Sunteți pe pagina 1din 15

Chemical Engineering Science,

Printed in Great Britain.

VOI.45, NO. 55 PP. 13w-1323,

THE

Department

1990

OOG9 2509/90
53.00 + 0.00
0 1990 Pergamon Press plc

RELATIVE
GAIN
MULTIVARIABLE
JIN-WEN

CHANG

FOR NON-SQUARE
SYSTEMS

and CHENG-CHING

YU+

of Chemical Engineering, National Taiwan Institute of Technology,


R.O.C.
(First

received

9 May

1989; accepted

in revised

form

25 August

Taipei, Taiwan 10772,

1989)

Abstract--Generally
speaking, chemical processes in nature are non-square systems with unequal numbers
of inputs and outputs. However, only limited tools, e.g. singular-value decomposition (SVD), are capable of
analyzing non-square muItivariable systems directly. In this paper, Bristols relative gain array (RCA) is
extended to non-square multivariable systems. The non-square relative gain array (NRC) is defined as the
ratio of the open-loop gain, when all loops are without any control, to the closed-loop gain, when all loops
other than the loop explored are under perfect control in the least-square sense. Properties of NRG are
rigorously derived. Similar to the square RCA, the NRC can be used to assess the performance of
non-square control systems based on steady-state information. Futhermore, the NRG can be served as
a criterion to choose a square subsystem from a non-square system if a square control system is preferred.
Two distillation examples are presented to illustrate the use of the NRC.

1. INTRODUCTION
Non-square

systems,

systems

with

unequal

numbers

of inputs and outputs, arise naturally in chemical


processes (Downs, 1984; Downs and Moore, 1981;
Lourtie, 1985; Treiber, 1984). However, most multivariable control system analysis and/or design tools,
e.g. the relative gain array (RGA), Niederlinski index
etc. (McAvoy,
1983), deal with square systems only,
except for the method of singular-value decomposition (SVD). Mathematically,
SVD has very good geometric interpretation (Klema and Laub, 1980, Downs,
1984, Lau et al., 1985). Downs and Moore (1981) used
SVD for measurement selection and variable pairings
in distillation
processes. Despite being meaningful
geometrically, the SVD method depends on input and
output scalings (Mijares et al., 1986; Nguyen et al.,
1988).
The RGA, known as an interaction measure for
multivariable
control systems (McAvoy,
1983), was
originally proposed
by Bristol (1966). Morari and
coworkers
(Grosdidier
et al., 1985; Skogestad
and
Morari, 1987) gave updated summaries of the RGA.
In the design of decentralized controllers (multi-loop
SISO controllers), the RGA is used to eliminate unworkable variable pairings (Yu and Luyben, 1986).
Manousiouthakis
et al. (1986) generalized the concept
of the RGA to block relative gain (BRG) which is
capable of handling partially decentralized
control
systems. Both the RGA and BRG deal with square
systems only.
The purpose of this work is to extend the RGA to
non-square systems. Systems having more outputs
than inputs are considered in this paper. Obviously, it
is not possible to control a system having more outputs than inputs perfectly, i.e. steady-state offsets in all
outputs are zero (Goodwin and Sin, 1984). However,

in the least-square sense, the control objective can be


stated as: to minimize the sum of square errors
(steady-state errors or steady-state offsets) of the outputs with fewer inputs. We call this kind of control
sense
control
in the least-square
law perfect
throughout this paper. Therefore, the concept of perfect control, e.g. for deriving closed-loop gain, can be
constructed for non-square systems. Following
the
definition of the non-square RGA (NRG), algebraic
properties of the NRG are derived. Generally, two
approaches can be taken to design a control system
for a non-square process. One approach is to design
a non-square
controller
to control
all outputs
(Treiber, 1984; Lourtie, 1985). A quantitative relationship between the NRG and the ultimate performance
of a least-square perfect controller is derived. Another
approach is to square the system down, i.e. to select
outputs such that the number of inputs and outputs
are equal (Downs and Moore, 1981). A square subsystem, e.g. measurements, selection criterion is also proposed based on the NRG. Quantitative relationship
can also be derived or conjectured from the NRG.
Two distillation examples (Doukas and Luyben, 1978;
Downs and Moore,
1981) are used to illustrate applications of the NRG.
2.

DEFINITIONS

Consider a process transfer function


with inputs u and outputs y:

matrix

(1)

Y(S) = %,%
where G,,, is an m x n process
matrix with M 2 n

Ccs,

transfer

function

y(., is an m x 1 output vector


UC,) is an n x 1 input

vector.

The system (Fig. 1) is under feedback control with


a controller, K. The relative gain between the ith

+Author to whom correspondence should be addressed.


1309

1310

JIN-WEN

CHANG

and CHENG-CHING

Yu

fore, we have

-2

Ldui1

=
Bji
(5)
Ykk+i
where gii is the jith element of G-i.
It should be emphasized that the only assumption
made about the controller K is that the controller has
integral action (can eliminate steady-state offset for
asymptotically
constant disturbances). No assumption is made about the controller structure. Therefore,
Aij becomes
aYi

Fig. 1. Concept of closed-loop gain in a feedback system.

output (yi) and jth input (u,.), A:, is defined as the


open-loop gain over the closed-loop gain. The openloop gain between yi and uj is described by the ijth
element of the process transfer function matrix. gii(.,.
Generally, the open-loop gain, gij(,,, is a function of
the frequency (let s = iw) and the definitions are the
same for both square (m = n) and non-square (m > n)
systems. However, the definition and interpretation of
the closed-loop gain are not that straightforward. The
closed-loop gain is defined separately for both square
and non-square systems.

2A. Square systems


For square systems, the closed-loop gain is defined
as the gain between yi and ui when all other outputs
are under perfect control, i.e. y,s are held constant for
all k # i (McAvoy,
1983; Grosdidier
and Morari,
1985). Figure 1 illustrates the concept of the closedloop gain provided that all other outputs are kept at
set points. Generally, it is not possible to keep all
other outputs, yks, under perfect control at all time if
G,,, contains dead-time and/or RHPT zeros. Therefore, only the steady-state aspect of perfect control is
meaningful for general systems. In this paper, only the
steady-state aspect (s = 0) of the closed-loop
gain is
discussed. Mathematically,
the relative gain can be
defined as

Aj

and the RGA,

Alternatively,

gain

A = G@(G-)=

out earlier, the open-loop


gain is the
gain of the process transfer function:

aY.

_-.A

[ auj

uk,k+j

(7)
multiplication.

2B. Non-square system


For a non-square system with more outputs than
inputs (m > n), it is not possible to keep all outputs at
their set points. Therefore, the sense of perfect control
in the definition of closed-loop gain should be modified. Perfect control in the least-square sense is proposed. That is, a controller K is designed such that the
steady-state offsets are minimized in the sense of leastsquare (i.e. the 2-norm of the steady-state error vector
is minimized). This type of controller is termed a leastsquare perfect controller. From the IMC structure
[Fig. 2 and Garcia and Morari (1982)] the output
vector, y(,,. can be expressed as
Y(si = %,[I,,

+ Ge,,, tG,,, -

&,)I - Gc,,, YE

and assuming no plant-model

error becomes

e(,) =

~7:;

(1, - G,,, GM

(10)

where e,,, is an m x 1 error vector


Since we are interested in the steady-state error, the
least-square perfect controller requires
= (GT,, Go,)-

GA, = G,.

