Output Feedback Sampled-Data Control of Nonlinear Systems in ...
Output Feedback Sampled-Data Control of Nonlinear Systems in ...
Output Feedback Sampled-Data Control of Nonlinear Systems in ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Output</strong> <strong>Feedback</strong> <strong>Sampled</strong>-<strong>Data</strong><br />
<strong>Control</strong> <strong>of</strong> <strong>Nonl<strong>in</strong>ear</strong> <strong>Systems</strong><br />
<strong>in</strong> <strong>Output</strong> <strong>Feedback</strong> Form<br />
Buzhou Wu, Zhengtao D<strong>in</strong>g<br />
<strong>Control</strong> <strong>Systems</strong> Centre, School <strong>of</strong> Electrical and Electronic Eng.,<br />
University <strong>of</strong> Manchester, P.O. Box 88, Manchester, M60 1QD,<br />
United K<strong>in</strong>gdom (email: Zhengtao.D<strong>in</strong>g@manchester.ac.uk).<br />
Abstract: This paper deals with sampled-data control <strong>of</strong> nonl<strong>in</strong>ear systems <strong>in</strong> the output<br />
feedback form. A sampled-data control strategy is proposed based on the exist<strong>in</strong>g control design<br />
<strong>in</strong> the cont<strong>in</strong>uous-time doma<strong>in</strong> via output feedback. The proposed control uses the sampled<br />
output and a discrete-time implementation <strong>of</strong> filters <strong>in</strong>volved. The overall stability <strong>of</strong> the system<br />
under the proposed control has been analyzed, and the semi-global asymptotic stability for the<br />
system with relative degree one and two is established by keep<strong>in</strong>g the sampl<strong>in</strong>g <strong>in</strong>terval with a<br />
specified range.<br />
Keywords: sampled-data control, nonl<strong>in</strong>ear systems, output feedback form<br />
1. INTRODUCTION<br />
Digital computers have been widely used for implementation<br />
<strong>of</strong> control strategies. For l<strong>in</strong>ear systems, computer<br />
implementation can be carried out based on the discretetime<br />
version <strong>of</strong> the control design, as there is always<br />
discrete time version which has the same structure as the<br />
orig<strong>in</strong>al cont<strong>in</strong>uous-time model. <strong>Sampled</strong>-data control <strong>of</strong><br />
l<strong>in</strong>ear systems has been <strong>in</strong>tensively studied <strong>in</strong> literature,<br />
see Chen and Francis [1995] for <strong>in</strong>stance.<br />
However most <strong>of</strong> the eng<strong>in</strong>eer<strong>in</strong>g systems are nonl<strong>in</strong>ear,<br />
and there have been tremendous progresses <strong>in</strong> control<br />
design for nonl<strong>in</strong>ear systems <strong>in</strong> the last decade, and a<br />
number <strong>of</strong> control design methods and control strategies<br />
have been proposed. To implement the nonl<strong>in</strong>ear control<br />
laws for nonl<strong>in</strong>ear systems <strong>in</strong> discrete time requires <strong>in</strong>vestigat<strong>in</strong>g<br />
the performance <strong>of</strong> nonl<strong>in</strong>ear sampled-data systems.<br />
Unfortunately the results on sampled-data control<br />
<strong>of</strong> l<strong>in</strong>ear systems can not be applied directly to nonl<strong>in</strong>ear<br />
sampled-data systems, due to the fact that unlike <strong>in</strong> the<br />
l<strong>in</strong>ear context, an exact, discrete-time model <strong>of</strong> a general<br />
nonl<strong>in</strong>ear system is hard to obta<strong>in</strong> and then not available<br />
for controller design. Furthermore, most <strong>of</strong> the cont<strong>in</strong>uoustime<br />
nonl<strong>in</strong>ear theories are not applicable to the sampleddata<br />
case, s<strong>in</strong>ce they rely on particular structures <strong>of</strong> nonl<strong>in</strong>ear<br />
systems, which are usually destroyed by sampl<strong>in</strong>g,<br />
for example, feedback l<strong>in</strong>earizability [Grizzle and Kokotovic,<br />
1988]. These facts strongly motivate the research<br />
on nonl<strong>in</strong>ear sampled-data control systems, to which a<br />
great deal <strong>of</strong> attention has been drawn recently, see Hou<br />
