Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Proceedings of the 10th Mediterranean Conference

on Control and Automation - MED2002


Lisbon, Portugal, July 9-12, 2002.

BINARY DISTILLATION COLUMN CONTROL


BASED ON STATE AND INPUT OBSERVABILITY
Addison Ros-Bolvar , Ferenc Szigeti

Universidad de Los Andes
Facultad de Ingeniera
Departamento de Control
Av. Tulio Febres Cordero
Merida 5101 - Venezuela
fax: +58 274 240 2846
e-mail: ilich@ula.ve, szigeti@ula.ve

Keywords: Nonlinear observers. Input observabil- 2. In [19], a FDI filter was designed for the same
ity. State elimination. Composition control of binary binary distillation columns, which was consid-
distillation columns. System Inversion. ered by J. Levine and P. Rouchon, in [10]. The
main idea to design a FDI filter, based on non-
Abstract linear observers was similar to the method ap-
plied in [6]. Linear dynamics was achieved by
In this paper, a method for the design of nonlin- injection of the output and output derivatives.
ear observers and input reconstruction is proposed.
The state and input observers is computed follow-
ing the Diops differential algebraic state eliminationNow, the following question arises: How can asymp-
approach. In the input reconstruction case the input totic observer be designed for a nonlinear control
observability is equivalent to system inversion con- system? Injection output and output derivatives are
dition. The elimination procedure is utilized for dual the elementary steps of the Diops state elimination,
composition control of binary distillation columns. [3]. T. Glad has considered the controller design
The particular observer allows us to recover the un- problem in a differential algebraic setting, [5]. The
measured feed composition, i.e., the input recon- construction of nonlinear observer is, traditionally,
struction, which is used in the controller design in based on differential geometry. The well-known
order to maintain desired compositions. The method Bracket Vanishing Algorithm, see [8, 9] is due to
is an alternative against the disturbance decoupling Krener, Isidori, and Respondek. The Bracket Span-
procedure. ning Algorithm is due to Bestle and Zeitz, in [1],
is just as hard computationally as Bracket Vanishing
1 Introduction Algorithm due to its requirements of Lie bracket cal-
culations. Phelps has improved those algorithms by
State elimination approach, via output derivatives, the substitution of the bracket vanishing condition
has been broadly applied in order to design full order by another one, [13].
state observers. For instance,
By another hand, the input reconstruction problem
have been studied in [17]. The idea is to reconstruct
1. In [6], D. Guillaume et al. have proposed sim- an unknown input from system output. In [7] input
ple observers for polymerization reactors. The observability and reconstruction, and fault detection
nonlinear dynamics was linearized by output in- and isolation (FDI) problem have been related. In
jection. Then linear observer was designed to [15, 20] the input observability and system inversion
the linear model. problem have been related to design FDI filters.
In this paper a nonlinear observer and the input ob- Now, let us eliminate x from the coupled system
servability condition are used for binary distillation
column control. The input observer and nonlinear x = f (x, u),
observer are designed from state elimination setting, x = f (x, u) + D(x x)
simultaneously. Thus, in the section 2 an observer y = g(x), y = g(x) + d(x x).
for nonlinear systems based on state elimination is
presented. In the section 3 the unknown input re- Hence, an implicit observer equation will be ob-
construction problem is considered. Conditions for tained
input observability are presented and input observers  
based on state elimination are proposed. In the sec- F x, x , u, u, . . . , y, y , . . . = 0,
u, . . . , y, y,
tion 4 the binary distillation column control problem
is studied. In this case, the composition of the feed with a non equality
flow is an unknown input. Thus, the column control  
proposed is based on state and input observers. G x, x , u, u, . . . , y, y , . . . 6= 0.
u, . . . , y, y,

