Anda di halaman 1dari 6

WP8 - 5:30

STABILIZATION OF FEEDBACK LINEARIZED NONLINEAR


PROCESSES UNDER BOUNDED PERTURBATIONS
Jean-Paul Calvet and Yaman Arkun
School of Chemical Engineering

Georgia Istitute of Technology, Atlanta, GA 30332-0100


where s(t) 61R", d(t) E RP, u(t) E R. f(z) and g(s) are
smooth vector fields and W(s) is a matrix of appropriate dimension with smooth salar function entries. Note that this
particular nonlinear system is linear with respect to its control input and disturbances.
Under the state and input transformations

ABSTRACT
In this paper, we consider the problem of desgniag a stabiling
liser controlle for a feedback lnearized nonlnear system under
the inlsuence of bounded perturbatios. The feedbak linearized
(and prturbed) nonlinear sysem is equivalent to a Quasi Linear
System. A suficet condition using Lyapunov stability theory is
presnted to obtain a linear state feedback controLler stabilizing the
Quasi Linew System and the nonear system a welL ktmethod
is gi to obtin th least conserative (smalest) stbiliing gain.
The resu are applied on a chemical reactor.

z = T(s)
u

i = Az + Bv + C(z)d

The theory of differential geometry occupies a major position in the field of control theory for nonlinarsystems [1].
In recent years, sveral control applications of differential geometry have appeared in aerospace engineering [2], robotics
[3], power system [41 and chem7ical eneerg 17. However, the control literature deling with differential geometry
as applied to nonline systems perturbed by modeled disturbances is sparse. In this paper, we give some new resuts
addressing this problem.
Our recent work has shown that under the feedback linearization transormations a perturbed nonlinear systm is
in general no longer tranformed to a linear system but to
a so called Quasi Linear System (QLS) 181. This QLS is
then affected by nonlinearities only due to a state dependent perturbation coupled with the external disturbances.

(3)
(4)

(5)

(A, B) is any controlable matrix pair usually given in the


Brunovaky Canonical Form (BCF). The state dependent perturbation s(z)d in (5) is given by

ax=
[(aT)
Ia.-t(s)
where (j) denotes the Jacobian matrix. Necessary and sufficient conditions gurnteeing the existence of these transformations can be found in [8].

STABILIZING A QLS

The problem we addres in this paper is to stabilize a nonlinear system under the influence of disturbances described by
(1) through its equivalent QLS. More specificaly, we attend
to solve the following problem.
We assume that the set of (time varying) disturbances is
bounded by a know value D as in

In this paper, we give a sufficient condition which guarantees that such a QLS can be stabilized by a linear controller

in an arbitrarily small neighborhood of the origin in the presence of bounded disturbances. Equivalently, this meas that

the original perturbed nonlinear system is stabilised around

its operating condition. This sufficient condition is derived


from Lyapunov theory and the stabili gains of the linear controller are obtained by solving an Algebraic Riccati
Equation. However, because of its sufficiency, the method is
inherently conservative. Therefore, a procedure to obtain the
least conservative result is given. As an example, we apply
this method to an unstable chemical reactor perturbed by
various disturbances.
The following notation is used throughout the paper: An
eigenvalue of a square matrix W is denoted by A(W). If W is
strictly positive definite and symetric, then V'(W) and A\(W)

d(t) E OD = {d E RP; IkdlI < D}


and we seek a linear controller which will guarantee 8stability of the QLS (and equivalently stability of the nonlinear system). By 6-stabilization, we require the states of the
QLS to be ultimately bounded in any arbitrary small neighborhood of the origin defined by a ball B(6)={z E Rn; lzll <

8}.

The following definition adopted fom [9] clarifies the concept of 6-stabflization for a Quas Linear System.
Definition: A QIS is said to be 6-stabilizable if there
exists a linear state feedback control law v,(t) = -Ksz(t)
such that for all admissible disturbances in ID there exists
a T > 0 for which z(t) 6 B(6) Vt > T. The control vg(t) is
then called a -stabiizable control.

are its maximum and minimum eigenvalues respectively. Also


= A*(W)/A*(W) is its condition number. The notation
1111 s the usual Euclidian norm.