et al. [1997], Teel et al. [1998], Dabroom and Khalil [2001],<br />
Khalil [2004] and Nesic and Teel [2004].<br />
Results on the problem <strong>of</strong> output feedback sampled-data<br />
control <strong>of</strong> nonl<strong>in</strong>ear systems have been reported <strong>in</strong> Dabroom<br />
and Khalil [2001] and Khalil [2004]. Both <strong>of</strong> Dabroom<br />
and Khalil [2001] and Khalil [2004] employed the emulation<br />
method to design sampled-data controllers for the<br />
same class <strong>of</strong> systems, while cont<strong>in</strong>uous-time controllers<br />
are actually state-feedback controllers us<strong>in</strong>g high-ga<strong>in</strong> observers.<br />
The difference is that <strong>in</strong> Dabroom and Khalil<br />
[2001] the high-ga<strong>in</strong> observer was designed <strong>in</strong> cont<strong>in</strong>uous<br />
time while <strong>in</strong> Khalil [2004], on discrete-time models. The<br />
results <strong>in</strong> Dabroom and Khalil [2001] and Khalil [2004]<br />
showed that, for some class <strong>of</strong> nonl<strong>in</strong>ear systems, the<br />
obta<strong>in</strong>ed sampled-data controller can recover the performance<br />
<strong>of</strong> the cont<strong>in</strong>uous-time state feedback controller,<br />
provided that the sampl<strong>in</strong>g period T → 0.<br />
In this paper, we will concentrate on sampled-data control<br />
<strong>of</strong> nonl<strong>in</strong>ear systems <strong>in</strong> the output feedback form. We<br />
will <strong>in</strong>vestigate the conditions, particularly those on the<br />
sampl<strong>in</strong>g period T, under which the sampled-data controller<br />
will stabilize the close-loop sampled-data system.<br />
The paper makes full use <strong>of</strong> exist<strong>in</strong>g methods for design<strong>in</strong>g<br />
cont<strong>in</strong>uous-time controllers for the nonl<strong>in</strong>ear systems <strong>in</strong><br />
the output feedback form Mar<strong>in</strong>o and Tomei [1995] and<br />
D<strong>in</strong>g [2003].<br />
2. PROBLEM STATEMENT<br />
We consider sampled-data control <strong>of</strong> one class <strong>of</strong> s<strong>in</strong>gle<strong>in</strong>put-s<strong>in</strong>gle-output<br />
nonl<strong>in</strong>ear systems which can be transformed<br />
<strong>in</strong>to the follow<strong>in</strong>g output feedback form<br />
with<br />
ẋ = A c x + φ(y) + bu<br />
y = Cx (1)<br />
⎡<br />
⎤<br />
0 1 0 · · · 0<br />
0 0 1 · · · 0<br />
A c =<br />
.<br />
⎢ . . . ..<br />
. ⎥<br />
⎣ 0 0 0 · · · 1 ⎦ ,<br />
0 0 0 · · · 0<br />
⎡ ⎤<br />
0 ... b = 0<br />
b ρ...<br />
⎢ ⎥<br />
⎣ ⎦<br />
b n
C = [ 1 0 · · · 0 ]<br />
where x ∈ R n is the state vector, u, y ∈ R the <strong>in</strong>put and<br />
the output respectively, b ∈ R n an known vector, while<br />
φ(y) is a smooth nonl<strong>in</strong>ear vector field satisfy<strong>in</strong>g φ(0) = 0.<br />
Assumption 1. The system is <strong>of</strong> m<strong>in</strong>imum phase, i.e., the<br />
polynomial B(s) = ∑ n<br />
i=ρ b is n−i is Hurwitz.<br />
The problem considered here can be described as follows:<br />
For system (1), if there exists a cont<strong>in</strong>uous-time output<br />
feedback controller u c given by<br />
˙ω(t) = ν(y,ω), u c = u c (y,ω) (2)<br />
which renders the orig<strong>in</strong> <strong>of</strong> system (1) asymptotically<br />
stable, then for any given <strong>in</strong>itial set, there exists a constant<br />
T ∗ > 0 such that for all 0 < T < T ∗ , the system will still be<br />
asymptotically stabilized by the sampled-data controller<br />
u d<br />
ω(m) = ν d<br />
(<br />
y(m − 1),ω(m − 1<br />
)<br />
)<br />
u d (t) = u c (y(m),ω(m)), ∀t ∈ [mT,mT + T) (3)<br />
where ν d is some discrete-time implementation <strong>of</strong> the<br />
system ω, T is the sampl<strong>in</strong>g period and ω(n) and y(n)<br />
denote the signals ω and y at the nth sampl<strong>in</strong>g po<strong>in</strong>t<br />
respectively, that is, ω(m) = ω(mT), y(m) = y(mT)<br />
3. CONTINUOUS-TIME CONTROL DESIGN<br />
In this paper we focus our discussion on the cases <strong>of</strong><br />
relative degree one and two.<br />
3.1 State transformation<br />