(11)

The superscripts T and + denote the transpose and


the Moore-Penrose
pseudo-inverse,
respectively
n

= &j(O).

Since only the steady-state aspect is considered, the


subscript (0) is dropped for clarity. The closed-loop
gain can be expressed as the element of G-i regardless of controller structure, e.g. multi-loop SISO controllers, block-diagonal
controllers
or full multivariable controller, as shown in Appendix A. There-

(8)

mismatch, we have

where Ge,,, is an n x m IMC controller transfer function matrix


is an m x n nominal
process transfer
%,
function matrix.

=(a,
As pointed
steady-state

(6)

where @ denotes element-by-element

(2)

ijj can be expressed as

Sijdji

A, is defined as

The closed-loop
closed-loop

Fig. 2. Non-square IMC structure.

1311

The relative gain for non-square multivariable systems


(Strang, 1980). Using the final-value theorem, the
steady-state offsets become [subscript (0) is dropped
for clarity]
i5 = (I - GG+),=.

Between the ith output and the jth input, the inverse
of the closed-loop gain is

auj

9ji

(14)

rayi

LauJoL=gij.g;
+ = rayi
-N

(15)

Thus, from the steady-state transfer matrix G and its


pseudo inverse G+, the non-square relative gain array, AN, can be promptly evaluated:
AN = G@(G+)T.

(16)

The definition is similar to the Bristol RGAs. Despite


the difference that the pseudo-inverse,
G+, is used
instead of the normal inverse, G-l, the simplicity of
the RGA remains.
OF THE

NON-SQUARE

RCA

(17)
From AN, we can compute the sums of all elements in
each row and in each column. The row sum vector
(RS) is defined as
RS=

j$liyj,
[
=

[m(l),

dE;lj,. .

j-1

1T

cs(2), . . . , es(n)]

(20)

and
es(i) = (9+

9)ii

(21)

where (g g)ii is the iith element of the n x n matrix


G+G.
Property 1: The sum of the elements in each column of
the NRG is always equal to unity. That is es(j) = 1.0
for all js.
Proof For an m x n matrix, G, with m > n, if GrG
non-singular, we have
G+ G = (GTG)-i GrG

= I,.

Since the jth column sum of the NRG,


element of G+ G, we have
es(j) =

is
(22)

es(j), is the jjth

F A; = 1.0.
i=r
QED

Property
1 says that the sum of NRG elements
along the longer side of the G matrix (the column sum
for m > n) is equal to unity. This property is consistent
with the result of square systems (RGA).
Property
2: The sum of the elements in each row of
the NRG
falls between zero and unity. That is:
0 < rs(i) < 1 for all is.
Proof: Appendix

C.

This property points out the fundamental difference


between the NRG and the RGA. Since it is not possible to keep all outputs perfect for a non-square system, the row sum of the NRG being less than unity
seems to indicate the deviation from perfect control
for each output. The next two sections will discuss the
row sum in greater details.
Corollary
1: Summation of all row sums, rs(i)s, is
equal to summation of all column sums, cs(j)s, and
the summation is equal to the column dimension.
That is
i$l rs(0 = j$l es(j)

= n

(23)

Proof!
. ,

t
j=

i-s(2), . . . .rs(m)]

A:jIT
1

(18)

where rs(i) is the sum of the ith row of the NRG.


Furthermore, it can be shown that (Appendix C)
us(i) = (99+ )ii

= [es(l),

CL

where gf is the jith entry of G+.


Notice that in order to achieve the control objective, controller structure, a full multivariable
leastsquare perfect controller, and steady-state gain of the
controller are assumed implicitly.
For a non-square system, G, the non-square relative gain (NRG) is defined as

3. PROPERTIES

(12)

where the superscript - denotes the deviation from


its nominal value at steady state. Geometrically,
the
matrix (I - GG)
projects g onto the orthogonal
complement of the column space of G. Subsequently,
the sum of the square of the steady-state offsets, ErB, is
minimized. Therefore, the steady-state closed-loop relationship, under the least-square perfect control, between ii and y becomes
,=G+y_
(13)

[ aYi

Similarly, the column sum vector (CS) is defined as

(19)

where (gg + )ii is the iith element of the m x m matrix


GG+.

This corollary indicates that the summation of row


sums [rs(i)]
is limited to the dimension of G (the
number of inputs). With n inputs, at best, we can keep
n output variables under perfect control. For an
m x n non-square system, the n degrees of perfection

JIN-WEN CHANG

1312

and CHENG-CHINO

are distributed among m output variables. This corollary clearly states the inherent limitation for systems
with fewer inputs.
Property 3: The NRG is invariant under input scaling
and is variant under output scaling. By input (output)
scaling we mean post (pre)-multiplication
of G by
a non-singular diagonal matrix S,(S, ).
Proof:
(1) Input scaling: For an n x n diagonal scaling
matrix S,, S, = diag (sl, sl,. . . , s,), it can be shown
that
(GS,)+
From

the definition,
A:

The NRG

= S;G+C.

A? can be expressed

as

= (GS,)@(S;G+)r.

for the system under input scaling is


(A&

IIJ

= gijsj

g;

= gijg;

= a&

Therefore 9 we have Af = AN.


(2) Output scaling: For the m x m diagonal scaling
matrix S,, S, = diag (sls sa, _ . . , s,), it can be shown
that
(S,G)+

Proof: Appendix

(1) the NRG is invariant under input scaling


(2) the NRG is variant under output scaling but the
scaling factors weight the relative importance of
the outputs.
Property 4: Any permutation of rows and columns in
a transfer function matrix G results in the same
permutation
in the NRG.
Mathematically,
let
AN = NRG(G)
and P, and P, are two square permutation
matrices.
If
(AN), = NRG(P,GP,),
then
(A),, = P,ANP,.
Proof: P, is an m x m permutation matrix and P, is an
n x n permutation matrix. Since P, and P, are orthogonal matrices, e.g. P; = PT and Pi 1 = P,, we
have
(P,GP,)+
The permutated

= P;G+

P,.

(24)

becomes

CP,GP,18 CVoGP,)+ I=

= CPoGP,l @ CP,(G+)P,l
= PoCW-46.3
(G+)=lP,
= P,ANP,.
QED
Similar to the RGA, any permutation of rows and
columns in a transfer function matrix G results in the
same permutation in the NRG.

= (S,G)O[(S,G)+]7T
[(c)l+

NRG

(AN), =

= (GTS,2G)-1GTS:.

= s,G@

D.