'w

TRANSFORMATION

We can now state a theorem which gives a ufficient condition for 6-stability of a QLS. First we require two key as

Consider a nonlinear system (NLS) of the form


i = f (x) + g(s)u + *()d
0 = f (zo) + g(o)uo

(4) +f(x)u

where T is a local diffeomorphism with T(xo) = O and (z) 5


O in a neighborhood Uo of so, the NLS can be transformed
to a Quasi Linear System described by

1 INTRODUCTION

sumptions.
A.1: The StAe dependent perturbation of the QLS lies in the
control input space, i.e. 3 X(z) s.t. s(z)d = BX(z)d

(1)

(2)

747

= HT > 0, R= RT much tha 2R1 -I > 0 and


> 0
[11]. However, this is not always true for the ARE'. As mentioned earlEir, the matrix pair (A, B) is usually the BCF.
With this particular form, Fig.l and Fig.2 give the plots of
the stabilizing gains for a 2- dinional ystem and for the

has a stattupper bound in a bal B(f):


A.2: X(z)
<'-+ IkII Vi E
3 (p,p) ER a.t.

!lIx()11'

B(E)

Theorem Given 6 suh that r > 6 > i;f thee exis a


P > 0 (strictly poitive definite) solving the Algebraic Riccati
Equation (ARE)

two Algebraic Riccati Equations (ARE and ARE'). For the


ARE we chose H = I and R = 1. However for the ARE'
the matrix P may not exist for low values of e. We thus had
to pick H = I and R = 0.1 which gave the largest range of
values of e for which the ARE' ha a solution. Note that, for
the ARE', the abilizing gains tend the infinity m e ten to
its lowest nonsero value (whih depends on R). However, for
the ARE, the gains tend to infinity - e tends to zr. These
two plots obtained will be umfu in the example. It is worth
noting that as e iacrcea, the stabilizngg decre. in a
mootonic way
I It quantiAssmption 2 is the mot cruial
fies in a way the 'mau of the nnliner peturbatio
through the value of two conan p and p. The challe
in this method is not to solve the Alpbraic Riccati Equation
which is usualy an easy tak; but rather, to find the constats p and p which are not unique so that Asumption 2 is

PA + AT P - PB(2R1C - I)BTP + He =0
for sm >0, H = Hr > 0 and R - fT > 0 satisying
a

>S

with

'-pp2D2
fA.(H)-p'OD2e

A(H) 9S2D2f' >0, the


-

(6)

the Contro IlW

9(t) - RR-'BPz(t)
(7)
is a 5-tabilizing control for the QLS for all initial conditin

B(r/1/j

with r- < .
Proof: See A edix.
it i importat to note that a F tends to infinity 5stabilion is asured for el initial oditiomsin R. Also,
similar to the one given
assumption 1 is a
inco
in 1101. However, this assmptoncan be relaxed. Indeed, if
Al is not stisfIe Coe ust vithe lwings
instead of 2:
in

saisfied.
In gral, assuming that the bound of the disturbances D
fied,
is Ib and the region of stabilization (ie. 5) is
thsmethod will give higher -n than nocem i order to
guarantee 5-stability doa QLS. enc, gien two ald pairs of
cosats (n, Pi) and (P2,p2), we will say that one.pair giVW
lee conservtive resl if fo a - set of p
Ho, R
and e,it give sl 5sabiliig gins tha the other -.
tib e
We illustae this fict in Fig.3: the cure gaiuine sos
gum wheras the '-curves C1 and
trend Of the s
C2 show the ted of V for twopuine of constanits (n.i) and
(n,pz). We -e that for c < (respectively e > ) the curve
cI
Cl (G) isbelowcurve.C (C1). Fort correpMdfi-toe
(e > r) curVe Ci (C) Wil give sl stabilizing - than
C (C1). Hence, we rcnzthat the curve below the other
results. In the example we will
one will gi e on
be able to detemine snag all the posible pairs of constat
p and p the one which will give the le conervaive -result.
According to the heorenem, for a specified Swe must have
this fixed value 5,
>5'. Fig.3shown that s a
. Therore the les consrvthe stabilizing gain d
S. We resume the
tive gai willb obtained as we let 6

