Professional Documents
Culture Documents
To cite this article: Mario E. Salgado & Arthur Conley (2004): MIMO interaction measure and controller structure selection,
International Journal of Control, 77:4, 367-383
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
INT. J. CONTROL, 10 March 2004, VOL. 77, NO. 4, 367–383
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 input–output 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 its insensitivity to delays and the fact that only one
Most industrial processes are multivariable in nature point of the process frequency response is considered.
and the control designer usually hopes that, with a After Bristol’s work was published, several research-
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
proper input–output pairing, there is no significant ers have studied the properties and usage of the RGA
channel interaction, since that makes a process variable (see, e.g. Witcher and McAvoy 1977, Skogestad and
hard to control without perturbing other variables of Morari 1987). Some others have proposed new measures
interest. This problem is embedded into the more gen- of interaction and criteria to choose a sensible input–
eral problem of controller structure selection (ranging output pairing. In 1971 Niederlinski suggested to use
from decentralized, up to full MIMO). However, before an index (the Niederlinski index, NI) which is also
dealing with this problem, the designer must decide on built upon the process d.c. gain matrix (Niederlinski
which inputs and which outputs will be used to build 1971). Although no additional information is obtained
any control architecture (van de Wal and de Jager 2001). regarding input–output pairing, this index provides
Throughout this paper we will assume that the decision direct information on the ability of a decentralized con-
has already been made regarding the input–output trol to stabilize a 2 2 MIMO system (Chiu and Arkun
selection problem. Among other things, this assumption 1991). A variation of the RGA was proposed by Zhu
will reflect upon the fact that we will be dealing only (1996); his proposal is known as the relative interaction
with square systems, i.e. equal number of inputs and array (RIA) and it is based on the concept of viewing
outputs. the interaction as an unmodelled term for a particular
Decentralized control, although it is a limited pairing (at d.c.).
flexibility choice, has several well-known advantages, The NI and the RIA do not encode more informa-
such as tuning, sequencing of loop closing and the tion on the process than the RGA, although they
possibility to use the knowledge and intuition built provide valuable alternative viewpoints. The common
around the control design of single input–single output feature of those indices is that they only use the
systems. system model at zero frequency. Dynamic features
In the decentralized (diagonal) architecture, a key were introduced in the relative dynamic gain array
issue is how inputs and outputs are paired. This issue (RDGA), (Witcher and McAvoy 1977, Bristol 1978).
has received a lot of attention over the last four decades. The RDGA also provides information of how inter-
The most significant result is the seminal work of Bristol action varies with frequency, suggesting bandwidths
(1966), who developed the idea of relative gain array for alternative pairings. That line of work was con-
(RGA). In the RGA, the channel interaction measure tinued by Gagnon and co-workers (Gagnon et al.
is built upon the d.c. gain of the MIMO process. 1999) through the generalized relative dynamic gains
Throughout the years the RGA has proved to be a (GRDG). More specifically, the frequency response of
useful tool. However it has limitations which have the process and that of a complementary sensitivity
been explored elsewhere, among them is the inability target are used to test different input–output pairings.
to cope with certain non-minimum phase structures, 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
Received 1 May 2003. Revised and accepted 13 January is used to denote the web of paths from the (vector)
2004.
* Author for correspondence. e-mail: msb@elo.utfsm.cl input to the (vector) output, independently of the dyna-
y Department of Electronic Engineering, Universidad mical order of each scalar subsystem. Hence, if we have
Técnica Federico Santa Marı́a, Chile. a p p MIMO system with transfer function HðsÞ, input
International Journal of Control ISSN 0020–7179 print/ISSN 1366–5820 online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/0020717042000197631
368 M. E. Salgado and A. Conley
vðtÞ ¼ Rp and output wðtÞ ¼ Rp , its structure will be sensitivity and the nominal complementary sensitivity
described by a symbolic array, as illustrated in (1) are given by
v1 v2 v3 vp1 vp So ðsÞ ¼ ðI þ Go ðsÞCðsÞÞ1
0 1 w ¼ diag fSo1 ðsÞ, So2 ðsÞ, . . . , Sop ðsÞg ð5Þ
0 0 1
1
B C w To ðsÞ ¼ Go ðsÞCðsÞðI þ Go ðsÞCðsÞÞ
B 0 0 C 2
B C ¼ diag fTo1 ðsÞ, To2 ðsÞ, . . . , Top ðsÞg ð6Þ
B C
B 0 0 0 0 C w3
B C where
structureðHðsÞÞ ¼ B .. .. .. .. .. .. C ..
B . . . . . .C
B C . So‘ ðsÞ ¼ ð1 þ G‘‘ ðsÞC‘ ðsÞÞ1 ð7Þ
B C
B 0 0 C
@ A wp1 To‘ ðsÞ ¼ G‘‘ ðsÞC‘ ðsÞð1 þ G‘‘ ðsÞC‘ ðsÞÞ 1
ð8Þ
0 0 0 wp for ‘ ¼ 1, 2, . . . , p.
ð1Þ We need now to assess the effect of our decision
to design a decentralized controller for a diagonal
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
where a -entry in the (i, j) position signals the presence plant model. When the controller (4) is used to control
of a subsystem with input vj(t) and output wi(t), and (2), the achieved sensitivity is given by
Hij 6¼ 0. On the contrary, a 0-entry in the (k, ‘) position
indicates that there is no direct dependency of the SðsÞ ¼ ðI þ GðsÞCðsÞÞ1 ð9Þ
output w‘ ðtÞ on input vk(t). Furthermore, we will refer which is, in general, a non-diagonal transfer function.
to the structure of HðsÞ1 as the inverse structure of HðsÞ, A key result is then given by the following lemma.
provided that HðsÞ is non-singular almost everywhere.
