design of p and pi controllers for quasi linear systems

© All Rights Reserved

7 vistas

design of p and pi controllers for quasi linear systems

© All Rights Reserved

- lesson3 multiplying by arrays
- maths
- c
- Heat Transfer in a Composite Wall
- IMP - An Algebraic Formulation of Static Isotropy and Design of Statically Isotropic 6-6 Stewart Platform Manipulators
- How to Solve Linear Equations_ _ CareerAnna
- Canonical Forms
- A DSP-Based Implementation of a New Nonlinear Control for a Three-Phase Neutral Point Clamped Boost Rectifier Prototype
- Pressure Data - Ultramarine
- Force
- Get Start CS
- 02
- GL10
- (14)-GravityDam1
- Linear Algebra Complete Book
- InTech-Dynamic Characterization of Ancient Masonry Structures
- pbis lesson- overview of pride1
- International Journal of Computational Engineering Research(IJCER)
- The matrix method produces the solution in a compact form.docx
- Here Are My Tips for Math Exam and i Am a Student Like You Only So

Está en la página 1de 12

415426,

Printed in Great Britain. All rights reserved

DESIGN

Copyright

0098-I 354/90

$3.00 + 0.00

(0 1990 Pergamon

Press plc

OF P AND PI STABILIZING

CONTROLLERS

FOR QUASI-LINEAR

SYSTEMS

J.-P.

School

1990

of Chemical

Engineering,

and Y.

CALVET

Georgia

Institute

ARKUN~

of Technology,

Atlanta,

GA 30332-0100,

U.S.A.

Abstract-The

systems studied in this paper are nonlinear systems perturbed by disturbances which are

feedback transformable

to quasi-linear systems [i.e. i = AZ + Bv + [(r)d with (A, B) controllable].

We

consider the problem of designing stabilizing controllers for perturbed nonlinear systems through their

equivalent quasi-linear systems. With the addition of integral action, we can guarantee not only ultimate

d-stabilization but also zero steady state offset for both the output of the quasi-linear system (u = Cz)

and for the equivalent output [y = h(x)] of the nonlinear system. Moreover,

when the so-called

disturbance matching condition is satisfied, it is shown that all the states of the quasi-linear system (and

the nonlinear system) will exhibit zero steady state offset. All the results presented here are for single

control input systems.

1.

INTRODUCTION

In recent years,

the use of differential

geometry

to

transform

nonlinear

systems

into linear systems

has

received

much

attention

in the control literature.

Several methods compete and differ in their definitions and the issues they address. One approach is to

transform a nonlinear system in an input/output

sense. This was first investigated by Gilbert and Ha

(1984) and Ha and Gilbert (1987). Another approach

is to transform the nonlinear state equations into an

equivalent linear system without specifying any output. This technique is known as the feedback linearization and has been investigated in particular by Su

(1982) and Hunt er al. (1983). The transformations

required are state and input transformations with a

nonlinear state feedback. Finally, feedback linearization with decoupling of outputs was addressed by

Isidori et al. (1981). Various control applications

have appeared in aerospace engineering (Meyer et al.,

1984), robotics (Tarn et al., 1984), power systems

(Marina, 1984) and chemical engineering (Hoo and

Kantor, 1987; Kravaris and Chung, 1987; Calvet and

Arkun, 1988a).

It should be noted that there is very limited knowledge about the robustness characteristics of these

techniques with respect to modeling errors and disturbances. For example, the pioneering work of

Kravaris and Palanki (1988a, b) shows that under

matching conditions it is possible to design robust

controllers for a transformed (in the input/output

sense) nonlinear system with modeling errors. However, the design of controllers guaranteeing stability

of the transformed nonlinear system influenced by

modeled disturbances has not yet been investigated.

Our recent work (Calvet and Arkun, 1988a, b) has

~__

___

tTo whom all correspondence should be addressed.

415

nonlinear system is in general no longer transformed

to a linear system but to a so-called quasi-linear

system, QLS. The QLS is then affected by nonlinearities only due to a state-dependent perturbation

coupled with the modeled disturbance. In this paper

we address the following problem:

Given a bound on the set of disturbances affecting the nonlinear system, we want to design a

proportional (P) controller for the QLS which will

ultimately stabilize the original nonlinear system.

Furthermore, if an output of the non-linear system

is specified, we want to design a stabilizing proportional integral (PI) controller for the QLS which

will guarantee zero steady state offset for the

output.

The paper is organized as follows. The origin of QLSs

is given in Section 2. In Section 3, a theorem gives the

sufficient condition for the design of (P) stabilizing

gains. The extension to PI stabilization is addressed

in Section 4. Because of the inherent conservation due

to the sufficiency of the theorem, a design procedure

to compute the least conservative stabilizing gains

possible is given in Section 5. In Section 6, a steady

state analysis shows that zero steady state offset can

be guaranteed for all the states of the original nonlinear system under some disturbance matching conditions. Finally, in Section 7 we apply the design

procedure to control an unstable reactor influenced

by various process disturbances.

2.

QUASI-LINEAR

SYSTEM

Definition

I-A

quasi-linear system (QLS) is a

system which is linear or affine with respect to

its control input and the modeled disturbances.

416

It is described

J.-P.

CALVE?

by:

tzId,

i=Az+Bv+

(1)

d E RP is a set of modeled disturbances and c(z) is an

n x p matrix with smooth

scalar entries. The pair

(A, B) is a controllable matrix pair. It is usually given

in the Brunovsky

canonical form (BCF):

and Y.