For system (1) with relative degree ρ = 2, we can <strong>in</strong>troduce<br />
the follow<strong>in</strong>g filter [Mar<strong>in</strong>o and Tomei, 1995]<br />
˙ξ = −λξ + u (4)<br />
where λ > 0 is the design parameter. With the vector ¯d = b<br />
∈ R n , the follow<strong>in</strong>g filtered transformation η = x − ¯dξ can<br />
transform system (1) <strong>in</strong>to<br />
˙η = A c η + φ(y) + dξ, y = Cη (5)<br />
where d = [A c ¯d + λ ¯d]. It can be shown that d1 = b ρ and<br />
n∑<br />
d i s n−i = B(s)(s + λ)<br />
i=1<br />
S<strong>in</strong>ce λ is positive, d is a Hurwitz vector with d 1 = b ρ = 1<br />
(here we assume b ρ = 1 without loss <strong>of</strong> generality). Therefore<br />
with ξ be<strong>in</strong>g the <strong>in</strong>put, system (5) is <strong>of</strong> m<strong>in</strong>imum<br />
phase and relative degree one. To extract the <strong>in</strong>ternal dynamics<br />
<strong>of</strong> (5), <strong>in</strong>troduced is the follow<strong>in</strong>g state transform<br />
z 1 = η 2 − d 2 η 1<br />
.<br />
z n−1 = η n − d n η 1<br />
y = η 1 (6)<br />
where z ∈ R n−1 . In the new coord<strong>in</strong>ates, system (5) can<br />
be written as<br />
ż = Dz + φ z (y)<br />
ẏ = z 1 + φ y (y) + ξ (7)<br />
where D is the companion matrix <strong>of</strong> d and is given by<br />
⎡<br />
⎤<br />
−d 2 1 · · · 0<br />
D = ⎣<br />
.<br />
.<br />
. ..<br />
. .. ⎦<br />
−d n 0 · · · 0<br />
and<br />
⎡<br />
d 3 − d 2 ⎤ ⎡<br />
⎤<br />
2 φ 2 − φ 1 d 2<br />
d 4 − d 3 d 2<br />
φ<br />
φ z (y) = y<br />
⎢<br />
. ⎥<br />
⎣<br />
d n − d n−1 d<br />
⎦ + 3 − φ 1 d 3<br />
⎢ . ⎥<br />
⎣<br />
2 φ n−1 − φ 1 d<br />
⎦<br />
n−1<br />
−d n d 2 φ n − φ 1 d n<br />
φ y (y) = φ 1 + yd 2<br />
where φ i denotes the ith component <strong>of</strong> the vector φ.<br />
F<strong>in</strong>ally the model for control design is the extended system<br />
consist<strong>in</strong>g <strong>of</strong> (7) and (4).<br />
3.2 <strong>Control</strong> for the case ρ = 1<br />
If system (1) is <strong>of</strong> relative degree one, then we are actually<br />
deal<strong>in</strong>g with the follow<strong>in</strong>g system<br />
ż = Dz + φ z (y)<br />
ẏ = z 1 + φ y (y) + u c1 (8)<br />
and the cont<strong>in</strong>uous-time control u c1 is designed as<br />
u c1 = −φ y − k 1 y − y −1 φ T z P 2 φ z (9)<br />
where k is a positive real and P is the symmetric positive<br />
def<strong>in</strong>ite solution <strong>of</strong> the Lyapunov equation<br />
D T P + PD = −(γ 1 + 2)I (10)<br />
with γ 1 be<strong>in</strong>g a positive real.<br />
3.3 <strong>Control</strong> for the case ρ = 2<br />
In the case <strong>of</strong> relative degree ρ = 2, backstepp<strong>in</strong>g technique<br />
will be employed to f<strong>in</strong>d the f<strong>in</strong>al control u c2 from the<br />
desirable value <strong>of</strong> ξ. Indeed, we have the follow<strong>in</strong>g control<br />
u c2 = −y + λˆξ + ∂ˆξ<br />
∂y (φ ∂ˆξ<br />
y + ξ) − ˜ξ(<br />
∂y )2 (11)<br />
where ˆξ = −φ y − k 1 y − y −1 φ T z P 2 φ z , and ˜ξ = ξ − ˆξ, which<br />
is governed by<br />
˙˜ξ = −λξ − ∂ˆξ<br />
∂yẏ + u (12)<br />
It is easy to verify that the stabiliz<strong>in</strong>g function ˆξ is a<br />
function <strong>of</strong> y and ˆξ(0) = 0, while the cont<strong>in</strong>uous-time<br />
control u c2 is a function <strong>of</strong> y and ξ and u c2 (0,0) = 0.<br />
3.4 Stability <strong>of</strong> the cont<strong>in</strong>uous-time system<br />
If ρ = 2, the time derivative <strong>of</strong> the follow<strong>in</strong>g Lyapunov<br />
function<br />
satisfies<br />
V c2 = z T Pz + 1 2 y2 + 1 2 ˜ξ 2 (13)
˙V c2 = −(γ 1 + 2)‖z‖ 2 + 2z T Pφ z<br />
+y(z 1 − k 1 y − y −1 φ T z P 2 φ z + ˜ξ)<br />
(<br />
∂ˆξ<br />
+˜ξ − λ˜ξ − y − ˜ξ<br />
∂y z 1 − ˜ξ<br />
( ∂ˆξ<br />
) ) 2<br />
∂y<br />
≤ −k 1 y 2 − γ 1 ‖z‖ 2 − λ˜ξ 2 − (‖z‖ − ‖Pφ z ‖) 2<br />
− 3 ( 1<br />
4 ‖z‖2 −<br />
2 z ∂ˆξ<br />
) 2<br />
1 + ˜ξ<br />
∂y<br />
≤ −k 1 y 2 − γ 1 ‖z‖ 2 − λ˜ξ 2 (14)<br />
which proves by the theorem 4.10 <strong>in</strong> Khalil [2002] that<br />
(y,z, ˜ξ) = 0 is an exponentially stable equilibrium po<strong>in</strong>t<br />
<strong>of</strong> the extended system <strong>of</strong> (7) and (4), which implies that<br />