The framework for the problem to be dealt with in Lemma 1: Consider the nominal loop defined by the pair
this paper is described next. ðGo ðsÞ, CðsÞÞ and the true loop defined by ðGðsÞ, CðsÞÞ,
We consider a plant with p inputs and p outputs where GðsÞ, Go ðsÞ and CðsÞ are given by (2), (3) and (4)
which can be modelled as a multivariable process with respectively. Then the nominal sensitivity So ðsÞ and the
matrix transfer function given by achieved sensitivity SðsÞ are related by
2 3 SðsÞ ¼ So ðsÞ½I þ HT ðsÞ1 ð10Þ
G11 ðsÞ G12 ðsÞ G13 ðsÞ G1p ðsÞ
6 G ðsÞ G ðsÞ G ðsÞ G ðsÞ 7 where the element (i, j) of the matrix HT ðsÞ is given by
6 21 22 23 2p 7
6 7 8
GðsÞ ¼ ½Gij ðsÞ ¼ 6 . .. .. .. .. 7 < Gij ðsÞ
6 .. . . . . 7 T ðsÞ i 6¼ j
4 5 ½HT ðsÞij ¼ Gjj ðsÞ oj ð11Þ
Gp1 ðsÞ Gp2 ðsÞ Gp3 ðsÞ Gpp ðsÞ :
0 i¼j
ð2Þ Proof: We first note that
Assume next that, using one of the known tools GðsÞ ¼ ðI þ Gl ðsÞÞGo ðsÞ
to decide on a good pairing of inputs and outputs, a ð12Þ
¼) Gl ðsÞ ¼ ðGðsÞ Go ðsÞÞðGo ðsÞÞ1
decentralized controller is designed. If the inputs
and outputs are re-labelled, when necessary, one only where Gl ðsÞ denotes the left multiplicative error. The
needs to design p independent SISO control loops, for ði, jÞ-element of Gl ðsÞ can be computed from (2) and (3)
G11 ðsÞ, G22 ðsÞ, . . . , Gpp ðsÞ. This approach implies that the to yield
plant is nominally described by a non-interacting model 8
and that interactions are captured in the modelling <0 i¼j
error. Thus the nominal plant model is ½Gl ðsÞij ¼ Gij ðsÞ ð13Þ
: i 6¼ j
Gjj ðsÞ
Go ðsÞ ¼ diag fG11 ðsÞ, G22 ðsÞ, . . . , Gpp ðsÞg ð3Þ
We next have that (Goodwin et al. 2001)
and the designed diagonal controller has transfer SðsÞ ¼ So ðsÞ½I þ Gl ðsÞTo ðsÞ1 ð14Þ
function
If we now use (6) and (12) in (14) the result (10)
CðsÞ ¼ diag fC1 ðsÞ, C2 ðsÞ, . . . , Cp ðsÞg ð4Þ and (11) is obtained. œ
The resulting closed loop performance, including In (11), Toj ðsÞ is the nominal complementary sensi-
tracking, regulation, noise immunity and robustness tivity of the jth loop. We observe that if the frequency
properties can be described using sensitivity functions responses of the off-diagonal terms, ½HT ðsÞij , are very
(Goodwin et al. 2001). In particular, the nominal small, then the resulting loops are well decoupled and
MIMO interaction measure and controller structure selection 369
almost behave like p independent SISO control loops. full MIMO control (where every process input depends
This is achieved if GðsÞ is strongly (column) diagonally on all process outputs), but more versatile than the
dominant, at least in the frequency bands of To1 , simple decentralized (diagonal) controller. These addi-
To2 , . . . , Top . tional architectures include block diagonal, triangular,
In this framework we can pinpoint several important sparse controllers, etc. Few, if any, of the known
elements, namely indices and interaction measures are useful to evaluate
alternative controller structures other than diagonal
. The (plant and loop) coupling characteristics,
controllers.
in general, are frequency dependent.
In this paper we propose an interaction measure
. The (loop) coupling characteristics, in general, which is based upon a dynamic model of the process.
are dependent on the choices made for the SISO (A preliminary version of these results was presented
control designs. at the CDC 2000 in Sydney (Conley and Salgado
. Different dominance characteristics may lead to a 2000).). This measure also quantifies interaction as a
different pairing of inputs and outputs. function of chosen channel bandwidths, gives criteria
We want to focus on the situation when the above for input–output pairing and helps to assess alternative
controller architectures. The proposed index, which is
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
continuous-time and discrete-time systems. We will state will be weakly observable from yðtÞ if and only if it
initially focus on continuous-time systems. coincides with the associated eigenvector.
Assume that a p p stable MIMO system has a state In summary, gramians quantify how hard it is to
space representation given by the 4-tuple (A 2 Rnn , control and to observe the system state, and the ranks
B 2 Rnp , C 2 Rpn , 0 2 Rpp ), then the controllability of P and Q are the dimensions of the controllable sub-
gramian, P 2 Rnn , and the observability gramian, space and observable subspace respectively. However,
Q 2 Rnn , are symmetric non-negative definite matrices gramians depend on the state space realization. To
which satisfy the Lyapunov equations (16) extract valuable information, the product PQ is formed
and its eigenvalues, i (i ¼ 1, 2, . . . , n), are computed. It
AP þ PAT þ BBT ¼ 0 AT Q þ QA þ CT C ¼ 0 ð16Þ can be proved that these eigenvalues are non-negative
Also, the matrices P and Q can be expressed as and that they do not depend on the particular realiza-
ð1 ð1 tion (see e.g. Glover 1984 and Kwakernaak and Sivan
T AT t T
1972). The system Hankel singular values (HSV) are
P¼ At
e BB e dt Q¼ eA t CT C eAt dt ð17Þ
0 0 defined as
ðiÞ
pffiffiffiffi
H ¼ i i ¼ 1, 2, . . . , n ð21Þ
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
2.2. Physical interpretation where the i s are ordered to obtain H ð1Þ ð2Þ
H
ðmÞ
Useful interpretations for gramians can be derived H 0.