ARKUN

to a QLS. The statedependent perturbation

term c(z)d in (1) is given by:

sary and sufficient conditions

guaranteeing

the existence of these transformations

can be found in Calvet

and Arkun (1988b).

3.

It is important

to realize that a QLS has the nice

property to be truly linear when free of disturbances

(d = 0). Also, note that the perturbation

term c(z)d

is state dependent and not parameter dependent [i.e.

<(q)d with 4 being a set of parameters] like it is often

the case in linear system theory. Though

there are

several techniques

to study the stability of linear

systems

with

parameter-dependent

perturbation

(Corless and Leitmann,

1981), the literature on the

stability of linear systems with state-dependent

perturbation

is sparse. It is the goal of this paper to

analyze and propose a method to study the stability

of such systems.

The origin of QLS appeared when the technique of

feedback

linearization

(Su, 1982) was implemented

on perturbed nonlinear systems. This was first investigated by the authors (Calvet and Arkun, 1988b), and

the pertinent results are summarized

in the sequel.

Consider

a nonlinear system (NLS)

of the form:

0 =f(xo)

+ g(x)u

+ Y (x)d,

+ g(xclh7

(2)

(3)

where x(t) E R, d(t) E RP, u(t) E R.J(x)

smooth vector fields and Y(x) is a matrix of appropriate dimension with smooth scalar function entries.

Also x0 is an equilibrium point for the control input

u = z+, and the set of disturbance inputs d = 0. Note

that this particular

nonlinear

system is also affine

with respect to its control input and disturbances.

Under the state transformation

and input transformation with feedback:

(4)

z = T(x),

u = a(x)

OF

QLS

we address in this paper is to

stabilize

a nonlinear

system

described

by (2, 3)

through its equivalent QLS. To do so, we assume that

the disturbances

are bounded,

and we look for a

suitable (linear) proportional

(P) control law under

which aN the states of the QLS will be ultimately

bounded.

The bound of the disturbances

is known

and it generates a closed set defined by:

i =.f(x)

P STABILIZATION

+ B(X)&

(5)

(not unique) with

T(x,) = 0 and p(x) # 0 in a neighborhood

U,, of x,,,

R,

= {d(t)

known bound. Note that the set of disturbances could

be time varying but should be bounded at all time by

D. We now give the definitions

of &-stabilization

(Schmitendorf,

1988)

and

ultimate

boundedness

(Corless and Leitmann,

1981) as adapted for QLS.

Definition 2-A

solution z(-):[t,,

co]+R,

z(t,) = z,,

(initial condition)

of a QLS is said to be ultimately

bounded

with

respect

to a closed

set B(S) =

{z E R; )lz[I <S}

if there

exists

a nonnegative

constant

(time)

T(z,,, S) < co possibly

dependent

and

6

but

not

on

t,

such

that

:;t, EZ&6) vr 2 t0 + T( z,, 6). We call B(6) a ball of

radius 6.

Definition &A

QLS is said to be &-stabilizable

if

there exists a linear state feedback

control

law

vg(t) = -&z(t)

such that for all admissible

(time

varying) disturbances

in R, the solution z(t) is ultimately bounded in B(6). The control v6 is then called

a 6-stabilizable

control.

In order to construct a linear d-stabilizable

control

law we require two key assumptions.

Assumprion

I-The

state-dependent

perturbation

term of the QLS c (z)d must lie in the control input

space:

is generated by the range of the matrix operator B.

This assumption

will be referred to as the disturbance matching

condition

or simply the matching

condition.

Assumption

Z-x(z)

has a state-dependent

upper

bound in a ball B(1):

3(p, p) E R2 (both

<p+~+II

VZEB(F).

This assumption is the most crucial one but can

always be satisfied in closed balls B(F) for a finite F.

However, if P tends to infinity, sometimes p and/or cc

may not be finite implying that Assumption 2 will not

be satisfied. In other words, this means that Assumption 2 can always be satisfied locally [in B(6)] but not

globaly (when i-00).

This choice of F is thus very

important and involves searching for the pair (p, p)

not in the whole space R but in an adequate ball B(F)

which must be specified judiciously. Also, this assumption introduces some inherent conservatism in

the design of 6 -stabilizing controllers. This point will

be discussed and illustrated later.

We now give an important theorem where the

following notation is used: an eigenvalue of a square

matrix W is denoted by 1( W). If W is strictly positive

definite, then 1 *(W) and 1 *(W) are its maximum

eigenvalues,

and

minimum

respectively.

Also

yw = E,*( W)/,l* ( W) is its condition number.

Theorem-Given

6 such that J 2 6 B 0; if there

exists a P > 0 (strictly positive definite) solving the

algebraic Riccati equation (ARE):

PA + ArP

for

some

satisfying

PB(2R-

6 > 0,

I)BTP

+ H/E

and

N=HT)O

=O,

R=Rr>O

417

6-stabilization is assured for all initial conditions

in R.

4. PI STABILIZATION

the notion of 6 -stabilization by adding integral action

in the control law. Indeed, with one manipulated

variable we can achieve zero steady state offset for

one output of the QLS. Note, however, that this type

of controller will guarantee zero steady state offset

only if the disturbances considered are ultimately

time invariant i.e.

3T* s.t. d(t) = constant

4.1.

Selection

v(l)=

law:

-R-B=Pz(t),

(7)

conditions in B(r/&)

with r -c F.

Proof-See

Calvet and Arkun (1989). The proof of

the theorem is based on the inspection of the sign of

the derivative of a candidate Lyapunov function

V(z) = zTPz; where P is the solution of an algebraic

Riccati equation whose form is different depending

on whether or not the disturbance matching condition is satisfied.

