You are on page 1of 17

MIMO interaction measure and controller structure selection

MARIO E. SALGADOy* and ARTHUR CONLEYy


In this paper, a MIMO interaction measure is described and its use in structure selection of multivariable controllers is
discussed. The proposal is built upon a gramian-based measure of interaction in multivariable plants and the results in
this paper apply to stable processes. This measure provides support for decentralized inputoutput pairing as well as for
a richer controller architecture selection in continuous and discrete-time time frameworks, including triangular, block
diagonal and sparse structures.
1. Introduction
Most industrial processes are multivariable in nature
and the control designer usually hopes that, with a
proper inputoutput pairing, there is no signicant
channel interaction, since that makes a process variable
hard to control without perturbing other variables of
interest. This problem is embedded into the more gen-
eral problem of controller structure selection (ranging
from decentralized, up to full MIMO). However, before
dealing with this problem, the designer must decide on
which inputs and which outputs will be used to build
any control architecture (van de Wal and de Jager 2001).
Throughout this paper we will assume that the decision
has already been made regarding the inputoutput
selection problem. Among other things, this assumption
will reect upon the fact that we will be dealing only
with square systems, i.e. equal number of inputs and
outputs.
Decentralized control, although it is a limited
exibility choice, has several well-known advantages,
such as tuning, sequencing of loop closing and the
possibility to use the knowledge and intuition built
around the control design of single inputsingle output
systems.
In the decentralized (diagonal) architecture, a key
issue is how inputs and outputs are paired. This issue
has received a lot of attention over the last four decades.
The most signicant result is the seminal work of Bristol
(1966), who developed the idea of relative gain array
(RGA). In the RGA, the channel interaction measure
is built upon the d.c. gain of the MIMO process.
Throughout the years the RGA has proved to be a
useful tool. However it has limitations which have
been explored elsewhere, among them is the inability
to cope with certain non-minimum phase structures,
its insensitivity to delays and the fact that only one
point of the process frequency response is considered.
After Bristols work was published, several research-
ers have studied the properties and usage of the RGA
(see, e.g. Witcher and McAvoy 1977, Skogestad and
Morari 1987). Some others have proposed new measures
of interaction and criteria to choose a sensible input
output pairing. In 1971 Niederlinski suggested to use
an index (the Niederlinski index, NI) which is also
built upon the process d.c. gain matrix (Niederlinski
1971). Although no additional information is obtained
regarding inputoutput pairing, this index provides
direct information on the ability of a decentralized con-
trol to stabilize a 2 2 MIMO system (Chiu and Arkun
1991). A variation of the RGA was proposed by Zhu
(1996); his proposal is known as the relative interaction
array (RIA) and it is based on the concept of viewing
the interaction as an unmodelled term for a particular
pairing (at d.c.).
The NI and the RIA do not encode more informa-
tion on the process than the RGA, although they
provide valuable alternative viewpoints. The common
feature of those indices is that they only use the
system model at zero frequency. Dynamic features
were introduced in the relative dynamic gain array
(RDGA), (Witcher and McAvoy 1977, Bristol 1978).
The RDGA also provides information of how inter-
action varies with frequency, suggesting bandwidths
for alternative pairings. That line of work was con-
tinued by Gagnon and co-workers (Gagnon et al.
1999) through the generalized relative dynamic gains
(GRDG). More specically, the frequency response of
the process and that of a complementary sensitivity
target are used to test dierent inputoutput pairings.
A major limitation is that the GRDG was devised
mainly for 2 2 systems.
Before proceeding any further it is necessary to
clarify that, in this paper, the expression system structure
is used to denote the web of paths from the (vector)
input to the (vector) output, independently of the dyna-
mical order of each scalar subsystem. Hence, if we have
a p p MIMO system with transfer function Hs, input
International Journal of Control ISSN 00207179 print/ISSN 13665820 online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/0020717042000197631
INT. J. CONTROL, 10 March 2004, VOL. 77, NO. 4, 367383
Received 1 May 2003. Revised and accepted 13 January
2004.
* Author for correspondence. e-mail: msb@elo.utfsm.cl
y Department of Electronic Engineering, Universidad
Te cnica Federico Santa Mar a, Chile.
vt R
p
and output wt R
p
, its structure will be
described by a symbolic array, as illustrated in (1)
v
1
v
2
v
3
v
p1
v
p
structureHs
0 0
0 0
0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0
0 0 0
0
B
B
B
B
B
B
B
B
B
B
B
@
1
C
C
C
C
C
C
C
C
C
C
C
A
w
1
w
2
w
3
.
.
.
w
p1
w
p
1
where a -entry in the (i, j) position signals the presence
of a subsystem with input v
j
(t) and output w
i
(t), and
H
ij
6 0. On the contrary, a 0-entry in the (k, ) position
indicates that there is no direct dependency of the
output w

t on input v
k
(t). Furthermore, we will refer
to the structure of Hs
1
as the inverse structure of Hs,
provided that Hs is non-singular almost everywhere.
The framework for the problem to be dealt with in
this paper is described next.
We consider a plant with p inputs and p outputs
which can be modelled as a multivariable process with
matrix transfer function given by
Gs G
ij
s
G
11
s G
12
s G
13
s G
1p
s
G
21
s G
22
s G
23
s G
2p
s
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
G
p1
s G
p2
s G
p3
s G
pp
s
2
6
6
6
6
6
4
3
7
7
7
7
7
5
2
Assume next that, using one of the known tools
to decide on a good pairing of inputs and outputs, a
decentralized controller is designed. If the inputs
and outputs are re-labelled, when necessary, one only
needs to design p independent SISO control loops, for
G
11
s, G
22
s, . . . , G
pp
s. This approach implies that the
plant is nominally described by a non-interacting model
and that interactions are captured in the modelling
error. Thus the nominal plant model is
G
o
s diag fG
11
s, G
22
s, . . . , G
pp
sg 3
and the designed diagonal controller has transfer
function
Cs diag fC
1
s, C
2
s, . . . , C
p
sg 4
The resulting closed loop performance, including
tracking, regulation, noise immunity and robustness
properties can be described using sensitivity functions
(Goodwin et al. 2001). In particular, the nominal
sensitivity and the nominal complementary sensitivity
are given by
S
o
s I G
o
sCs
1
diag fS
o1
s, S
o2
s, . . . , S
op
sg 5
T
o
s G
o
sCsI G
o
sCs
1
diag fT
o1
s, T
o2
s, . . . , T
op
sg 6
where
S
o
s 1 G

sC

s
1
7
T
o
s G

sC

s1 G

sC

s
1
8
for 1, 2, . . . , p.
We need now to assess the eect of our decision
to design a decentralized controller for a diagonal
plant model. When the controller (4) is used to control
(2), the achieved sensitivity is given by
Ss I GsCs
1
9
which is, in general, a non-diagonal transfer function.
A key result is then given by the following lemma.
Lemma 1: Consider the nominal loop dened by the pair
G
o
s, Cs and the true loop dened by Gs, Cs,
where Gs, G
o
s and Cs are given by (2), (3) and (4)
respectively. Then the nominal sensitivity S
o
s and the
achieved sensitivity Ss are related by
Ss S
o
sI H
T
s
1
10
where the element (i, j) of the matrix H
T
s is given by
H
T
s
ij

G
ij
s
G
jj
s
T
oj
s i 6 j
0 i j
8
<
:
11
Proof: We rst note that
Gs I G
l
sG
o
s
) G
l
s Gs G
o
s G
o
s
1
12
where G
l
s denotes the left multiplicative error. The
i, j-element of G
l
s can be computed from (2) and (3)
to yield
G
l
s
ij