(y,z,ξ) = 0 is globally asymptotically stable and so is the<br />
orig<strong>in</strong> (x,ξ) = 0. In case ρ = 1, we take V c1 = z T Pz + 1 2 y2<br />
as the Lyapunov function, and the same conclusion can be<br />
reached.<br />
4. STABILITY ANALYSIS<br />
The follow<strong>in</strong>g lemma is needed for stability analysis <strong>in</strong><br />
sampled-data case.<br />
Lemma 2. Let V : R n → R be a cont<strong>in</strong>uously differentiable,<br />
radically unbounded, positive def<strong>in</strong>ite function.<br />
Def<strong>in</strong>e D := {χ ∈ R n |V (χ) ≤ r} with r > 0. Suppose<br />
˙V ≤ −αV + βV m , ∀t ∈ (mT,(m + 1)T], (15)<br />
hold for all χ(mT) ∈ D, where α, β are any given positive<br />
reals with α > β, T > 0 the fixed sampl<strong>in</strong>g period and<br />
V m := V (χ(mT)). If χ(0) ∈ D, then the follow<strong>in</strong>g holds:<br />
lim χ(t) = 0 (16)<br />
t→∞<br />
Pro<strong>of</strong>. S<strong>in</strong>ce χ(0) ∈ D, then (15) holds for t ∈ (0,T] with<br />
the follow<strong>in</strong>g form<br />
˙V ≤ −αV + βV (χ(0)).<br />
Us<strong>in</strong>g comparison lemma [Khalil, 2002] it is easy to get<br />
from the above that for t ∈ (0,T],<br />
where<br />
V (χ(t)) ≤ e −αt V 0 + 1 − e−αt<br />
βV 0<br />
α<br />
= q(t)V 0 , (17)<br />
q(t) :=<br />
(<br />
e −αt + β )<br />
α (1 − e−αt )<br />
S<strong>in</strong>ce α > β > 0, then q(t) ∈ (0,1), ∀t ∈ (0,T]. Then we<br />
have<br />
V (χ(t)) < V 0 , ∀t ∈ (0,T]. (18)<br />
Particularly, lett<strong>in</strong>g t = T <strong>in</strong> (17) leads to V 1 ≤ q(T)V 0 ,<br />
which means that V 1 ∈ D. Therefore (15) holds for t ∈<br />
(T,2T]. By <strong>in</strong>duction, we have<br />
V (χ(t)) < V m , ∀t ∈ (mT,(m + 1)T]. (19)<br />
which states <strong>in</strong>tersample behaviour <strong>of</strong> the sampled-data<br />
system concerned, and <strong>in</strong> particular<br />
V m+1 ≤ q(T)V m (20)<br />
<strong>in</strong>dicat<strong>in</strong>g that V decreases at two consecutive sampl<strong>in</strong>g<br />
po<strong>in</strong>ts with a fixed ratio. From (20),<br />
V m ≤ q(T)V m−1 ≤ q m (T)V 0 (21)<br />
which implies that lim m→∞ V m = 0. The conclusion then<br />
follows from (19), which completes the pro<strong>of</strong>.<br />
4.1 Stability <strong>of</strong> the sampled-data system with ρ = 1<br />
In this case, the sampled-data system takes the follow<strong>in</strong>g<br />
form<br />
ż = Dz + φ z (y)<br />
ẏ = z 1 + φ y (y) + u d1 (22)<br />
where u d1 is the sampled-data controller and can be simply<br />
implemented by<br />
for all t ∈ [nT,nT + T).<br />
u d1 (t) = u c1 (y(n)) (23)<br />
For the stability <strong>of</strong> the closed-loop sampled-data system,<br />
we have the follow<strong>in</strong>g result.<br />
Theorem 3. For system (22) with the sampled-data controller<br />
u d1 shown <strong>in</strong> (23) and for any given <strong>in</strong>itial set conta<strong>in</strong><strong>in</strong>g<br />
the orig<strong>in</strong>, there exists a constant T 1 > 0 such that,<br />
for all 0 < T < T 1 , the overall system is asymptotically<br />
stable.<br />
Pro<strong>of</strong>. Def<strong>in</strong>e χ := [z,y] T . Let B r := {χ ∈ R n | ||χ|| ≤ r}<br />
with r any given positive real. We still choose V d1 (χ) =<br />
z T Pz + 1 2 y2 as the Lyapunov function candidate for the<br />
sampled-data system. Then we present some sets used<br />
throughout the pro<strong>of</strong>. Def<strong>in</strong>e c := max χ∈Br V d1 (χ) and<br />
the set Ω c := {χ ∈ R n |V d1 (χ) ≤ c}. There exist two<br />
K functions ψ 1 and ψ 2 such that ψ 1 (||χ||) ≤ V d1 (χ) ≤<br />
ψ 2 (||χ||). Let l := ψ1 −1 (c) + ν with ν an arbitrarily<br />
positive number. Def<strong>in</strong>e B l := {χ ∈ R n | ||χ|| ≤ l}. Then<br />
B r ⊂ Ω c ⊂ B l . The constants L u1 , L 1 , L 2 are Lipschitz<br />
constants <strong>of</strong> u c1 , φ y and φ z , respectively, chosen with<br />
respect to B l .<br />
These local Lipschitz conditions establishes that for the<br />
overall sampled-data system with χ(0) ∈ Ω c , there exists a<br />
unique solution χ(t) over some <strong>in</strong>terval [0,t 1 ). Notice that<br />
t 1 might be f<strong>in</strong>ite. However, later analysis shows that the<br />
solution does exist for every sampl<strong>in</strong>g <strong>in</strong>terval, and thus<br />