Note that when the QLS is a single control input

system, the parameter R is a scalar. Assumption 1 is

a disturbance matching condition similar to the one

given in Corless and Leitmann (1981). However, this

assumption can be relaxed. Indeed, if Assumption 1

is not satisfied, one must verify the following assumption instead of Assumption 2:

Assumption ?-The

perturbation term itself must

be bounded in a ball B(F):

3(p , p) E R2 (both finite) s.t.

lli(z)lj2 <p

+ ~l(z((~

Vz E B(P).

but one has to solve a different algebraic Riccati

equation (ARE) given by:

Vt > T*.

(8)

of the output

output of the original nonlinear system (2):

Y(l) = h]x(r)l,

(9)

the QLS, z,,

, z,. This has been recognized as a

disadvantage of the feedback linearization transformation (Kravaris and Chung, 1987). In the proposed

method, however, the output of the QLS must be a

linear combination of the states as:

y(t)=Cz(t)

with i,(H)

OF A QLS

CER~.

(10)

the unforced QLS observable. Hence without loss of

generality, we can set the output of the QLS to be Z, ,

i.e. C = [I, 0, . . , 01. Indeed, the unforced QLS can

be easily transformed to such a form through linear

transformation and feedback (Kailath, 1980).

The critical requirements for the output of the QLS

to satisfy simultaneously (9) and (10) can be tackled

in different ways:

Given a desired output (9) for the original

nonlinear system, choose a manipulated variable u which will admit feedback transformation

(4, 5) that gives a QLS with a linear output (10).

This is called an a priori output selection.

Apply the feedback linearization (4, 5) and then

select a linear output of the QLS (IO) which has

a physical meaning in the original nonlinear

system. This is called an a posteriori output

selection.

If these two approaches fail, one should resort to

partial linearization (Krener et al., 1984) or input/

output linearization (Kravaris and Chung, 1987)

which will not be addressed here. The option of

selecting an output a priori (before applying the

feedback transformation) or Q posteriori (after) has

been investigated in Kantor (1986) and applied by the

authors in Calvet and Arkun (1988a). In the reactor

application here, the selection of the output will be

done according to the procedure given above.

J.-P.

418

4.2.

The formulation

of the controller

CALVET

design problem

following:

Given the bound D of the set of disturbances, and

specifying 8, construct a control law which will:

1. a-stabilize the QLS (i.e. the norm of the states

will be ultimately bounded by 6).

2. Guarantee zero steady state offset of the output

of the QLS [under ultimately time invariant

perturbations (8)].

It is also important to realize that a-stability for a

QLS will automatically guarantee stability for the

original nonlinear system in a closed contour called

a 6-contour. This contour is obtained through the

inverse transformation [X = T-(z)] of the ball B(6).

Therefore, if stability of the original nonlinear system

is addressed through its equivalent QLS, the choice of

6 should be done judiciously so that it corresponds to

a desired region of stability, the 8-contour for the

original states x [see an application in Calvet and

Arkun (1989)].

The approach we suggest is to augment the dimension of the QLS and apply the P stabilization results

given in the previous section. First, we introduce an

additional state as:

z,+,(t)

z,(t)dz.

5 10

(11)

t = ;ir + Bv + c(z)d,

(12)

0

. . .

ERf+I,Xc+I>

0

i (jI

_ij=

0

0

(j

...

..I.

.

0

0;

and

0

0

0

and

constructed:

v(t) = -&Z(t)

= -[K,,

,=n

=

r-1

. . . , K,, K,+,]g(t)

c

I

Kjzi(t)

-K,+,

z, (t)dT,

.I@

(13)

and Y.

ARKIJN

where (K,, . . . , K,) and K,+,

proportional and integral gains.

The theorem given for P stabilization applies to PI

stabilization as well after replacing A, B, c(z) by

1, B, c(z). For example, if the disturbance matching

condition is satisfied by the augmented system (12);

then, one will have to solve the following algebraic

Riccati equation ARE:

P;i + ;i=P

PB(2R

Z)zP

+ H/e2

= 0.

be solved:

Pa + ;i=P

P(2BR

5.

DESIGN

-BT

Z)P + HJc2 = 0.

PROCEDURE

the search of the b-stabilizing gains. This procedure

applies for both P and PI stabilization of a QLS. It

is a graphical approach and is not analytical mainly

because a closed expression for the condition number

yp as a function of R, H and L is not available when

one solves the above algebraic Riccati equations.

5.1. Solving

ARE

and ARE

control K = R -BTP do not depend explicitly on the

constants p and ZJintroduced in Assumption 2 (Assumptions 2 or 2) but rather on the parameters t, H

and R of the algebraic Riccati equation. Assumption

1 will determine which Riccati equation needs to be

solved. Then all the possible gains can be computed

as a function of E (parametrized by R and H) and

independently of p and p. In other words, Assumption 2 need not to be checked before computing the

gains.

If the disturbance matching condition (Assumption

1) is satisfied, then, there always exists a unique P > 0

solving the ARE for all H = H > 0, R = RT with

2R --! - Z > 0 and t > 0. Indeed, one can recognize

that in this case the Riccati equation is similar to the

one encountered in the LQR optimal control problem (Kwakemaak and Sivan, 1972). As mentioned

earlier, the matrix pair (A, B) is usually in the BCF.

As an example, for a 2-D BCF and for R = 1 and

H = Z, we can plot the gains R -BTP

as a function

of L as shown in Fig. 1. Also, in Fig. 2, we give the

gains obtained when the QLS is augmented with the

integral action. Note that we can obtain the gains for

all values of es, and they tend to infinity as L tends

to zero.