A.?2: Te perturbation term itsf os be bounded in a ball


3 (p','W)E i as . tIj(z)IJ s p" +I11#12 Vi SE()
n, e abe Thom i hold but m
Wih this
has to hoe a diffeent Algebraic Rit Equatio (ARE')
given by
PA + ArP - P(2B7'B"- I)P + H/ =0
3.1 A Procedure to Deign 6Stabilizing Gains

BEcaus of the nuous number of p e , dening


6-sabiliin control may not be a eay tak. HOWever,
the
gains by
we can es the earch f es i
following d
First, one can se that the gains of the 5sabilizn control K = R1 BP do not deped explcitly on the co
p and p but rather on the parametersc, H and R Of the
Algebraic Riccuati Equt. Hence, ulb y one Wm take for
smity = I and then plot the gains K = R1B"P as a
ithm if necsay by R (mo
function of c and
often though we will alo take R = I so tha is the only
design parameter).
On the other hand, the constnts p and p materializndetly of the Aling the nolinearitie are -chosen i
gebraic Riccati Equation. Note that they ar not unique.
By calculating P from the Alebraic Rcati Equaton d
its condition number yp one can plot 5' a a function of f
parametrized by Rf. With this approah, the degner find
a lower bound on the size of the region of ultimate boundedness B(S') as a function of e. Then, for any 6 large than
V it is guaranteed that the gains obtained are 6-stabiliable.
This approach will be folkowed in the example.
It is kown that if the matrix pair (A, B) is contro
lable, then, there exists a saiqs P solving the ARE for all
a

procedure in four seps.


1. Solv the ARE (if 2 is

ld)

"tie) or ARE' (if A.2' is satis-

2. Fid the bet paip ad p whh gi

the least coaervativ

P'-crmve
ts. givea,pikP*so thatP60--5(but sWiUS> $).aad.setE
from the V-_curVe.
b

4. Caluae te

gais from the gain vase e cure

4 APPLICATION
As an illustzWaion, we now sow that the sandard CSTR
model perturbed by fuctuations in the feed temperatur and

the feed concentration falls into the category of nonlinearsystem descrbed by (1) and can be rnsformed into a QLS. We
will stabilize the CSTR under the influence of disturbance(s)
in an arbitrary neighborhood of an unstable operating coDi
tion by designing a linear stabilizing control for the equivalent QLS. The only knowledg required about the set of (time
varying) disturbace(s) will be its bound.

748

Transformation of a CSTR Model into a

4.1

time varying disturbances due to fluctuation of the feed temperature d1 and/or of the feed composition d2, can we design
a linear controller which will stabiize the QLS near the origin and henceforth the nonlinear CSTR model in an arbitrary
small neighborhood of the operating condition? More specifically, we want the state of the CSTR to remain bounded in
the domain

QLS

The dimenonles model of a first order exothermic ir.


versible reaction taking place in a CSTR is given by [12, 13].

f[2(X) ]I

[2]

][d2

:7

with

r f (Z) - [ -xi + Da(1

lf2(Z)

zi)ap[nez s]1

L-x + BD0(1-:l)cP [i.4

Z2

PI-(2-2.)

with
=
S(z)-[2

+(Z) [

4.2

(10)

,l

]-1

1+X2Iv)2Da cXp

D<;

1-la ca:p [*

T-(s)

(11)

where the inverse transformion - T-'(z) is obtained


through the one to one mapping of the state trandormation
x2

zi + x7

v n

fDt-+-1+-s

[22X-1-s
, I]/(
n

(13)

