Automatica 34 (2000) 867}877

Brief Paper
MIMO fuzzy internal model control夽
C. R. Edgar*, B. E. Postlethwaite
Process Cybernetics Group, Department of Chemical and Process Engineering, University of Strathclyde, Glasgow, Scotland, UK
Received 9 September 1996; revised 13 May 1998; received in "nal form 24 September 1999

Abstract

Model-based controllers are now beginning to gain widespread acceptance in industry. However, the majority of these controllers
are based on linear models and performance in controlling the non-linear processes common in the chemical industry is sub-optimal.
The use of a non-linear model could yield signi"cant improvements in control performance. In this study a relational model from
a fuzzy input space to a crisp output space is constructed by applying a least-squares identi"cation technique to past process data.
This model is termed a crisp-consequent fuzzy relational model (ccFRM) and is capable of giving an accurate representation of
a non-linear system. A novel inversion method is presented which allows the ccFRM to be inverted and used within the well-known
IMC structure. This new controller is termed a fuzzy internal model controller (FIMC) and test results are presented showing the
FIMC performing both servo and regulatory action on a multi-variable simulated pH system. This process is extremely non-linear
and exhibits severe interaction e!ects and is consequently a very di$cult system to control. The simulation is introduced in detail, as
are the tests carried out, and the performance of the FIMC in these tests is found to be encouraging. 2000 Elsevier Science Ltd. All
rights reserved.

Keywords: Fuzzy control; Fuzzy modelling; Model-based control; Multivariable control; pH control

1. Introduction Although linear IMC control schemes work well with

many linear processes, the performance achievable with
Model-based controllers have enjoyed rising popular- non-linear processes is not so good. Unfortunately, the
ity and recognition over the past few years. The ability of use of a linear model to describe a non-linear process
such controllers to handle process deadtime and MIMO introduces considerable inaccuracies and the control
systems combined with well-established structured de- achieved is sub-optimal. A non-linear model seems to
sign techniques has meant that they are being increasing- o!er the possibility of improved control performance and
ly accepted into industry. this has provided a motivation for research into non-
One type of model-based control scheme is internal linear modelling techniques.
model control (IMC) (Garcia & Morari, 1982), the gen- A fuzzy relational model (FRM) is one manner in
eral structure of which is shown in Fig. 1. Ideally, the which a non-linear system can be represented. A FRM
process model, PI , will be exactly the same as the actual utilises the tenets of fuzzy logic pioneered by Zadeh
process, P. However, even if this is not the case, controller (1977). In fuzzy set theory a value may belong to some
robustness can be obtained from careful tuning of the partial degree to a fuzzy set, whereas in traditional
"lter, Fe. The IMC structure allows model errors to be Boolean set theory only complete belonging or complete
handled explicitly and robustness to be designed in. exclusion are possible. This #exibility enables crisp nu-
These assets combined with a clear design methodology merical data to be represented in fuzzy, or vague, terms.
have made it an attractive control structure. For example, a human operator would tend to think in
terms of &low' #ow or &high' temperature; and fuzzy logic
allows a computer to cope with such vague linguistic
夽
This paper was not presented at any IFAC meeting. This paper was labels. A FRM can be de"ned as a mapping between
recommended for publication in revised form by Associate Editor A. a fuzzy input space and a fuzzy output space. In con-
Bagchi under the direction of Editor T. Basar.
* Corresponding author. Tel.: 0141-552-4400 ext. 2372; fax: 0141-552- structing a FRM one would "rst partition the ranges of
2302. all inputs and outputs using a collection of fuzzy sets.
E-mail address: [email protected] (C. R. Edgar) The FRM is an array containing the strength of the

0005-1098/00/$ - see front matter 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 9 9 ) 0 0 2 1 3 - 7
868 C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877

Fig. 1. IMC controller structure.