However, if the disturbance matching condition is

not satisfied, we may not be able to solve the ARE

for all the values of the parameters t, H and R.

Indeed, we can show that for a given H and R there

exists a value of e, say climbelow which ARE does not

have a positive solution P. The value of +,,, depends

on the parameters H and R. In Fig. 3, we give the

0;

A

0 urn

419

EPS

disturbancematchingcondition).

with

R = 0.1. Also, in Fig. 4 are the gains for the augmented QLS. The parameter R = 0.1 was chosen

because it gives the lowest value of c for which ARE

has a positive solution P with N fixed to the identity

matrix (the value of R was obtained by trial and

error). In comparison with the gains obtained from

the ARE, these gains tend to infinity as t tends to cltm.

no gains can be computed.

Below Ellrn

Clearly, it is more attractive to solve the classical

algebraic Riccati equation (ARE) than the ARE.

Therefore the matching condition (Assumption 1) is

a desired property. In the case that Assumption 1 is

not satisfied one may wonder if either a similarity

transformation (change of state coordinate a = Qz)

or another suitable difleomorphism [z = T(x)]

would render Assumption 1 satisfied. However, this

is not true as one can easily show that the matching

condition Assumption

1 is independent of the

disturbance matching condition).

in the z-domain.

5.2. Verifving Assumption 2

Assumption 2 is the most crucial assumption. In

fact, this assumption is the only one dealing with the

nonlinearities of the QLS. It quantifies in a way the

magnitude of the nonlinear state-dependent perturbation through the values of the two constants p

and p. Therefore, the challenge in our design method

is not to solve the algebraic Riccati equation (ARE

or ARE) which is usually an easy task; but rather to

find the constants p and p which are obviously not

unique and are problem dependent. In the application we will refer to an algorithm which computes

these constants.

5.3. Conservatism

In general, assuming that the bound of the disturbances D is available and the desired region of

200,

of

2

2

EL

with disturbance matching condition).

QLS without disturbance matching condition).

420

8l

values

Fig. 5. Graphical

procedure

to select the

will give higher stabilizing gains than it is necessary.

Hence, given two pairs of constants @,, p,) and

&, p2) satisfying Assumption 2, we will say that one

pair gives less conservative results, if for a given set

of parameters t, n and R, it gives smaller S-stabilization gains than the other pair. We then have the

following result that will become clearer later on. If:

a,:.,,(~,

give less conservative (smaller) gains than b2, pr). In

other words (for H and R fixed) if the curve SF,,,, vs

.Eis below the curve 6,:. p2; then, the pair @, , p, ) will

be preferred. The curve 6 * vs l will have the following

shape:

If the disturbance matching condition is satisfied, the 6* curve is strictly increasing with 6.

Also as L tends to zero, S* tends to zero as well

(see the schematic plots in Fig. 5).

If the disturbance matching condition is not

satisfied; then, the S* curve has a minimum

(3E s.t. as*/&),_,=

0) and the 6 l curve goes to

infinity as c tends to tlim(see schematic plots in

Fig. 6).

5.4. Procedure

In light of the above results, we give two different

procedures depending on whether or not the

procedures help the designer to get the smallest

stabilizing gains once 6 and D are specified. The gains

obtained will guarantee that the QLS will be ultimately stable in B(8), and with no steady sate offset

for the output if the QLS is augmented.

With matching condition:

This procedure is schematically illustrated in Fig. 5

where each step number is circled.

1. Plot the gain curve (K vs C) with usually Ei = I

and R = 1.

2. Find @&, , past) so that a&(e)

is the lowest

possible (this will avoid conservatism).

3. Locate S and pick 6 * so that S r 6 * (as required

by the theorem).

4. Get E (from 6* vs e curve).

5. Get the stabilizing gains K (from K vs L curves).

One can see in Fig. 5 that for 6 given, we get smaller

gains as 6* tends to 6. Also it is now clear that if

another pair @, JJ) gives a 6 * curve above the one

depicted in Fig. 5 it will give higher gains. Hence Step

2 in the procedure is the most important step which

avoids conservatism. It is important to notice that as

6 tends to zero, then 5* must tend to zero and

henceforth L. The stabilizing gains will then all tend

to infinity. As a result, asymptotic stability of all the

states of the QLS can be guaranteed as K tends to

infinity. This important result is not true when the

disturbance matching condition is not satisfied.

Design

of P

and

PI

stabilizing

controllers

421

gains

K

Fig.

6. Graphical

condition).

decreases and the amplitude of the overshoot (if it

exists) and the steady state offset will increase. This

can be seen from Fig. 1. Since the offset is related to

S, the steady state offset will decrease as e decreases

(the Kis increase). Also, the addition of integral

action in the control law will increase the values of the

gains. A comparison of Figs 1 and 2 shows that for

a given c, higher gains are obtained with the use of

integral action.

Without matching condition:

This procedure

is schematically

illustrated

in Fig. 6.

possible.

2. Find (Pbeat,~,_,) so that 6$,(s)

is the lowest

possible (thus will avoid conservatism).

3. Obtain 6 zin

as*

[from 3C s.t. X

~_~ = 0 then S&, = S*(C)].

4. Select 6* > ~5%(and 6 > 6* from the theorem).

5. Get L (from 6* vs c curve).

6. Get the stabilizing gains K (from K vs 6 curves).

Hence, for all S > 6 * z=-S tin such gains are 6 -stabilizable. Once again, when 6* tends to S, the gains will

be smaller. Also the existence of ~5% shows that 6

cannot be arbitrarily chosen as small as we wish. AS

a result asymptotic

QLS cannot be guaranteed contrary to results with

matching conditions.

