Documente Academic
Documente Profesional
Documente Cultură
Email: krishna.busawon@unn.ac.uk
AbstractIn this paper, we consider the issue of stabilizing the following class of systems:
a class of linear systems using irregular sampled output mea- {
surements. For this purpose, we design a standard linear state
x(t) = F x(t) + Gu(t)
(1)
feedback controller and an impulsive observer to provide an y(tk ) = Hx(tk )
estimate the non-measured states, which are subsequently fed
back in the control algorithm. We consider linear systems that where x(t) Rn is the state, u(t) Rm is the input and
can be decomposed, via a change of coordinates, into their y(tk ) Rp is the output. The matrices F , G and H are
respective measured and unmeasured dynamics. We consider
constant and of appropriate dimension. Moreover, the system
the two cases whereby the unmeasured subspace is stable and
unstable respectively. In the case where the unmeasured subspace is assumed to be detectable and stabilizable. We assume that
is stable, we employ a standard impulsive observer coupled with a there exists a linear transformation z = T x that permits to
continuous linear feedback control. Next, when the unmeasured transform system (1), into the following system:
subspace is unstable, we employ two cascaded observers - an
impulsive and a Luenberger observer - in conjunction with a
z1 (t) = A11 z1 (t) + A12 z2 (t) + B1 u(t)
linear feedback control. In order to prove the stability of the
overall closed-loop system we proposed a practical stability result z2 (t) = A22 z2 (t) + B2 u(t) (2)
y(tk ) = z1 (tk )
for a class of linear impulsive systems. Some simulation results
are presented to show the performance of the observer-based
control. Finally, some conclusions are drawn. where z(t) = (z1T , z2T )T with A11 Rpp , A12 Rp(np) ,
Index Termsobserver-based control, discrete measurement, A22 R(np)(np) , B1 Rpm , B2 R(np)m and A22
impulsive systems Hurwitz continuous.
We consider the two cases where the matrix A22 is stable
and unstable respectively. In the case where A22 is stable, we
I. I NTRODUCTION design an impulsive observer, which is basically a copy of
the system with an appropriate reset function. Additionally,
In many practical control applications, we are faced with we employ a continuous linear feedback in order to design
the issue of controlling a continuous-time system with mea- an observer-based control for the system. This effectively
surements available at discrete instants of time. When faced shows that a separation principle is possible in that case. In
with such a situation, one can attempt to discretise the system the case where A22 is not stable the observer design is not
if the output is sampled fast enough; so that the discretised trivial. For this, we employ two cascaded observers composed
differential equation describing the system gives a good rep- of a Luenberger and an impulsive observer. The reason for
resentation of the original continuous system. One can then including a Luenberger observer is to able to modify the
employ a discrete output feedback controller or a discrete eigenvalues of A22 via its output correction term.
observer-based controller to control the system. However, in The result obtained is also valid in the case where output
the case where the measurements are irregularly spaced such sampling frequency is smaller than the Nyquist-Shannon sam-
discretisation is not possible. Such is the case for chemical pling frequency; i.e. in the case of sparse measurements (see
processes in general, whereby the output measurements are eg. [1] [2]). Effectively, this work shows that it is possible to
available in discrete time and the time intervals between the guaranty closed loop stability under sparse measurement (see
samples are not necessarily regularly spaced or constant. In e.g. [3]).
such a case, one has to find alternative means to control these The paper is organized as follows: In the next section a sta-
systems. bility result with respect to impulsive systems is presented.This
In this paper, we propose an observer-based control design stability result is then used in the subsequent sections to prove
methodology for a class of linear systems with irregularly the stability of the overall closed-loop system. Additional
sampled output measurements. More precisely, we consider results on different types of stability of impulsive systems can
be found in [4], [5], [6], [7] and [2]. In Section 3, an observer- 1) A22 is a continuous Hurwitz matrix; that is its eigen-
based control is proposed for the case where the unmeasured values have strictly negative real part.
subspace is stable; i.e. A22 is stable. The case where the 2) ReA11 t 2 < 1, for all t [0, max [> 0.
unmeasured subspace is unstable is considered in Section 4. Then, > 0, the state of system (3) converges to a ball of
In Section 5, an example is given in order to highlight the radius (practical stability).
efficiency of the proposed observer based control scheme.
Finally, the paper ends with some conclusions and perspectives Proof: First, the solution of (3) for t [t+
k , tk+1 ] with the
for future research work. +
initial time t0 = tk is given by
II. P RACTICAL STABILITY FOR A CLASS OF IMPULSIVE t
A11 (ttk ) +
x1 (t) = e x1 (tk ) + eA11 (t ) A12 x2 ( )d (5)
tk
SYSTEM
and
In this section, we are going to give some conditions under
x2 (t) = eA22 (ttk ) x2 (t+
k) (6)
which a class of linear impulsive systems is practically stable.