We solve this problem in the following section by designing


a stabilizing controller for the equivalent QLS. It is worth
stressing here that desgning a stabilizing control which will
keep the state of a nonlinear system in a prespecified domain
like in (13) under the influence of any but bounded disturbance(s) is not a trivial problem. Without the help of the
Theorem, the designer could easily pick a set of unacceptable
stabling gains. These gains could either not keep the state
of the CSTR in the prespecified domain or worse, the CSTR
could exhibit a 'runaway' behavior.

where z and x2 are the dimensionles composition and temperature respectively. Da, B, v, , and z22f are the standard
dimensionles parameters 112] and di and d4 are the dimensionles feed temperature and feed composition fluctuations
-espectively 17]. The control variable u is the dimensionles
temperature of the cooling jacket unit. Note that this system is indeed linear with respect to the control variable and
the set of disturbances; thus the CSTR model is of the form
described by (1). In the squel, we wil select the dimensionles parameters to be: Da = 0.072, B = 8, P = 0.3,
u = 20 and z2go = 0. With these parameters, the CSTR
exhibits an ipition/extinction behavior. We will operate at
the unable operating condition us = -0.20, :7 = 0.5 and
7 = 3.03 satids g (2). Now, consider the state and input
transformations
Z = T(z) = X1 - zp
(8)
Zs = T,(X) = fi(z)
vJ< dTC,,f > +u < dT2,g >
(9)
where c .,. > denotes the inner product. Note also that this
state pce coordinate transformation maps z9' to the origin
in the s-state space (i.e. T(t') = 0). Then under these
transformations, the CSTR model is transformed to the QLS

[~~~~~

- AXI1
X< X + AXn
:1
-4:X2 <C X2 < :X 2+ 4n2

Stabilization of the CSTR Through its


Equivalent QLS

It is very important to realize that Sstability for a QLS does


not guarantee 6-stability for the equivalent nonlinear system.
In other words, z(t) E B(S) is not necesarily saisfied. Indeed, one has to realize that any ball B(r) in thes-tate space
is trandormed through the inverse transformation (12) to a
closed contour (we call it an r-contour) in the x-state space.
This contour is not n arily the same ball B(r) unle the
diffeomorphism T is the identity. We ilustrate this mapping
in Fig.4 for the CSTR model. For our application, the specification of the bound Axi and Ax2 should be done judiciously
so that it correponds to a contour which is mapped to a
ball in the -stat spe domain. For example one can pick
Ax = 0.2 and Ax, = 1.37 which correspond to a ball in
the s-state space of radius 0.2. By designing a 6-stabilizing
control (S = 0.2) for the equivalent QLS the designer can
guarantee that such a controller will keep the state of the
CSTR inside the 8-contour (corresponding to B(S)) in the xstate space coordinate. In the remaining we will tabilie the
CSTR under bounded disturbance(s) inside the region with
S = 0.2.
4.2.1 Selecting the best constant pair (p,s)
In order to apply the Theorem, we need to specify the initial
condition of the CSTR. We will assume that the states of
the CSTR were initialy at the unstable (and disturbance
free) operating point: utP = -0.20, :7: = 0.5 and x7 =
3.03 which is mapped through (8) to the origin in the zstate space. By referring to the proof in the Appendix (e.i.
first step after claim) we recognize that the initial condition
is insde B(6/1:yp). Hence, if a 6-stabilizing controller is
built, we can be sure that under the influence of bounded
disturbances the states of the QLS wil not leave at all time
the ball B(6) (and equivalently the state of the CSTR will
not lea've the 6-contour). Since we want 6-stabilization for
6 = 0.2, we restrict the investigation of Assumption 2 to the
ball B(r) with r = 0.201. In Fig.5 we give a curve which
identifies the suitable pairs of constants p and jA satisfying
A.2. In this figure, only feed temperature disturbance was
considered; then, the matching condition is satisfied and we
have to verify A.2 with

(12)