using energy concepts (see, e.g. Glover 1984). The HSV condense useful information regarding
system controllability and observability. For instance,
2.2.1. Controllability gramian. Consider the (stable) the number, r, of non-zero HSV corresponds to the
system described by the 4-tuple ðA, B, C, 0Þ. Find the dimension of the controllable and observable subspace,
optimal (in a quadratic sense) control uðtÞ in ð1Þ
S co . Also, the ratio H ðrÞ
=H is a measure of the skew-
t 2 ð1, 0Þ which steers the system state from the origin ness of the controllability and observability of the state
at time t ! 1 to a specified state xo at t ¼ 0 (we will of S co .
consider a normalized state, i.e. kxo k ¼ 1), i.e. solve the Another useful property of gramians is that they are
problem directly related to the system 2-norm. Specifically if the
ð0 system has a transfer function GðsÞ, then (Doyle et al.
min JðuÞ ¼ min uðtÞT uðtÞ dt 1992)
u2L2 ð1, 0Þ u2L2 ð1, 0Þ 1 ð18Þ
subject to xð0Þ ¼ xo with kxo k ¼ 1 kGk22 ¼ trace CPCT ¼ trace BT QB ð22Þ
The solution to this problem can be obtained using This property applies to continuous-time as well as
standard linear quadratic regulator theory (Goodwin to discrete-time systems.
et al. 2001). The optimal control, uo ðtÞ, is given by Gramian characteristics and properties will be
exploited in the coming sections to build and to apply
T
uo ðtÞ ¼ B eA t P1 xo ¼) Jðuo Þ ¼ xTo P1 xo ð19Þ an interaction measure.
Then the optimal cost is low for every normalized 3. Gramians and MIMO interaction
xo 2 Rn if and only if all eigenvalues of P are large
3.1. Elementary system
(since then, all eigenvalues of P1 are small). It is also
true that if one of the eigenvalues of P is small, then Consider the system state space description (A, B,
the optimal cost will reach its maximum for xo being an C, 0), where
eigenvector corresponding to the smallest eigenvalue
B ¼ b1 b2 bp ; C T ¼ c1 c2 cp
(recall that P and P1 share the same eigenvectors).
ð23Þ
2.2.2. Observability gramian. Assume that xð0Þ ¼ xo We can then associate with this MIMO system a
(with kxo k ¼ 1) and that uðtÞ ¼ 0 8t 0 then set of elementary (SISO) systems, each of them having
ð1 ð1
T a single input uj, j 2 f1, 2, . . . , pg, and a single output yi,
yðtÞT yðtÞ dt ¼ xTo eA t CT C eAt dt xo ¼ xTo Q xo i 2 f1, 2, . . . , pg, and a state space model given by
0 0
|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl} (A, bj , cTi , 0) with gramians Pj and Qi satisfying
Q
with input uj and output yi. Hence, the sets of HSV for 4. Interaction quantification
the elementary system and for the minimal system, char- The above analysis requires, to have practical
acterized by Gij ðsÞ differ only in a subset of zero HSV. interest, a way to quantify and to compare.
By definition, the products PQ and PjQi have non-
negative eigenvalues. However, the eigenvalues of a sum
3.2. Gramian decomposition
of products PjQi is not equal, in general, to the sum of
We next observe that a key property of the system the eigenvalues of each summand.
gramians is that they can be expressed as functions of the It turns out that the trace of the product PiQj is
gramians for the elementary systems. This is precisely a convenient basis to measure the interaction and the
stated in the following lemma. degrees of controllability and observability of different
Lemma 2 (Gramian decomposition): Let Pj and Qi controller structures. Recall that the trace of PjQi is the
be the controllability and observability gramians for the sum of the squared HSV for the elementary system with
elementary system (A, bj , cTi , 0) given by (23). input j and output i and also recall that the HSV associ-
Then, the original system controllability and observa- ated to the pair (Pj, Qi) quantify the combined abilities
bility gramians P and Q are given by of input uj and output yi to control and to observe the
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
coupling.
A potential weakness of the PM is its sensitivity
It is straightforward to prove that if the system
to input and output scaling. If, for instance, the entry
transfer function GðsÞ is diagonal, then Pj Qi ¼ 0 for
Gk1 k2 ðsÞ is multiplied by a gain K, then tracefPk2 Qk1 g is
all i 6¼ j.
multiplied by K2.
Obviously the maximum controllability and observa- One way to deal with this issue is to recall that (linear
bility is attained when a full MIMO controller architec- incremental) plant models connect (incremental) con-
ture is chosen, then all terms of the form Pj Qi have troller commands and (incremental) measured variables.
to be added to compute the system HSV. However if Both signals can be described in normalized ranges,
a restricted complexity controller is chosen, then only which arise from converting signal instrumentation
a subset of those terms is involved. ranges (such as 4 to 20 [mA], 0 to þ10 [V], etc.) to either
372 M. E. Salgado and A. Conley
Proof: We first note that dividing the frequency by This leads to tracefPQg ¼ 25=36.
is equivalent to perform a time scaling, by multiplying G2 ðsÞ
time by the same constant , i.e. substituting t by
where ¼ t. Thus, the state space representation of In this case we are dealing with a system of class CS 2 .
the scaled system is To verify that we perform a frequency scaling and we
compute the d.c. gain. Thus
dxðÞ A bj
¼ xðÞ þ uj ðÞ ð29Þ 2 1
d G2 ðsÞ ¼ 2
3 ðs=6Þ þ2ðs=6Þ þ 1
yi ðÞ ¼ ci xðÞ ð30Þ
We have thus that the gramians for the scaled system and then tracefPQg ¼ 1=6.
can be computed from G3 ðsÞ
1 ~ i A þ cTi ci ¼ 0
~iþQ This system is in class CS 3 , as can be verified using
AP~ j þ P~ j AT þ b bT ¼ 0 AT Q
j j frequency scaling and computing the d.c. gain
ð31Þ
1 6ðs=2Þ þ 1
The above equations imply that P~ j ¼ Pj = and G3 ðsÞ ¼
2 ðs=2Þ2 þ 52 ðs=2Þ þ 1
~ ~ i.