2 The nonlinear observer Theorem 2.1 Let us suppose that (1) is alge-
braically observable. Then, the plant state x can be
In this paper an algebraic observer design method is eliminated from (2) by the Diops algorithm.
proposed, based on the Diops state elimination al-
gorithm. The idea is the systematization of the naive Considering the differential algebraic setting of the
state elimination steps, mentioned in the introduc- definition of observability, that is, all elements of the
tion, into the celebrated state elimination in control. dynamics is (non differentially) algebraic over the
Let us consider an algebraic control system differential field < < u, y >, see [4].
Theorem 1 can be extended by factorization with re-
x = f (x, u), y = g(x); (1) spect to the zero dynamics of the unobservable sys-
tem.
n p
where x R is the state vector, u R is the con- The suggested design method, based on state elimi-
trol input, y Rm is the output vector. The functions nation, gives the same result for linear systems. For
f, g are of appropriate dimensions. f, g are algebraic simplicity, let us consider the linear system in canon-
(analytic) functions, respectively. ical form
The proposed design scheme is the following: The
construction of the nonlinear observer starts from the x 1 = x2 ,
equations ..
.
x n1 = xn ,
x = f (x, u) + D(x x), n
X
y = g(x) + d(x x), (2) x n = ai xi + u,
i=1
y = x1 .
for the observer states x, where the gains D, d are
algebraic functions, such that D(0) = 0, d(0) = 0, Then, the proposed observer can be computed from
and the system
e = D(e),
x 1

0 1 0 x1
.. .. . . . .. .
is an asymptotically stable algebraic differential . . . ..
= +D(xx)
equation. System (2) is considered as an observer, x n1 0 1 xn1

coupled with the plant states x. x n a1 a2 an xn
where the linear system e = D(x x) is asymptot- the resulting error equation is
ically stable. The states in terms of the output deriva-
tives are e 1 e1
e 2 = D e2 .
x1 = y, x2 = y, . . . , xn = y (n1) . e 3 e3

Let us denote the structure matrix of the system Hence, if the linear system e = De is asymptot-
T ically stable, observer (6) is convergent. The com-
by A, the vector y, y, . . . , y (n1) by Y , and
putable observer can be obtained by state elimina-
(0, . . . , 0, 1)T by en . Then, the observer is
tion. Indeed, using the expressions (3), (4), (5), for
x = AY + D(Y x) + en u x1 , x2 , x3 , the corresponding observers dynamics
are
= A x + en u + (A + D)(Y x). u1 u2 yu2 (u 1 +u2 u3 )y+u 1 (u1 u 3 u 1 u3 )y



x 1 2
u1 u 3 u 1 u3 u2 (u1 +u3 ) 2
The induced error equation for e = x x, is x2 = y


3 u2 u3 y+u2 (u 3 u1 u2 )yu 3 (u1 u 3 u 1 u3 )y
e = De, x u1 u 3 u 1 u3 u2 (u21 +u23 )
u3 y+(u1 u2 u 3 )y+u 2 2
3 (u1 +u3 )

u u u
u u (u 2 +u2 ) x
1
which is asymptotically stable. 1 3 1 3 2 1 3
+D y x2


u1 y(u 1 +u2 u3 )y+u 2 2
1 (u1 +u3 )y
2.1 Example u1 u 3 u 1 u3 u2 (u2 +u2 )
x3
1 3

Let us consider the bilinear system in order to illus- Next, simulations will be shown with u1 = cos (t),
trate the importance of the persistent excitation and u2 = 1, u3 = sin (t). The figure 1 shows the sys-
the universality of an input, for arbitrary universal tem output and the estimated output. The figure 2
input which distinguish all pair of initial states, see shows the output error. It is clear the potentiality of
[4]. The system the method proposed.

x 1 0 u1 u 2 x1 8

x 2 = u1 0 u3 x2 , y = x2 7

x 3 u2 u3 0 x3 5

4
Outputs

is observable. The state elimination can be achieved 3

by computing output derivatives in order to use, for 2

example, Gauss elimination. Hence, 1

u3 y + (u1 u2 u 3 )y + u3 (u21 + u23 )


1
0 1 2 3 4 5 6 7 8 9 10
Time t

x1 = , (3)
u1 u 3 u 1 u3 u2 (u21 + u23 )
Figure 1: Output and estimated output.
x2 = y, (4)
u1 y (u 1 + u2 u3 )y + u1 (u21 + u23 )y
x3 = .(5) Now, let us consider the non-universal controls u1 =
u1 u 3 u 1 u3 u2 (u21 + u23 )
u3 = 1, u2 = 0. The obtained system
Considering the observer in the form
x 1 = x2 , x 2 = x1 + x3 , x 3 = x2 , y = x2
x1