It is important to realize that the one to one mapping


between the nonlinear model and the equivalent QLS is restricted to a neighborhood of the operating condition t' in
x-tate space and to a neighborhood of the origin in s-state
space. We can find the boundaries of this neighborhood.
First, the input tranformation requires P(s) =< dT2,9 > to
be nonzero, this is equivalent to :, 1 (note that si = I
means ful conversion). In other words, in the tate space
we are restricted to -:7 <ci
z< 1-:.
, Second, the inverse
0. Theise
state transformation (12) requires Z2 + z1 + x
constraints generate three straight lines plotted in Fig.4 in
the s-state space where we pick :7l = 0.5.
We are now in a position to state the practical sabilization
problem: Assuming that the CSTR is affected by bounded

749

X(z

F(

++2/TJ]

Do cr4

(with the pair pu and P4t) a a function of c for various


values of the bound of the disturbance d1 (i.e. d4(t) < D Vt).
Note that a D increese incree as well. This is not
a surprisg result snce this implies higher iling gan
as one would expect for large disturbances. As mentimned
earlier, the design requirement is 8stabilization for 6 = 0.2.
According to the Therem and the procedure we should have
8 > 6' and we can choose 8' arbitrarily close to 6 = 0.2.
For illustration purpome, we pick ?' = 0.199 and D = 0.3.
Then frm Fig.6 we obtain e = 0.74. We then use this value
of in Fig.1 in order to get the stabilizing gains K1 = 1.35
and K2 = 2.12. Simua ion (see Fig.7l with various kinds of
disturbances (sep 4(t) = 0.3 and periodic d4(t) = 0.3 sin(t))
bounded by D = 0.3 show that under the stabilizing control
V = -Kx (with K = [1.35,2.12) the stae of -the CTR
model remai at el time in the 8contour (8 = 0.2). One
should interpret these result as follows:
1. In gene, the bond D Mthi disturbance(s) being avaiabl
and the region of sahata (is. S) b specid, the
pin obtained are the mimi_n gain tha ths method cam
yield whc wil guaratee 8-sbility of th CSTR model
under sag disturbances bonded by D. Hence any highe
gains WOud be stabliig a welL
2. Thee resu we deived oma suffict conditina (see the
Theom). Thefor one could et obtain - sabit
wi even manligri- than the ainizm' gainm found;
or, torate a higher bond on the disbance (with the
same minimum gain) and still have stability.
Disturbance In the feed cmpoitin: In Fig.8 we give
8' (with the best pair p..g and p...) as a fnctionofefor
vaious values of the bound ofthe disturban d2. We already
know from setion 3.1 that when the tching condition is
not stisfied (which is the ce here), the AREp Jshave solution for low value of e. Also as the value of i
d as, the biling ga and the condition number yp
both tend to infinit. Hence we should expect 60 to go to
infinity e dcre . This is shwn in FigS. Note that this
trend does not exist when the matching condition is satisf
(compare Fig.6 and Fig.8). Because of this fact, we rise
that we obtain an upper hound for the mgnitude bound of
the disturbance over which we cannot find gain guarteeing 8sabiity. For example, for 8' = 0.199 we cannot obtain
8stabilizing gain (8> 8V) when the boud of the disturbce is larger than D = 0.02. Hence, in general, when the
matching condition is not satied, the method is limited in
d only for lS
the sn that stabing pgn can be
values of lYs. In Fig.9, we give the response of the CSTR
model under the effect of variouIsind of disturbanc (stop
di(t) = 0.01 and peiodic 2(t) = 0.Oisin(t)) bounded by a
low vaue D = 0.01 so that biliing gans be obtained.
For 6' = 0.199 we obtain i = 3.7 and the sabilin gai
are K1 =0.69 and K2 = 1.06. As expeted, the states of the
CSTR model remain in the 5contour at aU time.
It is importnt to reaie that when the matching condi
tion (A.l) is satified we can parantee 8-stability of the QLS
in an srbstrsniy small neighborhood of the orig. Indeed as
i tends to sero tends to zero a well (see Fig6). Hence the
region of 8-stabiity can shrink to the origin and in this cae

=T-'(.)