Also, if the disturbance matching condition is

satisfied but is not detected by the designer, stabihzing controller gains can still be obtained through the

procedure without the matching condition, but in

general, the results will be more conservative.

6.

STEADY

STATE

ANALYSIS

OF PI STABILIZATION

control law is to achieve zero steady state offset of the

controlled output under ultimately time invariant

disturbances. In this section, we show that when the

disturbance matching condition is satisfied, the control law will lead to zero steady state offset, not only

for the output but also for the other states of the

QLS. Equivalently, by virtue of the state transformation [the diffeomorphism (4)], this means that UN the

states of the original nonlinear system will exhibit

zero steady state offset. We illustrate this by performing a steady state analysis. With the control law:

s

I

,=1

aa = -

Kiz,(t) - Km+,

r-l

21 (rW7,

t=af+r:<Z)d,

422

J.-P.

CALVET

and Y.

ARKUN

7. APPLICATION

with

CSTR model perturbed by fluctuations in the feed

temperature and the feed concentration falls into the

category of a nonlinear system described by (2, 3) and

can be transformed into a QLS by feedback linearization. We will then stabilize the CSTR under the

influence of these disturbances using PI control. We

will make use of the plots of the stabilizing gains

obtained in Section 5.1 and then follow the procedures given in Section 5.4. The only knowledge

required about the set of (ultimately time invariant)

disturbance(s) will be the bounds.

7.1. Transformafion

of a CSTR

model

into a QLS

The dimensionless model of a first-order exothermic irreversible reaction taking place in a CSTR

is given by Uppal ef al. (1976) and Ray (1981):

2 &_,&-)d,=O

zp+

k -2,.

with

. ..n.

i=l

i=n

A(x)

i=p

K.zPq

1

n+

,zz+

1 +

C LiWM

i=,

zp=o.

= 0,

[ h(x)

(14)

Since the disturbances are all ultimately time invariant, all the terms df, i = 1, . . . ,p are considered

constant in the steady state analysis. Note that the

last equation is a natural consequence of the fact that

the control was designed to guarantee zero steady

state offset of the output z, , The above equations do

not simplify further. In particular z: for i = 2, . . . , n

cannot be obtained explicitly in terms of the

disturbances dis. However, if we assume that c(z)d

satisfies the disturbance matching condition with

respect to B as:

3~ (z) E R xp s.t. <(z)d

= &

(z)d.

;;i,S + & (z)d = 0,

which gives simpler relations

z

i=l,...,n,

Under PI 6-stabilization and ultimately time invariant disturbances, all the states of the QLS (and

the original nonlinear system) will exhibit zero steady

state offset if the disturbance matching condition is

satisfied.

-x1

fDU(1

--z+BDa(l

-X,)eXP[&]

-x,)exp[~]-B(x,-x*~~)

and temperature respectively. Da, B, v, fi and xzcOare

the standard dimensionless parameters (Uppal et al.,

1976) and d, and d2 are the dimensionless feed

temperature and feed composition fluctuations, respectively (Calvet and Arkun, 1988a). The control

variable u is the dimensionless 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 (2). In the sequel, we will select the

dimensionless parameters to be: Da = 0.072, B = 8,

B = 0.3, v = 20 and x2 = 0. With these parameters,

the CSTR exhibits anignition/extinction behavior.

We will operate at the unstable operating condition

uOp= -0.2, x;p = 0.5 and x;p= 3.03 satisfying (3).

Now, consider the state and input transformations:

-xy.

z, =

T,(x)=x,

z2 =

T2(x) =f,(x)

116)

and

u = (dTzrf)

+ u(dTl>g).

(17)

that this state space coordinate transformation maps

xP to the origin in the z-state space [i.e. T(xP) = 01.

Then under these transformations, the CSTR model

is transformed to the QLS:

423

ARE where we picked n = I and R = I (see gain

plots in Fig. 2).

In order to verify Assumption 2, we inspect the

inequality in the ball B (f = 0.2). Among all the

with

-1

-1

x2

-Daexp

[

is obtained through the one-to-one mapping of the state

transformation:

~

1 +x2/v

2

II

(19)

X- T- 1(Z)

giving the lowest 6* curve is:

x, = Z, + x;

II

z2

+z,+xy . (20)

v

( [

I>

x2 = v In

2,

z, + x;)p

- ~7)

[ Da(1 -z,

In

Da(1 -z,

-x9)

Remark-We

will consider X, (i.e. the dimensionless composition) to be the output of the nonlinear

system. According to the state transformation (16)

such output corresponds to z,. Hence, the required

conditions (9) and (10) from Section 4.1 are satisfied

simultaneously and PI stabilization can he applied

with z, as output. However, if we would have chosen

x2 as output for the CSTR, then the residence time

should have been the new manipulated variable rendering the nonlinear system transformable to a QLS

with an output z, corresponding to the dimensionless

temperature. Here, the flexibility to choose a new

manipulated variable as mentioned in Section 4.1

indeed exists.