exists for all t ≥ 0 with the property that lim t→∞ χ(t) = 0.<br />
Consider the case when t = 0, χ(0) ∈ B r ⊂ Ω c . l<br />
can be chosen sufficiently large such that there exists a<br />
T 1 ∗ ∈ (0,t 1 ), and for all T ∈ (0,T1 ∗ ), the follow<strong>in</strong>g holds:<br />
χ(t) ∈ B l , ∀t ∈ [0,T],χ(0) ∈ Ω c (24)<br />
The existence <strong>of</strong> T ∗ 1 is ensured by cont<strong>in</strong>uous dependency<br />
<strong>of</strong> the solution χ(t) on the <strong>in</strong>itial conditions.<br />
Next we shall estimate |y(t)−y(0)| dur<strong>in</strong>g the <strong>in</strong>terval [0,T]<br />
forced by the sampled-data control u d1 . From the second<br />
equation <strong>of</strong> (22), the dynamics <strong>of</strong> y is<br />
It then follows that<br />
ẏ = z 1 + φ y (y) + u d1
∫ t<br />
y(t) = y(0) +<br />
∫ t<br />
0<br />
Then we have<br />
0<br />
∫ t<br />
|y(t) − y(0)| ≤<br />
∫ t<br />
z 1 (τ)dτ +<br />
(<br />
φ y (y) − φ y<br />
(<br />
y(0)<br />
) ) dτ +<br />
0<br />
‖z(τ)‖dτ<br />
} {{ }<br />
∫ t<br />
0<br />
0<br />
∫ t<br />
+<br />
0<br />
∆ 1<br />
u d1 (τ)dτ +<br />
∫ t<br />
L 1 |y(τ) − y(0)|dτ +<br />
0<br />
φ y<br />
(<br />
y(0)<br />
)<br />
dτ (25)<br />
L u1 |y(0)|dτ +<br />
∫ t<br />
0<br />
L 1 |y(0)|dτ(26)<br />
We first calculate the <strong>in</strong>tegral ∆ 1 . From the first equation<br />
<strong>of</strong> system (7), we obta<strong>in</strong><br />
z(t) = e Dt z(0) +<br />
∫ t<br />
0<br />
e Dt φ z<br />
(<br />
y(τ)<br />
)<br />
dτ (27)<br />
S<strong>in</strong>ce D is a Hurwitz matrix, there exist positive reals k 2 ,<br />
σ such that ‖e Dt ‖ ≤ k 2 e −σt . Thus,<br />
‖z(t)‖ ≤ k 2 e −σt ‖z(0)‖<br />
+<br />
+<br />
∫ t<br />
0<br />
∫ t<br />
0<br />
k 2 e −σ(t−τ) ‖φ z<br />
(<br />
y(τ)<br />
)<br />
− φz<br />
(<br />
y(0)<br />
)<br />
‖dτ<br />
k 2 e −σ(t−τ) ‖φ z<br />
(<br />
y(0)<br />
)<br />
‖dτ<br />
≤ k 2 e −σt ‖z(0)‖<br />
∫<br />
k 2 e −σ(t−τ) |y(τ) − y(0)|dτ<br />
+L 2<br />
t<br />
0<br />
∫<br />
+L 2<br />
t<br />
Then the follow<strong>in</strong>g <strong>in</strong>equality holds<br />
0<br />
k 2 e −σ(t−τ) |y(0)|dτ (28)<br />
∆ 1 ≤ k 2‖z(0)‖ (<br />
1 − e<br />
−σt ) + k 2L 2<br />
σ<br />
σ |y(0)|t<br />
+ k 2L 2<br />
σ<br />
∫ t<br />
0<br />
|y(τ) − y(0)|dτ (29)<br />
Now we are ready to compute |y(t) − y(0)|. In fact, from<br />
(26) and (29), we can see<br />
|y(t) − y(0)| ≤ A 1<br />
(<br />
1 − e<br />
−σt ) + B 1 t<br />
∫ t<br />
+H<br />
0<br />
|y(τ) − y(0)|dτ (30)<br />
where A 1 = σ −1 k 2 ‖z(n)‖, B 1 = L u1 |y(0) + L 1 |y(0)| +<br />
σ −1 k 2 L 2 |y(0)| and H = σ −1 k 2 L 2 +L 1 . Apply<strong>in</strong>g Gronwall-<br />
Bellman <strong>in</strong>equality Khalil [2002] to (30) produces<br />
(<br />
|y(t) − y(0)| ≤ A 1 1 − e<br />
−σt ) + B 1<br />
(<br />
e Ht − 1 )<br />
(<br />
H<br />
+A 1 σe Ht + He −σt<br />
−(H + σ) ) (H + σ) −1 (31)<br />
Sett<strong>in</strong>g t = T <strong>in</strong> the right side <strong>of</strong> (31) leads to<br />
where<br />
|y(t) − y(0)| ≤ δ 1 (T)|y(0)| + δ 2 (T)||z(0)|| (32)<br />
δ 1 (T) = H −1 (L u1 + L 1 + σ −1 k 2 L 2 ) ( e HT − 1 )<br />
δ 2 (T) = σ −1 k 2<br />
(<br />
σe HT + He −σT − (H + σ) ) (H + σ) −1<br />
+σ −1 k 2<br />
(<br />
1 − e<br />
−σT ) (33)<br />
Next we shall study the behaviour <strong>of</strong> the sampled-data<br />
system us<strong>in</strong>g the chosen Lyapunov function candidate V d1 .<br />
When t ∈ (0,T), its time derivative satisfies<br />
˙V d1 = −(γ 1 + 2)‖z‖ 2 + 2z T Pφ z + y(z 1 + φ y + u d1 )<br />
≤ −k 1 y 2 − γ 1 ||z|| 2 + |y||u d1 − u c1 |<br />
≤ −k 1 y 2 − γ 1 ||z|| 2 + L u1 |y − y(0)|y|<br />
≤ − ( k 1 − L u1<br />
2 (δ 1(T) + δ 2 (T)) ) y 2 − γ 1 ||z|| 2<br />
+ L u1<br />
2 δ 1(T)|y(0)| 2 + L u1<br />
2 δ 1(T)||z(0)|| 2<br />
≤ −α 1 (T)V d1 + β 1 (T)V d1 (χ(0)) (34)<br />
where<br />
{<br />
}<br />
γ 1<br />
α 1 (T) = m<strong>in</strong> 2k 1 − L u1 δ 1 (T) − L u1 δ 2 (T),<br />
λ max (P)<br />
{<br />
β 1 (T) = max L u1 δ 1 (T), L }<br />
u1δ 2 (T)<br />
(35)<br />
2λ m<strong>in</strong> (P)<br />
where λ max (·) and λ m<strong>in</strong> (·) denote the maximum and<br />
m<strong>in</strong>imum eigenvalues <strong>of</strong> a matrix, respectively.<br />
It then can be shown that there exists a T ∗ 2 > 0 so that<br />