0 u1 u 2 x1 is not observable. After the state transformation
x2 = u1 0 u3 x2 +
x3 u2 u3 0 x3 x4 = x3 x1 , x5 = x 3 + x 1 ,

x1 x1 the unobservable state can be easily computed from
+D x2 x2
(6)
x3 x3 x 2 = x4 , x 4 = 2x2 , x 5 = 0.
1

0
If a system is input observable, the input reconstruc-
1 tion problem consists in the synthesis of a device or
2
a mechanism which has as input the measured out-
Output error

3
puts, and it should take place as output a signal that
4

5
should converge to the observable input. One can
6 easily notice that the input reconstruction problem is
7
closely linked to the problem of system inversion.
8
0 1 2 3 4 5 6 7 8 9 10
Time t

Definition 3.1 Consider the system (1). The input


Figure 2: Output error. u(t) is said observable if that input can be distin-
guished from zero by the output y(t), i.e., if y(t) = 0
The elimination procedure computes the states x2 = for t 0, implies u(t) = 0 for t > 0.
y, x4 = y, however the input-output relation does not
allows us to compute the state variable x5 ; y = x 4 = The input observability of LTI systems is almost
2y. The unobservable state is not asymptotically equivalent to their left invertibility, see [7]. Indeed,
stable, hence our system is not detectable. By using the input observability of
the general procedure by state elimination must fail. x(t)
= Ax(t) + Bu(t)
Indeed, from the observer equation y(t) = Cx(t) + Du(t),

x2 x4 x2 x2 is equivalent to the following property: The outputs
x 4 = 2x2 + D x4 x4 ,
y1 ,y2 , corresponding to the initial states x1 , x2 and
x 5 0 x5 x5 inputs u1 , u2 , respectively, are equal, if and only if
with asymptotically stable gain matrix-D, the state x1 and x2 are undistinguishable and u1 = u2 , [15].
x5 can not be eliminated. More direct relationship can be established between
these concepts, if the input set will be restricted to
Remark 2.1 1. Rigorous proof of theorem re- the smooth inputs
quires the consideration of the bad, the non-  (n1)

universal inputs, which constitutes the real ob- U = u; u(0) = 0, . . . , u (0) = 0 .
stacle to the existence of observers, see the sec- Indeed in this case input observability and system
ond part of Example. invertibility are equivalent. Thus, if the input is
observable, we can reconstruct it. This properties
2. Let us also distinguish the initialized and non-
have been used for fault detection and isolation,
initialized observers.
[14, 15, 21].
The non initialized observers work for any ini-
For nonlinear systems, the input observability and
tial state of the observer, while the initialized
system invertibility have been related, [20]. The sys-
observers only work for some initial state of the
tem inversion implies the input observability. Sys-
observer.
tem inverse can be obtain by state elimination.
The dependence of the observer on its initial
states originates another difficulty, see [12]. Let us consider a dynamic system
x(t)
= f (x, u) + g(x, u)(t), x(0) = x0
3 Input observability y(t) = h(x); (7)
In the problem of input reconstruction, the first task where the functions f, g, h are of appropriate dimen-
consists in evaluating the input observability, distin- sions. f, g, h are algebraic (analytic) functions, re-
guishing whether the changes of the input of a dy- spectively. is an unknown input. In this context,
namic system are reflected as changes at the output, the detectability (observability) of can be defined
[17]. by:
Definition 3.2 The input is said to be non de- 4 Binary distillation column control
tectable if for 6= 0 the relation case
y(x0 , x, u, 0) = y(x0 , x, u, ) A distillation column is one of the most impor-
is satisfied; otherwise the input is said detectable. tant example in process control theory and prac-
tice. Columns handle multicomponent feeds, how-
Consequently, if the input is detectable, then it is ever, many can be approximated by binary mixtures.
possible to construct an input observer. A single feed stream is fed as saturated liquid at
its bubblepoint onto the (m-th) feed tray. But it is
Input observer, which can be based on state elimina-
very difficult to obtain accurate, continuous mea-
tion, is an other dynamic system, with inputs u, y,
surements of the composition of nearly pure streams.