On the other hand, if only feed composition disturbance is


considered, the matchirng condition is not satisfied and we
thus have to verify assumption A.2' with

(z)= [1

DaexP[i']

In Fig.5, the area marked by

Is=T-1(.)

aU the

"yes' (no') identifies

values (p,p) for whih A.2 is (not) satisfied (a similar curve


can be plotted if -only feed compoition disturbance is considered). Also, for pairs on the line itself, the asumption is
still satisfied. This line

was

the state

obtained by

concentric circles. We first fixed pu. knowing


that necessrily puns > p,, with p., IIx(z)tII=o, and

space with

then obtained the suitable pj,,

the

sig; i.e.

esetp

the

B(

ball

i.e.

p =

0.201)

(p,gi)

eshow

line.

have

cons

wi

corr

p*

cure

by

p-r

any

similr argument any pair


mo

pair wh

to find the

Then the

>pg

that the

than

give

found

2 in the

To do so, we first select a p


Next idetify the
ph,pu) and pick any othe pa above

t
the curvep
HEnc, we realize that a

l1

Having

As

resul

below

we

(pu

theline(p,I

satfyin

wewne

PI >

on the line
of

satisfied with

is checed rplae X(z) above by g(z)).


psible pair

the let conrvative

give

A.2

that

I)iS)112

(if A.2'

so

vatie

othe
on

the

plot

of

afction
a

(mi,PuA)

by
(,) for all value

cated

on

the line

is

ofi.

is alway

Pair above the line. By the


ie

and -on the

result than any othe

lft

of point

on

the

We thus restrict our study to the segment AB where we


=
2 + fl-2
=epd..
0u

(14)

In order to find the beat pair, we need to minimize the


L
expresion of ? (withc fixed) as afuact
given by (14). We get easiy

of4pLwith

"Id82)) = 0 * A.(H) -

Yp = 0

then 2% << and thus as L


Hec,if D >
imereses, 8' decrease..Tefo,th best -ai for which 8'
is minimum is the point B = (42Q,). On the other hand,
then the bes pair for whichi * minimumi
if DYS < A
the point A = (pu.', p,'0). in our applcatio, the point
A will be the best pair (i.e. we will always have D'S <
The value of these constants are: (i) for d only
(Pn,Yent) = (0.45,1.51); (ii) for d4 only ( ) =
(2.40,13.22). A similar study was performed when disturbances in feed temperature and feed composition ue considered siimaneosly. In this cam, one has to verify awumption A.2' with j(z) given in (11). Note that the matching
condition is not satisfed. The values of the best constants
are: for d1 asdd0 (pn, ka)=(2.40,18.80).
4.2.2 Selecting the stabiling gains
Disturbance In feed temperatue: In Fig.6 we plot 8'

the sabilizing gains tend to infinity. Indeed, we -m in Fig.l


that as - 0, K' - oo. Note however that if the matching condition is not satisfied then 8-sbility in an arbitrarily
small neighborhood of the origin is sot guaranteed. Indeed,
in FigS we we that there is no value of for which 8* can
tend to zero.

750

L(z, t) < zT(PA+ATP PB(2R- 1 I)BTP)Z+p2D2+M 21IZII12rP


and from the Algebraic Riccati Equation
L(z, t) c< _zs f s +9D2 +pA21Z!12D2
< -4A.(H)1z112 +2D2 + x1212D2
In the second inveqality above, we ued the Rayleigh Principl [14].
Hence, the derivative of V inside B(r) is bonded by

CONCLUSION

This paper olves the problem of stabilizing a nonlinea aye


tem in a neighborhood of its operating condition under the
influence of any but bounded disturbance. The nonlinear
system is first wmmed to be feedback linearizable to a socalled QLS. Then a Lsiar controller is built stabilizing the
QLS and the nonlinear system as well. Two key asumptions are required: (i) a disturbance mtching condition and
(ii) a bound on the nonlinear state dependent perturbation.
Then we are able to deign a linear controller stabilzing the
nonlinear system in an arbitrarily small neighborhood of its
operating condition. The first asumption can however be
relaxed but the method, in our application, shows that we
can tolerate only disturbances with small magnitude and we
canot guarantee 8stability in any rbitrarily small neighborhood. We also give a procedure which leads to the lest
conservatwe results with this approach. Finally, this method
can be extended to 6-stabilization with no steady sate offset
of one output by adding integral action in the control law.
Results are available in [15].