More precisely, we consider the following class of linear Replacing (6) in (5), we obtain, at t = tk+1 :
impulsive systems:
x1 (tk+1 ) = eA11 k x1 (t+ )
tk+1 k
x 1 (t) = A11 x1 (t) + A12 x2 (t); t = tk
eA11 (tk+1 ) A12 eA22 ( tk ) x2 (t+
x 2 (t) = A22 x2 (t) + k )d
(3)
x1 (t+ k ) = Rx1 (tk )
tk
+
x2 (tk ) = x2 (tk ) where k = tk+1 tk .
Consider the following Lyapunov function:
where x(t) Rn , A11 Rpp , A12 Rp(np) , A22
R(np)(np) and R Rpp . The sampling sequence tk V (x) = V1 (x1 ) + V2 (x2 )
T = {ti : i N} R. We assume that there exists max and with
min with max > min > 0 such that i > 0
2
V1 (x1 ) = xT1 x1 = x1
ti + min < ti+1 < ti + max. (4) 2
V2 (x2 ) = xT2 x2 = x2 .
The above strict inequality permits to exclude the Zeno phe-
nomena since system (3) is essentially a particular class of Starting with V2 (x2 ), we get
hybrid systems [8]. V2 (x2 (t+ +
( )
V2 (x2 (tk )) = k+1 )) V2 x2 (tk )
Additionally, we define:
V2 (x2 (tk+1 )) V2 x2 (t+
( )
= k)
x(t+
k) , lim x(tk + h) = eA22 k x2 (t+ )2 x2 (t+ 2
k )
h0
( A22 2 k ) +
x(t
k) , lim x(tk h) = x(tk ) e k
1 V (x2 (tk ))
h0
Throughout the paper, we assume that the system (3) verifies Since the matrix A22 is Hurwitz continuous, this implies that
the sufficient conditions for the existence of solution [7]. eA22 k is a discrete Hurwitz matrix and therefore
Before stating the main stability result of this section, we eA22 k 2 < 1
recall some important results and definitions. First, recall
that if A Rnn is a discrete Hurwitz matrix (that is, all Consequently,
eigenvalues are strictly inside the unit circle of the complex V2 (x2 (tk )) < 0 (7)
plane) then limk Ak = 0. Additionally, there exist at least This inequality ensures the uniform asymptotic convergence
a matrix norm , such that A < 1. of x2 to zero (uniformly with respect to x1 ).
Next, we recall that: Therefore, > 0 there exists h such that:
Definition 1. [4] The system (3) is said to be:
1) Practically stable with respect to (, ) if given (, ) k > h , x2 (tk ) < . (8)
with 0 < < , we have that x(t0 ) < implies Now, for the Lyapunov function V1 , we have
x(t) < , t > t0 for some t0 R.
2) Uniformly practically stable with respect to (, ), if 1) V1 (x1 (tk )) = V1 (x(t+ +
k+1 )) V1 (x(tk ))
holds for every t0 R. = V1 (Rx1 (tk+1 )) V1 (x1 (t+
k ))
3) Practically asymptotically stable with respect to (, ),
if 1) holds and limt x(t) = 0. By replacing the above expression of x1 (tk+1 ), we get:
V1 ReA11 k x1 2 x1 2
t
We can now state the following: +R tkk+1 eA11 (tk+1 ) A12 eA22 ( tk ) d x2 2
Theorem 1. Assume that system (3) verifies the following (ReA11 k 2 1)V1 (x1 )
t
conditions: +A12 2 x2 2 tkk+1 ReA11 (tk+1 ) 2 eA22 ( tk ) 2 d
Since A22 is Huzwitz continuous and that from the second is practically asymptotically stable.
condition ReA11 (tk+1 ) 2 < 1 for all [tk , tk+1 [, we
have: t Proof: By setting e1 = z1 z1 , e2 = z2 z2 , we obtain:
V1 (ReA11 k 2 1)V1 (x1 ) + A12 2 x2 2 tkk+1 d ( ) ( )( )
e 1 A11 A12 e1
(ReA11 k 2 1)V1 (x1 ) + k A12 2 x2 2 =
e 2 0 A22 e2
So, from (8) we have
> 0, h > 0, k > h : x2 (tk ) < . On the other hand, by expanding the closed-loop system (12)
Now, choose > 0 and define V1,min = min V1 (x1 ). Then, we obtain
x1 = z = (A BK) z + BKe
(ReA11 k 2 1)V1 min
for all x1 (k) > and k < , we where
A12
have:
( )
A11 B1 K1 A12 B1 K2
V1 (x1 (tk )) < 0 A pp = (A BK) =
B2 K1 A22 B2 K2
( )
Owing to the fact that > 0 can be arbitrarily chosen, the B K
2 1 B K
1 2
Bpp = BK =
only constraint on k is ReA11 k 2 < 1. Consequently (7) B 2 K1 B 2 K 2
and the previous inequality imply that h such that k > h The reset function becomes:
we have V (x1 (tk ), x2 (tk )) < 0 if x1 > .