Despite being variant under output scaling, the


physical meaning of output scaling is quite clear. With
respect to input/output scalmgs, the NRG has the
following properties:

The scaling matrix (S,) cannot be factored out as the


previous case except for the following
case. For
s1=s2=
... =s,=si,wehave
A:

Yu

= AN.

Generally
A,N # AN.

QED

Similar to the RGA, the NRG is invariant under


input scaling. However, the NRG is variant under
output scaling. It should be emphasized that property
3 holds only for systems with more outputs than
inputs (for systems with more inputs than outputs, the
opposite is true). Since the concept of least-square
perfect control is used in the definition of the closedloop gain, the output scaling is equivalent to the
weighted least square [Lawson and Hanson (1974, p.
183)]. The magnitude of an output scaling factor, i.e.
si, weights the relative importance of the ith output.
The next corollary
gives the interpretation
of the
scaling factor, si, under asymptotic conditions.
:Corollary 2: If the ith scaling factor of the output
scaling matrix is zero, i.e. sI = 0, then the ith controlled variable is disregarded in the context of leastsquare perfect control and the corresponding
row
sum, rs(i), is zero. If the ith scaling factor, si, goes to
infinity, i.e. si + co, then the ith controlled variable is
always held constant and the corresponding row sum
approaches unity.

Property 5: The definition of the NRG is consistent


with the definition of the RGA when the number of
input variables is equal to the number of output
variables. Mathematically,
if G is a square matrix,
then AN = A.
ProoE

For a square and non-singular

matrix G:

G+ = (G=G)-G=
= G-(G=)-lGr
= G-.

QED
Property 5 shows that the RGA is a special case of
the NRG. Properties l-5 state fundamental properties of the NRG.
Property 6: For an m x 1 system,
the NRG is
. . . , g,ilT,
AN = [I,,,
with

A,,,

. . . , &,7=

G = [gii,

gsi,

relative gain for non-square multivariable systems

The

For a one-column matrix, the NRG is a vector with


all positive elements. Each NRG element is the ratio
of the square of the open-loop
gain to the sum of
square of all elements.
Property 7: The elements of the NRG approach infinity as the non-square system matrix G becomes nearly
singular.

Proof: From eqs (C3) and (C4) of Appendix


be expressed as
gij 5

det Gj[IJ

C, Ji,can

det G [Zti,]

1 = 1

A$ =

det (GTG)

As G becomes nearly singular


elements of the NRG approach

[det (GrG)
infinity.

4 01, the

QED
Consider a nearly singular non-square system G [G
is singular if and only if det(GTG)
= O]:
a
G=

The NRG

becomes

A=

a(1 + 6)

1+b

-1

26
l+S
-

2s
-1
~
26
~1+6

Let S = 0.001:

2s
-1
6

500.5
500.5
-1000
= [

I.

500
500
1001
-

5000.5
5000.5
10,000

-5000
-5cOo.
10,001

This special case implies a nearly singular system


matrix results in large elements in the NRG. For this
special case, the smaller 6 is, the larger the values of
n: are. However, the degree of singularity for nonsquare systems is less critical than square systems,
since it is not very likely to have a singular (or nearly
singular) system, especially for cases with m + n.
Property 8: Consider an m x n transfer matrix G with
its pseudo-inverse
G+ and its associate NRG, AN.
Relative changes in the gijs and 25s are related by the
following expressions.

g;

da:
nc

det [ (G)T(Gij)]
lgdet(GTG)

+ (1 - 2a;)

2.

B.

It seems that the analytical expression, eq. (25), does


not directly imply relative sensitivity between 1: and
.gr,, since 3,;s appear in both the numerator and the
denominator of the RHS of eq. (25). However, the first
term in the RHS of eq. (25) is implicitly related to the
1; as we can see from eq. (BS) of Appendix B. Let us
take the ill-conditioned
G in the previous case as an
example,
the numerical
results for the first (the
element in brackets) and the second term of the RHS
of eq. (25) for all ijs with 6 = 0.001 are
[ZOOO] + ( - 1000)
[2000] + ( - 1000)
[ [ - lOOO] + (2001)

[ - 20001+ (1000)
[ - 20001 + (1001)
[999] + ( - 2000)

With 6 = 0.0001, we have


[20,000] + ( - 10,000)
[2O,M30] + ( - 10,000)
[ [ - lO,OOO] + (20,001)

4. NON-SQUARE

and for 6 = 0.0001:

AN=

Appendix

1
1

[ - 2O,ooo] + (10,001)
[ - 2O,OOO] + (10,001)
L-99991 + ( - 20,001)

The results show that the premultiplication


factor in
the RHS of eq. (25) grows in the magnitude which is
proportional
to the size of dc. Therefore, we can
conjecture that the relative sensitivity between ,%gand
gij is proportional to the size of 1;. For a system with
large value of nzs, it becomes more sensitive to errors
in gvs.

[ a

Proof:

1313

(25)

CONTROL

SYSTEM

PERFORMANCE

As pointed out earlier, two approaches


can be
taken to handle non-square systems. The first approach is to design a non-square controller for a given
non-square system (Treiber, 1984, Lourtie, 1985). Implications of the NRG to the ultimate performance of
non-square controllers are discussed in this section.
Since it is not possible to keep every output at its set
point with fewer inputs (notice that the process transfer function is an m x n matrix with m > n) in the face
of disturbances, the ultimate steady-state performance
measure is, therefore, the steady-state offset (error) in
each output. In most cases, a least-square perfect
controller [eq. (1 l)] is designed such that the 2-norm
of the steady-state offset is minimized. The NRG
measures, quantitatively,
the ultimate steady-state
performance of a non-square control system.
Theorem 1: Consider a stable non-square system with
more outputs than inputs. If a least-square perfect
controller [eq. (1 l)] is employed and the closed-loop
system is stable, then the deviation of the ith row sum
of the NRG
from unity [l - rs(i)] is equal to the
steady-state offset in the ith output when a unit step
set point change is made in the ith output.
Proof: Based on the IMC structure (Fig. 2), if the
least-square perfect controller (Section 2B) is used, the
closed-loop relationship becomes
B = (I - GG+ ),*.

(26)

1314

JIN-WEN CHANG

If a unit step set point

change

and CHENG-CHING

is made on the ith

output, i.e.
-Set = (0, . . . , 0, 1, 0, . . . , o)=
Y
the steady-state

error of the ith output is


ei = 1 - (gg+)ii

where (gg+)ii is the iith entry of GG+.


From eq. (19), we have
ci = 1 - w(i).

QED
The closeness of each row sum, n(i), to unity has
a strong implication
in non-square control system
performance. In the case when one output is more
important than the other, we can simply weight that
particular output heavier than the other, e.g. rescale

40

.-----

Yv

the process transfer function matrix with output


scaling matrix provided with a larger scaling factor on
that output. The row sums of output resealed NRG
change accordingly and the row sum of that particular output becomes closer to unity. Therefore, the
physical meaning of output scaling on the NRG becomes more transparent.
4A. Example: DL column (Doukas and Luyben,
1978)
Consider a side-stream distillation example (Fig. 3),
separating benzene, toluene and xylene, originally
studied by Doukas and Luyben (1978). Table 1 gives
the process transfer function matrix. The concentrations of four impurities in three product streams are
controlled by three manipulated
variables: reboiler
duty, reflux ratio and side stream flow rate. Doukas
and Luyben (1978) squared up the system by adding

Raflux

----87

Sidestrram

Feed

Steam

Bottoms
Fig. 3. DL distillation column.