-Q [11Z2
L(z,[t)

A4,H)

(15)

D2

pg2D2E2j

where Q _ [A lffD2eZ]. Beore we complete the proof, we


claim the folkming.
Claim: Let IIz(O)HI 5 r < .-E, then under the control law (7)
we have lIz(t)I] <c Vt > 0. Moreover, IIz(t)II d(r) Vt > { with
d(r)

{oyt-/pi
r,~--pj

if

if

r
< ro
r >ro

(16)
(8

and
Hp
p2D2

a6

(17)
a~ ~,1-1
,A.(H) - p2.02L2 -v5F
Proof Clahim: l I = maz(r,ro) then Ij](0)]j S P and
d(r) = f- .'p
2 f (7p > 1); thus IIz(0)II < 1 5 d(r). Suppose
3t4 8t. I1(hts)Il > d(r) but still I]z4jj < f
(18)
since (-) is contiuow and Ix(0)II < <
f 5 d(r) < IIz(ts)II, then
3t2 < ts a.t- I 2(t2)li= fand II(t)II . PYr E [t2, t5]
(19)
^I

REFERENCES
[1] Isidori A., Nonlinear Control Systems: An Introduction,'
(Lecture Notes in Control and Informatin Sceknces, VoL 72). New
York: Springer-Verag, 1985.
[2J Meyer, G., Su, R., Hunt, H. R., Automatica, VoL 20, No. 1,
pp. 103-107, 198.

[5] Tun, T. J., Beicy, A. K., Iidori, A., Chen, Y., Proc. 23rd
IEEE CDC, Las Vegas, NV, pp. 738-71, December 1984
[4] Marino, R., IEEE b-ans. Autom. Control, Vol. AC-29, No. 3,

But using the Rayleigh Principle and bom (19) we have:

A.(P)Iz(t8)112 < V(z(ts)) = V(X(t2)) + J7$ L(z(r),r)dr. Sice


IIXt)II < t OI 1t, t,] one has

Marh 1984.
[5] Hoo, K. A., andKantor, J. C., Chem. Eug. Commun., VoL 48,

385, 1987.

A.(P)iIs(ts)112 S AX(P)12

[8] Kravaris, C., and Chung, C. B., AIChE Jonal, VoL 33, pp.
592-603, 1987.
[7] Calvet, J.P., and Arkun, Y., Industrial and Chemistry Research, VoL 27, No. 10, pp. 1822-1831, 1988.
[8] Calvet, J-P., and Arkun, Y., Int. J. Control, Vol. 48, No. 4,
pp. 1551-1559, 1988.
[9] Schmitendorf, W. E., Int. J. Control, VoL 47, No. 1, pp. 8595,

f Q [IIX(r)112 -_

] dr

(2)

SinCe on [t2,t3j N(t)]


>. rO it mplies that Iz(fl20
2- 0
and thu on [t2,t3l we have Q [1II(r)Il2 -r] 2 0. Thefore the
integ in (20) is non negative; hence )t.(P)IIz(ts)]e . A'(P)f
or 1[4t3)]P < -YPf2 = [d(r)12 This contradict (18) and thu the

claim is proved.
We no'w complete the proof of the Theorem in two stps.
First; let 110()11 < r S /,f, and 6 < . From the claim we
have Ijl(t)]] < d(6/V9F) butt > g = ro, hence d(8/ffy;)
V z(O) E
/ =
we have
< SVt.
Second; let 110x()]] < r and r/t//i > r > 6/1V/IF, then asme

1988.