(t) = F (, u, u,
. . . , y, y,
. . .) The large holdup of liquid in the column, reboiler
(t) = H(, u, u, . . . , y, y,
. . .) (8) and reflux drum tend to make distillation control sys-
tems sluggish. Besides, there is also an effective
where in the limits (t ), 7 . The observer time delay of up several minutes before a change in
can be an inverse system, [15, 20]. The reconstruc- the feed or reflux stream is noticed at the intermedi-
tion of may take place in different information ate trays. Feed flow rate (mol/min) is measured, and
granulation, for example, : 1) is zero or not (bi- its composition is an unmeasured disturbance. Many
nary granulation); 2) is reconstructed as functions. efforts have been made to robustly control distil-
Nevertheless, each of these tasks can be solved by lation process, against the unmeasured disturbance.
system inversion. Unfortunately, the decoupling matrix of the standard
Thus, the detectability notion of can be formulated model of an idealized binary distillation column, is
based on the invertibility condition for the system (7) ill-posed, see [11]. Its decoupling matrix is almost
considering the state elimination: singular, becaused the compositions of successive
trays are closed to each other. Dual composition
Definition 3.3 is said detectable if control was proposed in [2]. Singular perturbation
there exists a differential polynomial technics lead to aggregated models, see [10]. Taka-
P (u, u,
. . . , y, y,
. . . , , ,
. . .) such that system matsu et al. have solved the disturbance attenuation
(7) with input and output y, is left invertible. Then, problem, using differential geometric methods, [22].
there exists a dynamic system (8), the inverse system The qualitative description of the stability of the bi-
from (u, y) 7 , which allows to reconstruct , nary distillation columns can be found in [16]. Our
supposing that (u, y, ) satisfies the non equality purpose is to design nonlinear observers, in order
to recover the unmeasured disturbance for the dual
P (u, u,
. . . , y, y, . . .) 6= 0.
. . . , , ,
composition control as it has been proposed in [18].
Remark 3.1 The role of the differential polyno- In our case, a binary system (two components) is
mial P (u, u,
. . . , y, y,
. . . , , , assumed with constant relative volatility throughout
. . .) is the invertibil-
ity condition: certain determinant of a functional the column, that is, the vapour leaving the tray is in
matrix is not zero. The inversion algorithm have equilibrium with the liquid on the tray. The rela-
been showed in [20]. tive volatility is a measure of the ease of separation.
This is the ratio between the tendency to vaporize
Thus, the existence of an input observer in order two components. The greater value means easier
to estimate an unknown input is based on a nec- separation (see Figure 3). A simple vapour-liquid
essary and sufficient condition for the invertibility equilibrium relationship can be used
of systems,[15]. We can notice that the observer x
can depend on the derivatives of the output. These k(x) = .
1 + ( 1)x
derivatives are also required to system invertibility,
[14]. On the other hand, the relative volatility is effected
Thus, the nonlinear state equations are
1
H1 x 1 = v(t)x1 + v(t)k(x2 ),
k(x), mole fraction in vapor

=10
0.8
=5
Hi x i = l(t) (xi1 xi ) +
0.6
=2
=1 +v(t) (k(xi+1 ) k(xi )) ,
0.4
i = 2, . . . , m 1,
0.2
Hm x m = l(t) (xm1 xm ) f (t)xm +
0

0 0.2 0.4 0.6 0.8 1 +v(t) (k(xm+1 ) k(xm )) + f (t)z,


x, mole fraction in liquid
Hi x i = l(t) (xi1 xi ) + f (t)(xi1 xi ) +
Figure 3: Relative Volatility. +v(t) (k(xi+1 ) k(xi )) ,
i = m + 1, . . . , n 1,
Hn x n = l(t) (xn1 xn ) + f (t)(xn1 xn ) +
+v(t)xn v(t)k(xn ),
by the pressure. The relative volatility reduces as
the pressure is raised. Besides, the relative volatility where y1 = x1 , z1 = xn are the top and the bottom
approaches unity as the pseudo-critical point of the components, respectively. Our purpose is to design
mixture is reached. a nonlinear observer and input observer to estimate
Any time delay will be neglected. We assume that the unmeasured states and recover the unmeasured
the vapour rate v(t) through all trays of the column disturbance z(t), in order to control robustly, the top
is the same, both at transient time and steady state and bottom compositions y1 and z1 .
as well. The liquid flow rate leaving the trays is
l(t). The measured feed flow rate is f (t), and the 4.1 Observer design
unmeasured composition is z(t) at time t, which is
considered as an unknown input. Figure 4 shows a The observer is based on the elementary state elimi-
schematic representation of the nomenclature used. nation procedure. In this case, due to the distillation
column nonlinear model, is not necessary to select
the gain D as we will show next.
Condenser
We notice that from the last equation
Tray1

