipsoid
h
D(t) := 2 F1 x(t) + F2 ẋ(t) + F3 x(t − h1 )χ + x(t − h2 )(1 − χ) E(0, Q1 ) determined by
Z t Z t−h1 ⊺ α −1
+F4 χ ẋ(s)ds + (1 − χ) ẋ(s)ds + F5 ξ(t) × Q1 = Ŷ (22)
t−h1 t−h2
β

G11 x(t) + G21 x(t − h1 ) χ(t) + G12 x(t) + G22 x(t − h2 ) × is an attractive ellipsoid for the closed-loop system (13).
Moreover, corresponding dynamic controller matrices are
(1 − χ(t)) + ξ(t) − ẋ(t) , given by h i
Θ1 = X̂1 Ŷ1−1 X̂2 Ŷ3−1 . (23)
O(t) := ξ ⊺ (t) [χ(t)Qf1 + (1 − χ(t))Qf2 ] ξ(t).
Proof: Introduce the auxiliary matrices:
Using the quasi-Lipschitz condition (7), we next get the
final estimation Ψ :=diag (Y1 , Y2 , Y3 , Y3 , Y4 , Y4 , Y5 ) = F −1 ,
dV X1 = K1 Y1 , X2 = K2 Y3 , H1i = Si−1 , H2i = Ri−1
(x(t), ẋ(t)) ≤ η ⊺ (t)Wη(t) − aV (x(t), ẋ(t)) + D(t) ± O(t)
dt
≤ η ⊺ (t) [χ(t)W1 + (1 − χ(t)W2 )] η(t) − aV (x(t), ẋ(t)) The matrix constrains Wi (P, S1 , S2 , R1 , R2 , Θ1 , F ) ≤ 0 in

(20) can be rewritten in the form of LMIs
+ δ̂1 χ(t) + (1 − χ(t))δ̂2 .
Theorem is proved. W i (Y, H11 , H12 , H21 , H22 , X1 , X2 , Ψ) :=
We have a immediate result from (15)-(19). Ψ [Wi (P, S1 , S2 , R1 , R2 , Θ1 , F )] Ψ⊺ ≤ 0
Corollary 3.4: Let W1 (P, S1 , S2 , R1 , R2 , Θ1 , F ) and
W2 (P, S1 , S2 , R1 , R2 , Θ1 , F ) be negative semidefinite ma- W i are defined in the Appendix.
trices. Then the composed function V(t) = V (x(t), ẋ(t)) is Relaxation leads to the auxiliary optimization prob-
a storage function from Lemma 3.1, and lem set up in Theorem 3.5 and provide a good frame-
work for numerical solution of the problem of feedback-
dV control synthesis. Following the method developed above,
(x(t), ẋ(t)) ≤ −αV (x(t), ẋ(t)) + β
dt we derive practical stability conditions under u(t) :=
Proof: We put α = a and β := max(δ̂1 , δ̂2 ). We w2 (x(t), q(t), hq(t) ). This strategy can get a better per-
formance in the case of specific control design. For this
add and subtract β in (19), and select some appropriate purpose now consider the following LKFs
matrices P, S, R1 , R2 , Θ1 , F such that W1 and W2 be
negative semidefinite matrices. Zt
Vi (x(t), ẋ(t)) = x⊺ (t)Pi x(t) + eai (s−t) x⊺ (s)Si x(s)ds
Since the invariance and attractive properties of the
t−hi
E(0, Q1 ) are guaranteed by the condition (16), we get a (24)
constructive interpretation of (14) Z0 Zt
ai (s−t) ⊺
β + hi e ẋ (s)Ri ẋ(s)dsdθ
minimize tr P −1 −hi t+θ
α
subject to (20) where Pi , Si , Ri for i = {1, 2} are positive definite
P, Si , Ri > 0, P = P ⊺ , Si = Si ⊺ , Ri = Ri ⊺ , i = 1, 2 symmetric matrices, hi > 0 are constant delays and ai > 0.
Let
Wi (P, S1 , S2 , R1 , R2 , Θ1 , F ) ≤ 0, Θ1 ∈ Γ2 (x(·)). 
Zt
Nevertheless, due to nonlinearities in the restrictions, Di (t) := 2 F1i x(t) + F2i ẋ(t) + F3i x(t − hi ) + F4i ẋ(s)ds
the selection of variables (P, S1 , S2 , R1 , R2 , Θ1 ) and free
t−hi
parameters F is a non trivial computational task. Note i⊺ h i
that nature of this optimization problem is heavily non- +F5i ξ(t) × G1i x(t) + G2i x(t − hi ) + ξ(t) − ẋ(t) ,
linear (the nonlinear objective functional and bilinear Oi (t) := ξ ⊺ (t)Qfi ξ(t) − ξ ⊺ (t)Qfi ξ(t).
constraints). Motivated from that facts, let us now refor-
mulate the problem (20) in a relaxed form that contains In order to summarize the procedure we only present
the Linear Matrix Inequalities (LMI). Let Y := P −1 . the final result associated with the second controller.
Theorem 3.5: Assume that the following optimization
problem Theorem 3.6: Assume that the following optimization
problem
minimize tr Y
subject to minimize tr Yi
W 1 < 0, W 2 < 0, (21) subject to (25)
Mπ > 0, Qr > 0, π = 1, 11, r = 1, 12, i = 1, 2 Ωi < 0, Mπi > 0, Qπi > 0; π = 1, 7, i = 1, 2
Y > 0, Y5 > 0, ρi > 0, a > 0, H1i > 0, H2i > 0 Yi > 0, Y5i > 0, ρi > 0, ai > 0, H1i > 0, H2i > 0