0 i j
G
ij
s
G
jj
s
i 6 j
8
<
:
13
We next have that (Goodwin et al. 2001)
Ss S
o
sI G
l
sT
o
s
1
14
If we now use (6) and (12) in (14) the result (10)
and (11) is obtained.
In (11), T
oj
s is the nominal complementary sensi-
tivity of the jth loop. We observe that if the frequency
responses of the o-diagonal terms, H
T
s
ij
, are very
small, then the resulting loops are well decoupled and
368 M. E. Salgado and A. Conley
almost behave like p independent SISO control loops.
This is achieved if Gs is strongly (column) diagonally
dominant, at least in the frequency bands of T
o1
,
T
o2
, . . . , T
op
.
In this framework we can pinpoint several important
elements, namely
. The (plant and loop) coupling characteristics,
in general, are frequency dependent.
. The (loop) coupling characteristics, in general,
are dependent on the choices made for the SISO
control designs.
. Dierent dominance characteristics may lead to a
dierent pairing of inputs and outputs.
We want to focus on the situation when the above
analysis suggests that no decentralized design yields
a satisfactory performance. This failure usually arises
from the conict between having a suitable (nominal)
bandwidth and achieving a robust performance.
The natural next step is to increase the controller
complexity in such a way that one (or more) plant
input is made dependent on more than one plant meas-
urement, i.e. the controller is no longer a diagonal con-
troller. The structure selection might need to proceed.
The question is then: how do we gradually increase the
control structure complexity, beyond a diagonal controller
structure, such that a satisfactory performance is achieved
without necessarily going into a full MIMO controller.
One possible answer would be to use the dynamic
RGA (DRGA) developed by Witcher and McAvoy
(1977), which is basically a natural extension of
Bristols RGA, namely
DRGAG G jo G jo
T
15
where the operation symbol denotes element by
element multiplication. Equation (15) applies to square
MIMO systems; however, it can be extended to non-
square systems using the Penrose pseudo-inverse.
Although this measure usually reveals some hidden
interaction features, it can be hard for the designer
to make decisions regarding the controller structure
based upon the comparative analysis of p p frequency
responses. In particular, the interpretation of the phase
of those frequency responses is very hard. Furthermore,
this type of interaction measure does not provide
answers for certain plant model structures such as
those belonging to the class of triangular models, since
in those cases, DRGAG is a diagonal matrix, leading
to a unique inputoutput pairing, which is sensible only
if there exists a diagonal controller which delivers a
satisfactory performance.
Control of industrial interacting process is not only
connected to decentralized architectures, since there are
other control architectures which are simpler than the
full MIMO control (where every process input depends
on all process outputs), but more versatile than the
simple decentralized (diagonal) controller. These addi-
tional architectures include block diagonal, triangular,
sparse controllers, etc. Few, if any, of the known
indices and interaction measures are useful to evaluate
alternative controller structures other than diagonal
controllers.
In this paper we propose an interaction measure
which is based upon a dynamic model of the process.
(A preliminary version of these results was presented
at the CDC 2000 in Sydney (Conley and Salgado
2000).). This measure also quanties interaction as a
function of chosen channel bandwidths, gives criteria
for inputoutput pairing and helps to assess alternative
controller architectures. The proposed index, which is
built on the system gramians, can also provide a meas-
ure of the achievable performance of a given controller
architecture with respect to either the full MIMO case
or to another controller architecture. We consider
mainly stable square MIMO systems in the continuous
time domain. The extension to discrete-time systems is
straightforward and it will be covered succinctly and
illustrated with examples.
The main advantage of the proposed index is that
it captures the relevant process dynamics in a reduced
set of numbers. This is similar to the HII developed by
Wittenmark and Salgado (2002). However, it has not the
limitation of being applicable only to decentralized con-
troller inputoutput pairing. Indeed, the interaction
measure proposed in this paper allows to select more
complex controller structures.
The paper focuses on interaction quantication and
alternative controller structures. This topic is strongly
connected to issues such as control synthesis, robust
stability, integrity and xed modes. Although they are
outside the scope of this paper, brief comments will be
made regarding some of them.
This paper is organized as follows: in } 2 control-
lability and observability gramians are revisited, in } 3 the
connection between gramians and MIMO interaction
is built, in } 4 the new interaction measure is developed,
} 5 shows how the previous ideas are extended to
discrete-time systems, } 6 shows the application of the
new measure to the choice of controller structure and
in } 7, the connection between this new measure and the
control design problem is explored. Conclusions and
further research directions are presented in } 8.
2. Gramian fundamentals
2.1. Denitions
Gramians, in control theory, are matrices which
describe controllability and observability properties of a
given stable linear system. They can be computed for
MIMO interaction measure and controller structure selection 369
continuous-time and discrete-time systems. We will
initially focus on continuous-time systems.
Assume that a p p stable MIMO system has a state
space representation given by the 4-tuple (A 2 R
nn
,
B 2 R
np
, C 2 R
pn
, 0 2 R
pp
), then the controllability
gramian, P 2 R
nn
, and the observability gramian,
Q 2 R
nn
, are symmetric non-negative denite matrices
which satisfy the Lyapunov equations (16)
AP PA
T
BB
T
0 A
T
QQA C
T
C 0 16
Also, the matrices P and Q can be expressed as
P

1
0
e
At
BB
T
e
A
T
t
dt Q

1
0
e
A
T
t
C
T
Ce
At
dt 17
2.2. Physical interpretation
Useful interpretations for gramians can be derived
using energy concepts (see, e.g. Glover 1984).
2.2.1. Controllability gramian. Consider the (stable)
system described by the 4-tuple A, B, C, 0. Find the
optimal (in a quadratic sense) control ut in
t 2 1, 0 which steers the system state from the origin
at time t ! 1 to a specied state x
o
at t 0 (we will
consider a normalized state, i.e. kx
o
k 1), i.e. solve the
problem
min
u2L
2
1, 0
Ju min
u2L
2
1, 0

0
1
ut
T
ut dt
subject to x0 x
o
with kx
o
k 1
18
The solution to this problem can be obtained using
standard linear quadratic regulator theory (Goodwin
et al. 2001). The optimal control, u
o
t, is given by
u
o
t Be
A
T
t
P
1
x
o
) Ju
o
x
T
o
P
1
x
o
19
Then the optimal cost is low for every normalized
x
o
2 R
n
if and only if all eigenvalues of P are large
(since then, all eigenvalues of P
1
are small). It is also
true that if one of the eigenvalues of P is small, then
the optimal cost will reach its maximum for x
o
being an
eigenvector corresponding to the smallest eigenvalue
(recall that P and P
1
share the same eigenvectors).
2.2.2. Observability gramian. Assume that x0 x
o
(with kx
o
k 1) and that ut 0 8t ! 0 then

1
0
yt
T
yt dt x
T
o

1
0
e
A
T
t
C
T
Ce
At
dt
|{z}
Q
x
o
x
T
o
Qx
o
20
Thus every x
o
2 R
n
(with kx
o
k 1) will be highly
observable from yt if and only if all eigenvalues of Q
are large. Also, if Q has a small eigenvalue, an initial
state will be weakly observable from yt if and only if it
coincides with the associated eigenvector.
In summary, gramians quantify how hard it is to
control and to observe the system state, and the ranks
of P and Q are the dimensions of the controllable sub-
space and observable subspace respectively. However,
gramians depend on the state space realization. To
extract valuable information, the product PQ is formed
and its eigenvalues, z
i
(i 1, 2, . . . , n), are computed. It
can be proved that these eigenvalues are non-negative
and that they do not depend on the particular realiza-
tion (see e.g. Glover 1984 and Kwakernaak and Sivan
1972). The system Hankel singular values (HSV) are
dened as
o
i
H

z
i
p
i 1, 2, . . . , n 21
where the z
i
s are ordered to obtain o
1
H
! o
2
H
! !
o
m
H
! 0.
The HSV condense useful information regarding
system controllability and observability. For instance,
the number, r, of non-zero HSV corresponds to the
dimension of the controllable and observable subspace,
S
co
. Also, the ratio o
1
H
,o
r
H
is a measure of the skew-
ness of the controllability and observability of the state
of S
co
.
Another useful property of gramians is that they are
directly related to the system 2-norm. Specically if the
system has a transfer function Gs, then (Doyle et al.
1992)
kGk
2
2
trace CPC
T
trace B
T
QB 22
This property applies to continuous-time as well as
to discrete-time systems.
Gramian characteristics and properties will be
exploited in the coming sections to build and to apply
an interaction measure.
3. Gramians and MIMO interaction
3.1. Elementary system
Consider the system state space description (A, B,
C, 0), where
B b
1
b
2
b
p