This completes the proof of Theorem 1. e1 (t+
k ) = Re1 (tk )
III. C ASE 1: STABLE UNMEASURED SUBSPACE The overall state dynamics and observation error dynamics can
Consider system (2) and assume that the matrix A22 is be written as
( ) ( )( )
Hurwitz continuous. e A 0 e
=
In a compact form system (2) is written as: z Bpp App z
{
= Az(t) + Bu(t)
z(t) with the following reset function:
(9)
y(tk ) = Cz(tk )
e1 (t+
k) e1 (tk )
e2 (t+ )
( ) ( ) [ ]
A11 A12 B1 R 0 e2 (tk )
where A = , B = and C = k
z1 (t+ ) = 0 I2np z1 (tk )
0 A22 B2 k
z2 (t+
( )
Ip 0 with Ip being the p-dimensional identity matrix. k) z2 (tk )
Remark 1: Note that the above assumption that there exists The observation error dynamics has exactly the same form
a linear transformation z = T x that permits to transform as system (3) and is independent of zi , so e is practically
system (1), into (2) is not too restrictive, because if ker{H} asymptotically stable. Now, again from the theorem 1 for all
is included on the stable subspace of F , it is always possible > 0 h > 0 such that ei (k) < . Now setting V3 (z) ,
to select a part of the output y(tk ) as the considered output, 2
z T z = z and proceeding in the same way as for the proof
in order to obtain the system (2) under the transformation of Theorem 1 , we obtain:
z = T x.
V3 (z(tk )) = V3 eApp k z(tk )
(
We design an observer for system (2) as follows: tk+1 )
App (tk+1 )
z1 (t) = A11 z1 (t) + A12 z2 (t) + B1 u(t) + e Bpp e( )d
tk
z2 (t) = A22 z2 (t) + B2 u(t) (10) V3 (x1 (tk ))
+
1 (tk ) = R
z z1 (tk ) + (Ip R)z1 (tk )
As k > h : e2 < , we have:
y(tk ) = z1 (tk )
where R Rmn is the reset matrix, such that V3 eApp k z2
tk+1
R = diag{r1 , .., rp }, with ri ] 1, 1[, for i = 1, .., p. + eApp (tk+1 ) Bpp d 2 z2
As a result, a linear observer based feedback is given by: tk
(
z1 (t)
) (ReApp k 2 1)V3 (x1 )
u(t) = K z , (K1 , K2 ) (11) tk+1
z2 (t)
+ ReApp (tk+1 ) Bpp eA22 ( tk ) 2 d x2 2
where K Rmn is chosen such that (A BK) is stable. tk
(eApp k 2 1)V3 (z(k))
tk+1
Theorem 2. Assume that there exists max and min with
+Bpp 2 eApp (tk+1 ) 2 d
max > min > 0 such that i > 0, ti + min < ti+1 < tk
ti + max. Then, the closed-loop system:
This ensures the practical stability and completes the proof of
= Az(t) BK z(t) = (A BK) z(t) + BKe(t) (12)
z(t) Theorem 2.
IV. C ASE 2: UNSTABLE UNMEASURED SUBSPACE that for any initial condition (0) such that (0) < max ,
the original systems state z converges to z < and the
Consider again system (2) and assume that the matrix A22 2
+
is unstable. extended state at each sampling time tk converge into a set
Remark 2. As in the previous section, that there exists a B , { : < }, ifthe following conditions are verified:
linear transformation z = T x that permits to transform system A22 M 0 M
(1), into (2) is not too restrictive in this case either. This is i) the matrix 1 = 0 A11 L1 A12 is a
because it is always possible to consider only one part of the 0 L2 A22
output y as a considered output, in order to obtain the system continuous Hurwitz matrix.
(2) under the transformation z = T x. However, in some cases, ( ii) the control )matrix 2 =
A11 B1 K1 A12 B1 K2
this output restriction can destroy the system detectability and is a continuous Hurwitz
B2 K1 A22 B2 K2
then the basic assumption for considering only detectable and
matrix.
stabilizable system is violated.