Qi ¼ Qi . Thus Pj Qi ¼ P~ j Q œ
Therefore tracefPQg ¼ 0:7025.
Lemma 3 says that what matters in the quantifica-
tion of interaction are certain dynamic features such A careful analysis of the results in table 1 provides
as the damping coefficient in resonant modes, relative valuable insight. Part of that insight is summarized in
location of poles and zeros, presence of non-minimum the following observations.
MIMO interaction measure and controller structure selection 373
tracefPQg tends to 0:25, as expected. the composed system S 1 can be described by the 4-tuple
ðA1 , B1 , C1 , 0Þ, where
(iv) Systems in class CS 3 have a zero at s ¼ 1=. If
jj , i.e. the zero is much closer to the A0 0 B0
A1 ¼ ; B1 ¼ ; C1 ¼ 0nn 1
imaginary axis than the poles, then tracefPQg C0 0 0
becomes very large. If < 0, we are in the pre- ð36Þ
sence of non-minimum phase systems and then Let us now express the gramians P1 and Q1 as
tracefPQg is even larger. This discussion is con-
P11 P12 Q11 Q12
sistent with known design criteria which say that P1 ¼ ; Q1 ¼ ð37Þ
small zeros, and especially, small positive zeros PT12 P22 QT12 Q22
are significant dynamic features (see, e.g. Good- then those gramians can be computed from (33) and
win et al. 2001). (34), leading to
The functions considered in table 1 are deceivingly P11 ¼ P0 P12 ¼ A0 P0 CT0 P22 ¼ C0 P0 CT0 ð38Þ
simple. However, in different mixtures, they are enough Q11 ¼ Q0 Q12 ¼ 0 Q22 ¼ 1 ð39Þ
to model a vast number of MIMO plants. Time delays,
absent in table 1, are a dynamic feature highly signifi- We then form the product P1 Q1 and compute its
cant in process control, and they will be included in the trace, this yields
analysis of the discrete-time case. tracefP1 Q1 g ¼ tracefP0 Q0 g þ C0 P0 CT0 ð40Þ
The applicability of previous results to discrete-time Then, the result (35) follows by induction and
systems is highly relevant for two reasons: firstly, modern on using (22). Note that a key in this proof is that
process control is based on sampled-data models to be Ckþ1 ¼ ½ 0ðnþkÞðnþkÞ 1 for k ¼ 0, 2, . . . , ‘ 1. œ
used in conjunction with digital technology and, sec- Assume that Gij is the transfer function of S ‘ , then
ondly, in this framework we can deal with systems having Lemma 4 says that the value of ij in the PM will
time delays (recall that in the continuous-time case, consistently increase as we increase ‘.
delays yield state space models of infinite dimensions), To illustrate Lemma 4 we consider a plant model
which are very common in industrial processes. Time which is a classic in process control.
delays are not only common but they may also play a
key role when deciding the controller structure. This Example 3: A continuous-time plant has a transfer
latter feature is formally stated in the following lemma. function given by
Lemma 4: Consider the scalar stable system S ‘ in es
GðsÞ ¼ K ; 2 Rþ
figure 1, composed by a scalar system S 0 and a pure sþ
374 M. E. Salgado and A. Conley
The corresponding sampled-data model, with zero- Furthermore, a large ij indicates that the control uj
order hold, sampling period and ¼ N, N 2 Z, is has strong authority upon output yi. Thus, the complex-
given by ity of a controller structure should be traded off against
1a the closeness to one of the sum of the chosen ij
Gd ðzÞ ¼ K ; where a ¼ e elements. For future reference we will denote this sum
zN ðz aÞ
by S.
To apply Lemma 4 we make ‘ ¼ N and Go ½z ¼ Note that, in singling out the most significant ij
zN Gd ðzÞ, with elements, the PM suggests an associated nominal plant
Ao ¼ a; Bo ¼ Kð1 aÞ; Co ¼ 1; Do ¼ 0 model. Then, it is plausible that the parameter S quan-
tifies the combined controllability and observability of
Then, using (33) and (34), we have that that associated nominal model.
1a 1 Consider the plant output y‘ . Then, the analysis of
Po ¼ K 2 ; Qo ¼ the ‘th row, ‘ , in the PM is required. If the plant has p
1þa 1 a2
inputs and p outputs, then the most significant elements
Thus
in ‘ are those which exceed the average value 1=p 2 .
K2 Those are the primary candidates to be considered to
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
tracefPN QN g ¼ tracefPo Qo g þ NkGo k22 Example 4: A plant with three inputs and three out-
puts has the PM given by
K2 1a
¼ 2
þ NK 2 2 u1 u2 u3 3
ð1 þ aÞ 1þa 0:1400 0:0645 0:0233 y1
(¼ 4
0:1633 0:1485 0:0661 5 y2
1 þ Nð1 e2 Þ 0:0739 0:0591 0:0863 y3
¼ K2
ð1 þ e Þ2
where we have highlighted all those elements which
exceed the average value which, in this case, is
1=9 ¼ 0:1111.
6. Application to controller architecture selection If we first evaluate a (square) decentralized structure,
The key idea presented in } 4 is that the PM high- we have to choose a pattern with three rows and three
lights those elements in the transfer function matrix, columns where only one element per row and one
which are more significant in the description of the element per column are allowed. We then observe that
MIMO system. At the same time, the PM suggests a the pairing (u1 , y2 ), corresponding to the largest element
path to gradually increase the complexity of the nominal in the row 2 , should not be chosen in the selected
model Go(s) and, as a consequence, to gradually increase controller structure, since then the maximum attainable
the controller structure complexity. S would be
Thus, to determine the structure of the controller Sa ¼ 12 þ 21 þ 33 ¼ 0:3141
we must first obtain a nominal model Go ðsÞ. In a non-
discriminant (very unlikely) scenario, all 0 < ij < 1 However, if we discard the pairing (u1 , y2 ) and,
have the same value, i.e. they are all equal to 1=p 2 . instead, we choose the pairing (u1 , y1 ), a higher S can
However, in a highly discriminant PM, i.e. when there be achieved with
exists a pair (i, j) for which ij 1=p 2 , one can say that
Sb ¼ 11 þ 22 þ 33 ¼ 0:3748
all states in the elementary system with output yi and
input uj are either hard to control or hard to observe. Hence the best pairing for a decentralized controller
This suggests that there are no significant benefits is (u1 , y1 ), (u2 , y2 ) and (u3 , y3 ).