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 52
978-1-4673-2169-3/12/$31.00 ©2012 IEEE
2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control
Mexico City, Mexico. September 26-28, 2012

has optimal
solution Υ̂i := Ŷi , Ψ̂i , Q̂πi , Ĥ1i , Ĥ2i , X̂1i , X̂2i
ρˆi , âi . Let αi = âi and β = δi . The ellipsoid E(0, Q2i ) 0.8

determined by
αi 0.6
Q2i = Ŷi−1 (26)
βi 0.4

is an attractive ellipsoid for the closed-loop system Σ2 (13). x2

0.2
The corresponding dynamic controller matrices are given
0
by
−1 −1

X̂11 Ŷ11 X̂21 Ŷ31 −0.2
Θ2 = −1 −1 (27)
X̂12 Ŷ12 X̂22 Ŷ32 −0.4
.
−0.6
−0.6 −0.4 −0.2 0 0.2 0.4
Theorem 3.6 can be proved by the same way as the x
1

previously obtained result.

1200
−3
x1
IV. Numerical Aspects 1000
5
x 10
x2

Now we apply the technique developed previously to 800

0
a computational example. Optimization problem deter- 600
mined in Theorems 3.5 and 3.6 is solved. We apply some x(t)
400 −5
algorithms based on the standard MATLAB packages 15 20 25 30 35 40

(Sedumi/Yalmip, Ellipsoidal Toolbox, Simulink). The nu- 200

merical procedure is illustrated on the following example. 0

−200

sin(x1 (t)) − 4x2 (t) + x1 (t − h1)
f1 = −400
x2 (t) − x2 (t − h1 ) 0 5 10 15 20 25 30 35 40 45 50
time (s)

x2 (t) + 0.5x1 (t − h2 )
f2 =
−x1 (t) − 2sign(x2 (t)) + 0.2x1 (t − h1 ) + 0.1x2 (t − h2 ) Fig. 1. State variables under w1

1 −4 1 0 2
A1 = A1 = B1 = h1 = 0.3
0 1 0 −1 −1

0 1

0.5 0

0.4 stability properties. The implementable computational al-
A2 = A2 = B2 = h2 = 1
−1 −2 0.2 0.1 0.1 gorithms that generates the corresponding linear feedback
controllers are presented and their effectiveness is demon-
The computed gain matrices are
strated by a numerical example. The results obtained can
K1 = [−2.3592 1.5588] K2 = [−2.6098 1.0769] be extended to some classes of output-based feedback
K11 = [−728.0435 − 409.9273] K21 = [−6.060 11.112] control problems.
K12 = [−2922.25 688.33] K22 = [−0.699 1.410]