Table 1. Process transfer function matrix for the DL column


Reboiler duty (u,)
Toluene in bottom (yl)

Toluene in bottom (y2)

Benzene in side draw (ys)

Benzene in side draw (yl)

- g g*le-

1.59s

11.36s + 1
5.984e-2.24

Reflun ratio (21,)

0.374e- 7.7S

(22.2s

(21.74s

+ 1)2

1.986e-0~7

14.298 + 1

(66.67& + l)*

2.38e-0,42

0.0204e-0~5g"

(1.43s
-

+ 1)2

11.67e-~9S
12.19s + 1

Side draw (I(~)

(7.145 + l)*
- 0.176e-0.485
(6.9,s + l)*

11 3e-3.79
+ 1)

5.24eC60
(400s + I)2
- 0.33e-0.68
(2.388 + I)*
4.48e-0.525
(11.11s

+ 1)2

1315

The relative gain for non-square multivariable systems


a fourth manipulated variable, the sidedraw location.
Treiber (1984) treated the DL column as a 4 x 3
system. A non-square controller should be design for
the non-square system. The NRG
can be used to
assess control system performance a priori.
The NRG for the 4 x 3 system is

1.2

0.8
0.6

0.241
0.006
0.028
0.726

- 0.103
1.094
0.000
0.008

0.861
- 0.100
0.002
0.237

cc
-5

0.4

0.2

0.0

with

cs =

[l.OOo

RS = CO.998

1.000

1.000

l.Ooo]~

0.030

-0.2

I-

The controller is equivalent to the non-square version


of SPMC
[simplified
model
predictive
control
(Arulalan and Deshponde, 1987)]. Simulation results
(Figs 4-7) show that little steady-state offsets in the
first, second and fourth output when unit step set
point changes are made in these variables. However,
a significant offset is observed in y3 for a unit step set
point change in yy (Fig. 6). Table 2 summarized the
steady-state offsets in all outputs which are exactly the
results predicted by the NRG. Treiber (1984) also
pointed out that the third output, y3, is poorly controlled from the results of closed-loop
simulation.
Clearly, the ultimate non-square control system performance can be assessed from the NRC directly.

Fig.

200

.o

0.4
co
-s

0.2

0.0

-0.2

oc

responses for a unit step set point


change in y, with a non-square IMC controller.

600

Fig. 6. Closed-loop
responses for a unit step set point
change in y, with a non-square IMC controller.

0.6

400

Time

0.8

Fig. 4. Closed-loop

SbO

5. Closed-loop
responses for a unit step set point
change in yz with a non-square IMC controller.

Time

460

Time

From the row sum of the NRG, it is clear that the


third output will be controlled poorly if a non-square
controller is used. A non-square IMC controller is
designed to verify this prediction by the NRG. Since
our objective is to illustrate the steady-state offsets in
the controlled
variables, a simple non-square IMC
controller is designed:

-02C

260

0.971]=.

-0.4

2
0

200

400

500

Time

Fig. 7. Closed-loop
responses for a unit step set point
change in y, with a non-square IMC controller.

1316

JIN-WEN

CHANG

and

Table 2. Steady-state offsets of the DL column for a unit


step set point change in each variable with the static IMC
least-square controller
SCl
Yl

Yl
Y2
Y3
Y4

0.0016
-0.0001
- 0.0388
-0.0067

5. CONTROL

Se
Y2

o.ooo1

0.0000
-0.0025
- 0.0004

STRUCTURE

Se,

Y3
-

0.0388
-0.0025
0.9699
- 0.1664

CHENG-CHING

Yu

Consider an m x n process transfer function matrix


G with m > n. If we choose n outputs for control, the
system can be partitioned into

SC

Y4

-0.0067
- 0.0004
- 0.1664
0.0286

SELECTION

An alternative approach to designing a control


system for a non-square process is to square the
system down. That is to select the same number of
outputs as inputs and a square controller can be
designed for the square subsystem (Fig. 8). This is
a common
practice in chemical process control
(Joseph and Brosilow, 1978; Downs and Moore, 1981;
Downs, 1984; Morari and Stephanopoulos,
1980) and
is generally termed selection of secondary measurements. However, only the method of SVD proposed
by Downs and Moore (1981) is capable of handling
a non-square system in a direct fashion. Despite having good geometric interpretation, the SVD method is
both input and the output scaling dependent and the
secondary measurement selection method proposed
by Downs and Moore (1981) may fail in some occasions, e.g. the largest element in the different column
vector of U may happen to be the same output. The
NRG
provides an alternative for the selection of
a square subsystem (secondary measurements).
Consider a non-square process, G, with a square
controller as shown in Fig. 8. The non-square system
G can be partitioned into a square subsystem G, and
the complementary
(remaining) subsystem G,. The
control objective is to minimize the sum of square
error (SSE) of uncontrolled outputs when the square
subsystem is under perfect control. In the least-square
sense, this is equivalent to choosing n out of m (with
m > n) data sets and solve the n linear algebraic equations directly for the coefficient such that SSE for the
complete m data sets is minimized. Notice that there
are C(m, n) combinations to choose from. In terms of
tubular reactor control, this means to choose some
temperature measurement location(s) such that when
the chosen temperatures are under integral control
the temperature profile in the reactor can be held
most constant for any variation in the controlled
temperature. In terms of distillation control, this is to
select temperature control tray(s) such that the temperature profile within the distillation column can be
kept most constant. The row sum of the NRG provides some information in this regard.

[,-;-I

= [ ;I-]u

(27)

where ys is an n x 1 output vector for the selected


(controlled) outputs and yR is an (m - n) x 1 output
vector for the remaining
(uncontrolled)
outputs.
Clearly, there are C(m, n) choices for ys. The objective
is to minimize the SSE for the uncontrolled outputs,
yn, for any variation in the controlled outputs, ys.
Notice that only steady-state error is considered. The
closed-loop
gains for the square subsystem are described by
6 = G, 1r:_

(28)

When the square subsystem is under perfect control,


the steady-state error for all outputs is
e = (I, x n - GG;

)y:=

(29)

where I, X,, is an m x IZ matrix with unity in the


diagonal zero elsewhere, and e is an 111x 1 steadystate error vector. For a particular choice of the
square subsystem, G,, the SSE is defined as
SSE =

k IIs
i=1

II: =

II(ImX,, - GG,- )y::

)I:

(30)

i=l

where ytft 1s
. an n x 1 vector with unity in the ith entry
and zero elsewhere, and e(i) is an m x 1 error vector
corresponding to the specific input 7:::. The row sum
of the NRG provides optimal (exact) solution to the
problem for two special cases, namely, case of n = 1
and case of m = n + 1, and suboptimal
(approximated) solution for other cases.
5.4. Case ofn = 1
This case is quite common in chemical process
control: to choose a measurement location for singleloop control, e.g. to choose a temperature control tray
in a distillation control. In this case G is a one-column
matrix. From property 5, the row sum is
2

t-s(i) = m

9il

&

= 22.