to consider the transfer function Gij ðsÞ in the nominal If we are prepared to use a controller which is more
model. The converse is also true, i.e. when ij is much complex than a decentralized controller, we could make
larger that 1=p 2 , then some states in the elementary y2 to be commanded not only from u1, but also from u2,
system with output yi and input uj are easy to control this sparse controller would yield
and easy to observe. Hence the transfer function Gij ðsÞ is
a strong candidate to be included in the nominal model. Sc ¼ 11 þ 22 þ 33 þ 21 ¼ 0:5381
MIMO interaction measure and controller structure selection 375
Once the PM has been computed and Go ðsÞ has been an indication of a dynamics dominance in the system.
chosen, we then know which plant inputs are affected More will be said when closed loop design specifications
by which plant outputs. This determines the web of are discussed in } 7.
dependencies of y on u. We then need to determine the Some additional examples are presented next for
controller structure, i.e. the web of dependencies of u further illustration of the proposed ideas.
on y (feedback controller structure). This can be done
Example 5 (Continuous-time system): Consider a 3 3
by inspection from the PM if Go(s) is either a diagonal
system with transfer function
or block diagonal system with blocks having symmetric
2 3
structure, but it is not usually so for those models which 0:4 4ðs þ 3Þ 2
include sparse or triangular structures. In the general 6 ðs þ 1Þ2 ðs þ 2Þðs þ 5Þ sþ4 7
6 7
case, one has to mirror the dependencies in such a way 6 2 2 1 7
6 7
that the dependency of u from y (feedback controller GðsÞ ¼ 6 2 7
6 ðs þ 2Þðs þ 1Þ ðs þ 2Þ sþ2 7
structure) be the inverse of that in the nominal model 6 7
4 6ðs þ 1Þ 4 8 5
structure (dependency of y on u). The rationale of that
ðs þ 5Þðs þ 4Þ 2 ðs þ 2Þðs þ 5Þ
mechanism can be appreciated from the discussion ðs þ 3Þ
below.
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
ð43Þ
Given Go(s), we require that C(s) has a structure
The PM for this system is
which, at least, can achieve a diagonal complementary
sensitivity, so that we are able to decouple the channels. u1 u2 u3
We recall that 2 3
0:0370 0:2018 0:0385 y1
(¼ 6 7 ð44Þ
To ðsÞ ¼ Go ðsÞCðsÞðI Go ðsÞCðsÞÞ1 4 0:2226 0:0578 0:0385 5 y2
¼ ðCðsÞ1 Go ðsÞ1 IÞ1 0:2193 0:0457 0:1389 y3
Hence, to be able to have a diagonal To ðsÞ it is Note that this system has non-minimum phase zeros
necessary that the structure of C(s) be the inverse of at 0:1134
j1:8216.
the structure of Go(s). We observe that Case 1. Decentralized controller: In this case we need
1 to select three elementary systems, to pair every input
(i) C(s) is not required to be equal to Go ðsÞ .
(ii) C(s) will usually be unstable to ensure integral to a different output. The PM matrix suggests that the
action and, in general, it should exhibit high best pairing (the reader can verify that any other pairing
gains in the frequency bands where reference yields a lower S) would be (u1 , y2 ), (u2 , y1 ) and (u3 , y3 ),
and disturbances have significant energy. this yields
(iii) The choice structureðCðsÞÞ ¼ structureðGo ðsÞ1 Þ Sd ¼ 12 þ 21 þ 33
ð45Þ
gives the simplest controller structure to achieve ¼ 0:2018 þ 0:2226 þ 0:1389 ¼ 0:5633
a diagonal To(s). However with that controller
structure, other than diagonal structures for The nominal plant model would be
To(s) can also be achieved. 2 3
4ðs þ 3Þ
6 0 0 7
For large scale systems, where the number of inputs 6 ðs þ 2Þðs þ 5Þ 7
and outputs precludes a selection by inspection, it is 6 2 7
6
Go ðsÞ ¼ 6 0 0 7
7
possible to solve the controller structure problem using 6 ðs þ 2Þðs þ 1Þ 7
a computer program, since the selection problem can be 4 8 5
0 0
set in a dynamic programming framework. In those high ðs þ 2Þðs þ 5Þ
complexity systems, the heuristic limits suggested above ð46Þ
tend to go down.
In general, there is no simple theory yet to say what Then
is the minimum S which makes a given nominal model 2 3
0 0
acceptable. However, in the authors’ experience, the 6 7
structureðGo ðsÞÞ ¼ 4 0 0 5 ¼) structureðCðsÞÞ
nominal model should yield S > 0:70 to be on the safe
side. When S < 0:5, it is highly likely that no controller 0 0
0 1
with the selected structure yields satisfactory perform- 0 0
ance. If a high S is achieved mainly by the contribution B C
¼ structureðGo ðsÞ1 Þ ¼ @ 0 0A
of a reduced (in comparison with the total number of
0 0
entries, p p) number of elements in the PM, then
useful insight is being obtained, since that situation is ð47Þ
376 M. E. Salgado and A. Conley
Therefore, in this case, the controller will have the to increase S, can be formed by making u3 depending
structure not only on y3 but also on y2, this yields
2 3
0 C12 ðsÞ 0 Sb ¼ 12 þ 21 þ 23 þ 31 þ 33
6 7
CðsÞ ¼ 64 C21 ðsÞ 0 0 7 5 ð48Þ ¼ 0:2018 þ 0:2226 þ 0:0385 þ 0:21930 þ 0:1389
0 0 C33 ðsÞ ¼ 0:8211
One can observe that the selected structure yields ð53Þ
a value for S which is just over 50% of the overall For this choice, the plant nominal model is
combined measure of controllability and observability. 2 3
In the next case, to increase S, we add an additional 4ðs þ 3Þ
6 0 0 7
term in the nominal plant model and in the controller. 6 ðs þ 2Þðs þ 5Þ 7
6 2 1 7
Case 2. Sparse controller: Assuming the initial pairing 6 7
Go ðsÞ ¼ 6 0 7
chosen in Case 1, the next largest contribution to S 6 ðs þ 2Þðs þ 1Þ sþ2 7
6 7
would be obtained on making u1 driving not only y2 4 6ðs þ 1Þ 8 5
0
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
the small d.c. gain in G31 ðsÞ, equal to 0:3 when compared no indication in the RGA on how to select a more
to the d.c. gain of the main competitor, G21 ðsÞ, which has complex model (sparse or block diagonal).