In this case we used x(0) = [1000 1000]⊺. The obtained References

results are shown in Fig. 1 (first strategy control), and
Fig.2 (second control law). [1] A.D. Ames, Y. Or, “Stability and completion of Zeno equilibria
in Lagrangian hybrid systems”, IEEE Transactions on Auto-
While both strategies are able to stabilize the system matic Control, vol. 56, 1322-1336 pp., 2011.
(6), in the sense of practical stability, there is a better [2] D. Bainov, P. Simeonov, “Integral inequalities and applica-
estimate of the convergence zone for the trajectories with tions”, Mathematics and Its Applications, v. 215, Dordrecht :
Kluwer Academic Publishers, 1992.
the control law (10). Ellipsoids of figure 2 are inside of
[3] F. Blanchini, S. Miani, “Set-Theoretic Methods in Control”,
ellipsoid of figure 1. Birkhauser, Boston, 2008.
Conservatism can arise in the estimation of exponential [4] C. Corduneanu, O. Ignatyev, “Stability of invariant sets of
convergence bounds derived from (16). functional differential equations with delay”, Nonlinear Funct.
Anal. & Appl., Vol. 10, No. 1, pp. 11–24, 2005.
[5] R.A. Decarlo, M.S. Branicky, S. Pettersson and B. Lennartson,
V. Concluding Remarks “Perspectives and results on the stability and stabilizability of
hybrid systems”, in Proc. IEEE: Special Issue Hybrid Systems,
In this paper, we study a robust control design for a vol. 88, pp. 1069–1082, 2000.
class of switched systems with time-delays in the presence [6] Daoyi Xu, Zhichun Yang, “Attracting and invariant sets for a
of bounded uncertainties. The practically stabilizing tech- class of impulsive functional differential equations”, Journal of
Mathematical Analysis and Applications, Volume 329, Issue 2,
nique is based on an specific extension of the invariant pp. 1036–1044, 2007.
ellipsoid method and incorporates a LMI-constrained op- [7] E. Fridman, “Descriptor discretized Lyapunov functional
timization problem. We construct two control strategies: method: analysis and design”, IEEE Transactions on Automatic
a static one and a switching one. Both strategies lead to Control, vol. 51, no. 5, pp. 890–897, May, 2006.
[8] E. Fridman, M. Dambrine, “Control under quantization, satu-
characterizations of the corresponding invariant ellipsoids ration and delay: An LMI approach”, Automatica, vol. 45, no.
with minimal size that can be interpreted as practical 10, pp. 2258–2264, October, 2009.

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 53
978-1-4673-2169-3/12/$31.00 ©2012 IEEE
2012 9th International Conference on Electrical Engineering, Computing Science and Automatic Control
Mexico City, Mexico. September 26-28, 2012

0.1 [20] A.S. Poznyak, “Advanced Mathematical Tools for Automatic

Control Engineers: Deterministic Technique”, Elsevier, Amster-
0.08
dan, 2008.
0.06 [21] R. Villafuerte, S. Mondie and A.S. Poznyak, “Practical stability
0.04 of time delay systems:LMIs approach”, in Proceedings of the
0.02 47th IEEE Conference on Decision and Control, pp. 4807-4812,
x2 Cancun, Mexico, 2008.
0