; C
T
c
1
c
2
c
p

23
We can then associate with this MIMO system a
set of elementary (SISO) systems, each of them having
a single input u
j
, j 2 f1, 2, . . . , pg, and a single output y
i
,
i 2 f1, 2, . . . , pg, and a state space model given by
(A, b
j
, c
T
i
, 0) with gramians P
j
and Q
i
satisfying
AP
j
P
j
A
T
b
j
b
T
j
0 A
T
Q
i
Q
i
A c
i
c
T
i
0
24
A key observation is that G
ij
s is the transfer func-
tion of the minimal realization of the elementary system
370 M. E. Salgado and A. Conley
with input u
j
and output y
i
. Hence, the sets of HSV for
the elementary system and for the minimal system, char-
acterized by G
ij
s dier only in a subset of zero HSV.
3.2. Gramian decomposition
We next observe that a key property of the system
gramians is that they can be expressed as functions of the
gramians for the elementary systems. This is precisely
stated in the following lemma.
Lemma 2 (Gramian decomposition): Let P
j
and Q
i
be the controllability and observability gramians for the
elementary system (A, b
j
, c
T
i
, 0) given by (23).
Then, the original system controllability and observa-
bility gramians P and Q are given by
P
X
p
j1
P
j
Q
X
p
i1
Q
i
25
Proof: Firstly, Lyapunov equations in (24) are built
for all elementary systems, i.e. for i, j 1, 2, . . . , p
1, 2, . . . , p, then the equations for the P
j
s are added
and the result is the Lyapunov equation for P. The
same procedure is applied to build the equation for
the observability gramian. To achieve those results,
we use the fact that BB
T

P
p
j1
b
j
b
T
j
and C
T
C
P
p
i1
c
i
c
T
i
.
Remark 1 (Gramian decomposition interpretation):
From the gramian decomposition introduced in
Lemma 2, it can be seen that the product PQ for the
multivariable process is given by (26)
PQ
X
p
j1
P
j
!
X
p
i1
Q
i
!

X
p
i1
X
p
j1
P
j
Q
i
26
Then, the product PQ can be computed as the
sum of the corresponding products P
j
Q
i
associated to
each of the p p single-input single-output elementary
systems.
Also, if in some sense (to be dened later) the
products P
i
Q
j
and P
j
Q
i
, i 6 j, are much smaller
than maxfP
i
Q
i
, P
j
Q
j
g, then channels i and j have little
coupling.
It is straightforward to prove that if the system
transfer function Gs is diagonal, then P
j
Q
i
0 for
all i 6 j.
Obviously the maximum controllability and observa-
bility is attained when a full MIMO controller architec-
ture is chosen, then all terms of the form P
j
Q
i
have
to be added to compute the system HSV. However if
a restricted complexity controller is chosen, then only
a subset of those terms is involved.
4. Interaction quantication
The above analysis requires, to have practical
interest, a way to quantify and to compare.
By denition, the products PQ and P
j
Q
i
have non-
negative eigenvalues. However, the eigenvalues of a sum
of products P
j
Q
i
is not equal, in general, to the sum of
the eigenvalues of each summand.
It turns out that the trace of the product P
i
Q
j
is
a convenient basis to measure the interaction and the
degrees of controllability and observability of dierent
controller structures. Recall that the trace of P
j
Q
i
is the
sum of the squared HSV for the elementary system with
input j and output i and also recall that the HSV associ-
ated to the pair (P
j
, Q
i
) quantify the combined abilities
of input u
j
and output y
i
to control and to observe the
system state. This choice has the following interrelated
properties:
(i) the trace of any P
j
Q
i
is non-negative and state
realization independent;
(ii) the trace of any sum of terms P
j
Q
i
is the sum of
the traces of the individual terms;
(iii) the trace of any sum of terms P
j
Q
i
increases (or,
at least, it does not decrease) when new terms are
added;
(iv) the trace of P
j
Q
i
is equal to the sum of the
squared HSV for the system with transfer func-
tion G
ij
s;
(v) the trace of PQ is larger than, or at least
equal to, the trace of any sum of terms P
j
Q
i
(i 1, 2, . . . , p; j 1, 2, . . . , p).
The information encompassed in the above results
can be organized in a matrix (
ij
2 R
pp
, which
we will call the participation matrix (PM), dened by

ij

trace P
j
Q
i

trace PQ
27
Note that the trace measure has been normalized by
tracePQ. This implies that 0 -
ij
- 1 and that
X
p
i1
X
p
j1

ij
1 28
A potential weakness of the PM is its sensitivity
to input and output scaling. If, for instance, the entry
G
k
1
k
2
s is multiplied by a gain K, then tracefP
k
2
Q
k
1
g is
multiplied by K
2
.
One way to deal with this issue is to recall that (linear
incremental) plant models connect (incremental) con-
troller commands and (incremental) measured variables.
Both signals can be described in normalized ranges,
which arise from converting signal instrumentation
ranges (such as 4 to 20 [mA], 0 to 10 [V], etc.) to either
MIMO interaction measure and controller structure selection 371
percentage or per unit values. This leads to models
which must be interpreted as in the following example.
Example 1: Assume that an elementary system with
input u
j
(t), in the 010 V range, and (measured) output
y
i
(t), in the 420 mA range has a linearized model
given by
G
ij
s
b
1
s b
0
s
2
a
1
s a
0
; a
1
> 0, a
0
> 0
Then a 2 V change in u
j
corresponds to 20%
(0.2 p.u.) of full scale and yields a (steady state) change
in y
i
equal to
b
0
a
0
20%
b
0
a
0
3.2 mA
since the full scale (100%) in the measured output
corresponds to 16 mA.
This framework is similar to that proposed elsewhere
(e.g. in Glad and Ljung 2000), and avoids the ambiguity
of models described in physical or engineering units. For
a more detailed discussion on scaling see Maciejowski
(1989) and Glad and Ljung (2000).
Another signicant feature of this interaction
quantier is that it is insensitive to frequency scaling.
This can be appreciated in the following lemma.
Lemma 3 (Gramians and frequency scaling): Consider
a stable scalar system with transfer function G
ij
s and
state description given by A, b
j
, c
i
, 0. Let us denote its
gramians by P
j
and Q
i
. Consider also the frequency scaled
system with transfer function G
ij
s,, with 2 R

, and
gramians
~
PP
j
and
~
QQ
i
. Then P
j
Q
i

~
PP
j
~
QQ
i
.
Proof: We rst note that dividing the frequency by
is equivalent to perform a time scaling, by multiplying
time by the same constant , i.e. substituting t by t
where t t. Thus, the state space representation of
the scaled system is
dxt
dt

A

xt
b
j

u
j
t 29
y
i
t c
i
xt 30
We have thus that the gramians for the scaled system
can be computed from
A
~
PP
j

~
PP
j
A
T

b
j
b
T
j
0 A
T
~
QQ
i

~
QQ
i
A c
T
i
c
i
0
31
The above equations imply that
~
PP
j
P
j
, and
~
QQ
i
Q
i
. Thus P
j
Q
i

~
PP
j
~
QQ
i
.
Lemma 3 says that what matters in the quantica-
tion of interaction are certain dynamic features such
as the damping coecient in resonant modes, relative
location of poles and zeros, presence of non-minimum
phase zeros, and so on. This, in conjunction with the
magnitude scaling property, has several useful applica-
tions. For instance, we can compute tracefPQg for
simple systems, as shown in table 1.
The utility of this table can be appreciated in the
following example.
Example 2: Compute tracefPQg for the systems
G
1
s
5
s 3
; G
2
s
24
s 6
2
;
G
3
s 2
3s 1
s 1s 4
32
G
1
s
We observe that this system belongs to class CS
1
. This
can be appreciated by performing a frequency scaling
and computing the d.c. gain. Thus
G
1
s
5
3

1
s,3 1
This leads to tracefPQg 25,36.
G
2
s
In this case we are dealing with a system of class CS
2
.
To verify that we perform a frequency scaling and we
compute the d.c. gain. Thus
G
2
s
2
3

1
s,6
2
2 s,6 1
and then tracefPQg 1,6.
G
3
s
This system is in class CS
3
, as can be veried using
frequency scaling and computing the d.c. gain
G
3
s
1
2

6 s,2 1
s,2
2

5
2
s,2 1
Therefore tracefPQg 0.7025.
A careful analysis of the results in table 1 provides
valuable insight. Part of that insight is summarized in
the following observations.
System class Gs tracefPQg
CS
1
K
1
s 1
K
2
1
4
CS
2
K
1
s
2
s 1
K
2
2
2
4
2
CS
3
K
s 1
s
2
s 1
K
2
2
2
2
2
2
4
2
Table 1. Simple transfer functions and their combined
controllability and observability (o 2R