(l(t) + f (t)) xn1 = Hn z1 + (l(t) + f (t)) z1


Reflux Reflux
v(t)z1 + v(t)k(z1 ), (9)
Tray2 v(t) Drum
hence xn1 can be expressed by z1 . Let us de-
Traym-1 v(t) l(t)
fine new, computed outputs z2 , z3 , . . . , znm+2 and
Traym l(t) y2 , y3 , . . . , ym+1 in order to express the unmeasured
v(t) x1 states x2 , x3 , . . . , xn1 , similarly to expression (9):
Trayn-2
n
Trayn-1 v(t) l(t) X
Reboiler
l(t)
zI = Hi x i +
i=nI+2
Trayn +(l + f v)xn , 2 < I n m + 2,
Bottoms y2 = (H1 x 1 + vx1 ),
xn yI = l(x1 xI1 ) vx1
XI1
Hi x i , 2 < I m + 1.
Figure 4: Distillation column.
i=1
Then, recursively Here, formula (10) is expressed in terms of the out-
1 puts, and xm+1 is also computed recursively from
xI = (znI+1 + vk(xI+1 )) , z1 , . . . , znm . Analogously, for i = m+1, . . . , n1,
l+f
I = n 1, n 2, . . . , m, the corresponding xi , xi+1 can be computed from
1 outputs z1 , z2 , . . . ,:
xm1 = (znm+2 + vk(xm ) f z) .
l
Hi x i = (l + f + 1) xi + vk(xi+1 ) + zni+2 +
On the other hand, 1
yI + (zni+1 + vk(xi+1 )) ,
xI = , I = 2, . . . , m. l+f
v + ( 1)yI i = m + 1, . . . , n 1,
Let us consider the sum of all equations of the Hn x n = (v l f )
xn + z2 + z1 xn .
dynamics using the expressions from ym+1 and
znm+1 :
The observer must be stable, i.e. the error e(t) =
znm+1 ym+1 lxm = f z (10) x(t) x(t) must tend to zero as t goes infinity. The
The equation (10) allows to reconstruct the unknown error equation is
input z (input observer).
H1 e 1 = (v + 1)e1 ,
Then, the observer is Hi e i = (l + 1)ei , i = 2, . . . , n,

H1 x1 = (v + 1) x1 +
  Hi e i = (l + f + 1)ei , i = m, m + 1, . . . , n.
y2
vk + y1
v + ( 1)y2
   In this particular case, it is easy to recognize that the
y i+1 observer gains are as follow
Hi x i = (l + 1) xi + v k
v + ( 1)yi+1
d1 = (v + 1),
 
yi
k
v + ( 1)yi di = (l + 1), i = 2, . . . , n,
lyi1 yi di = (l + f + 1), i = m, m + 1, . . . , n.
,
v + ( 1)yi1 v + ( 1)yi
i = 2, . . . , m 1, The vapour rate v(t), liquid flow rate l(t), feed

flow rate f (t), and the physical constants Hi , i =
Hm x m = (l + f + 1) xm + v k(xm+1 )
1, . . . , n are positive and slowly variants in the time,
hence the error equation is asymptotically stable.
 
ym
k
v + ( 1)ym
lym1 4.2 Controller design

v + ( 1)ym1 The dual composition problem supposes the mea-
(l 1)ym surement (modelling) of the quality of the top or bot-
+z
v + ( 1)ym tom product.