(31)

g5

From eqs (30) and (31) it can be shown that, for


a particular choice of G,, i.e. the ith subset of S(l, m)
(Appendix B), the SSE becomes
m
c sj:
j=1

SSE(i)

(32)

Sfl

and
m

c si:
Fig.8. Square control structure for a non-square system.

SSE(i)

= +
9il

-_1=1--1.
n(i)

(33)

1317

The relative gain for non-square multivariable systems


Notice that SSE(i) is for the case when the ith output
is chosen as the controlled
variable. Equation (33)
shows that choosing the largest row sum of the NRG
can lead to the smallest SSE subject to any perturbation in the controlled variable. From property 6, we
know that a large of giI corresponds to a large value
of AZ and rs(i). Therefore, choosing the element with
the largest magnitude from a transfer function matrix
is equivalent to choosing the largest row sum from the
NRG in this case. This is physically understandable
for the single-loop control, since the heuristic in common practice is to choose a system that the manipulated variable has a large effect on the controlled
variable.
SB. Caseofm=n+
1
For this almost square system, we can square down
the system by choosing the first n outputs to form
a square subsystem, G,, and to derive the relation
between
the
SSE
and
rs.
Since
there
are
k [k = C(n + 1, n) = n + l] choices to select a set of
n-tuple from n + 1 possible elements, the results for
the remaining (k - 1) combinations
can also he
similarly derived. For this particular choice, we have

1, a small row sum in the RHS of eq. (39) indicates


a small SSE in the corresponding
square subsystem.
Now we can justify that the small row sum of the
NRG in the complementary
(remaining) system indicates a small SSE in the square subsystem. Therefore,
the criterion for the selection of a square subsystem is
to eliminate the controlled variable with the smallest
row sum in the NRG.
SC. Other cases
Besides the two special cases, the cases with two or
more input variables and many output variables also
occur frequently. For example, we often try to keep
the temperature profile in a distillation column with
reflux flow, steam flow, etc. Unfortunately, no direct
link between the SSE and row sum can be derived
analytically. The closest relationship between the SSE
and the row sum of the NRG for two input variables
is (Appendix E)

.9nz
. ...
.9nn
.1.
G=
-__
=9.1
911

912

..

91.

G,

G,

(34)

______________

[l[

9ml

4m2

...

9mn

This particular choice corresponds to the first subset


from S(n, m). Therefore, SSE( 1) can be calculated from
eq. (30):
SSE(l)

11
e(i) 11:=

i=l

I/I - GG,

y::

II;-

(35)

i=l

Since G, = G[l]

(Appendix

B), we have

n+l

SSE(l) =

,g2(detGci)= det(G=G)
(det G[l])

- (detG[l])Z

(det G[lJ)
(36)

and from Appendix


rs(m)

det (GTG) -

@etG Cll)

det(GTG)

(37)

From eqs (36) and (37), we get


SSE(l)

Mm)
1 - w(m)

(38)

or in a more general form


SSE(j)

rs(n + 2 - j)
1 - rs(n + 2 - j)

where thej in SSE(j)


denotes the sum of square error
when the jth subset of S(n, m) is chosen to form
a square subsystem.
Since the value of the row sum is between zero and

SSE(j)

det(GTG)

(det

G[

ii1

*s(i))

(40)

j])z

where i is the ith element of the jth 2-tuple subset of


S(2, m).
Despite the subsystem selection criterion generated
by previous two cases, choosing the square subsystem
by eliminating outputs with smaller row sums leads to
the minimum SSE [as defined in eq. (30)]. Equation
(40) bears no direct link to this criterion. Counterexamples can be established such that by choosing the
n largest of row sums is not equivalent to the minimization of the SSE.
A computer experiment is conducted to test the
heuristic generated from previous two cases. Ten
thousand random matrices were generated for a given
dimension. The optimal and suboptimal square subsystems are identified (by optimal we mean that the
particular square subsystem results in the minimum
SSE and by suboptimal we mean that the subsystem
results in the one next to the minimum SSE) and the
subsystem selected by the NRG criterion is checked
against the optimal and suboptimal subsystem. The
results (Table 3) show that the NRG subsystem selection criterion leads to the optimal subsystem in
almost 77% of the cases tested and results in the optimal or/and a suboptimal subsystem in almost
88% of all cases tested. The reason for the NRG
subsystem selection criterion fails to produce the optimal subsystem can be seen from eq. (40). The row sum
(which the criterion based on) only accounts for part of
the contribution to the SSE. The other factor comes
from the determinant of the selected square system.
Generally, the NRG subsystem selection criterion ensures a large determinant, but it does not guarantee
the largest determinant. However, for a square subsystem with competitive row sums, determinants of
square subsystems play a deciding role for the values
of the SSE. This explains the observation that, when
the NRG subsystem selection criterion fails to produce the optimal square subsystem, the second largest
row sum generally resultsin the minimum SSE. From the

JIN-WEN

1318

3. Numbers

Table

NRG

CHANG

and

CHENG-CHING

Yu

of optimal and suboptimal cases generated by the


square subsystem selection criterion

m
n

9532:
46X0

906 1
530
9144
420

9307
582
8356
1176
7974
1228

8830
479
7593
1091
6302
923
7785
1461

8350
818
7564
1081
5478
2060
6446
1696
7788
762

7185
1353
7353
1518
6182
1485
6357
541
7362
1158
7x10
1042

3
4
5
6
7

+Number of cases tested for a eiven dimension = 10,000.


*Number of optimal cases.
$Number of suboptimal cases.
results

of the

conjecture

computer

that

the

experiment,

NRG

we can

subsystem

further

selection

cri-

subsystem in most cases. After having selected the variables


to square down the system, we can go on to design
controllers for the square subsystem.
will result in the optimal

terion

5Cl.

or suboptimal

Example: DL column.From the RS

informa-

4), we can eliminate the third output to


form a square subsystem. Thus the rest of AN becomes
tion

(Section

(AN) =

0.241
0.006
CO.7261
[

0.103
Cl.0947
0.008

CO.8611
0.100
0.237

.
I
0

If the variables are paired according to the NRG as


shown in the (A) (those in brackets), the controller
structure becomes (u,, y,; u2, y2; z+, y,}, The RGA,
A, for the square subsystem is

A=

0.245
0.006
co.7491
C

0.103
[ 1.0941
0.009

[0X58]
- 0.100
0.242

Fig.