, [ 2 R, K 2 R.
372 M. E. Salgado and A. Conley
(i) The result for system class CS
1
says that the
combined controllability and observability for
rst-order systems (as measured by tracefPQg,
i.e. by the sum of the squared HSV) is completely
determined by its d.c. (static) gain.
(ii) Systems in class CS
2
are resonant systems when
0 - o - 2. The damping factor is then equal to
0.5o and tracefPQg grows as o decreases. This
suggests that highly resonant systems are strong
candidates to be considered when deciding on the
MIMO controller structure.
(iii) Systems in class CS
2
have two real poles for
o > 2. We also note that, as o grows, the system
tends to a rst-order system, since one of the
poles is much faster than the other. For o ) 2,
tracefPQg tends to 0.25, as expected.
(iv) Systems in class CS
3
have a zero at s 1,[. If
j[j ) o, i.e. the zero is much closer to the
imaginary axis than the poles, then tracefPQg
becomes very large. If [ <0, we are in the pre-
sence of non-minimum phase systems and then
tracefPQg is even larger. This discussion is con-
sistent with known design criteria which say that
small zeros, and especially, small positive zeros
are signicant dynamic features (see, e.g. Good-
win et al. 2001).
The functions considered in table 1 are deceivingly
simple. However, in dierent mixtures, they are enough
to model a vast number of MIMO plants. Time delays,
absent in table 1, are a dynamic feature highly signi-
cant in process control, and they will be included in the
analysis of the discrete-time case.
5. The discrete-time case
The theory presented in the previous sections applies,
mutatis mutandis, to discrete-time systems. In particular,
the gramians are computed from the equations
P APA
T
BB
T
0 33
QA
T
QA C
T
C 0 34
The applicability of previous results to discrete-time
systems is highly relevant for two reasons: rstly, modern
process control is based on sampled-data models to be
used in conjunction with digital technology and, sec-
ondly, in this framework we can deal with systems having
time delays (recall that in the continuous-time case,
delays yield state space models of innite dimensions),
which are very common in industrial processes. Time
delays are not only common but they may also play a
key role when deciding the controller structure. This
latter feature is formally stated in the following lemma.
Lemma 4: Consider the scalar stable system S

in
gure 1, composed by a scalar system S
0
and a pure
delay , where 2 N. Further assume that S
0
is described
by the 4-tuple A
0
, B
0
, C
0
, 0, with A
0
2 R
nn
, gramians
P
0
and Q
0
and with transfer function G
o
(z). If we denote
the gramians of S

by P

and Q

, then
tracefP

g tracefP
0
Q
0
g C
0
P
0
C
T
0
tracefP
0
Q
0
g kG
o
k
2
2
35
Proof: We will rst consider the case when 1. Then,
the composed system S
1
can be described by the 4-tuple
A
1
, B
1
, C
1
, 0, where
A
1

A
0
0
C
0
0
!
; B
1

B
0
0
!
; C
1
0
nn
1

36
Let us now express the gramians P
1
and Q
1
as
P
1

P
11
P
12
P
T
12
P
22
!
; Q
1

Q
11
Q
12
Q
T
12
Q
22
!
37
then those gramians can be computed from (33) and
(34), leading to
P
11
P
0
P
12
A
0
P
0
C
T
0
P
22
C
0
P
0
C
T
0
38
Q
11
Q
0
Q
12
0 Q
22
1 39
We then form the product P
1
Q
1
and compute its
trace, this yields
tracefP
1
Q
1
g tracefP
0
Q
0
g C
0
P
0
C
T
0
40
If we now let 2, we can similarly prove that
tracefP
2
Q
2
g tracefP
1
Q
1
g C
1
P
1
C
T
1
41
where we can use (37) and (38) to obtain
C
1
P
1
C
T
1
C
0
P
0
C
T
0
. This leads to
tracefP
2
Q
2
g tracefP
0
Q
0
g 2C
0
P
0
C
T
0
42
Then, the result (35) follows by induction and
on using (22). Note that a key in this proof is that
C
k1
0
nknk
1 for k 0, 2, . . . , 1.
Assume that G
ij
is the transfer function of S

, then
Lemma 4 says that the value of
ij
in the PM will
consistently increase as we increase .
To illustrate Lemma 4 we consider a plant model
which is a classic in process control.
Example 3: A continuous-time plant has a transfer
function given by
Gs Ko
e
st
s o
; o 2 R

Figure 1. Discrete-time system with time delay.


MIMO interaction measure and controller structure selection 373
The corresponding sampled-data model, with zero-
order hold, sampling period and t N, N 2 Z, is
given by
G
d
z K
1 a
z
N
z a
; where a e
o
To apply Lemma 4 we make N and G
o
z
z
N
G
d
z, with
A
o
a; B
o
K1 a; C
o
1; D
o
0
Then, using (33) and (34), we have that
P
o
K
2
1 a
1 a
; Q
o

1
1 a
2
Thus
tracefP
o
Q
o
g
K
2
1 a
2
kG
o
k
2
2
C
o
P
o
C
T
o
P
o
K
2
1 a
1 a
from where
tracefP
N
Q
N
g tracefP
o
Q
o
g NkG
o
k
2
2

K
2
1 a
2
NK
2
1 a
1 a
K
2
1 N1 e
2o

1 e
o

2
6. Application to controller architecture selection
The key idea presented in } 4 is that the PM high-
lights those elements in the transfer function matrix,
which are more signicant in the description of the
MIMO system. At the same time, the PM suggests a
path to gradually increase the complexity of the nominal
model G
o
(s) and, as a consequence, to gradually increase
the controller structure complexity.
Thus, to determine the structure of the controller
we must rst obtain a nominal model G
o
s. In a non-
discriminant (very unlikely) scenario, all 0 -
ij
- 1
have the same value, i.e. they are all equal to 1,p
2
.
However, in a highly discriminant PM, i.e. when there
exists a pair (i, j) for which
ij
( 1,p
2
, one can say that
all states in the elementary system with output y
i
and
input u
j
are either hard to control or hard to observe.
This suggests that there are no signicant benets
to consider the transfer function G
ij
s in the nominal
model. The converse is also true, i.e. when
ij
is much
larger that 1,p
2
, then some states in the elementary
system with output y
i
and input u
j
are easy to control
and easy to observe. Hence the transfer function G
ij
s is
a strong candidate to be included in the nominal model.
Furthermore, a large
ij
indicates that the control u
j
has strong authority upon output y
i
. Thus, the complex-
ity of a controller structure should be traded o against
the closeness to one of the sum of the chosen
ij
elements. For future reference we will denote this sum
by S.
Note that, in singling out the most signicant
ij
elements, the PM suggests an associated nominal plant
model. Then, it is plausible that the parameter S quan-
ties the combined controllability and observability of
that associated nominal model.
Consider the plant output y

. Then, the analysis of


the th row,

, in the PM is required. If the plant has p


inputs and p outputs, then the most signicant elements
in

are those which exceed the average value 1,p


2
.
Those are the primary candidates to be considered to
dene the nominal model G
o
(s). However, it is not
always true that the largest element in

has to be
included in the array dening the nominal model and
the associated controller structure. That and other issues
are illustrated in the following example.
Example 4: A plant with three inputs and three out-
puts has the PM given by
(
u
1
u
2
u
3
0.1400
0.1633
0.0739
2
4
0.0645
0.1485
0.0591
0.0233
0.0661
0.0863
3
5
y
1
y
2
y
3
where we have highlighted all those elements which
exceed the average value which, in this case, is
1,9 0.1111.
If we rst evaluate a (square) decentralized structure,
we have to choose a pattern with three rows and three
columns where only one element per row and one
element per column are allowed. We then observe that
the pairing (u
1
, y
2
), corresponding to the largest element
in the row
2
, should not be chosen in the selected
controller structure, since then the maximum attainable
S would be
S
a

12

21

33
0.3141
However, if we discard the pairing (u
1
, y
2
) and,
instead, we choose the pairing (u
1
, y
1
), a higher S can
be achieved with
S
b

11

22

33
0.3748
Hence the best pairing for a decentralized controller
is (u
1
, y
1
), (u
2
, y
2
) and (u
3
, y
3
).
If we are prepared to use a controller which is more
complex than a decentralized controller, we could make
y
2
to be commanded not only from u
1
, but also from u
2
,
this sparse controller would yield
S
c