7.2. PI stabilization

tions (d, = 0)

under feed

temperature perturba -

where

(K,, K2) and K3 are, respectively, the proportional

and integral gains to be determined. In this case, the

state-dependent perturbation term satisfies the disturbance matching condition with:

(l-

(1 +&Iv>2

x) Da exp[*]]_

law show that as expected both states of the CSTR

exhibit zero steady offset under a step disturbance of

d,(t) = 0.3. As a performance criterion we can also

obtain the integral below the curve z, (t) = x,(t) - xyp

vs time. Indeed according to the construction of the

augmented QLS (1 I, 12) and the steady state equations (15) we have:

z? = lim C z, IT) dr = xod,

Consider that the CSTR is subject to feed temperature perturbations only. With a PI stabilizing controller we then have the augmented QLS described

by:

X(z)=

by the authors and is available in Calvet and Arkun

curves as a function of L and

(1989). a:__,,.,,

parametrized by various bounds D of the disturbance

d, are displayed in Fig. 7. According to the procedure

(with matching condition) we can now compute the

stabilizing gains of the PI controller. Let d = 0.2 and

6* = 0.199; also we consider ultimately time invariant

disturbance bounded by D = 0.3. The a-stabilizing

gains are then:

r_,(Z1

I-CC

Jo"'

JG

= 0.048

Remark-Note

that under the condition of disturbance matching condition, it is not necessary to

here dimensionless composition or temperature).

Indeed, integral action on z, {no matter what its

relationship with the original nonlinear system may

mean) will guarantee zero steady state offset of all the

states of the QLS and henceforth of the states of the

original nonlinear system (i.e. here the CSTR).

424

00

0.24

0.20

0.16

0.12

0.06

0.520

0.515

OH0

z

z

3.06

3.10

3.16

3.12 3.14

c!

x

0.505

k

a495

I

2

matching condition).

7.3. PI stabilization

tions (d, = 0)

under feed

composition

perturba-

Fig.

8. PI stabilization

admissible

are:

perturbation

only,

then, with a PI stabilizing controller,

the augmented

QLS is described by:

law is again u(t)=

-%I:

K,z,(t).

One can easily see that the disturbance

matching

condition is not satisfied. Then the algebraic Riccati

equation to solve is ARE where we picked n = I and

R = 0.1 (see gain plots in Fig. 4). With the algorithm

given in Calvet and Arkun (1989) we get the pair

(p. p) satisfying

Assumption

2 in B (P = 0.2) that

gives the lowest 6* curve (for n = I and R = 0.1) vs

L. The pair is:

(P bcs,, p~brrt)= (2.4,

3D2

1

a

Time

EPS

Fig. 7. 6 * curves

I

6

1

4

(with disturbance

tion).

value of 0.006.

[K,, K,,

Then

= [2.44,

&]

matching

condi-

the stabilizing

gains

1.86,

1.21.

A simulation

in Fig. 11 with these gains implemented in the control law shows that, as expected,

only the output of the QLS i.e. z, will exhibit zero

steady state offset. However,

the other state z2 will

exhibit

a steady

state offset.

By virtue of the

diffeomorphism

T, this corresponds

to zero steady

13.22).

dFbt._

curves as a function of L and parametrized by

various bounds D of the disturbance d2 are displayed

in Fig. 9. One can see that, as a result of the absence

of disturbance matching condition,

all the 6* curves

parametrized

by D have a minimum S&(D).

Therefore, 6 cannot be as small as we may wish. For

example, a-stabilization

with 6 = 0.2 cannot be guaranteed for disturbance

having a bound larger than

0.0065. This can be seen in Fig. 10 where we plotted

S,&(D)

as a function

of D. Indeed,

such curve

gives regions where a-stabilization

can or cannot be

implemented.

According

to the procedure

(without

matching

condition),

we can now compute the stabilizing gains

of the PI controller.

Let d = 0.2 and 6 = 0.199. As

a bound

D for the disturbance

dZ we picked an

a30

00

t

0.25

0.20

0.15

0.10

Fig. 9.

Ek

I

6

curves vs c parametrized

by D (without

bance matching condition).

I

7

distur-

425

0.30 -

0.25 8 -stabilization

posaibls

0.00

8-stabilization

not

possible

IILLI

0.01

0.001

D

Fig. 10. Region of d-stability (without matching condition).

Fig.

composition

x,

and a steady state offset for x2 the dimensionless

temperature.

BCF)

f(x), g(x) = Smooth vector fields in R (infinitely

differentiable i.e. C)

Y(x) E WXp = Perturbation matrix associated with the

disturbances

z = T(x) = A

nonlinear

one-to-one

mapping

(diffeomorphism)

dT/ax = Jacobian matrix of T

u = S(.X, U) = A (single) input nonlinear transformation with state feedback

B(6) = Ball of radius 6

r, P = Real positive constants denoting the

radii of balls E(r), B(T)

6,6* = Real positive constants denoting the

radii of balls B(d), B(d *)

p. p = Real positive constants satisfying Assumption 2

D = Real positive number, bound of the

disturbance(s)

R, c RP = Set of disturbances

c(z) = aT/ax Y(x)

= Perturbation of the QLS

X(Z)ER rp = Row vector satisfying Assumption I

L = Real positive number (parameter of the

ARE)

R, W = Matrices of the ARE (parameters)

,I(~) = Eigenvalue of a square matrix

,I*(.), A* (~) = Maximum and minimum eigenvalues

of a positive definite matrix

v(.) = 1 *(-)/A * (.) = Condition number

P z 0 = Solution of an algebraic Riccati equation (ARE or ARE)

8.

CONCLUSION

The theory and application of the P and PI stabilization of quasi-linear systems (QLS) is presented. The

origin and practical importance

of QLSs is introduced, and the concept of ultimate boundedness

and

&-stabilization

is adapted

for

such

systems.

A

theorem and a procedure

are given to compute the

stabilizing gains in the least conservative sense for the

proposed methodology.

The results permit to stabilize the class of nonlinear

systems with bounded

disturbances

which are feedback

transformable

to

QLSs.