Hm x m = (l + f + 1) xm + v k(xm+1 ) The quality is measured by analyzers, for example,
  by using a final point analyzer. The final point tem-
ym perature may depend on the states, and the distur-
k
v + ( 1)ym bance terms. This dependence can be identified, and
lym1 it is supposed to be functions of all variables

v + ( 1)ym1
y1 (t) = gTop (x(t), v(t), l(t), f (t), z(t)),
(l 1)ym
+ znm+1 . z1 (t) = gBottom (x(t), v(t), l(t), f (t), z(t)).
v + ( 1)ym
The purpose is to control the top and bottom com- is obtained.
position. We suppose that top and bottom desired Considering the coupled system (11), (12), (13), x3 ,
composition are y1 (t) = C1 and z1 (t) = C2 , respec-
and x4 can be expressed algebraically in terms of l,
tively. Let us show the case n = 5, as an example.
v, f . For example,
Thus, the dynamics are written by
 
H1 x 1 = vx1 + vk(x2 ), v v
x4 = 1 C2 + k(C2 ). (14)
H2 x 2 = l(x1 x2 ) + v(k(x3 ) k(x2 )), l+f l+f
H3 x 3 = l(x2 x3 ) + v(k(x4 ) k(x3 ))
and
f x3 + f z,  
d v
H4 x 4 = (l + f )(x3 x4 ) + v(k(x5 ) k(x4 )), x 4 = (k(C2 ) C2 ). (15)
dt l+f
H5 x 5 = (l + f )(x4 x5 ) + vx5 v(k(x5 )),
y1 = x1 z1 = x5 . Similarly for x3 , k(x3 ), and k(x4 ).
Hence the computed outputs are The substitution of all terms is elementary but rather
tedious computation. The obtained equation is the
y2 = (H1 x 1 + vx1 ) = (H1 y 1 + vy1 ), implicit form of the control law with one free control
y2 term. For example choosing l(t), v(t) is computed.
x2 = ,
v + ( 1)y2
y3 = l(x1 x2 ) vx1 H1 x 1 H2 x 2 , For this example, let us suppose that C1 = 0.98,
C2 = 0.05. Thus, in the steady state, if Hi = 100,
y4 = l(x1 x3 ) f x3 vx1 F = 100, z = 0.04, = 5, then, v = 0.0501 and
H1 x 1 H2 x 2 H3 x 3 , l = 0.55. Let us choose v = 0.0501 and l(t) is
z2 = H5 x 5 + (l + f v)x5 , computed. Figures 5, 6, and 7 show the simulation
z3 = H4 x 4 + H5 x 5 + (l + f v)x5 , results:
z4 = H3 x 3 + H4 x 4 + H5 x 5 + (l + f v)x5 .
Hence,
f z = z3 y4 lx3 .
Now, let us suppose that the desired output com-
positions y1 = C1 , z1 = C2 are constant. Then,
x 1 = 0, x 5 = 0, and from v 6= 0 and 0 =
C1
H1 0 = vC1 + vk(x2 ), x2 = v(1)C 1
, obviously

C1 6= 1 is supposed. x2 is also a constant. There-
fore, the system dynamics are reduced to
  0.98

C1
0 = l C1 + v(k(x3 ) C1 ),
( 1)C1
(11) 0.97

0 = (l + f )(x4 C2 ) + vC2 v(k(C2 )) (12)


By substituting the unmeasured term
0.96

f z = H3 x 3 + H4 x 4 + (l + f v)C2 + (v l)C1
into the third equation
0.95 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

C1
H4 x 4 = (l + f )x3 l + (l v)C1
( 1)C1 Figure 5: x1 -vs- t.
v(k(x4 ) k(x3 )) + (v l f )C2 (13)
0.07

0.8
0.06

0.05
0.7

0.04

0.6
0.03

0.02
0.5

0.01

00 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

Figure 6: x5 -vs- t. Figure 7: l -vs- t.