11

22

33

21
0.5381
374 M. E. Salgado and A. Conley
Once the PM has been computed and G
o
s has been
chosen, we then know which plant inputs are aected
by which plant outputs. This determines the web of
dependencies of y on u. We then need to determine the
controller structure, i.e. the web of dependencies of u
on y (feedback controller structure). This can be done
by inspection from the PM if G
o
(s) is either a diagonal
or block diagonal system with blocks having symmetric
structure, but it is not usually so for those models which
include sparse or triangular structures. In the general
case, one has to mirror the dependencies in such a way
that the dependency of u from y (feedback controller
structure) be the inverse of that in the nominal model
structure (dependency of y on u). The rationale of that
mechanism can be appreciated from the discussion
below.
Given G
o
(s), we require that C(s) has a structure
which, at least, can achieve a diagonal complementary
sensitivity, so that we are able to decouple the channels.
We recall that
T
o
s G
o
sCsI G
o
sCs
1
Cs
1
G
o
s
1
I
1
Hence, to be able to have a diagonal T
o
s it is
necessary that the structure of C(s) be the inverse of
the structure of G
o
(s). We observe that
(i) C(s) is not required to be equal to G
o
s
1
.
(ii) C(s) will usually be unstable to ensure integral
action and, in general, it should exhibit high
gains in the frequency bands where reference
and disturbances have signicant energy.
(iii) The choice structureCs structureG
o
s
1

gives the simplest controller structure to achieve


a diagonal T
o
(s). However with that controller
structure, other than diagonal structures for
T
o
(s) can also be achieved.
For large scale systems, where the number of inputs
and outputs precludes a selection by inspection, it is
possible to solve the controller structure problem using
a computer program, since the selection problem can be
set in a dynamic programming framework. In those high
complexity systems, the heuristic limits suggested above
tend to go down.
In general, there is no simple theory yet to say what
is the minimum S which makes a given nominal model
acceptable. However, in the authors experience, the
nominal model should yield S > 0.70 to be on the safe
side. When S - 0.5, it is highly likely that no controller
with the selected structure yields satisfactory perform-
ance. If a high S is achieved mainly by the contribution
of a reduced (in comparison with the total number of
entries, (p p) number of elements in the PM, then
useful insight is being obtained, since that situation is
an indication of a dynamics dominance in the system.
More will be said when closed loop design specications
are discussed in } 7.
Some additional examples are presented next for
further illustration of the proposed ideas.
Example 5 (Continuous-time system): Consider a 3 3
system with transfer function
Gs
0.4
s 1
2
4s 3
s 2s 5
2
s 4
2
s 2s 1
2
s 2
2
1
s 2
6s 1
s 5s 4
4
s 3
2
8
s 2s 5
2
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
5
43
The PM for this system is
(
u
1
u
2
u
3
0.0370
0.2226
0.2193
2
6
4
0.2018
0.0578
0.0457
0.0385
0.0385
0.1389
3
7
5
y
1
y
2
y
3
44
Note that this system has non-minimum phase zeros
at 0.1134 j1.8216.
Case 1. Decentralized controller: In this case we need
to select three elementary systems, to pair every input
to a dierent output. The PM matrix suggests that the
best pairing (the reader can verify that any other pairing
yields a lower S) would be (u
1
, y
2
), (u
2
, y
1
) and (u
3
, y
3
),
this yields
S
d

12

21

33
0.2018 0.2226 0.1389 0.5633
45
The nominal plant model would be
G
o
s
0
4s 3
s 2s 5
0
2
s 2s 1
0 0
0 0
8
s 2s 5
2
6
6
6
6
6
6
4
3
7
7
7
7
7
7
5
46
Then
structureG
o
s
0 0
0 0
0 0
2
6
4
3
7
5 ) structureCs
structureG
o
s
1

0 0
0 0
0 0
0
B
@
1
C
A
47
MIMO interaction measure and controller structure selection 375
Therefore, in this case, the controller will have the
structure
Cs
0 C
12
s 0
C
21
s 0 0
0 0 C
33
s
2
6
6
4
3
7
7
5
48
One can observe that the selected structure yields
a value for S which is just over 50% of the overall
combined measure of controllability and observability.
In the next case, to increase S, we add an additional
term in the nominal plant model and in the controller.
Case 2. Sparse controller: Assuming the initial pairing
chosen in Case 1, the next largest contribution to S
would be obtained on making u
1
driving not only y
2
but also driving y
3
, this yields
S
s

12

21

33

31
0.2018 0.2226 0.1389 0.2193 0.7826
49
For this choice, the plant nominal model is
G
o
s
0
4s 3
s 2s 5
0
2
s 2s 1
0 0
6s 1
s 4s 5
0
8
s 2s 5
2
6
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
7
5
50
Then
structureG
o
s
0 0
0 0
0
0
B
B
@
1
C
C
A
) structureCs
structureG
o
s
1

0 0
0 0
0
0
B
B
@
1
C
C
A
51
and the corresponding controller is
Cs
0 C
12
s 0
C
21
s 0 0
0 C
32
s C
33
s
2
6
4
3
7
5 52
In this case, the selected structure yields a value for S
which is over 75% of the overall combined measure of
controllability and observability.
Case 3. Block diagonal controller: Assuming the sparse
structure chosen in Case 2, a block diagonal structure,
to increase S, can be formed by making u
3
depending
not only on y
3
but also on y
2
, this yields
S
b

12

21

23

31

33
0.2018 0.2226 0.0385 0.21930 0.1389
0.8211
53
For this choice, the plant nominal model is
G
o
s
0
4s 3
s 2s 5
0
2
s 2s 1
0
1
s 2
6s 1
s 4s 5
0
8
s 2s 5
2
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
5
54
Note that this model has non-minimum phase zeros
at 0.3119 j1.7905.
Then
structureG
o
s
0 0
0
0
0
B
B
@
1
C
C
A
) structureCs
structureG
o
s
1