If the so-called

disturbance

matching

condition is satisfied, we show that PI stabilization

will

guarantee zero steady state offset, not only for the

output but also for all the other states of the QLS

(and henceforth

for the original nonlinear system).

Simulation

results on an open-loop

unstable (and

perturbed) CSTR mode1 illustrate and agree with the

theory.

NOMENCLATURE

R = Set of real n vectors

norm for x E R, IIxII =

I/x II = Euclidean

R"

Xl_, xf

m = All n x m real matrices

I E R x = Identity matrix

x E R = States of a system (usually nonlinear)

z E R = States of a system (usually linear or

quasi-linear)

y = h(x) = (Single) output of a system

d E RP = Disturbance vector

24= (Single) control input of a system

(usually nonlinear)

zi= (Single) control input of a system

(usually linear or quasi-linear)

K E R = Stabilizing gains

Il.

PI stabilization (without

condition).

disturbance

matching

Abbreviations

ARE

BCF

QLS

P

PI

=

=

=

=

=

Brunovski canonical form

Quasi-linear system

Proportional

Proportional integral

REFERENCES

Calvet J.-P. and Y. Arkun, Feedforward and feedback

linearization of nonlinear systems and its implementation

using IMC.

Znd. Engng

Chem.

Rex. 27, 1822-1831

(1988a).

Calvet J.-P. and Y. Arkun, Feedforward and feedback

linearization of nonlinear systems with disturbances. Znt.

J. Control 48, 1551~1559 (1988b).

J.-P. CALVET

426

Calvet J.-P. and Y. Arkun, Stabilization of feedback linearized nonlinear processes under bounded perturbations.

1989 ACC, Pittsburgh (1989).

Corless M. and G. Leitmann, Continuous state feedback

guaranteeing uniform untimate boundedness for uncerain

dynamic systems. IEEE

Trans. Automat.

Control

26,

1139-1144 (1981).

Gilbert E. and I. J. Ha, An approach to nonlinear feedback

control with application to robotics. IEEE Trans. System.

Man. Cyhernet.

14, 879 (1984).

Ha I. J. and E. Gilbert, Robust tracking in nonlinear

systems. IEEE Trans. Automat. Control 24, 763 (1987).

Hoo K. A. and J. C. Kantor, Linear feedback equivalent

and control of an unstable biological reactor. Chem.

Engng

Commun.

46,

385 (1987).

systems. D&$erential Geometric

Control

Theory (R. W.

Brockett,. R. S. Millman, and H. J. Sussman, Eds), p. 268.

Birkhauser, Boston (1983).

Isidori A., A. J. Krener, C. Giri-Giorgi and S. Monaco,

Nonlinear decoupling via feedback: a differential geometric approach. IEEE Trans. Auromar. Control 26, 331

(1981).

Kailath T., Linear Systems. Information and System Sciences

Series. Prentice-Hall.

Englewood Cliffs, New Jersey

(1980).

Kantor

J. C.,

Proc.

ACC,

Seattle, pp.

1014-1016

(1986).

Kravaris C. and C. B. Chung, Nonlinear state feedback

and Y. ARKUN

synthesis by global input/output linearization. AIChE

33, 592-603

Jl

(1987).

under structured uncertainty. AIChE JI 34, 1119-I 125

(1988a).

Kravaris C. and S. Palanki, A Lyapunov approach for

robust nonlinear state feedback synthesis. IEEE Trans.

Automat.

Control 33, 118&l 191 (1988b).

Krener A., A. Isidori and W. Respondek, Proc. 23rd IEEE

CDC. Las Vegas, pp. 126130

(1984).

Kwakemaak J. N. and R. Sivan, Linear Optimal Control

Systems.

Wiley, New York (1972).

Marino R., An example of nonlinear regulator. IEEE Trans.

Automat.

Control 29, 000 (1984).

Meyer G., R. Su and H. R. Hunt, Application of nonlinear

transformations to automatic flight control. Automatica

20,

103-107

(1984).

Sea Island, Georgia (1981).

Schmitendorf W. E., Stabilizing controllers for uncertain

linear systems with additive disturbances. ht. .J. Control.

47, 85-95 (1988).

Su R., On the linear equivalents of nonlinear systems.

System Control Letrs 2, 48-52 (1982).

Tam T. J., A. K. Bejczy, A. Isidori and Y. Chen, Nonlinear

feedback in robot arm control. Proc. 23rd IEEE CDC,

Las Vegas, pp. 736751

(1984).

Uppal A., W. H. Ray and A. B. Poore, Chem. Engng Sci.

31, 205 (1976).

- lesson3 multiplying by arraysCargado porapi-302026639
- mathsCargado porgurpreet3131
- cCargado porvvajaitu3250
- Heat Transfer in a Composite WallCargado porravidahiwala
- IMP - An Algebraic Formulation of Static Isotropy and Design of Statically Isotropic 6-6 Stewart Platform ManipulatorsCargado porHareesha N G
- How to Solve Linear Equations_ _ CareerAnnaCargado porKat Gupta
- Canonical FormsCargado porBrahadeesh Perinkolam Murali
- A DSP-Based Implementation of a New Nonlinear Control for a Three-Phase Neutral Point Clamped Boost Rectifier PrototypeCargado porThái Vĩnh Thành
- Pressure Data - UltramarineCargado pornaseemtyche
- ForceCargado porMANOJ M
- Get Start CSCargado porAnjireddy Thatiparthy
- 02Cargado porNandu Mahesh
- (14)-GravityDam1Cargado porRazvan
- GL10Cargado porsuyogbhave123
- Linear Algebra Complete BookCargado porSyed Muhamad Jahanzeb
- InTech-Dynamic Characterization of Ancient Masonry StructuresCargado porFrancisco Calderón
- pbis lesson- overview of pride1Cargado porapi-233954409
- International Journal of Computational Engineering Research(IJCER)Cargado porInternational Journal of computational Engineering research (IJCER)
- The matrix method produces the solution in a compact form.docxCargado porAlex Tkr Tkr
- Here Are My Tips for Math Exam and i Am a Student Like You Only SoCargado porAashishSethi
- Validation of FACTS ModelsCargado porGerman Ruita
- Lecture 11Cargado porBharathwaj Sreedhar
- Further Partial Differential EquationsCargado poralexandrosmf
- CCargado porOTorresGonzalez
- ScienceCargado porvanderlei.medeiros
- Cilamce2017 0805 Papic CamilaufpeCargado porMarlos José Ribeiro Guimarães
- Calculator TricksCargado porKristine Rose Vergara
- Vector 2Cargado porstefan
- mathematics_syllabus.pdfCargado porKushlendra Sinh
- Aa547 Lecture Lec14Cargado porBabiiMuffink