a unit d.c. gain. To go deeper into this discussion we
To complement the illustration of the interaction
consider two cases.
quantifier and the controller structure selection we will
Change 1: The d.c. gain of G31 ðsÞ is changed from 0.3 next consider two examples in discrete-time.
to 0.45.
Example 6 (Sampled-data control): We assume that
Change 2: The zero is shifted from 1 to 0.5, keeping the the plant in Example 5 is going to be under digital con-
d.c. gain at its original value, i.e. equal to 0.3 trol, using a sampling period of 0.1 s and a zero-order
hold. We then need to choose a controller structure
Change 1
based upon a sampled data model. That model has a
In this case transfer function given by
2 3
9ðs þ 1Þ 0:001872z þ 0:001751 0:3307z 0:2451 0:1648
G31 ðsÞ ¼ 2 6 z2 1:81z þ 0:8187
6 z2 1:425z þ 0:4966 z 0:6703 7
7
s þ 9s þ 20 6 0:009056z þ 0:008194
6 0:008762z þ 0:007668 0:09063 7
7
Gd ðzÞ ¼ 6 2 7
u u2 u3 6 z 1:724z þ 0:7408 z2 1:637z þ 0:6703 z 0:8187 7
2 1
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
3 6 7
4 0:3604z þ 0:3993 0:01642z þ 0:01344 0:03184z þ 0:02522 5
0:0290 0:1584 0:0302 y1
¼) ( ¼ 6 z2 1:277z þ 0:4066 z2 1:482z þ 0:5488 2
z 1:425z þ 0:4966
7
4 0:1747 0:0454 0:0302 5 y2 ð58Þ
0:3873 0:0358 0:1090 y3 The PM for this system is
The change in the d.c. gain of G31 ðsÞ yields a signifi- 2 u1 u2 u3 3
cant increase in the value of 13 , making it mandatory 0:0319 0:2289 0:0461 y1
(¼ 4
to have u1 depending on the measurement of y3 in any 0:1943 0:0514 0:0389 5 y2
controller structure. In particular, for a decentralized 0:2394 0:0418 0:1273 y3
controller, the pairing ðu1 , y3 Þ makes it possible to If a decentralized controller is to be used, the pairing
reach the maximum value for S, which is 0:5759. Note yielding the highest S is the same as in the continuous
that the d.c. gain of G31 ðsÞ is still less than half the d.c. case, i.e. ðu1 , y2 Þ, (u2, y1) and (u3, y3). A highly relevant
gain of G21 ðsÞ. feature in this example is that it illustrates the fact that
Change 2 the largest ij must not necessarily be included to form
the controller structure. In this example the largest ij
In this case is 31 ; however, the pairing (u1, y3) would yield at most
6ð2s þ 1Þ S ¼ 0:5072.
G31 ðsÞ ¼
s2 þ 9s þ 20 The next example highlights the ability of PM to
u u2 u3 assess the significance of time delays in MIMO inter-
2 1 3
0:0251 0:1370 0:0262 y1 action and therefore in the controller structure selection.
¼) ( ¼ 6 7 This example also verifies the validity of Lemma 4.
4 0:1512 0:0392 0:0262 5 y2
0:4698 0:0310 0:0943 y3 Example 7 (Sampled-data system with time delays):
The sampled-data transfer function of a 2 2 plant
We have moved the non-minimum phase zero has a transfer function given by
closer to the origin, while maintaining the d.c. gain at 2 3
0:5 0:15
its original value. Now the predominance of 13 is even 6
GðzÞ ¼ 6
ðz 0:5Þ ðz 0:8Þzl 7
7 ð59Þ
stronger than when the first change was introduced, 4 0:1 0:3 5
forcing the dependence of u1 on y3 in any controller
ðz 0:5Þðz 0:8Þ z 0:7
structure. In particular, for a decentralized controller,
the pairing ðu1 , y3 Þ makes it possible to reach the maxi- where ‘ is a non-negative integer, quantifying a pure
mum value for S, which is 0:633. time delay in the path from u2 to y1. The participation
A final point to note is that the RGA for this plant matrix is next computed for ‘ ¼ 0, for ‘ ¼ 3, and for
is given by ‘ ¼ 10. The results are
2 3
0:3171 0:1239 0:2797 0:2272
0:0831 0:9111 0:1720 (j‘¼0 ¼ ; (j‘¼3 ¼ ;
RGAðGÞ ¼ 4 1:3809 0:2745 0:1064 5 ð57Þ 0:3122 0:2469 0:2754 0:2177
0:2979 0:3634 0:9345 0:2193 0:3941
(j‘¼10 ¼
It can be seen that the RGA suggests the same 0:2159 0:1707
diagonal structure as the PM does. However, there is ð60Þ
378 M. E. Salgado and A. Conley
When ‘ ¼ 0, the PM indicates that a decentralized from decentralized to full MIMO control, through
controller should be based upon the pairings (u1, y1) sparse control, triangular and block diagonal. In a prac-
and (u2, y2), leading to S ¼ 0:3171 þ 0:2469 ¼ 0:5571. tical situation, when a MIMO plant with many inputs
However, when ‘ ¼ 3, the information provided by the and many outputs is to be controlled, one would aim,
PM is ambiguous. But, when ‘ ¼ 10, the PM indicates an at least, to break the control design problem in several
anti-diagonal controller, i.e. the pairings (u1, y2) and simpler MIMO control problems, i.e. we are aiming to
ðu2 , y1 Þ, yielding S ¼ 0:3941 þ 0:2159 ¼ 0:61. The quali- describe our process using a block diagonal nominal
tative nature of these results is in agreement with intui- model. The advantage of doing this is manyfold: design
tion, since as the delay increases, the authority of u2 on simplicity, robust performance, integrity, online tuning,
y1 grows in importance in the overall system dynamics. maintenance, and so on. The PM is one tool to support
Note that the four static gains in GðsÞ are comparable. this approach.
