Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Control of Multivariable Systems

I
M. D. MESAROVIC

Introduction case; however, in the multivariable system the kind of speci-


fication must be stated at the beginning of the study. The first
The control theory of multivariable systems has been regarded kind of specification does not lead towards a unique solution in
mainly as an extension or generalization of the control theory the synthesis of a system; the second kind of specification could
of single variable systems; it is, however, logically inconsistent determine the terminal behaviour but the structural behaviour
to consider it so. Between the single and the multivariable remains indeterminable, while the third kind of specification
systems there are basic differences which are more than just the uniquely determines the system.
difference in the number of variables associated with the
system; consequently new problems appear in the considera- Principal Objectives of a Control Theory of MuItivariable
tion of multivariable systems that either do not exist or are Systems
trivial in the single variable systems. However, a successful
The development of a control theory for multi variable
treatment of these problems is sine qua non in any serious
systems has been divided into three parts:
attack on multi variable systems and should present a base for
(1) Analysis and synthesis of multivariable systems: The non-
a new control theory of such systems.
existence of this theory greatly handicapped the control theory.
There are many good reasons why the control of multi-
(2) Synthesis of the control part of the system: After the
variable systems is of great importance for the further develop-
system behaviour is determined the control elements can be
ment of the science and art of control in general. Let us
synthesized in a straightforward way.
mention here two of them.
(3) The analysis of the performance of the system: For a
(a) Consideration of the control of physical systems using
given controlled process and a given controller configuration
signal models has led to the cybernetical approach in which
the particular characteristics of the controller should be
one is interested exclusively in the quality of control; however,
determined according to the improvement in the stability or
the qualitative and quantitative attributes of a system are
optimal performance of the system. The methods developed
inseparable. This becomes apparent in any study of an actual
for this purpose in single variable control are applicable
physical system rather than of a simplified model of the system;
analogously in multivariable control.
consequently in the present state of the art of control one has a
The first two parts are specific for the multi variable control
rather sophisticated approach when studying the quality of
having no counterpart in single variable systems. The basic
control, but when one returns to design, trial and error
difference between the single variable systems and multi variable
techniques, intuition and experience only are of help. However,
systems shows here at its best.
any single variable signal model (except in some very special
The principal objective of the treatment that follows is to
cases) is nothing more than an abstraction of the actual multi-
prove the existence of this basic difference and to show some
variable system (with at least two conjugate variables), thus a
practical consequences of it.
model based on the latter should present the basis for a much
improved theory that will combine the quantitative attributes Principle of Uncertainty in Structure
with the rational design approach.
(b) Even if the signal model of a physical system is accepted In dealing with a system, we are interested in the problem of
as the basis for the analysis, a multivariable system very often determining its behaviour from the outside. The problem can
results. be formulated by stating the question: To what extent can the
behaviour of a system be determined through changing and
Terminal and Structural Behaviour of a System simultaneously measuring inputs and outputs? Only external
variables are dealt with, i.e. the system should not be disturbed
In dealing with multi variable systems one is ultimately by measuring its internal variables. The answer to the problem
interested in two kinds of behaviour: (a) the relation between a can be found in the following theorem.
system and its environment is usually expressed in terms of In a multivariable system the structural behaviour is indeter-
relationships between the outputs and inputs which define the minable from the outside of the system.
terminal behaviour of the system; and (b) the way in which Because all our knowledge of the physical world comes from
terminal influences are transferred through the system define external experimentation upon various physical systems one
structural behaviour of the system. can say that the principle of the uncertainty in structure offers
The specification of the terminal behaviour of a system from a limit up to which the behaviour of the physical world can be
the control point of view can be made in the following ways: determined. In the author's opinion the uncertainty in structure
(a) by a given set of inputs and outputs; (b) by a given set of together with the uncertainty in measurements (known as the
outputs regardless of inputs; and (c) by a definite demand Heisenberg's principle of uncertainty) present the basic
regarding the behaviour of the system but regardless of the principles which regulate our relation to the physical world and
terminal variables. therefore condition our knowledge of it.
The difference in the kind of specification of the terminal The significance of the principle of uncertainty in structure,
behaviour of a system becomes trivial in the single variable for the engineer, is a decisive one: it justifies the purposeful