- CONTROL SYSTEMS from The Electronics Engineers' Handbook.pdfCargado porAlda England
- KSA Driver's ManualCargado porAhmad Rahan
- Entrepreneurial Actions Innovation and Appropriability (SEJ 2007) Burgelman & HittCargado porAhmad Rahan
- General Technical Engineering TermsCargado porAhmad Rahan
- Local Invariant Set TheoremCargado porAhmad Rahan
- The Complete Guide to Simple OEECargado porWan Sek Choon
- Large Dictionary EnglishCargado porAhmad Rahan
- Model Predictive Heuristic Control: Applications to Industrial ProcessesCargado porAhmad Rahan
- CFBC Boiler - A SurveyCargado porgjanklesaria
- 1997_Verbanck_Development of a MathematicalCargado porSelva Ram
- sliding mode in control and optimizationCargado porAhmad Rahan
- Stochastic Modeling of Indoor Air TemperatureCargado porAhmad Rahan
- STABILIZATION OF FEEDBACK LINEARIZED NONLINEAR PROCESSES UNDER BOUNDED PERTURBATIONSCargado porAhmad Rahan
- Sliding Mode Control_SlotineCargado porAhmad Rahan
- Boiler Efficiency CalcCargado porsugandaraj522
- RCER 2004 RegulationsCargado porStephen Le
- Modeling and optimization of the NOx emission characteristics of a tangentially firedboiler with artificial neural networksCargado porAhmad Rahan
- SE 517 Lecture_03(2)Cargado porAhmad Rahan
- Introduction to LabVIEWCargado porRamaDinakaran
- non linear system for DC motorCargado porAhmad Rahan
- MATLAB Function BlockCargado porAhmad Rahan
- Softsensor for Estimation of Steam Quality in Riser Tubes of BoilersCargado porAhmad Rahan
- Journal Paper-Influence of CombustionCargado porAhmad Rahan
- Drum Boiler ModelingCargado porAhmad Rahan
- Sequential Algorithms for Multiplication and DivisionCargado porAhmad Rahan
- IC Proposal PresenationCargado porAhmad Rahan
- Doc 820114Cargado porbaskhem
- Matrixcookbook WikiCargado porAhmad Rahan
- How to Calculate OEE – The Real OEE Formula with Examples _ Lean Execution - Intelligent MetricsCargado porAhmad Rahan

- Lipschutz-LinearAlgebraCargado porAndres Tejedor
- scilab5bCargado portamuztech
- UHoensch Differential Equations and Applications Using Mathematica.pdfCargado porJhoan Joaqui
- Geometric Algebra and Its Application to Mathematical PhysicsCargado porhomiya88
- mathsiCargado porapi-236544093
- Nineteen Dubious Ways to Compute the Exponential of a MatrixCargado porsportingdan
- 1Cargado porMohd Nasim
- palm_ch2Cargado porA Raheem Siddiqui
- 4024_y04_sw_10Cargado porRida Snob
- Teaching General Relativity by WaldCargado porjpjk2128
- Lecture2, MatricesCargado porHisham Mostafa
- Inverse MatrixCargado porΈνκινουαν Κόγκ Αδάμου
- math110s-hw2solCargado porankit
- CG–UNIT-VCargado porSri Sri
- Rudiments of Io AnalysisCargado porJay-ar Mira
- MIT2_29F11_samplequiz1Cargado porgokuler137
- Notes on the equivalence of normsCargado porEdgar Marca
- elk-20-3-8-1003-413Cargado porSai Siva
- mc graw hill ryerson calculus and vectors 12, chapter 7 answers with stepsCargado portrachea459
- hw1_solsCargado porJose guiteerrz
- diffgeometry.pdfCargado porerlandfilho
- kpk-9th-maths-ch01.pdfCargado pormr. khan
- ModalMassCargado porDeana White
- EigenValues.pdfCargado porSayan Bhattacharyya
- Helmholtz DecompositionCargado porSebastián Felipe Mantilla Serrano
- 1pdf.net Angular Momentum University of Notre DameCargado porLam Sin Wing
- 04 Matrix OperationsCargado porValan Pradeep
- Inverse Laplace Transform Lecture-3Cargado porSingappuli
- BC0043 COMPUTER ORIENTED NUMERICAL METHODS PAPER 2Cargado porSeekEducation
- Lecture Notes Banach Spaces Chapter 1Cargado porKitch Lesner

## Mucho más que documentos.

Descubra todo lo que Scribd tiene para ofrecer, incluyendo libros y audiolibros de importantes editoriales.

Cancele en cualquier momento.