The values for S in all cases are similar. To increase Except when dealing with a decentralized nominal
its value, a triangular structure in the controller should model, which breaks the problem into p-SISO designs,
be used. In each case, the structure of the controller is in all other cases we need to use the available tools
determined by building the structure of Go ðsÞ1 . provided by the MIMO control theory and practice.
When ‘ ¼ 0, a sensible triangular controller is The proposal of a tool like the PM does not pose
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
obtained if we make u1 depending on y1 and if, at the problems that are different to those which are
same time, u2 depends on y1 and on y2, this yields a pair currently research topics in the field.
(u1 , y1 ) and a triad (u2 , y1 , y2 ) leading to S ¼ 0:3171þ Our preferred choice to synthesize a controller can
0:2469 þ 0:3122 ¼ 0:8693. When ‘ ¼ 3, a triangular be explained considering the fact that its structure is the
controller is built with the triad (u2 , y1 , y2 ) and the pair inverse of the structure of the nominal model. We can
(u1 , y2 ); in this case S ¼ 0:2797 þ 0:2754 þ 0:2272 ¼ use the Youla parametrization of all stabilizing control-
0:7823. The same controller structure should be chosen lers for a stable plant (see, e.g. Maciejowski 1989, Glad
when ‘ ¼ 10, leading to S ¼ 0:3941 þ 0:2159 þ 0:2193 ¼ and Ljung 2000, Goodwin et al. 2001). In this method-
0:8293 ology, the controller satisfies
At this stage, the fact that the model Go(s) was We next illustrate these ideas with Cases 1 and 2 in
obtained using either PM or other tool is irrelevant, Example 5, where a significant difference exists in the
since the issues to face now are classical MIMO design values for S. The nominal models, diagonal and sparse,
issues, some of which are discussed below. given by (46) and (50) are denoted by God ðsÞ and Gos ðsÞ,
respectively, and their inverses are
7.1. Stability God ðsÞ1
2 3
The controller Q(s) is designed using the nominal 0 0:5ðs þ 1Þðs þ 2Þ 0
model Go(s), thus stability of Q(s) is necessary and suffi- 6 0:25ðs þ 2Þðs þ 5Þ 7
¼6
4 0 0 7
5
cient for the stability of the nominal control loop. Since sþ3
Go(s) is open loop stable, there is always a controller of 0 0 0:125ðs þ 2Þðs þ 5Þ
the given structure, that stabilizes the loop. This can be ð63Þ
proved either using root locus arguments or fixed mode
ideas (see, e.g. Anderson and Clements 1981, Tarokh Gos ðsÞ1
2 3
1985). However, since this controller must be able to 0 0:5ðs þ 1Þðs þ 2Þ 0
control the real plant, we must guarantee that, at least, 6 0:25ðs þ 2Þðs þ 5Þ 7
6 0 0 7
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
(iii) A quadratic measure of the performance is The cost function (66) has been also computed for
computed on line, as both cases and it is shown in figure 5.
A brief analysis of the above results shows that:
ðt
J¼ eðtÞT eðtÞ dt ð66Þ (i) Both designs are robust stable. However, it can
0 be verified that this does not hold if 1.
(ii) The choice of ¼ 0:1 has been made only for the
where eðtÞ is the vector error rðtÞ yðtÞ.
sake of illustration of the control methodology.
The performance of both loops are shown in In a more realistic environment, there are other
figures 3 and 4. requirements (such as noise, input saturation and
3
Plant outputs
2
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
1 y1(t)
0
y (t)
3
−1
y2(t)
−2
0 5 10 15 20 25 30
Time [s]
Figure 3. Plant outputs. Diagonal controller. ¼ 0:1.
2
Plant outputs
1
y1(t)
0
y (t)
3
−1
y2(t)
−2
0 5 10 15 20 25 30
Time [s]
Figure 4. Plant outputs. Sparse controller. ¼ 0:1.
1
Jdiag(t)
Control error measure
0.8
0.6
0.4 Jsparse(t)
0.2
0
0 5 10 15 20 25 30
Time [s]
Figure 5. Accumulated control errors. ¼ 0:1.
MIMO interaction measure and controller structure selection 381
transient performance) likely to suggest a much where Suo ðsÞ is the control sensitivity (Goodwin et al.
larger . This issue is also connected to the 2001). If we assume no constraints on the reference,
discussion below. we thus observe that the plant input spectrum is shaped
(iii) The sparse controller provides a better perfor- by a filter, where the filtering characteristics are deter-
mance than the diagonal controller. However mined by the control sensitivity.
those results change with the reference direction- A simple yet effective way to introduce the band-
ality. For instance, if we use r1 ðtÞ ¼ r2 ðtÞ ¼ width limitation is to choose a diagonal transfer func-
r3 ðtÞ ¼ ðt 1Þ, i.e. the direction is ½1 1 1T and tion FðsÞ which captures essential features of Suo ðsÞ.
the three references change simultaneously, then Every diagonal entry in FðsÞ is associated to the intended
the performance of the sparse controller would closed loop bandwidth. In this strategy the (i, i) element
be even better than that of the diagonal case. in FðsÞ is then defined by
(iv) Several other references might be tried. However Ao ðsÞ
½Fii ðsÞ ¼ ð67Þ
in every trial, the references in all channels should Ai ðsÞ
have the same magnitude; otherwise we would
be allocating arbitrary weights to the different where Ao(s) and Ai(s) are stable polynomials, with Ao(s)
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
We next compute the PM for the filtered function is not a wide-band signal. To overcome this latter prob-
which gives lem, one could also model the reference as the output of
an additional filter with white noise input; however, that
2 u1 u2 u3 3
is beyond the scope of this paper.