100

110
CONTROL OF MULTiV ARIABLE SYSTEMS

approach in dealing with engineering systems. For any given of the output (not necessarily all) of others. If a change in one
purpose we can decide which of the structural patterns suits us of the sub-systems does not affect the others, the sub-systems
best in manipulations with external variables but we cannot are to be considered as non-interacting (independent).
determine the actual structure of the system.
The proof of the above theorem is almost self evident. r------------,I
I
The relationships between the inputs Xl,. . . X" and outputs I I Y1
I
Vb • . . Yn can be described in the general form: I I
I I
YI = F).ry(xl ,· . . X,,, Yl'· .. Yi- l , Yi~l'· .. Y1II) (I) I
I
I
I
I I
The problem can now be formulated in the following way: I II
All variables xl(t) . .. xJt); Yl(t) . .. YII,(t) are known ~~~LJ

functions of time. Can the functions F l . . • F", be uniquely E I


I
determined? Further, how adequately do the functions I
F1 • •• Fm describe (or reflect) the structure of a system? I
I
All that we have at our disposal for solving the problem is a : Yn
set of equations (equation I). However, from this set one cannot
uniquely determine which of the (11 + m) variables influences
any particular output. We can transform this set of equations
by eliminating some (at most m) of the variables from every Figure 1. P canonical representation ola system
one of the equations. For example, we can eliminate all output
variables and obtain a new set - - - - , XI
I--'-+--+--i
Yi = F j ,,(x1 • • • x,,) (2)
This set of equations is very often regarded as describing the
actual behaviour of the system; however, by reducing equation
1 to equation 2 we are putting a definite constraint upon the
E H
structural behaviour of the system under consideration with
no physical just{{ication for doing this. In linear cases this
constraint means that the sub-systems are supposed to be
non-interacting which best shows how restrictive this Xn I
I
assumption is. I
I
Instead of eliminating all output variables we can eliminate I
IL ___________ IL _________ J
some of the input variables from the right side of some of the
equations, e.g. we can derive the following set of equations Figure 2. V canonical representation of a system
Yi = F jXll(Xb· .. xn-m+l, Yl'· .. Yi-l, Y;+I'· .. Ym) (3)
To this set of equations corresponds quite a different kind of
structural pattern but with identical terminal behaviour.
Once the structural pattern is assumed the terminal behaviour
can be determined. 11
11
Canonical Representation of Systems
M YJ
From an infinite number of equivalent representations one
chooses some that can best be suited for dealing with the system H
externally. These representations should be: (a) simple in
manipulation with the external variables; and (b) of such a E
xj + I Yj+l
structure that can reflect appropriately the interactions as the
most prominent structural attributes of any multivariablesystem.
11
The above conditions are satisfied by representations with the
summing points located only at the very beginning or end of ,I
the system and therefore they will be referred to as canonical 11
representation. There are three kinds of canonical representa- 11
tion: (a) non-interacting or P representation [FIgure 1, I
equation 2]; (b) interacting or V representation [FJ8"ure 2,
equation 3]; and (c) mixed or H representation [FJ8"ure 3, Yn
equation 1].
Interactions in a System Figure 3. H canonical representation of a system
When using the term interaction, one should clearly define
to which elements this term refers. In our case the interactions In P representation, the system has been divided into sub-
refer to the sub-systems or more precisely their outputs. The systems each of them corresponding to one of the outputs
definition of 'interaction' is as follows: (FJ8"ure 1). The basic characteristic of this kind of representa-
The sub-systems should be considered as interacting if tion is that the sub-systems are non-interacting. If any property
changes in the properties in any of them influence the behaviour whatsoever of say the jth sub-system changes, only one of the