[10] Corla, M., and Leitmann, G., IEEE frana. Autom. Control,
Vol AC-26, No. 5, pp. 1139-1144, 1981.
[111 Kwakernaak, 3J N., and Sivan, R., 'Linear Optimal Control
Systems'. Wilsy-Interacience, 1972.
[12] Uppal, A., Ray, W. H., and Poore, A. B., Chem. Eng. Sci.,
VoL 31, pp. 205-, 1976.
[13] Ray, W. H., Proc. of Found. Coaf. on Chemical Proces
ControL Sea Island, Georgia, 1981.
[141 Frnklin, J. N., 'Matrix Theory'. Englewood CUiftr PrenticeHall, pp. 141-, 1988.
[15] Calvet, 3-P., and Arhun, Y., submitted to Compute and

IiI(t)I

B(6/8v')

TThus

IIz(t)jI> 6/,A/i

but IIz(t)1 c f v t E 10, T1

(21)

with

T A(p)(r2 - 62/7p2)
Q(62/yp -42)

Chemical Engineering.

(22)

Note th T is positive. Then


A.(P)]Iz(T)]P < V(X(T)) = V(z(O)) + if L(z(r), r)dr
)'(P)Ij_(0)lI2 - ifQ []Iz(r)lI' p.2] dr
Since
> 6// iTp we have -Qll(r)ll2 < -Q$!

APPENDIX
Proof: Consider a candidate Lyapunov function V(z) = TP:.
With the control law given in (7) and according to A.1 the cloedloop sysem is desribe by i = Az-BR'BTPs + BX(z)4.
Then the deivative of V is given by L(x,t) = V(x(t)) =
2TP(As - BT1BYPx + Bx(z)d)
But 2tNrPAx = x(PA + ATP)z ad inside B(r) we have

IIs(t)jI

A.(P)Iz(T112

,X (P) l(T)11 2

I7p

<

and thus

),*(p)Ij Z() 2 fW
f7 Q [03[44 0d-r or after inflPflx()]]2l

tegrting k.(P)Il$T)112

< A.(P)r2-[_ -] T Acc


ing to the definition of T in (22) and ater some algebra we get
/9p =
A.(P)IIz(T)112 < 2*(P)r2 - Y(P)(r2 - 62/2P)=X(P).

2TPBX(x)d < 21BPSz| tIx(?dll C 21IrPBII l Ix(z)I D


< IIBrPzII2 +IIx(z)IrD2 < s-PBBTPZ + (+p211z 12)D2
In the thd ineqaality we ued the inequality 21zy1 S x + u2

A2(P)62A*(P) or II(T)l[2<

~ T~

lip

'I-p

whichon1tra-

dicts (21), thu 3t2 E [0,Tj s.t. I]z(t2)jI < 6/y5/. Thea from the
caim we obtin Ilx(t)II < d(6/ifj) = 6 yVt t2. This conclude
the proof.

whereas in the forth inequality we used 2. Thus

751

MC

80

j. o
J
t0-

.-Jr,

Figure 1: Sbz

fi the Al%

Figue 6: Plos oi r

by i_(p from the ARE).

ti

'.

14

.4-

a
-L-

N1

I
l

is

Rtt

.W"

gue
v

':

is

.o i-tasof*ts

Lb

Yi

timt
of

han

de&d

di.

d-SM

FigureS;

pxthe

ata

/o

to

-'

8 OW
FigUM.NPILNt

05-

(7

by

i*
/

44

S/

t; th

0|

ARC).

0X74

Fiur 4:

Figure

15: Satging Am

0.201).

into 6-0

M*

_pto 2: A.2 'With di onY

j.5-s
a

*
Si:S-^#

im

Figure

t.) 0
3

9: Stabiliation of the CSTR inside the

various kinds of bounded disturbances

752

d2.

5-contour for

Anda mungkin juga menyukai