0:0094 0:2279 0:0167 y1
(¼ 4
0:0645 0:0236 0:0112 5 y2
0:5740 0:0300 0:0427 y3
8. Conclusions
If a decentralized controller structure is chosen, then In this paper a new measure of dynamic channel inter-
the pairing should be (u1, y3), ðu2 , y1 Þ and ðu1 , y3 Þ leading action in stable MIMO systems has been proposed. This
to S ¼ 13 þ 21 þ 32 ¼ 0:8091. An interesting feature measure is based upon the system controllability and
of this result is that the entry 13 , corresponding to observability gramians. It thus makes use of the ability
G31 ðsÞ, is now the predominant element. This is due not of the gramians to describe the difficulty (or otherwise)
to its d.c. gain, which is small, but to its non-minimum to observe and to control the system state. In particular,
phase zero. The strong presence of this feature is due to it has been shown that the sum of the squared Hankel
the choice of the dominant closed loop pole, which is singular values reveals several relevant control design
Downloaded by [Texas State University - San Marcos] at 12:45 08 May 2013
sþ1 sþ2 sþ2 outputs. Some criteria have been associated to the pro-
FðsÞ ¼ diag 8 ,4 ,4 posed measure to pair inputs and outputs in a decentral-
sþ8 sþ8 sþ8
ized control architecture. However, the key feature of
leading to the PM is its unique ability to provide support for a
richer controller architecture selection in continuous
2 u1 u2 u3 3
and discrete-time time frameworks, including triangular,
0:0046 0:2210 0:0139 y1
(¼ 4 block diagonal and sparse structures. Furthermore,
0:0315 0:0124 0:0060 5 y2
0:6695 0:0187 0:0225 y3 preliminary results show that this measure helps the
designer to comparatively assess the benefits of those
By making the closed loop even faster, an even more richer controller structures. The PM also sheds light
predominant role is played by the non-minimum phase on fundamental design limitations and highlights
zero in G31 ðsÞ. dominant dynamic features (small stable zeros, non-
minimum phase zeros, time delays and resonant
The use of this type of filtering can be made part of
peaks). It has been also shown that the PM, when
the iteration process which normally takes place in any
applied to a suitable modified plant model, can provide
control design problem. One has to keep in mind that
information regarding interaction as a function of a
the effect of the filter is meaningful only if the frequency
projected closed loop velocity. In this usage, the PM
response of GðsÞ is significantly different to that of
also helps to spot design conflicts. Those are the main
GðsÞFðsÞ. Also, the benefit of this filtering strategy does
contributions of the ideas proposed in this paper.
not lie in numerical accuracy but in that it reveals an
Further research should include extension of the PM
interplay of the different interactions as the projected
ideas to unstable processes (which would require a new
closed loop bandwidth changes. As seen in Example 8,
definition of gramians (e.g. as in Zhou et al. 1999) and
when the closed loop speed is increased, some entries
the building of more precise numerical guidelines to
in the PM become more and more significant. Those
select the controller structure. Optimality with restricted
particular elements are associated to dynamic features
structure controllers and input-saturation effects are
which are usually hard to deal with in any control design
also challenges for future research.
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 specified closed loop bandwidth is
unsuitable for the given plant and that, consequently, Acknowledgements
a design trade-off has to be achieved by choosing a Helpful discussions with Professor Pedro Albertos
different bandwidth. from Valencia Polytechnic University, Spain, are
The limitations of this filtering strategy as an accu- gratefully acknowledged. The authors are also grateful
rate numerical tool are originated in the simplicity of for the support received from CONICYT-Chile
the filter (67) and in the fact that usually the reference through grant FONDECYT-1040313.
MIMO interaction measure and controller structure selection 383
Gagnon, E., Desbiens, A., and Pomerleau, A., 1999, interaction. ISA Transactions, 16, 35–41.
Selection of pairing and constrained robust decentralized Wittenmark, B., and Salgado, M., 2002, Hankel norm
PI controllers. In Proceedings of the American Control based interaction measure for input-output pairing. In 15th
Conference, pp. 4343–4347. IFAC World Congress Conference Proceedings, Barcelona,
Glad, T., and Ljung, L., 2000, Control Theory. Multivariable
Spain.
and Nonlinear Methods (London: Taylor and Francis).
Wolovich, W., and Falb, P., 1976, Invariants and canonical
Glover, K., 1984, All optimal Hankel norm approximations
of linear multivariable systems and their l1-error bounds. forms under dynamic compensation. SIAM Journal of
International Journal of Control, 39, 1115–1193. Control and Optimization, 14, 996–1008.
Goodwin, G. C., and Sin, K., 1984, Adaptive Filtering Zhou, K., and Doyle, J., 1998, Essentials of Robust Control
Prediction and Control (Englewood Cliffs, NJ: Prentice- (Englewood Cliffs, NY: Prentice Hall).
Hall). Zhou, Z., Salomon, G., and Wu, E., 1999, Balanced realiza-
Goodwin, G., Graebe, S., and Salgado, M. E., 2001, tion and nodel reduction for unstable systems. International
Control System Design (Upper Saddle River, NJ: Prentice- Journal of Robust and Nonlinear Control, 9, 183–198.
Hall). Zhu, Z.-X., 1996, Variable pairing selection based on indivi-
Havre, K., 1998, Studies on controllability analysis and dual and overall interaction measures. Industrial Engineering
control structure design. PhD thesis, Department of in Chemical Research, 35, 4091–4099.