−0.02
Appendix
−0.04
i
−0.06
Matrices W i (Y, H11 , H12 , H21 , H22 , X1 , X2 , Ψ) = [ωjk ]
−0.08 are symmetric. Here i = 1, 2 and {j, k} = 1, 7.
−0.1
1
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05 ω11 = aQ1 + Q2 + Q9 + G11 Y1 + Y1⊺ G⊺ 11
x 1
1 ω22 = h21 Q5 + (h2 − h1 )2 Q6 − Y1 − Y1⊺
1
ω33 = e−ah1 (Q3 − Q4 ) + G21 Y3 + Y3 G⊺ 21 + Q11
1 1
2500 ω44 = −e−ah2 Q3 ω55 = −h1 e−ah1 Q7
−3
x 10 x1 1 1
ω66 = −(h2 − h1 )e−ah2 Q8 ω77 = (2 − ρ1 )Y5
2000 x2 1 1
2 ω12 = Q1 − Y1 + Y1⊺ G⊺ 11 ω ⊺ ⊺
13 = G21 Y3 + Y1 G11
1 1
1500
ωj4 = 0, j = {1, 3, 5, 7} ω15 = Y1⊺ G⊺ 11
0 1 1
ωj6 = 0, j = {1, 5, 7} ω17 = Y5 + Y1⊺ G⊺ 11
1 ⊺ 1 ⊺
1000 −2 ω23 = G21 Y3 − Y1 ω25 = −Y1
1 ⊺ 1 ⊺ ⊺
x(t) 10 20 30 40 ω27 = Y5 − Y1 ω35 = Y3 G21
500 1 1
ω37 = Y5 + Y3⊺ G⊺ 21 ω57 = Y5
0
G11 Y1 := A1 Y1 + B1 X1 G21 Y3 := Ad1 Y3 + B1 X2
2
ω11 = aQ1 + Q2 + Q10 + G12 Y1 + Y1⊺ G⊺ 12
2
−500 ω22 = h21 Q5 + (h2 − h1 )2 Q6 − Y1 − Y1⊺
2 ⊺ ⊺
ω44 = −e −ah2
Q3 + G22 Y3 + Y3 G22 + Q12
−1000 2 2
ω33 = e−ah1 (Q3 − Q4 ) ω55 = −h1 e−ah1 Q7
1 1
−1500
ω66 = −(h2 − h1 )e−ah2 Q8 ω77 = (2 − ρ2 )Y5
2 2
0 5 10 15 20 25 30 35 40 45 50 ω12 = Q1 − Y1 + Y1⊺ G⊺ 12 ωj3 = 0, j = {1, 2, 4, 7}
time (s) 2 2
ω14 = G22 Y3 + Y1⊺ G⊺ 12 ωj5 = 0, j = {1, 4, 6, 7}
2 2
ω16 = Y1⊺ G⊺ 12 ω17 = Y5 + Y1⊺ G⊺ 12
Fig. 2. State variables under w2 2
ω24 = G22 Y3 − Y1⊺ ω262
= −Y1⊺
2 ⊺ 2 ⊺ ⊺
ω27 = Y5 − Y1 ω46 = Y3 G22
2 2
ω47 = Y5 + Y5⊺ G⊺ 22 ω67 = Y5
G12 Y1 :=
A 2 Y 1 + B 2 X 1 G 22 Y
3 := Ad2 Y3 B2 X2
[9] S. Gonzalez-Garcia, A. Polyakov, A. Poznyak, “Output linear Q1 Y1 ⊺ Q2 Y1 ⊺ Q3 Y3 ⊺
M1 = M2 = M3 =
controller for a class of nonlinear systems using the invari- Y
1 Y Y 1 H 11 Y3 H12
ant ellipsoid technique”, American Control Conference 2009 , Q4 Y3 ⊺ Q5 Y1 ⊺ Q6 Y1 ⊺
M4 = M5 = M6 =
pp.1160–1165, 10-12 June 2009. Y3 H11 Y1 H21 Y1 H22
Q7 Y4 ⊺ Q8 Y4 ⊺
[10] W. Haddad, V. Chellaboina, “Nonlinear Dynamics Systems and M7 = M8 =
Control”, Princeton University Press, Princeton, 2008. Y4 H21 Y4 H22
Q9 + Q10 Y1 ⊺
[11] H.K. Khalil, “Nonlinear Systems”, 3rd edition, Prentice Hall, M9 = −1
Y 1 δ 11 Q δ11
+ δ12 Qδ12
Upper Saddle River, New Jersey, 2002.
Q11 + Q12 Y3 ⊺