measured feed composition estimation has been pro-


5 Conclusions posed. Unmeasured composition (unknown input) is
estimated by using the state elimination procedure.
The reduction to the state elimination of the observer
This procedure is simpler than the disturbance de-
design for nonlinear systems with prescribed error
coupling. The difficulty of the method is due to
dynamics and error outputs is an easy to use new
the presence of measured composition derivatives.
method. The difficulties are really propres ones of
However, low order dynamics allows us to use low
the nonlinear systems: the excitation and the univer-
order derivatives.
sality of the inputs, the dependence of the observer
convergence on its initial states.
References
The computational complexity is also hard, accord-
ing to the hard complexity of the elimination algo- [1] D. Bestle and M. Zeitz. Canonical form ob-
rithm. However, as in several example is shown, server design for non-linear time-variable sys-
if the elimination can be reduced into, for exam- tems. Int. J. Control, 38(2):419431, 1983.
ple Gauss elimination (output injection), then the
[2] Y. Creff, J. Levine, and P. Rouchon. Qualita-
method is effective. For linear systems, our method
tive behaviour of distillation columns and their
gives the same extended Luenbergers observers.
control. In 1th European Control Conference,
Input observability for nonlinear systems based on pages 10691075, Paris, 1991.
state elimination have been defined. The state elim-
ination procedure allows to establish an invertibility [3] S. Diop. Elimination in control theory. Math.
condition, which is lied to input observability condi- Control, Signals and Systems, 4(1):1732,
tion. 1991.
As applications example, the dual composition con- [4] S. Diop and M. Fliess. Nonlinear observ-
trol of a binary distillation columns based on the un- abilitty, identifiability, and persistent trajecto-
ries. In 30th IEEE Conf. on Decision and Con- to 15th IFAC World Congress (IFAC 2002),
trol, pages 714719, Brighton, 1991. Barcelona - Spain, 2002.

[5] S.T. Glad. Inonlinear regulators and ritts re- [16] H.H. Rosenbrock. A lyapunov function with
mainder algorithm. In Proc. Colloque Interna- applications to some nonlinear physical sys-
tional sur lAnalyse des Systemes Dynamiques tems. Automatica, 1(1):3153, 1962.
Controles, Lyon, 1990.
[17] H.J. Sussmann. Single input observability of
[6] D. Guillaume, P. Rouchon, and J. Rudolph. continuous time systems. Math Systems The-
Two simple observers for a class of polymer- ory, 12(11):371393, 1979.
ization reactors. In 4th European Control Con-
[18] F. Szigeti, A. Ros-Bolvar, and L. Rennola.
ference, pages TUEF5, Brussels, 1997.
Dual composition control of a binary distilla-
[7] M. Hou and R.J. Patton. Input observabil- tion columns based on nonlinear state observer.
ity and input reconstruction. Automatica, In IEEE Conference on Decision and Control
34(6):789794, 1998. 2000 , pages 45564557, 2000.

[8] A.J. Krener and A. Isidori. Linearization by [19] F. Szigeti, A. Ros-Bolvar, and L. Rennola.
output injection and nonlinear observers. Sys- Fault detection and isolation in the presence
tems & Control Letters, 3:4755, 1983. of unmeasured disturbance: Application to bi-
nary distillation columns. In 4th IFAC SAFE-
[9] A.J. Krener and W. Respondek. Nonlinear ob- PROCESS Conference , pages 10161019, Bu-
servers with linearizable error dynamics. SIAM dapest - Hungary, 2000.
J. Control & Optimization, 23:197216, 1985.
[20] F. Szigeti and A. Rvar-Bolvar. System inver-
[10] J. Levine and P. Rouchon. Quality control of sion and fault detection: The failure affine non-
binary distillation columns based on nonlinear linear case. In Proposed to MED02 Control
aggregated models. Automatica, 27(1):463 Conference , 2002.
480, 1991.
[21] F. Szigeti, C. Vera, J. Bokor, and A. Edelmayer.
[11] W. L. Luyben. Process Modeling, Simu- Inversion based fault detection and isolation. In
lation and Control for Chemical Engineers. 40th IEEE Conference on Decision and Con-
McGraw-Hill Publishing, New York, 1990. trol , pages 10051010, 2001.

[12] F. Olivier. Some theoretical problems in ef- [22] T. Takamatsu, I. Hashimoto, and Y. Nakai.
fective differential algebra and their relation A geometric approach to multivariable system
to control theory. In Proc. NOLCOS92, vol- design of a distillation column. Automatica,
ume 1, pages 301306, Barcelona, 1992. 15(2):387402, 1979.

[13] A.R. Phelps. On constructing nonlinear ob-


servers. SIAM J. Control & Optimization,
29:516534, 1991.

[14] A. Ros-Bolvar. Sur la Synth`ese de Filtres de


Detection et Diagnostic de Defauts dans les
Syst`emes Dynamiques. PhD thesis, Universite
Paul Sabatier, Toulouse, France, 2001.

[15] A. Ros-Bolvar, F. Szigeti, and G. Garcia. On


fdi filters and system invertibility. In Accepted

You might also like