101

111
,
M. D. MESAROVIC

outputs, viz. the jth output, will be influenced. What is more, order of the vanishing minor of the system matrix. Every
the sub-systems could have inherem feedback and become infinitely coupled system of the jth order can be reduced to a
unstable, but only one of the outputs becomes unstable while all system with n inputs and (n - j) independent outputs. The j
other outputs remain stable and in fact are unchanged. This dependent outputs can be represented by an additional cascad-
is the basic drawback of P representation. No really inter- ing matrix.
related system can be successfully represented by this scheme. Which of the outputs should be regarded as dependent is, in
Even the most elementary problem, such as that of stability in general, an arbitrary choice.
an inter-related system, cannot be considered if the system is For an n variable case one obtains
P represented. On the other hand if V canonical representation
(6)
is used the sub-systems are inter-related by the very structure
of the adopted scheme of representation. If some of the sub-
systems are not inter-related, the corresponding interaction where N"j does not depend upon controllers characteristics,
elements in V representation naturally become zero. and M;. depends upon controllers characteristics and errors in
One can decide upon the strength of the ,interactions from the the outputs.
possibilities of their change from outside of the system. As The possibility of a change in interactions directly depends
infinitely strong inter-related sub-systems we shall define the again upon the system determinant, which reflects the strength
system satisfying either of the following two conditions: (a) of the actual interactions in the process itself.
both outputs move together, i.e. their dynamical (transient)
Minimizing Conditions for the Synthesis of the Multivariable
behaviour is identical; and (b) one of the outputs (in fact
Systems
everyone) can be expressed as depending only on the other out-
puts. In other words, one of the outputs does not depend on The synthesis procedure is related to the class of input
the inputs explicitly but only through the other outputs. functions . In single variable systems if a system behaves ideally
Consider a two variable system . Assume that the system is with respect to a given input (with sufficient general character-
canonically represented and that the controllers are forward istics) it will behave ideally with respect to the whole class of
acting. We want to determine the possibilities of changing the inputs (e.g. to all real temporal functions). In multivariable
relation between Yl and Y2 from outside the system. To this systems, a system can behave ideally with respect to a given
end one can obtain for the output )'1 set of inputs but very poorly with respect to some other sets;
therefore in the synthesis of multivariable systems the terminal
Fn 1
)'1 = F. Y2 + D p- (5 2 )e1 + 5 22e2 ) (4) behaviour should be uniquely specified.
21 21 If the specification of the system is of the first or second kind,
where: Fn , F21 are transfer functions of the controlled process; the optimization procedure should be performed by mini-
5 21 , 5 22 are transfer functions of the controllers; e1' e 2 are mizing some function of the output errors. The success of the
errors in both outputs; and D = FnF22 - F12 F21 . synthesis depends ultimately upon freedom in configuration.
Assume that an incremental change in controller occurs . Let us take as the criterion of performance of a system the
The change in the output Yl is then definite integral of a function depending on all the outputs and
time
-LlYl = -Fn
LlY2 F21
+ D [I- (
F21
e1 + 5 21 -Lle-1 + 5 22 -Lle-2 )]
Ll521 Ll521
Ll5
--
LlY2
21 (5) e= f - F [Y1"
t,.
. . )',," tl dt (7)
t,
From this expression one can see that a change in the relation The function F is often non-linear presenting the measure
between the outputs Yl and Y2 may result only from the term of the errors in the system.
multiplied by the determinant of the system; therefore the In the single variable system one can achieve optimal or
smaller the determinant the larger the change in the controller ideal behaviour of the system only if there is complete freedom
should be if the inter-relations between the sub-systems are to of configuration. For the multivariable systems, however, the
be changed. In the limiting case ~ --+ 0, one must make an following theorem is valid :
infinite change in the controller characteristic which is equiva- In a multivariable linear system the optimal performance as
lent to stating that the inter-relations between the sub-systems defined by expression 7 can be obtained by using anyone of the
cannot be changed. transfer functions upon which the output under consideration
In the limiting case the elements interconnecting two sub- depends.
systems in P canonical representation are finite, while the For the determination of the optimal behaviour of a system
corresponding elements in V representation are infinite. Assume there is only one minimizing equation. If the optimali,z4?g
that we have a system whose strength of inter-relation has been function is, say, Vj the minimizing equations a r e '
changed until eventually the infinitely strong coupling has been
achieved. Which representation, P or V, is more suitable to of
__
1 ,J. D +

represent the system under all circumstances? In P repre- oM; 1 .

sentation the change of interaction can be obtained merely by


changing one of the sub-system elements while the inter- . + of'''i
CM D",
C J
= 0 (8)
connecting elements themselves remain unchanged. On the
other hand, in V representation the change of the strength in cM
-=0
j
(9)
interaction reflects directly (and only) on the elements inter- r; V j
connecting two sub-systems.
The results could be extended to systems with more than two The second equation becomes the identity for the linear case
variables. At first, however, the order of the infinite coupling while the remaining equation does not depend explicitly upon
between the sub-systems must be defined. the optimal function Vj' This basic theorem is valid also for
A system is infinitely coupled of the (n -jth) order if j is the many classes of non-linear systems.