[12] A.B. Khurzhanski, P. Varaiya, “Ellipsoidal techniques for reach- M10 = −1
Y 3 δ Q
21 δ21 + δ
22 δ22Q
ability under state constraints”, SIAM Journal on Control and
Qf1 + Qf2 In
Optimization, 45, pp. 1369–1394, 2006. M11 = In 1
ρ +ρ Y5
1 2
[13] V. Lakshmikantham, S. Leela, A.A. Martynyku, “Practical sta-
i
bility of nonlinear systems”, New York: World Scientific Pub- Matrices Ωi (Yi , H1i , H2i , X1i , X2i , Ψi ) = [ωjk ] are sym-
lishing Co., 1990.
[14] D. Liberzon, “Switching in systems and control”, Systems &
metric. Here i = 1, 2 and {j, k} = 1, 5. The components
i
control. Foundations & applications, Boston, MA: Birkhauser, ωjk are given by:
2003. i ⊺ ⊺
ω11 = ai Q1i + Q2i + G1i Y1i + Y1i G1i + Q6i
[15] Lin H. and Antsaklis P. J., “Stability and Stabilizability of i ⊺ ⊺
ω33 = −e−ahi Q4i + G2i Y3i + Y3i G2i + Q7i
Switched Linear Systems: A Survey of Recent Results”, IEEE i ⊺ ⊺ i ⊺ ⊺
ω12 = Q1i − Y1i + Y1i G1i ω13 = G2i Y3i + Y1i G1i
Transactions on Automatic Control, vol. 54, n0. 2, February, i ⊺ ⊺ i
ω14 = Y1i G1i ω15 = Y5i + Y1⊺ G⊺
2009. i 2
ω22 = hi Q3i − Y1i − Y1i ⊺ i
ω23 = G2i Y3i − Y1i
11
⊺
[16] J. Lunze, F. Lamnabhi-Lagarrigue, “Handbook of Hybrid Sys- i
ω24 = −Y1i ⊺ i
ω25 = Y5i − Y1i⊺
tems Control: Theory, tools and applications”, New York: Cam- i
ω34 = Y3i⊺ ⊺
G2i i
ω35 = Y5i + Y3⊺ G⊺21
bridge University Press, 2009. i
ω44 = −hi e−ahi Q5i i
ω45 = Y5i
[17] M. Mera, A. Poznyak, V. Azhmyakov, “On the robust control i
ω55 = (2 − ρi )Y5i
design for a class of continuous-time dynamical systems with a G1i Y1i =
Ai Y1i +⊺ B
i X1i Y3i = Ai⊺Y3i + Bi X2i
G2i
sample-data output”, in Proceedings of the 18th IFAC World Q1i Y1i Q2i Y1i Q3i Y1i⊺

Congress, Milano, Italy, 2011. M1i = M2i = M3i =

Y Yi Y1i H1i Y1i H2i
1i
[18] B.T. Polyak, A.V. Nazin, M.V. Topunov, S.A. Nazin, “Rejection Qfi In
⊺ ⊺
Q4i Y3i Q5i Y4i
M4i = M5i = M6i = I 1
of Bounded Disturbances via Invariant Ellipsoids Technique”,  3i
Y H1i Y4i H2i n ρ 1
Y 5i
Decision and Control, 2006 45th IEEE Conference on , pp.1429– ⊺ ⊺ 
Q6i + Q7i Y1i Y3i
1
1434, 13–15 Dec. 2006. M7i = 
 Y1i δ Qδ
1i
−1
1i
0 

[19] A.S. Poznyak, “Deterministic output noise effects in sliding Y3i 0 1
δ2i Q−1
δ
2i
moded observation”. In book “Variable Structure Systems: from
principles to implementation” (ed. By A.Sabanovic, L.Fridnman
and S.Spurgeon), Chapter 3, pp.45–80, IEE-Press, London,
2004.

IEEE Catalog Number: CFP12827-CDR

ISBN: 978-1-4673-2168-6 54
978-1-4673-2169-3/12/$31.00 ©2012 IEEE