0
0 0
0
0
B
B
@
1
C
C
A
55
The corresponding controller is then
Cs
0 C
12
s C
13
s
C
21
s 0 0
0 C
23
s C
33
s
2
6
4
3
7
5 56
In this case, the selected structure implies that we use
a full MIMO controller for the 2 2 subsystem with
inputs u
1
and u
3
and outputs y
2
and y
3
. The controller
structure is completed by a SISO controller for the
scalar subsystem with input u
2
and output y
1
. This
arrangement yields a value for S which is only slightly
larger than that obtained for the sparse structure; this
suggests that there is little incentive for designing and
using a block diagonal controller instead of a sparse
controller.
6.1. Discussion of the results
The PM for the system in this example seems to
contradict the analysis of the results in table 1 in one
respect: the presence of a non-minimum phase zero in
G
31
s does not make
13
large enough to force the pair-
ing u
1
, y
3
into a decentralized controller. This is due to
376 M. E. Salgado and A. Conley
the small d.c. gain in G
31
s, equal to 0.3 when compared
to the d.c. gain of the main competitor, G
21
s, which has
a unit d.c. gain. To go deeper into this discussion we
consider two cases.
Change 1: The d.c. gain of G
31
s is changed from 0.3
to 0.45.
Change 2: The zero is shifted from 1 to 0.5, keeping the
d.c. gain at its original value, i.e. equal to 0.3
Change 1
In this case
G
31
s
9s 1
s
2
9s 20
) (
u
1
u
2
u
3
0.0290
0.1747
0.3873
2
6
4
0.1584
0.0454
0.0358
0.0302
0.0302
0.1090
3
7
5
y
1
y
2
y
3
The change in the d.c. gain of G
31
s yields a signi-
cant increase in the value of
13
, making it mandatory
to have u
1
depending on the measurement of y
3
in any
controller structure. In particular, for a decentralized
controller, the pairing u
1
, y
3
makes it possible to
reach the maximum value for S, which is 0.5759. Note
that the d.c. gain of G
31
s is still less than half the d.c.
gain of G
21
s.
Change 2
In this case
G
31
s
62s 1
s
2
9s 20
) (
u
1
u
2
u
3
0.0251
0.1512
0.4698
2
6
4
0.1370
0.0392
0.0310
0.0262
0.0262
0.0943
3
7
5
y
1
y
2
y
3
We have moved the non-minimum phase zero
closer to the origin, while maintaining the d.c. gain at
its original value. Now the predominance of
13
is even
stronger than when the rst change was introduced,
forcing the dependence of u
1
on y
3
in any controller
structure. In particular, for a decentralized controller,
the pairing u
1
, y
3
makes it possible to reach the maxi-
mum value for S, which is 0.633.
A nal point to note is that the RGA for this plant
is given by
RGAG
0.0831 0.9111 0.1720
1.3809 0.2745 0.1064
0.2979 0.3634 0.9345
2
4
3
5
57
It can be seen that the RGA suggests the same
diagonal structure as the PM does. However, there is
no indication in the RGA on how to select a more
complex model (sparse or block diagonal).
To complement the illustration of the interaction
quantier and the controller structure selection we will
next consider two examples in discrete-time.
Example 6 (Sampled-data control): We assume that
the plant in Example 5 is going to be under digital con-
trol, using a sampling period of 0.1 s and a zero-order
hold. We then need to choose a controller structure
based upon a sampled data model. That model has a
transfer function given by
G
d
z
0.001872z 0.001751
z
2
1.81z 0.8187
0.3307z 0.2451
z
2
1.425z 0.4966
0.1648
z 0.6703
0.009056z 0.008194
z
2
1.724z 0.7408
0.008762z 0.007668
z
2
1.637z 0.6703
0.09063
z 0.8187
0.3604z 0.3993
z
2
1.277z 0.4066
0.01642z 0.01344
z
2
1.482z 0.5488
0.03184z 0.02522
z
2
1.425z 0.4966
2
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
5
58
The PM for this system is
(
u
1
u
2
u
3
0.0319
0.1943
0.2394
2
4
0.2289
0.0514
0.0418
0.0461
0.0389
0.1273
3
5
y
1
y
2
y
3
If a decentralized controller is to be used, the pairing
yielding the highest S is the same as in the continuous
case, i.e. u
1
, y
2
, (u
2
, y
1
) and (u
3
, y
3
). A highly relevant
feature in this example is that it illustrates the fact that
the largest
ij
must not necessarily be included to form
the controller structure. In this example the largest
ij
is
31
; however, the pairing (u
1
, y
3
) would yield at most
S 0.5072.
The next example highlights the ability of PM to
assess the signicance of time delays in MIMO inter-
action and therefore in the controller structure selection.
This example also veries the validity of Lemma 4.
Example 7 (Sampled-data system with time delays):
The sampled-data transfer function of a 2 2 plant
has a transfer function given by
Gz
0.5
z 0.5
0.15
z 0.8z
l
0.1
z 0.5z 0.8
0.3
z 0.7
2
6
6
4
3
7
7
5
59
where is a non-negative integer, quantifying a pure
time delay in the path from u
2
to y
1
. The participation
matrix is next computed for 0, for 3, and for
10. The results are
(j
0

0.3171 0.1239
0.3122 0.2469
!
; (j
3

0.2797 0.2272
0.2754 0.2177
!
;
(j
10

0.2193 0.3941
0.2159 0.1707
!
60
MIMO interaction measure and controller structure selection 377
When 0, the PM indicates that a decentralized
controller should be based upon the pairings (u
1
, y
1
)
and (u
2
, y
2
), leading to S 0.3171 0.2469 0.5571.
However, when 3, the information provided by the
PM is ambiguous. But, when 10, the PM indicates an
anti-diagonal controller, i.e. the pairings (u
1
, y
2
) and
u
2
, y
1
, yielding S 0.3941 0.2159 0.61. The quali-
tative nature of these results is in agreement with intui-
tion, since as the delay increases, the authority of u
2
on
y
1
grows in importance in the overall system dynamics.
Note that the four static gains in Gs are comparable.
The values for S in all cases are similar. To increase
its value, a triangular structure in the controller should
be used. In each case, the structure of the controller is
determined by building the structure of G
o
s
1
.
When 0, a sensible triangular controller is
obtained if we make u
1
depending on y
1
and if, at the
same time, u
2
depends on y
1
and on y
2
, this yields a pair
(u
1
, y
1
) and a triad (u
2
, y
1
, y
2
) leading to S 0.3171
0.2469 0.3122 0.8693. When 3, a triangular
controller is built with the triad (u
2
, y
1
, y
2
) and the pair
(u
1
, y
2
); in this case S 0.2797 0.2754 0.2272
0.7823. The same controller structure should be chosen
when 10, leading to S 0.3941 0.2159 0.2193
0.8293
7. Control design and the participation matrix
The PM idea is not tied to any particular design
methodology, since its role is to provide support to the
designer in the preliminary and crucial stage of building
a nominal model for a multivariable plant starting from
a full MIMO linear, stable model.
In any linear MIMO control design problem we
assume that we have a calibration model such as the
transfer function (2), which is already an approximate
description of the process to be controlled. When a
decentralized approach is pursued, what we are doing
is to approximate the calibration model by a nominal
model of the form (3). We know that, in doing so, addi-
tional modelling errors are introduced; however we are
prepared to accept that (i.e. its deleterious eects on the
control performance) in exchange for simplicity in the
controller design and other practical advantages. In this
case, the role of the PM (and other alternative tools) is
to help us to choose a suitable inputoutput pairing to
build a decentralized description of the plant.
If no decentralized controller is able to deliver an
acceptable performance, we must rene our nominal
model, by including one or more signicant interactions.
This is a stage where the contribution of the PM
becomes unique: it helps us to decide which of the inter-
actions are the most signicant (and to provide a sensi-
ble denition of signicant interactions). The rening of
the nominal model may involve several iterations going
from decentralized to full MIMO control, through
sparse control, triangular and block diagonal. In a prac-
tical situation, when a MIMO plant with many inputs
and many outputs is to be controlled, one would aim,
at least, to break the control design problem in several
simpler MIMO control problems, i.e. we are aiming to
describe our process using a block diagonal nominal
model. The advantage of doing this is manyfold: design
simplicity, robust performance, integrity, online tuning,
maintenance, and so on. The PM is one tool to support
this approach.
Except when dealing with a decentralized nominal
model, which breaks the problem into p-SISO designs,
in all other cases we need to use the available tools
provided by the MIMO control theory and practice.
The proposal of a tool like the PM does not pose
problems that are dierent to those which are
currently research topics in the eld.
Our preferred choice to synthesize a controller can
be explained considering the fact that its structure is the
inverse of the structure of the nominal model. We can
use the Youla parametrization of all stabilizing control-
lers for a stable plant (see, e.g. Maciejowski 1989, Glad
and Ljung 2000, Goodwin et al. 2001). In this method-
ology, the controller satises
Cs I QsG
o
s
1
Qs QsI G
o
sQs
1
61
where Qs is any stable proper transfer function matrix.
We recall that the Youla-parameterized control loop has
the form shown in gure 2.
It is then straightforward to verify that
structureCs structureG
o
s
1
if structure
Qs structureG
o
s
1
and if Q(s) has the same
structure and if I G
o
sQs is either diagonal or has
also that structure. Therefore, if the PM leads to a block
diagonal G
o
s, then C(s) must be block diagonal, and
that is achieved if Q(s) is chosen to be block diagonal
of consistent dimensions. For instance, in Case 3 in
Example 5, the design problem breaks into a set of
one SISO design problem and one 2 2 full MIMO
design problem. In that case Q(s) must be chosen
(modulo a column permutation) as a diagonal matrix
with a 1 1 block and a 2 2 block. A similar sym-
metry can be built for triangular, block triangular and
sparse models.
Figure 2. Control loop with Youla parameterization.
378 M. E. Salgado and A. Conley
At this stage, the fact that the model G
o
(s) was
obtained using either PM or other tool is irrelevant,
since the issues to face now are classical MIMO design
issues, some of which are discussed below.
7.1. Stability
The controller Q(s) is designed using the nominal
model G
o
(s), thus stability of Q(s) is necessary and su-
cient for the stability of the nominal control loop. Since
G
o
(s) is open loop stable, there is always a controller of
the given structure, that stabilizes the loop. This can be
proved either using root locus arguments or xed mode
ideas (see, e.g. Anderson and Clements 1981, Tarokh
1985). However, since this controller must be able to
control the real plant, we must guarantee that, at least,
the designed Q(s) stabilizes the calibration model G(s).
This is a robust stability issue and has to be dealt with
using the existing theory which mainly relies on the
singular values of the modelling error. A simulation
of the control loop with the calibration model can
anticipate whether the design is robust stable. If the
simulation result is unsatisfactory, the designer can
either modify the bandwidths or he/she can use the
PM to rene the nominal model, incorporating new
interactions to diminish the modelling error. It is not
known so far a quantitative relationship between the
PM and the degree of robustness. However it is sensible
to assume that the larger S is, the more robust is the
control design. This assumption is consistent with the
results obtained in the case studies.
7.2. Performance
The Youla approach leads to a nominal complemen-
tary sensitivity function which is given by
T
o
s G
o
sQs; where Ys T
o
sRs 62
This highlights another known feature of the Youla
parametrization, namely the controller Q(s) should be
chosen as close as possible to the inverse of the nominal
plant model. In other words, the control design problem
can always be formulated as the building of a stable and
proper plant inverse which, at the same time, satises
a web of constraints (linear and non-linear) and funda-
mental limitations (Goodwin et al. 2001). In the limit,
i.e. when Qs G
o
s
1
, perfect tracking would be
achieved, although one of the trade-os implies to
negotiate a good inverse against requirements such as
having a proper and stable Q(s).
A key issue is that the underlying inversion require-
ment in the Youla approach is consistent with the
requirement, derived from the PM, that C has the
same structure as the inverse of the nominal model
G
o
(s), since this requires that Q(s) has the same structure
as G
o
s
1
.
We next illustrate these ideas with Cases 1 and 2 in
Example 5, where a signicant dierence exists in the
values for S. The nominal models, diagonal and sparse,
given by (46) and (50) are denoted by G
od
s and G
os
s,
respectively, and their inverses are
G
od
s
1