102

112
CONTROL OF MULTIVARIABLE SYSTEMS

In proof of this theorem the following assumptions are made: This result will be used in the next section for the study of
(a) Dependence of an output, say jth, upon some particular some practical systems.
characteristic function of the system Vh and upon other
outputs, YI' .. Y,,, and time f, is known, Interacting versus Non-interacting Control

Yi = MlvI" . . Yj- V Yh I,' .. y"" Vj, t) (10) The studies in the previous sections lead towards interacting
control. Much attention has been paid to achieving non-
(b) The relations between all the outputs generally expressed interacting control that is advantageous in some practical
by a set of (n - 1) equations are also known . situations. It is of great importance to compare interacting
and non-interacting control in the case where the non-inter-
Gl,j = FU(YI ' .. YIII) = 0
action is not an end in its own but is merely supposed to be a
means for improving the quality of control.
We shall consider here only a relatively simple example of a
two variable system in P canonical representation. For a given
( 11) set of inputs zero errors could be achieved giving

FI2 = - rI2Fn : F21 = - -r


1
F22 ; r12 = -Xl
X
(16)
l2 2

Assume that the inputs differ only in magnitude and that both
proper subsystems are the same Fn = F 22 • The mean error is
Gm,; = F,",/YI ' .. YIII) = 0
then
If errors of all outputs are to be minimized we obtain a set of

J(-or. dl'lJ' oc d1'2J '" dt p(I'I, 1'2)(' + ~) (1'1 -


the minimizing equations: r I2 1,)2
00 - 00 - 'X: 12
8Fl,j 8F j _ I. ; 8F j + l,j (,F"" j (17)
8 M DI
.,
+ ... + aM Dj _
J
I + bM D H
J
, · · · BM D",
J
=
(12)
0
[L"'",FUI)X(t - tI) dtl] [1'" F(t2)X(t - (2) dt2]

This set could be satisfied with m transfer functions. Because For non-interacting control one has
the linear system is defined with nm transfer functions, (nlll - m)
of them remain undetermined after the synthesis procedure; J~"'Cl:dVI I_oo",dV I - : dt p(I'I' 1'2)(1/ + 1'/)
2
therefore additional constraints must be given . There are (18)
many different kinds of synthesizing procedures of the multi-
variable systems, depending upon the kind of additional con- [L~t(tI)X(t - fI ) dtl] [LOO", F(t2)X(t - (2 ) df2]
straints. Some of them will be discussed in the following
section. If the events associated with the variations of amplitudes are
of equal likelihood in a given region and have zero probabilities
Synthesis of Multivariable Systems with many Sets of Inputs outside, the result is given in Figure 4. For the case where the
The additional conditions that the synthesized system must
satisfy can be obtained by taking additional sets of inputs. One e
of the convenient ways to deal with such a large number of sets
of inputs is to consider their temporal mean. If f I . . . t k are
the time instants when the input sets occur, the time mean is
then
----r;" --)--........
L~ao dfAP(tI" f OCI) dt[~Yit, t
/ VI
{X)",dfI . . . .. tA.) j )] (13)
/
/
where P(tI' . . f k ) is the joint probability that the particular
sets of inputs will occur at 11 , . . fA" 2
V1
If the system behaviour is fairly stable and the time instants
11 , . . Ik are far enough apart, one can consider the responses
Figure 4. Dependence of error upon amplitude ratios of both
inputs: full line, non-interacting control; dash-dot line,
of the systems as independent. The possibility then exists that interacting control
the various sets of inputs differ only by the value of some
particular parameters. The mean time response is then amplitude ratio appears as random the curves in Figure 5 are
obtained. For a more general two-dimensional case with
I_~j)dl'I' .. J~70", dL'AJ:,,,dt p(l'I,' .. 1'1),I)t, I ' ]> • • • rl) (14) exponentially decaying probability distributions, the results are
given in Figures 6 and 7. The ultimate advantage of the inter-
For example, if various outputs differ only in magnitude and acting or non-interacting control depends upon the amplitude
the system is P canonically represented, the time mean is distribution. However, there is a region of amplitude distri-
bution in which the interacting control is under all circumstances
better than non-interacting. This will be referred to as inter-
I_oo",dV I . . . • [ "'", dl'l/ .[":rodt p(r l , · . . I'd action region. For the two variable system under consideration
(15) the interaction region is a conic intersection with the equation
(Figure 8)
1'1 2 + l}r l2 - 21'11'2r12(1 + rill = 0 (19)

103

113
I
M. D. MESAROVIC

,-
···•
I
,
I

,
I

\ ,'el,
I
I
I
I
I
I
I
I
, , I

Figure 5. Plot of error versus amplitude ratio of inputs for Figure 7. Dependence of error for interacting control and
non-interacting (en) and interacting (er) control exponentia/ly decaying probability distribution of the input
amplitudes

r 12 =2

__~==~~~~~~~~~~~rI2=5
v1

Figure 6. Dependence of error for non-interacting control


and exponentia/ly decaying probability distribution of the input
amplitudes Figure 8. Interaction regions for a two variable system