α 1 (T) > β 1 (T) > 0, for all T ∈ (0,T ∗ 2 ). Note from (33)<br />
that both δ 1 (T) and δ 2 (T) are actually the cont<strong>in</strong>uous<br />
functions <strong>of</strong> T with δ 1 (0) = δ 2 (0), which implies β 1 (0) = 0.<br />
It can then be established from (35) that the function<br />
e 1 (T) := α 1 − β 1 is a decreas<strong>in</strong>gly cont<strong>in</strong>uous function<br />
<strong>of</strong> T with e 1 (0) > 0, which asserts by the cont<strong>in</strong>uity <strong>of</strong><br />
e 1 (T) the existence <strong>of</strong> T ∗ 2 .<br />
F<strong>in</strong>ally, set T 1 = m<strong>in</strong>(T ∗ 1 ,T ∗ 2 ). From Lemma 2 we have<br />
V 1 ∈ Ω c . Thus the above analysis can be repeated for next<br />
<strong>in</strong>terval [T,2T), and for each <strong>in</strong>terval [mT,mT +T). Then<br />
the conclusion follows from Lemma 2.<br />
4.2 Stability <strong>of</strong> the sampled-data system with ρ = 2<br />
We first give out the implementation <strong>of</strong> the sampled-data<br />
controller u d2 . Indeed u d2 is given as follows<br />
u d2 (t) = u c2<br />
(<br />
y(m),ξ(m)<br />
)<br />
, ∀t ∈ [mT,mT + T) (36)<br />
where y(m) can be obta<strong>in</strong>ed by sampl<strong>in</strong>g y(t) at each<br />
sampl<strong>in</strong>g <strong>in</strong>stant, while ξ(m), as one part <strong>of</strong> the controller,<br />
is obta<strong>in</strong>ed digitally by
ξ(m) = e λT ξ(m − 1) +<br />
λ −1 (1 − e −λT )u d2<br />
(<br />
y(m − 1),ξ(m − 1)<br />
)<br />
(37)<br />
Then we have the follow<strong>in</strong>g theorem:<br />
Theorem 4. For the system <strong>of</strong> (7) and (4) with the<br />
sampled-data controller shown <strong>in</strong> (36) and (37), and for<br />
any given <strong>in</strong>itial set conta<strong>in</strong><strong>in</strong>g the orig<strong>in</strong>, there exists a<br />
constant T 2 > 0 such that for all 0 < T < T 2 , the controller<br />
u d2 asymptotically stabilizes the system.<br />
Pro<strong>of</strong>. In this case the analysis is more <strong>in</strong>volved than that<br />
for the case ρ = 1, as the effect <strong>of</strong> the dynamic filter <strong>of</strong> ξ<br />
has to be dealt with.<br />
Notice from (37) that ξ(n) is the exact, discrete-time<br />
model <strong>of</strong> the system<br />
˙ξ = −λξ + u d2 , (38)<br />
s<strong>in</strong>ce u d2 rema<strong>in</strong>s constant dur<strong>in</strong>g each <strong>in</strong>terval and the<br />
dynamics <strong>of</strong> ξ shown <strong>in</strong> (38) is l<strong>in</strong>ear. This <strong>in</strong>dicates that<br />
we can use (38) <strong>in</strong>stead <strong>of</strong> (37) for the stability analysis<br />
<strong>of</strong> the sampled-data system, as (37) and (38) are virtually<br />
equivalent at each sampl<strong>in</strong>g <strong>in</strong>stant.<br />
Next def<strong>in</strong>e χ = [z,y, ˜ξ] T . Then a Lyapunov function<br />
candidate is chosen as the same as for cont<strong>in</strong>uous-time<br />
case, that is, V d2 (χ) = z T Pz + 1 2 y2 + 1 2 ˜ξ 2 . Similar to the<br />
case <strong>of</strong> ρ = 1, the sets B r , Ω c and B l can also be def<strong>in</strong>ed<br />
such that B r ⊂ Ω c ⊂ B l , and there exists a T ∗ 3 > 0 such<br />
that for all T ∈ (0,T ∗ 3 ), the follow<strong>in</strong>g holds<br />
χ(t) ∈ B l , ∀t ∈ (0,T],χ(0) ∈ Ω c . (39)<br />
Consider the case when t = 0, χ(0) ∈ B r ⊂ Ω c . Then we<br />
shall compute the estimates <strong>of</strong> |ξ(t)−ξ(0)| and |y(t)−y(0)|<br />
for t ∈ (0,T). From (38), we have<br />
ξ(t) = e λt ξ(n) +<br />
∫ t<br />
0<br />
e λ(t−τ) u d2 dτ (40)<br />
Then, us<strong>in</strong>g the Lipschitz property <strong>of</strong> u c2 and the fact that<br />
u d2 (0,0) = 0, it can be shown that<br />
|ξ(t) − ξ(0)| ≤ (1 − e −λt )|ξ(0)|<br />
∫t<br />
( )<br />
+|u d y(0),ξ(0) | e −λ(t−τ) dτ<br />
≤ δ 3 (T)|y(0)| + δ 4 (T)|ξ(0)| (41)<br />
where δ 3 (T) = λ −1 L u2<br />
(<br />
1 − e<br />
−λT ) , δ 4 (T) = (λ −1 L u2 +<br />
1) ( 1 − e −λT) and L u2 is a Lipschitz constant <strong>of</strong> u c2 with<br />
respect to the set B l . As for |y − y(0)|, we have from (7)<br />
∫ t<br />
y(t) = y0) +<br />
∫ t<br />
+<br />
+<br />
0<br />
∫ t<br />
It can then be shown that<br />
0<br />
0<br />
0<br />
∫ t<br />
z 1 (τ)dτ +<br />
0<br />