600

9. Closed-loop
responses for a unit step set point
change in Yt CyBis the uncontrolled output).

0.8

D YI
* Yn
- Ya

0.6 -

- Y4

gains and reset times from the


BLT method
Controller
K

{UI_Y,I loop
{W-Y, 1 loop
{Ua-Y, I loop

400

Time

Notice that there is little difference between A and A


for this example. A multi-loop PI controller is used to
control the square subsystem. The controller gains
and reset times (Table 4) are determined used the BLT
method (Luyben, 1986). Steady-state offsets exist for
the uncontrolled variable, y,, for unit step set point
changes in the controlled variables (Figs 9-l 1). Summation of the square of these steady-state errors (SSE)
Table 4. Controller

200

gain

0.0986
10.2330
- 0.1159

Reset time
*, (min)
25.33
107.48
143.73

-0.4

Fig.

200

600

600

10. Closed-loop
responses for a unit step set point
change in Y, (JJ, is the uncontrolled output).

The relative

gain for non-square

multivariable

1319

systems

consistent with the result of the NRG


selection
method by choosing controlled variables 1, 2 and 4.
But if the G matrix is under input scaling, i.e. the first
column is multiplied by 3, then the results of SVD

1.5

1 .o

0.5

v)

0.0

- 0.611
0.370
0.141

G, =

-__
x
-0.5

-1

- 0.649
0.271
- 0.096

[ - 0.6861
.cI-

-1.5

co.7051
0

49.969
I-

1O

400

200

600

Time
Fig. 11. Closed-loop responses for a unit step set point
change in y, (ya is the uncontrolled output).

can be predicted

In

used the NRG:

rs(3)
= 1 ;;;*f
1 - rs(3)

SSE =

= ( - 0.0391)2

+ ( - 0.0024)= + ( - 0.1723)2.
The NRG
square
subsystem
selection
criterion
handles non-square system directly. In some cases, it
provides quantitative information about the SSE for
the uncontrolled outputs.
Another technique to handle the non-square system
is the SVD method. Unlike the NRG,
the SVD
method depends on both input and output scalings.
Mathematically,
SVD decomposes G into three component matrices:
G = UZVT

0
0

[
-

G=

18.10
0
0

x [

[
-

- 0.423
0.139
- 0.135
[0.885]
0

0
11.095

1.614
0

0.452
[O.SSS]
- 0.033
0.072

0.859

- 0.512

- 0.024

0.509
0.060

0.858
- 0.053

0.077
- 0.997 1

0.070

0
12.055

1.614
0

- 0.116

- 0.008

0.115
0.167

0.991
- 0.075

- 0.077
- 0.997

this case, the SVD method chooses controlled


variable 4, 4 and 2 to form a square subsystem. This
example shows that the controlled variables (marked
by brackets in U matrix) change as the system is
(input) scaled differently. On the contrary the NRG
subsystem selection criterion is invariant under input
scaling.
5C2. Example:
DM
column (Downs
and Moore,
1981) Another distillation column example (Fig. 12)
studied by Downs and Moore (1981) is used to illustrate the selection of sensors locations using the NRG
method. The azeotropic distillation column is under
R-Q control and the steady-state gains are given in
Table 5. If only one manipulated variable is used, e.g.
R or Q, to hold the temperature
profile, the onecolumn NRG
should be calculated (Table 5). The

v
t
R

_---__ 6

-----_ L)

=
.

The SVD subsystem selection criterion chooses the


largest element from u(i)s [u(i) is the ith column vector
of U]. For the DL column, the result of SVD is

0.993

where W is an m x n orthonormal
matrix,
Z is an n x n diagonal matrix with singular values
in the diagonal.
VT is an n x n orthonormal
matrix.
The transfer matrix function of the DL column can be
decomposed into
[ - 0.7843
0.438
0.104
- 0.427

0.452
[0.889]
0.032

Fig. 12.

DM

distillation

column.

1320

JIN-WEN

and CHENG-CHING

CAANG

Yu

Table 5. Steady-state gains and the NRGs for the cases with a single
manipulated variable (DM column)
NRC
Tray j

(2)~
- 0.0773
- 0.2399
- 2.5042
- 5.9973
- 1.6773
0.0217
0.1977
0. I 290
0.0646

Table 6. NRG

Tray j
9
8
:
5
4
3
2
1

(2)~

R only

0.0135
0.2379
2.4223
5.7838
1.6582
0.0259
- 0.1587
- 0.1069
- 0.0539

0.0000
0.0013
0.1388
CO.79633
0.0623
0.0000
0.0009
0.0004
0.0001

and row sums for the R-Q structure


NRC

- 0.0085
- 0.2665
- 1.2364
12.1885
- 11.5669
0.1876
1.1761
0.4214
0.1018

rs(i)
0.0154
0.2747
1.3750
11.3720
11.8991
0.2320
0.9764
0.3611
0.0878

0.007
0.008
0.141
CO.8171
0.333
CO.4201
0.200
0.060
0.024

quoted brackets in Table 5 are the selected temperature control trays. As shown in Table 5, the NRG
selection criterion matches our physical intuition to
choose the temperature control trays with the largest
steady-state gains. The temperature control tray in
both cases is tray 6 (Table 5).
It is quite common to have more than one manipulated variable to control a distillation column. The
NRG and row sums for the R-Q control structure are
shown in Table 6. The NRG
selection criterion
chooses trays 4 and 6 to form a square subsystem.
This choice is not obvious if we analyze the steadystate gain matrix only (Table 5). It is interesting to
note that the SVD method also chooses trays 4 and
6 as the temperature control trays (Downs and
Moore, 1981).
In terms of control structure selection (to choose
a square subsystem from a non-square process), the
NRG method has several advantages:

(1) the NRG method handles a non-square system


directly,

(2) the NRG method is output scaling dependent

and the scaling factors weigh the relative importance of the outputs,
(3) the NRG method is invariant under input scaling,
(4) the physical interpretation of the NRG method
is straightforward.

Q only

0.0000
0.0013
0.1392
co.79331
0.0652
0.0000
0.0006
0.0003
0.0001

6. CONCLUSIONS

Bristols RGA is extended to non-square systems.


In this paper, an m x n non-square system with more
outputs than inputs (m > n) is considered. The NRG
is defined with the notion of least-square perfect control. The properties of the NRG are rigorously derived. Similarities and differences between the NRG
and the RGA are also pointed out. The results on the
properties of the NRG can be summarized as follows:

(1) the sum of elements in each column adds up to 1;


(2) the sum of elements in each row falls between
(3)
(4)
(5)

(6)
(7)
(8)

0 and 1;
NRG is input scaling independent;
NRG is output scaling dependent;
permutation of rows and columns in the transfer function matrix results in the same permutation in the NRG;
for square systems, the NRG is reduced to the
RGA;
a nearly singular transfer function matrix results in large elements in the NRG;
large NRG elements indicate the system is sensitive to modeling error (error in gij).