Other Kind of Synthesis (b) Very attractive additional conditions from the practical
There are many additional conditions that have been shown point of view are related to the size of the hardware used, its
as valuable in the synthesizing procedure. We shall mention cost, efficiency, etc. To this end the dissipated and stored power
here only two of them: in the particular elements has been shown as useful. It is
(a) The principle of the uncertainty allows many different interesting to note that under these conditions every single
procedures in reticulation (subdivision) of systems. The way of variable signal model should be converted into a two variable
reticulating the system can be used as an additional condition model.
in the synthesis. In this, one should have in mind what will be
the circumstances under which the system will work. The
Reference
I
constraining condition can be related to the power transferred MESAROVIC, M. D. Foundations for a Multivariahle Control Theory.
through the interacting elements (requiring its minimum), etc. M.T.T. Tech. Press, 1960. Wiley & Sons

Summary
The control of multivariable systems essentia lly differs from the in respect to many sets of inputs the probability distribution of the
control of the single variable systems due to the presence of inter- characteristic parameters should be introduced. The benefit of
actions. The terminal behaviour of a multi variable system can be using interacting or non-interacting control depends upon these
determined if enough sets of inputs are given, but the structural probabilities.
behaviour of a system cannot be determined from outside the system. The synthesis of the multi variable system is not uniquely defined
An inter-related system can best be represented by V canonical if the integral measure of the performance is used, and therefore
representation. additional constraint must be introduced. The kind of synthesis
When the performance of a linear system is measured by an integral depends upon the kind of constraint. The way of sub-dividing the
of some function of the output, the number of the minimizing equa- system and the dissipated power offer very useful and practical
tions is equal to the number of the outputs. If synthesis is performed constraints.

Sommaire
La commande des systemes multi variables differe, de celle des peut etre represente de la fa"on la meilleure par la representation
systemes it simple variable it cause de la presence des interactions. canonique.
Le comportement final d'un systeme multi-variable peut etre deter- Quand la performance d'un systeme lineaire est mesuree par
mine si on se donne suffisamment de groupes de valeurs des donnees l'integrale d'une certaine fonction de la sortie le nombre des equations
d'entree, mais le comportement structural du systeme ne peut etre de minimisation est egal au nombre de sorties. Si on accomplit la
determine de l'exterieur de ce systeme. Un systeme it inter-relations synthese par rapport it un grand nombre de groupes de valeur

104

114
CONTROL OF MULTIV ARIABLE SYSTEMS

d'entree, on doit introduire la distribution probabiliste des para- univoque si on emploie la mesure integrale de la performance, et par
metres caracteristiques. Le benefice qu'on peut tirer de l'emploi de consequent on doit introduire une condition supplementaire. Le
la commande interactive ou non interactive depend de ces proba- genre de synthese depend du genre de condition. La fac;on de
bilites. reticuler le systeme et la puissance dissipee fournissent des conditions
La synthese du systeme multi variable n'est pas definie de fac;on tres utiles et tres pratiques.

Zusammenfassung
Mehrfachregelungen unterscheiden sich von Regelungen in Systemen Gruppierungen der Eingangsgrof3en durchgefiihrt wird, sollte die
mit einer geregelten Variablen durch die gegenseitigen Beeinflus- wahrscheinliche Verteilung der charakteristischen Parameter beruck-
sungen der geregelten Grof3en. Das Verhalten der Ausgangsgrof3en sichtigt werden. Die Vorteile, die man aus der Verwendung von sich
eines Mehrfachsystems kann durch eine geniigende Anzahl von gegenseitig beeinflussenden oder sich nicht beeinflussenden Rege-
Gruppierungen der Eingangsgrof3en bestimmt werden, aber die lungen ziehen kann, hangt von diesen Wahrscheinlichkeiten ab.
Struktur des Systems kann nicht von auf3en ermittelt werden. Ein Die Synthese des Mehrfachsystems ist nicht eindeutig, wenn das
vermaschtes System kann am besten durch die sogenannte kano- IntegralmaI.l der AusgangsgroI.le angewendet wird, daher mussen
nische V-Schaltung wiedergegeben werden. Zusatzbedingungen eingefiihrt werden. Die Art der Synthese hangt
Wenn die Arbeitsweise eines linearen Systems durch das Integral von der Art der Zusatzbedingungen ab. Der Weg, das System zu
einer bestimmten Funktion des Ausgangs bewertet wird, ist die Zahl unterteilen und die Untersuchung der Energieverluste liefern nutzliche
der zum Minimum zu bringenden Gleichungen gleich der Zahl der und praktische Zusatzbedingungen.
Ausgange. Wenn die Untersuchung mit Riicksicht auf mehrere

105

115

You might also like