ξ(τ)dτ<br />
(<br />
φ y (y) − φ y<br />
(<br />
y(0)<br />
) ) dτ<br />
φ y<br />
(<br />
y(0)<br />
)<br />
dτ (42)<br />
∫ t<br />
|y(t) − y(0)| ≤<br />
0<br />
∫ t<br />
+<br />
0<br />
∆ 1<br />
‖z(τ)‖dτ<br />
} {{ }<br />
∫ t<br />
0<br />
‖ξ(τ)‖dτ<br />
} {{ }<br />
L 1 |y(τ) − y(0)|dτ +<br />
∆ 2<br />
+<br />
∫ t<br />
0<br />
L 1 |y(0)|dτ(43)<br />
where ∆ 1 is already shown <strong>in</strong> (29). It can be obta<strong>in</strong>ed from<br />
(40) that<br />
∆ 2 ≤ |ξ(0)| (<br />
1 − e<br />
−λt ) + L u<br />
( )<br />
|y(0)| + |ξ(0)| t (44)<br />
λ<br />
λ<br />
With (29), (43) and (44), an analysis similar to that for<br />
the case ρ = 1 can be performed to produce<br />
|y(t) − y(0)| ≤ δ 5 (T)|y(0)| + δ 6 (T)||z(0)|| + δ 7 (T)||ξ(0)||(45)<br />
where<br />
δ 5 (T) = H −1 (L 1 + λ −1 L u2 + σ −1 k 2 L u2 ) ( e HT − 1 )<br />
δ 6 (T) = σ −1 k 2<br />
(<br />
σe HT + He −σT − (H + σ) ) (H + σ) −1<br />
+σ −1 k 2<br />
(<br />
1 − e<br />
−σT )<br />
δ 7 (T) = λ −1 (1 − e −λT ) + λ −1 L u2 (e HT − 1)<br />
+λ −1( λe HT + He −λT − (H + λ) ) (H + λ) −1<br />
The time derivative <strong>of</strong> V d2 dur<strong>in</strong>g the <strong>in</strong>terval [0,T)<br />
satisfies<br />
˙V d2 = −(γ 1 + 2)‖z‖ 2 + 2z T Pφ z + y(z 1 + φ y + ξ)<br />
+˜ξ ( − λξ + u c2 − ∂ˆξ<br />
∂y (z 1 + φ y + ξ) )<br />
+˜ξ(u d2 − u c2 )<br />
≤ −k 1 y 2 − γ 1 ||z|| 2 − λ˜ξ 2<br />
+|˜ξ||u d2 − u c2 | (46)<br />
In addition, it is known that ξ = ˜ξ + ˆξ and ˆξ (i.e., u c1 ) is a<br />
function <strong>of</strong> y with a Lipschitz constant L u1 . Then we have<br />
|ξ(0)| ≤ |˜ξ(0)| + |ˆξ(y(0))| ≤ |˜ξ(0)| + L u1 |y(0)| (47)<br />
From (41), (45) and (47), it can be verified that<br />
|˜ξ||u d2 − u c2 | ≤ ε 1 |y(0)| 2 + ε 2 ||z(0)|| 2 + ε 3 |˜ξ(0)| 2 + ε 4 |˜ξ| 2<br />
where ε 1 = L u2<br />
(<br />
δ3 + δ 5 + L u1 (δ 4 + δ 7 ) ) , ε 2 = L u2 δ 6 ,<br />
ε 3 = L u2 (δ 4 + δ 7 ), ε 4 = L u2<br />
( ∑ 7<br />
i=3 δ i + L u1 (δ 4 + δ 7 ) ) .<br />
From above we then have<br />
where<br />
˙V d2 ≤ −k 1 y 2 − γ 1 ||z|| 2 − (λ − ε 4 )˜ξ 2<br />
+ε 1 |y(0)| 2 + +ε 2 ||z(0)|| 2 + ε 3 |˜ξ(0)| 2<br />
△<br />
= −α 2 (T)V d2 + β 2 (T)V d2 (χ(0)) (48)<br />
{<br />
α 2 (T) = m<strong>in</strong> 2k 1 ,<br />
β 2 (T) = max<br />
{<br />
2ε 1 ,<br />
}<br />
γ 1<br />
λ max (P) ,2(λ − ε 4)<br />
}<br />
ε 2<br />
λ m<strong>in</strong> (P) ,2ε 3<br />
(49)<br />
Note that ε i (0) = 0 (i = 1,2,3,4). Follow<strong>in</strong>g the same<br />
approach as shown <strong>in</strong> the analysis for the case ρ = 1, we
can establish that there exists a T ∗ 4 so that for all 0 < T <<br />
T ∗ 4 , α 2 (T) > β 2 (T) > 0. F<strong>in</strong>ally, set T 2 := m<strong>in</strong>(T ∗ 3 ,T ∗ 4 ),<br />
which completes the pro<strong>of</strong> by Lemma 2.<br />
Remark 1. Both Theorem 3 and Theorem 4 only declare<br />
the existence <strong>of</strong> a certa<strong>in</strong> upper limit <strong>of</strong> sampl<strong>in</strong>g period,<br />
but states no <strong>in</strong>formation <strong>of</strong> the affects <strong>of</strong> control parameters<br />
and <strong>in</strong>itial sets <strong>of</strong> the system on the upper limit. The<br />
way the <strong>in</strong>itial set and the controller parameters affect the<br />
upper limit is still subject to further <strong>in</strong>vestigation.<br />
Remark 2. If ρ = 1, the controller reduces to a static<br />
feedback controller. If ρ = 2, the dynamic controller<br />
uses a particular l<strong>in</strong>ear filter, a convenience <strong>in</strong> control<br />
implementation. However, stability analysis for ρ ≥ 3 is<br />
more complicated and rema<strong>in</strong>s a subject <strong>of</strong> further studies.<br />
5. SIMULATION<br />
Consider the follow<strong>in</strong>g system with relative degree ρ = 2<br />
ẋ 1 = x 2 + y 2<br />
ẋ 2 = u, y = x 1<br />
The filter ˙ξ = −λξ + u is <strong>in</strong>troduced so that the filtered<br />
transformation η 1 = x 1 and η 2 = x 2 − ξ, and the state<br />
transformation z = η 2 − λη 1 can render the system <strong>in</strong>to<br />
the follow<strong>in</strong>g form<br />
ż = −λz + y 2 − λ 2 y − λy 2<br />
ẏ = z + (λ + y)y + ξ<br />
F<strong>in</strong>ally, the stabiliz<strong>in</strong>g function ˆξ = −ky − (λ + y)y −<br />
1<br />