The theoretical development of the NRG has the


following practical implications:
(1) The NRG can be used to assess the ultimate
performance of non-square controllers. quantitative
results between the steady-state errors for a leastsquare-based non-square controller and the row sum
of the NRG are also established. The result indicates
that the deviation of the ith row sum from unity is
eqtial to the steady-state offset in the ith output when
a unit step set point change is made in that output.
(2) A square subsystem (secondary measurements
or sensor location) selection criterion is also proposed, based on the NRG.
This criterion forms
a square subsystem by eliminating outputs with
a small row sum. The NRG selection criterion assures
that the profile of all outputs can be maintained at its
best in most cases when the square subsystem is under
integral control.

The relative gain for non -square multivariable systems


examples are used to illustrate apof the NRG.
The results show that the
NRG is a useful tool in analyzing a non-square system
in a direct manner.
Two

distillation

plications

Acknowledgements-After
the completion of this manuscript, Reeves and Arkun (1989) published a paper on the
non-square block relative gain. Working toward a similar
goal, interestingly enough, different definition, properties

and implications are developed. This research is supported


by National Science Council of the R.O.C. under contract
NSC 79-O402-E01 l-03.

output orthonormal
matrix by SVD
input vector
ith element of u vector
ith column vector of U
input orthonormal
matrix by SVD
output vector
jth element of y vector
output vector for the corresponding
G,
output vector for the corresponding
G,

U
u

ui

4)
VT
Y
Yi
Ys
YR
Greek
6

letters

A
NOTATION

SSE
SSE(i)

block relative gain array


column sum vector of the NRG
ith element of column sum vector
determinant of a matrix
closed-loop error vector
ith element of closed-loop
error vector
steady-state error vector corresponding
to
the ith specific input vector y::
process transfer function matrix
process transfer function matrix employed
in calculation of the IMC controller
Moore-Penrose
pseudo-inverse of G
inverse of G
IMC controller
ijth element of G
ijth element of G-l
jith element of G+
G matrix with ith row and jth column removed
square submatrix with rows from the jth
subset of S(n, m)
submatrix with rows from Ith subset of
S(n, m) such that the ith row is always included
submatrix with rows from Ith subset of
S(n, m) such that the ith row is always excluded
complementary
(remaining) subsystem of G
square subsystem of G
iith element GG* matrix
iith element of G+G matrix
matrix of the product of G and G
matrix of the product of G+ and G
internal model control
controller transfer function matrix
non-square relative gain array
square permutation matrix
relative gain array
right-hand side
row sum vector of the NRG
ith element of row sum vector
Laplace transformation variable
diagonal scaling matrix
ith diagonal element of scaling matrix
sum of square error defined in eq. (30)
SSE when the ith subset of S(m, n) is chosen

SVD

singular-value

BRG
cs
es(i)

det
e
ei
e(i)
G
e
G+
G-1
Gc
Sij
*
9 ij
$7;

G[I]

GR
GS
99, + hi

@Ggjii
G+G
IMC
K

NRG
P

RGA
RHS
RS
i-s(i)
:
si

decomposition

1321

iij

AN
n;
I:
:
0

small number
RGA
ijth element of A
NRG
ijth element of AN
singular-value matrix by SVD
time constant
determinant of the GTG matrix
frequency

Subscripts
CL

I
0

OL
P

closed-loop
input
output
open-loop
permutation

Superscripts
N
non-square

set
T
+

system

set point

transpose
Moore-Penrose
pseudo-inverse
deviation from nominal value at steady state
REFERENCES

Arulalan, G. R. and Deshponde, P. B., 1987, Simplified


model predictive control. Ind. Engng
- - Chem. Rex 26, 347.
Bristol, e. H., 1966, On a new measure of interaction of
multivariable process control. IEEE Trans. autom. Control
AC-11,
133. Doukas, N. and Luyben, W. L., 1978, Control of sidestream
columns separating ternary mixtures. Anafyt. Instrum.
16,
485.
Downs, J. J., 1984, Paper presented at Distillation Control
Short Course, Lehigh University, Bethlehem, PA.
Downs, J. J. and Moore, C. F., 1981, Steady state gain
analysis
for
azeotropic
distillation.
Proc.
JACC,
Charlottesville. VA.
Gantmacher, F. R., 1977, The Theory of Matrices, p. 9.
Chelsea, New York.
Garcia, C. E. and Morari, M.. 1982. Internal model control:
1. A unifying review and some new results. Ind. Engng
Chem. Process
Des. Dev. 21. 308.
Goodwin, G. G. and Sin, K: S., 1984, Adaptive
Filtering,
Prediction
and Control,
p. 141. Prentice-Hall, Englewood
Cliffs, NJ.
Grosdidier, P., Morari, M. and Holt, 3. R., 1985, Closedloop properties from steady state gain information. Ind.
Engng Chem. Fundam. 24, 221.
Joseph, B. and Brosilow, C., 1978, Inferential control of
processes. I: steady state analysis and design. A.1.Ch.E.
J.
24,485.
Klema, V. C. and Laub, A. J., 1980, The singular value
decomposition: its computation and some applications.
IEEE Trans. aurom. Control
AC-25
164.

1322

JIN-WEN

and CHENG-CHING

CHANG

Lau, H., Alvarez, J. and Jensen, K. F., 1985, Synthesis of


control
structure
by singular value analysis: dynamic
measures of sensitivity and interaction.
A.1.Ch.E.
J. 31,
427.
Lawson, C. L. and Hanson, R. J., 1974, Solving Least Square
Problems,
p. 183. Prentice-Hall,
Englewood Cliffs, NJ.
Lourtie,
P. M. G., 1985, The inverse Nyquist array for
non-square systems. Proc. American Control Conference,
Boston, MA.
Luyben, W. L., 1986, Simple method for tuning SISO controllers in multivariable
systems. Ind. Engng Chem. Process Des. Dev. 25, 654.
Manousiouthakis,
V., Savage, R. and Arkun, Y., 1986, Synthesis of decentralized process control structures using the
block relative gain. A.1.Ch.E.
J. 32, 991.
McAvoy, T. J., 1983, Interaction
Analysis.
ISA Monograph,
Research Triangle Park, NC.
Mijares, G., Cole, J. D., Naugle, N. W. and Holland, C. D.,
1986, A new criterion for the pairing of control and manipulated variables. A.1.Ch.E.
J. 32, 1439.
Morari,
M. and Stephanopoulos,
G., 1980, Studies in the
synthesis of control structures for chemical processes III.
A.I.Ch.E.
J. 26, 247.
Nguyen, T. C.,Barton,
G. W., Perkins, J. D. and Johnston, R.
D., 1988, A condition number scaling policy for stability
robustness analysis. A.I.Ch.E.
J. 34, 1200.
Reeves, D. E. and Arkun, Y., 1989, Interaction
measure for
nonsquare
decentralized
control structures.
A.1.Ch.E.
J.
35, 403.
Skogestard,
S. and Morari, M., 1987, Implication
of large
RGA elements on control performance.
Ind. Engng Chem.
Res. 26, 2323.
Strang, G., 1980, Linear Algebra
and Its Applications,
2nd
Edition. Academic Press, New York.
Treiber, S., 1984, Multivariable
control of non-square
system. Znd. Engng Chem. Process Des. Deu. 23, 854.
Vidyasagar,
M., 1985, Control
System Synthesis:
a Factorization Approach.
MIT Press, Cambridge,
MA.
Yu, C. C. and Luyben, W. L., 1986, Design of multiloop
SISO controller
for multivariable
processes.
Ind. Engng
Chem. Process Des. Deu. 25, 498.
Yu, C. C. and Luyben, W. L., 1987, Robustness with respect
to integral controllability.
Ind. Engng Chem. Res. 26, 1043.