0 0.5s 1s 2 0
0.25s 2s 5
s 3
0 0
0 0 0.125s 2s 5
2
6
6
4
3
7
7
5
63
G
os
s
1

0 0.5s 1s 2 0
0.25s 2s 5
s 3
0 0
0 0.375
s 2
2
s
2
1
s 4
0.125s 2s 5
2
6
6
6
6
4
3
7
7
7
7
5
64
We immediately observe that in none of these cases
we can achieve the ideal Q(s), i.e. that leading to
T
o
s I, since the inverses in (63) and (64) are impro-
per. A simple strategy to overcome this problem is to
choose Q(s) as
Qs G
od
s
1
Hs 0 0
0 Hs
2
0
0 0 Hs
2
2
6
4
3
7
5;
Qs G
os
s
1
Hs 0 0
0 Hs
3
0
0 0 Hs
2
2
6
4
3
7
5
65
for the diagonal and the sparse cases, respectively. In
(65) Hs 1,ts 1, and t is a small positive number.
We thus ensure that the complementary sensitivity T
o
(s),
for both cases, is close to the identity matrix I in a
frequency band determined by the choice of t. This
setting, although simple, due to the lack of other con-
straints, allows us to evaluate the eectiveness of the
PM for comparable nominal designs. To perform the
evaluation we use the loop shown in gure 2, where
now the calibration model (43) substitutes the block
named Plant, and the nominal model G
o
(s) is replaced
by either G
od
(s) or by G
os
(s). The conditions for the
simulation are:
(i) The loop is initially at rest.
(ii) Step references are applied at dierent instants,
in particular r
1
t jt 1, r
2
t jt 10
and r
3
t jt 20 where jt is the unit step.
The reference levels give the same weight to the
performance in each of the three channels. Also,
their time proles have been chosen to facilitate
the display of the results.
MIMO interaction measure and controller structure selection 379
(iii) A quadratic measure of the performance is
computed on line, as
J

t
0
et
T
et dt 66
where et is the vector error rt yt.
The performance of both loops are shown in
gures 3 and 4.
The cost function (66) has been also computed for
both cases and it is shown in gure 5.
A brief analysis of the above results shows that:
(i) Both designs are robust stable. However, it can
be veried that this does not hold if t ( 1.
(ii) The choice of t 0.1 has been made only for the
sake of illustration of the control methodology.
In a more realistic environment, there are other
requirements (such as noise, input saturation and
0 5 10 15 20 25 30
2
1
0
1
2
3
4
Time [s]
P
l
a
n
t

o
u
t
p
u
t
s
y
1
(t)
y
2
(t)
y
3
(t)
Figure 3. Plant outputs. Diagonal controller. t 0.1.
0 5 10 15 20 25 30
2
1
0
1
2
3
Time [s]
P
l
a
n
t

o
u
t
p
u
t
s
y
1
(t)
y
2
(t)
y
3
(t)

Figure 4. Plant outputs. Sparse controller. t 0.1.
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
Time [s]
C
o
n
t
r
o
l

e
r
r
o
r

m
e
a
s
u
r
e
J
diag
(t)
J
sparse
(t)
Figure 5. Accumulated control errors. t 0.1.
380 M. E. Salgado and A. Conley
transient performance) likely to suggest a much
larger t. This issue is also connected to the
discussion below.
(iii) The sparse controller provides a better perfor-
mance than the diagonal controller. However
those results change with the reference direction-
ality. For instance, if we use r
1
t r
2
t
r
3
t jt 1, i.e. the direction is 1 1 1
T
and
the three references change simultaneously, then
the performance of the sparse controller would
be even better than that of the diagonal case.
(iv) Several other references might be tried. However
in every trial, the references in all channels should
have the same magnitude; otherwise we would
be allocating arbitrary weights to the dierent
channels.
In a general framework, the performance of the
nominal control loop is limited by the presence of non-
minimum phase zeros in G
o
(s). In this situation, Q(s)
can never be made equal to G
o
s
1
, since then Q(s)
would be unstable. It is also known that non-minimum
phase zeros limit the bandwidth of the closed loop
(Goodwin et al. 2001). This is the situation for Case 3
in Example 5 where, as noted, there are two zeros
located at 0.3119 j1.7905. A general situation can be
dealt with using sophisticated techniques such as factor-
ization with zero-interactors (Wolovich and Falb 1976,
Goodwin and Sin 1984, Goodwin et al. 2001) or equiva-
lent innerouter factorization (Havre 1998, Zhou and
Doyle 1998). Going into a design of that complexity
is worthwhile only if it improves signicantly the loop
performance. This can be analyzed comparatively using
the value of S for the competing structures.
7.3. Closed loop bandwidth specication
In the derivation of the participation matrix and
in the development of its applications to controller
structure selection no design specication has been
considered. A natural question is how to modify the
proposed tools when a closed loop bandwidth constraint
is introduced. To address that issue we consider a
standard MIMO nominal control loop, which is
shown in two equivalent (assuming a stable nominal
model) structures in gure 6.
We observe from gure 6 that
Ys G
o
sUs G
o
sS
uo
sRs T
o
sRs
where S
uo
s is the control sensitivity (Goodwin et al.
2001). If we assume no constraints on the reference,
we thus observe that the plant input spectrum is shaped
by a lter, where the ltering characteristics are deter-
mined by the control sensitivity.
A simple yet eective way to introduce the band-
width limitation is to choose a diagonal transfer func-
tion Fs which captures essential features of S
uo
s.
Every diagonal entry in Fs is associated to the intended
closed loop bandwidth. In this strategy the (i, i) element
in Fs is then dened by
F
ii
s
A
o
s
A
i
s
67
where A
o
(s) and A
i
(s) are stable polynomials, with A
o
(s)
containing, as roots, the poles (at least the dominant
ones) in the ith column of the plant transfer function.
Also, A
i
(s) having roots suggested by the intended band-
width in the ith channel. Note that when the closed loop
is specied to have a larger bandwidth than the plant,
then (67) describes a high-pass lter; when the situation
is the other way around, then (67) describes a low-pass
lter. In every case, F(s) is normalized at d.c., i.e. to
have F0 I; this avoids the introduction of arbitrary
scaling.
Once a diagonal Fs is chosen, we compute the PM
for the ltered plant G
f
(s), which is dened as
G
f
s G
o
sFs
The introduction of this (or a similar) type of
ltering is necessary to capture some essential design
trade-os. For instance, we know that a delay or a
non-minimum phase zero is model-relevant depending
on the closed loop bandwidth. Something similar applies
to slow poles, resonant modes and so on. To illustrate
this approach we next consider an example.
Example 8: Consider the same plant as in Example 5.
We observe that the dominant system poles are located
at 1, 2 and 2 for the rst, second and third column
respectively. Further assume that the dominant closed
loop poles in the three channels are at 4. Hence a
suitable choice could be
Fs diag 4
s 1
s 4
, 2
s 2
s 4
, 2
s 2
s 4
& '
where the static gain in all entries has been normalized
to 1.
Figure 6. Equivalent loop.
MIMO interaction measure and controller structure selection 381
We next compute the PM for the ltered function
which gives
(
u
1
u
2
u
3
0.0094
0.0645
0.5740
2
4
0.2279
0.0236
0.0300
0.0167
0.0112
0.0427
3
5
y
1
y
2
y
3
If a decentralized controller structure is chosen, then
the pairing should be (u
1
, y
3
), u
2
, y
1
and u
1
, y
3
leading
to S
13