2 λ2 (y + λ) 2 y 2 and the control u c can be obta<strong>in</strong>ed us<strong>in</strong>g<br />
(11). For simulation, we choose λ = 3, k = 4.<br />
Simulations are carried out by Simul<strong>in</strong>k us<strong>in</strong>g zero-order<br />
hold blocks for the case where the <strong>in</strong>itial values are x 1 (0) =<br />
1 and x 2 (0) = 200. Results shown <strong>in</strong> Fig.1 and Fig.2<br />
<strong>in</strong>dicate that the sampled-data system is asymptotically<br />
stable when T = 0.0001s. Further simulations show that<br />
the overall system is unstable if T = 0.0005s. In summary,<br />
the example illustrates that for a range <strong>of</strong> sampl<strong>in</strong>g period<br />
T, the sampled-data controller designed <strong>in</strong> the former<br />
sections can asymptotically stabilise the sampled-data<br />
system.<br />
x 1<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
−0.5<br />
0 0.5 1 1.5 2 2.5 3<br />
Fig. 1. The time response <strong>of</strong> x 1 for T = 0.0001<br />
t (s)<br />
x 2<br />
200<br />
150<br />
100<br />
50<br />
0<br />
−50<br />
0 0.5 1<br />
t (s)<br />
1.5 2 2.5 3<br />
Fig. 2. The time response <strong>of</strong> x 2 for T = 0.0001<br />
6. CONCLUSION<br />
We have presented an analysis for output feedback control<br />
<strong>of</strong> one class <strong>of</strong> nonl<strong>in</strong>ear sampled-data systems. It has<br />
been shown that <strong>in</strong> the case <strong>of</strong> relative degree ρ = 1,2,<br />
the sampled-data version <strong>of</strong> cont<strong>in</strong>uous-time controllers<br />
will still asymptotically stabilise the system for any given<br />
<strong>in</strong>itial sets, provided that the sampl<strong>in</strong>g period T is small.<br />
REFERENCES<br />
T. Chen and B.A. Francis. Optimal <strong>Sampled</strong>-<strong>Data</strong> <strong>Control</strong><br />
<strong>Systems</strong>. Communications and <strong>Control</strong> Eng<strong>in</strong>eer<strong>in</strong>g<br />
Series. Spr<strong>in</strong>ger-Verlag, London, 1995.<br />
A.M. Dabroom and H.K. Khalil. <strong>Output</strong> feedback<br />
sampled-data control <strong>of</strong> nonl<strong>in</strong>ear systems us<strong>in</strong>g highga<strong>in</strong><br />
observers. IEEE Transactions on Automatic <strong>Control</strong>,<br />
46:1712–1725, 2001.<br />
Z. D<strong>in</strong>g. Adaptive output regulation <strong>of</strong> a class <strong>of</strong> nonl<strong>in</strong>ear<br />
systems with completely unknown parameters. In<br />
Proceed<strong>in</strong>gs <strong>of</strong> the American <strong>Control</strong> Conference, pages<br />
1566–1571, USA, 2003.<br />
J.W. Grizzle and P.V. Kokotovic. <strong>Feedback</strong> l<strong>in</strong>earization<br />
<strong>of</strong> sampled-data systems. IEEE Transactions on Automatic<br />
<strong>Control</strong>, 3(9):857–859, 1988.<br />
L. Hou, A.N. Michel, and H. Ye. Some qualitative properties<br />
<strong>of</strong> sampled-data control systems. IEEE Transactions<br />
on Automatic <strong>Control</strong>, 42:1721–1725, 1997.<br />
H.K. Khalil. <strong>Nonl<strong>in</strong>ear</strong> <strong>Systems</strong>. Prentice-Hall, Inc., 3rd<br />
edition, 2002.<br />
H.K. Khalil. Performance recovery under output feedback<br />
sampled-data stabilization <strong>of</strong> a class <strong>of</strong> nonl<strong>in</strong>ear systems.<br />
IEEE Transactions on Automatic <strong>Control</strong>, 49:<br />
2173–2184, 2004.<br />
R. Mar<strong>in</strong>o and P. Tomei. <strong>Nonl<strong>in</strong>ear</strong> <strong>Control</strong> Design:<br />
Geometric, Adaptive and Robust. Prentice-Hall, Inc.,<br />
1995.<br />
D. Nesic and A.R. Teel. A framework for stabilization<br />
<strong>of</strong> nonl<strong>in</strong>ear sampled-data systems based on their approximate<br />
discrete-time models. IEEE Transactions on<br />
Automatic <strong>Control</strong>, 49:1103–1122, 2004.<br />
A.R. Teel, D. Nesic, and P.V. Kokotovic. A note on <strong>in</strong>putto-state<br />
stability <strong>of</strong> sampled-data nonl<strong>in</strong>ear systems. In<br />
Proceed<strong>in</strong>gs <strong>of</strong> the 37th Conference on Decision and<br />
<strong>Control</strong>, pages 2473–2478, USA, 1998.