The observer-based control for system (2) is quite different Proof: One can check that
to the one proposed in the previous section as far as the
observer part is concerned. In fact, we employ two cascaded e1 (t+
k ) = Re1 (tk )
observers - an impulsive and a Luenberger observer. More
2
precisely, Now, we define a Lyapunov function V1 (e1 ) , eT1 e1 = e1
, we obtain
V1 (e1 (t+ k+1 ))
z (t) = A z (t) + A z (t) + B u(t) = V1 (ReA11 k e1 (t+
tk+1 A (t )
k ) + tk e 11 k+1 A12 e2 ( )d )
1 11 1 12 2 1
z2 (t) = A22 z2 (t) + M (2 (t) z2 (t)) + B2 u(t) + +
Assuming that [tk , tk+1 [, e2 ( ) < max (this will
1 (t) = A11 1 (t) + A12 2 (t) + L1 ( z1 (t) 1 (t)) + B 1 u(t) be proved later in the proof), with the same arguments as for
2 (t) = A22 2 (t) + L2 ( Theorem 1, we have:
k
(13) t
+R tk eA11 (tk+1 ) A12 eA22 ( tk ) max d 2
k+1
where K Rmn , with R = diag{ r1 , .., rn1 } and 1 < (ReA11 k 2 )V1,max
ri < 1, for i = 1, . . . , n1 and M , L1 and L2 are matrices of t
+A12 2 max tkk+1 ReA11 (tk+1 ) 2 eA22 ( tk ) 2 d
appropriate dimensions.
As for the controller, we employ a linear control as before: with V1,max , max {V1 (e1 )}. It is always possible to
e1 =max
find R such that:
( )
z1 (t)
u(t) = K z , (K1 , K2 ) (14)
z2 (t) ( )2
V1 (x1 (t+ k+1 )) <
or (tk+1 ) < (16)
whereby the estimated state variables z1 and z2 are issued from 4
the generalized cascaded impulsive observer. Note that the above inequality is verified by R = 0 and, by
Denote , (eT1 , eT2 , T1 , T2 , z1T , z2T )T with ei = zi zi and continuity, also by a neighborhood of R = 0. Setting 2 ,
i = z . Then, one can show that: (eT2 , T1 , T2 , z1T , z2T )T and using the Lyapunov function V2 ,
(
e 1
) (
A11 A12
)(
e1
) (
0
) T2 P2 2 , with P2 a symmetric positive definite (SPD) matrix,
= + we have:
e 2 0 A22 M e2 M 2
V 2 = A2 2 + B 2 e 1 T P2 2
[ ]
and
+T2 P2 A2 2 + B 2 e1
( ) ( )( ) ( ) [ ]
1 A11 L1 C A12 1 L1 e 1
= +
2 L2 C A22 2 L2 e 1
with
Additionally, the the closed-loop system via the observer is [
1 0
]
given by: A 2 =
3 2
( ) ( )( )
z1 B 2 K 1 B 1 K2 e1
= (15) where
z2 B 2 K 1 B 2 K2 e2 (
1 K2 0 0
B
)
(
A11 B1 K1 A12 B1 K2
)(
z1
) 3 = 2 K2 0 0
+ B
B2 K1 A22 B2 K2 z2
and B 2 = [0T , LT , LT , (B1 K1 )T , (B2 K2 )T ]T . Conditions i)
1 2
We are now able to state the main result of this section: and ii) of Theorem 3 ensure that A2 is a continuous Hurwitz
Theorem 3. Assume that system (2) satisfies Assumption matrix. Consequently, there exists an SPD matrix P2 such that
2 and consider the closed-loop system (15). Then, for all AT P2 + P2 A2 = Id . Consequently,
2
max > 0 and > 0 with max > , there exist a maximum
sampling period max > 0 and at least a reset matrix R such V 2 = 2 2 + 2B2 e1 P2 2
Considering the fact that 2 and for e1 ( )2
2 4
by continuity of e1 between [tk , tk+1 ], then there exists max
such that [tk , tk + max ], we have:
5 10 15 20 25 30 35 40
V 2 = 2 2 + 2B2 e1 P2 2
Therefore, X2 converges to and e1 (tk ) converges to ( )2
2( 4
)2 5 10 15 20 25 30 35 40
V. S IMULATION RESULTS
Fig. 1. System states with sampling period k = 0.4
Let us consider the following linear triangular system:
x 1 = 2x1 + x2 + 3x3 + u
x 2 = x2 + x3
(17)
x 3 = x3 u
y(tk ) = x1 (tk )
5 10 15 20 25 30 35 40
Here, we have
( )
A11 = 2, A12 = 1 3 5 10 15 20 25 30 35 40
( )
1 1
A22 =
0 1
It can be seen that the matrix A22 is unstable. We are 5 10 15 20
Time (s)
25 30 35 40
5 10 15 20 25 30 35 40
10 20 30 40 50 60
Time (s)