APPENDIX

A. CLOSED-LOOP

From the IMC


pressed as

structure

GAIN FOR SQUARE

(Fig. 2), the input

a(,) = CI + Gc,,, (G,,, -

be ex-

Let

gi

denote

*m and

the

ith

row

vector,

that

is

P. 9)l
det (CG)
The NRG

can be further

i (det G[z~).
i=*

expressed

(Bt)

as

032)
where @ = det (GTG), and GTG is assumed to be nonsingular (@ # 0).
Since GTG is nonsingular and positive definite, its determinant is always positive. Therefore

(B3)

034)

and
d2@
dg2. = 2 det [ (Gij)T(Gij)]
J

(B5)

where Gj is the G matrix with the ith row and ith column
removed. The sensitivity of the NRG can be expressed as

gifidet

s_
2;

[(Gj)T (GJ)]

A det (GTG)

+ (1 -

21;)

1
2.

(B6)

The first term in the RHS can not be reduced further for the
non-square system. For the square system (m = n), however:
~., = gij( I,

l)+jdet

Gi

det G

(Yu and Luyben,

and

g$ det [(GU)r(GU)]
Equation

than a fuil multivariable

Gr = [g&

. . , gin). In order to square down G in its


minimum rank n, we have k = C(m, n) choices about gi from
row vectors of G matrix. For notational
convenience,
let
S(n, m) be the collection
of all strictly increasing n-tuples
{i,, i,, . . . , in} with the elements in order [Vidysagar (1985,
p. 391)] That is 1 4 i, < i, d . _ . < i, <m. For example,
S(3, 5) = ((1. 2, 31, (1>2,4), (1,2, 5), &3,4).
(1, 3, 51, (L4, 5),
(2, 3,4),
(2,3, 51, (2,4, 5), (2,3, 51, (2,4, 51, (3,4,5) 1 and
S(n, n) = {(1,2, 3,. , . , n)}. Let C;,, be an n x n square
matrix with the rows from I [1~S(n,
m)] and the columns
from J. Since JE S(n, n) is a singleton set, let G,,[i]
denote
the submatrix composed from row vectors belonging to the
ith subset of S(n, m). The subscript IJ is dropped for clarity.
From
the Binet-Cauchy
formula
[Gantmacher
(1977,

gij( -

l)i+j det Gj
det

det (GTG)
If the controller structure is other
controller (e # G), we also have

we have

gi = (Yil, Bi2.

(-42)

Consider the stea_dy-state aspect. If a full multivariable


controller is used (G = G), from the final-value theorem, we
have

x , then

G = [yij],

GTG=[j,gki(IkjInXm

(Al)

where the controller structure is characterized


by e:, e.g. for
multi-loop SISO controllers
we have e = diag (gii) and for
a full multivariable controller
we have e = G. In order to
eliminate the steady-state
offset, we require
Gc (0) = %:

APPENDIX
Assume

SYSTEM

can

%,)I - Cc,,,Y;:,

Yv

(B6) can be reduced

further

1987)

(B7)

1 1, w3)
= 112.

to

dL..
2
= (1 - I,,)?.

(B9)

ij

Therefore,
equivalent

the inverse of the closed-loop


gain [which
to the change in tij over a change in jjy J is

is

APPENDIX
(I) *s(i) = (99 + Iii

The diagonal

element

of (GG+)

can be expressed

as

tW
(gg+)ii
where 4 ji is the jith element

of G,;:.

2 .9,kgz
L=I

2
j=1

nc = rs(i)

(Cl)

The relative

gain for non-square

(2) 0 d rs(i) < 1


From Appendix 3, we have
AN

Cl

systems

1323

APPENDIX

For an tn x 2 system:

lg..d@

_A!-

multivariable

(C2)

2 Q, dgij

(El)

with
@ = det (GrG)
where k = C(m, n).
Take the derivative

f: (det G[i])2
i=,

(C3)
we specify

of Q and express Lz as

gij 5
A; _

det Gj[I;,,]

*=I

det CC&,]

to derive the relationship between the SSE and RS of the


NRG. Other combinations
can be derived from the same
induction:

(C4)

C (detGCI])
I=,
where G[Z] is an n x n matrix with rows from the Ith subset
of S(n, m)
G[I;,,]
is an n x n matrix with rows from the Ith
subset of S(n,m) such that the ith row is always included.
k is equal to C(m - l,n ~ 1).
The ith row sum of the NRC is
rs(i) =
Substituting

i: $.
,= 1

(det G

j-s(j) = I=:

detG,

-..
det G,

det G,

(C5)

eq. (C4) into eq. (C5), we have

W)

and from Appendix

Ch,l)
(Cc)

1
7

rs(1) =

C (det G[r])
I=1
Since all possible G[ f&Is are subsets of all possible G[I]
and (det G[Z])2 is greater than zero, it is obvious that
0 d t-s(i) < 1.
APPENDIX

From

Appendix

(C7)

i-s(2) =

C, we know that
*
gij c det Gj[ I;,,] det G [ I;,,]

2; =

=I

(C4)

(detG[l])*

SSE(1) =

I=1
If the output scaling matrix
S, = diag(1,

has the form


1,.

. , si,

. , 1,l)

PI)

i lie(i) II: =
det(&G){
i=1
+I:::

;::I+

1:::
. +I,,:

iz:(
;,:I

the ,?c under output scaling becomes


sFgii 5

i= 1

(l,N)ij =

det G[1;,,]

det G[I;,,]

(D2)

sf 5 (det G[I;,,])
I= 1

where G[I)]
is an n x n matrix with rows from the Ith
subset of S(n, m) such that the ith row is always excluded and
k = k - k, and the ith row sum takes the form

Crs(r310= 5 (nbv)ij
j=I

r-s(l) + W(2) =

1
det (GrG)
+ SSE(l)

SSE(1) =

ST i
=

(E3)

5 (det G[I])
r=l

(det G,)]

det (CT G)
(det G,)

(E4)

[rs(l)

+ rs(2)]

- 2.

(E5)

- 2

(E6)

det(G[l;,,]

I=1

sf 2 (det G[l;i,J)Z
I= 1

[Z(det GS)

.
+

(D3)

5 (det GII(il])Z
I*=1

Asymptotically,
if si is equal to zero, then [r-s(i)], is equal to
zero. If si goes to infinity, then [rs(i)],
approaches unity.

For general cases eq. (ES) becomes


det(GTG)
SSE(J)
= (det
G[j])

{ 2 W)}
i=,

where i is the ith element of the jth 2-tuple subset of S(2, m).

S-ar putea să vă placă și