21

32
0.8091. An interesting feature
of this result is that the entry
13
, corresponding to
G
31
s, is now the predominant element. This is due not
to its d.c. gain, which is small, but to its non-minimum
phase zero. The strong presence of this feature is due to
the choice of the dominant closed loop pole, which is
much faster than the non-minimum phase zero; this
describes one of the design conicts (see Goodwin
et al. 2001). To investigate this issue a bit further we
make the dominant closed loop pole equal to 8. We
thus choose
Fs diag 8
s 1
s 8
, 4
s 2
s 8
, 4
s 2
s 8
& '
leading to
(
u
1
u
2
u
3
0.0046
0.0315
0.6695
2
4
0.2210
0.0124
0.0187
0.0139
0.0060
0.0225
3
5
y
1
y
2
y
3
By making the closed loop even faster, an even more
predominant role is played by the non-minimum phase
zero in G
31
s.
The use of this type of ltering can be made part of
the iteration process which normally takes place in any
control design problem. One has to keep in mind that
the eect of the lter is meaningful only if the frequency
response of Gs is signicantly dierent to that of
GsFs. Also, the benet of this ltering strategy does
not lie in numerical accuracy but in that it reveals an
interplay of the dierent interactions as the projected
closed loop bandwidth changes. As seen in Example 8,
when the closed loop speed is increased, some entries
in the PM become more and more signicant. Those
particular elements are associated to dynamic features
which are usually hard to deal with in any control design
process. In those cases a high S is obtained by including
only a couple of elements in the PM. That is an
indication that the specied closed loop bandwidth is
unsuitable for the given plant and that, consequently,
a design trade-o has to be achieved by choosing a
dierent bandwidth.
The limitations of this ltering strategy as an accu-
rate numerical tool are originated in the simplicity of
the lter (67) and in the fact that usually the reference
is not a wide-band signal. To overcome this latter prob-
lem, one could also model the reference as the output of
an additional lter with white noise input; however, that
is beyond the scope of this paper.
8. Conclusions
In this paper a new measure of dynamic channel inter-
action in stable MIMO systems has been proposed. This
measure is based upon the system controllability and
observability gramians. It thus makes use of the ability
of the gramians to describe the diculty (or otherwise)
to observe and to control the system state. In particular,
it has been shown that the sum of the squared Hankel
singular values reveals several relevant control design
trade-os.
The gramian-based measure, which we have called
the participation matrix (PM), is a matrix of numbers, it
is built upon a dynamic plant model and it has no
limitations regarding the number of plant inputs and
outputs. Some criteria have been associated to the pro-
posed measure to pair inputs and outputs in a decentral-
ized control architecture. However, the key feature of
the PM is its unique ability to provide support for a
richer controller architecture selection in continuous
and discrete-time time frameworks, including triangular,
block diagonal and sparse structures. Furthermore,
preliminary results show that this measure helps the
designer to comparatively assess the benets of those
richer controller structures. The PM also sheds light
on fundamental design limitations and highlights
dominant dynamic features (small stable zeros, non-
minimum phase zeros, time delays and resonant
peaks). It has been also shown that the PM, when
applied to a suitable modied plant model, can provide
information regarding interaction as a function of a
projected closed loop velocity. In this usage, the PM
also helps to spot design conicts. Those are the main
contributions of the ideas proposed in this paper.
Further research should include extension of the PM
ideas to unstable processes (which would require a new
denition of gramians (e.g. as in Zhou et al. 1999) and
the building of more precise numerical guidelines to
select the controller structure. Optimality with restricted
structure controllers and input-saturation eects are
also challenges for future research.
Acknowledgements
Helpful discussions with Professor Pedro Albertos
from Valencia Polytechnic University, Spain, are
gratefully acknowledged. The authors are also grateful
for the support received from CONICYT-Chile
through grant FONDECYT-1040313.
382 M. E. Salgado and A. Conley
References
Anderson, B. D. O., and Clements, D. J., 1981, Algebraic
characterization of xed modes in decentralized control.
Automatica, 17, 703712.
Bristol, E. H., 1966, On a new measure of interaction for
multivariable process control. IEEE Transactions on
Automatic Control, 11, 133134.
Bristol, E. H., 1978, Recent results on interaction in multi-
variable process control. In 71

AIChE Conference.
Chiu, M.-S., and Arkun, Y., 1991, A new result on relative
gain array, Niederlinski index and decentralized stability
condition: 2 2 plant case. Automatica, 27, 419421.
Conley, A., and Salgado, M. E., 2000, Gramian based
interaction measure. In 34th CDC Conference Proceedings,
Sydney, Australia.
Doyle, J., Francis, B., and Tannenbaum, A., 1992,
Feedback Control Systems (New York: Macmillan
Publishing Co.).
Gagnon, E., Desbiens, A., and Pomerleau, A., 1999,
Selection of pairing and constrained robust decentralized
PI controllers. In Proceedings of the American Control
Conference, pp. 43434347.
Glad, T., and Ljung, L., 2000, Control Theory. Multivariable
and Nonlinear Methods (London: Taylor and Francis).
Glover, K., 1984, All optimal Hankel norm approximations
of linear multivariable systems and their l
1
-error bounds.
International Journal of Control, 39, 11151193.
Goodwin, G. C., and Sin, K., 1984, Adaptive Filtering
Prediction and Control (Englewood Clis, NJ: Prentice-
Hall).
Goodwin, G., Graebe, S., and Salgado, M. E., 2001,
Control System Design (Upper Saddle River, NJ: Prentice-
Hall).
Havre, K., 1998, Studies on controllability analysis and
control structure design. PhD thesis, Department of
Chemical Engineering, Norwegian University of Science
and Engineering.
Kwakernaak, H., and Sivan, R., 1972, Linear Optimal
Control Systems (New York: Wiley Interscience).
Maciejowski, J., 1989, Multivariable Feedback Design
(Wokingham, England: Addison Wesley).
Niederlinski, A., 1971, A heuristic approach to the design of
linear multivariable interacting control systems. Automatica,
7, 691701.
Skogestad, S., and Morari, M., 1987, Implications of
large RGA elements on control performance. Industrial
Engineering in Chemical Research, 26, 23232330.
Tarokh, M., 1985, Fixed modes in multivariable systems
using constrained controllers. Automatica, 21, 495497.
vande Wal, M., and de Jager, B., 2001, A review of methods
for input/output selection. Automatica, 37, 487510.
Witcher, M. F., and McAvoy, T. J., 1977, Interacting
control systems: Steady-state and dynamic measurement of
interaction. ISA Transactions, 16, 3541.
Wittenmark, B., and Salgado, M., 2002, Hankel norm
based interaction measure for input-output pairing. In 15th
IFAC World Congress Conference Proceedings, Barcelona,
Spain.
Wolovich, W., and Falb, P., 1976, Invariants and canonical
forms under dynamic compensation. SIAM Journal of
Control and Optimization, 14, 9961008.
Zhou, K., and Doyle, J., 1998, Essentials of Robust Control
(Englewood Clis, NY: Prentice Hall).
Zhou, Z., Salomon, G., and Wu, E., 1999, Balanced realiza-
tion and nodel reduction for unstable systems. International
Journal of Robust and Nonlinear Control, 9, 183198.
Zhu, Z.-X., 1996, Variable pairing selection based on indivi-
dual and overall interaction measures. Industrial Engineering
in Chemical Research, 35, 40914099.
MIMO interaction measure and controller structure selection 383

You might also like