relationships between all the di!erent combinations of (Sing & Postlethwaite, 1996). The least-squares identi"ed
output and input fuzzy sets. These relationship strengths FRM may be described algebraically as
can either be obtained from expert knowledge or numer-
k(y )"R k(x ) 2 k(x ), (1)
ically from past process data. With expert knowledge L
often unavailable or di$cult and time consuming to where y is the predicted output, R the fuzzy relational
obtain there has been a great deal of interest in the model, x the model input, n the number of model inputs,
automatic identi"cation of FRMs from past process data k(x ) the vector of degrees of membership to fuzzy sets for
and in recent years a considerable amount of research input x , k(y ) the vector of degrees of membership to
has been conducted in this "eld. fuzzy sets for predicted output y , and o the fuzzy com-
There have been many di!erent fuzzy relational identi- positional operator.
"cation algorithms described in the literature, both The fuzzy compositional operator used in the least-
numerical optimization methods and &once-through' squares FRM is the summated product. This operator uses
identi"cation techniques. A number of numerical tech- the algebraic product of input degrees of membership
niques were investigated by Sing and Postlethwaite and their corresponding entry in R. The outputs from
(1996) and were found to produce very accurate models, di!erent fuzzy relationships are summated.
albeit slowly. &Once-through' methods, such as that sug- Consider
gested by Ridley, Shaw and Kruger (1988) are much
faster but also considerably less accurate. K"kron(k(x ),2, k(x )), (2)
L
This paper details a novel controller formulation where kron is the Kronecker-tensor product. Now if S is
which utilises a non-linear fuzzy model within the IMC the number of reference sets used to de"ne the variable
controller structure. This new scheme is termed a range of y and N is the number of elements in K, then
fuzzy internal model controller (FIMC) and retains the )
R will be a S;N matrix and
features and advantages of the linear IMC scheme but is )
able to produce much improved control on non-linear ,)
k (y )" R ;K for j"1,2, S. (3)
processes due to the non-linear modelling capabilites of H HG G
the fuzzy model. The remainder of the paper describes G
a least-squares identi"cation technique for FRMs before The crisp predicted output, y , can be found by using
introducing a reduced model format, termed a crisp-con- the fuzzy mean method of defuzzi"cation
sequent FRM (ccFRM). An analytical technique to in- 1 k (y );y(
vert models of this type is then presented and the FIMC y " H H H, (4)
1 k (y )
structure is introduced. The FIMC has been tested using H H
a multi-variable simulated pH process and the system where y( is the characteristic value of fuzzy set j.
H
and test results are detailed in this paper. Finally, the Now, one of the restrictions that is placed on the model
relevance of the test results in discussed and the en- structure for a least-squares identi"ed FRM is the use of
couraging performance of the FIMC scheme is com- fully overlapping triangular reference sets to de"ne the
mented upon. variable ranges, therefore in this case y( is the centre
H
point of set j. The use of fully overlapping triangular
reference sets means that (Lowe & Webb, 1991)
2. The least-squares identi5ed fuzzy relational model
1
k (y )"1. (5)
H
Recent work (Postlethwaite, Brown & Sing, 1997) has H
seen the development of a linear least-squares technique Therefore, from (4)
to build FRMs. This technique produces accurate models
and is capable of constructing these models much faster 1
y " k (y );y( (6)
than alternative direct search optimisation techniques H H
H
C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877 869

and from (3) have a major problem associated with them, computer
processing time. The direct search mechanism employed
1 ,) in FRMBC (Postlethwaite et al., 1997) works well for
y " R ;K ;y( . (7)
HG G H SISO systems but is processor intensive and speed prob-
H G
lems are likely upon extension to MIMO systems. There
Now for any given model, provided set de"nitions are
was a need to develop a control technique which could
not adjusted on-line, R and y( are both constants
HG H cope with MIMO systems and which was not computer
de"ned during the model identi"cation phase, i.e.
processor intensive.
R ;y( "C . (8) Using a model inverse to "nd suitable controller action
HG H HG has been a reasonably popular method in linear control.
Therefore (7) becomes Unfortunately, the inversion of non-linear models is
very di$cult and little work has been done in this area.
1 ,) Attempts have been made to utilize an arti"cial neural
y " C ;K . (9)
HG G network (ANN) as a non-linear model within the IMC
H G
structure (Hunt & Sbabaro, 1991). This controller formu-
Now expanding this expression yields
lation "rst trains an ANN as a non-linear input}output
,) model representing the process. Another ANN is then
y " (C #C #2#C );K . (10) trained as an inverse mapping of the process model and
G G 1G G
G this second ANN is used as the controller block in the
If IMC structure. This scheme works reasonably well but
the controller block is not a true model inverse and thus
R "(C #C #2#C ). (11) the zero-o!set property of the IMC scheme is sacri"ced.
"G G G 1G
The inversion of a FRM poses di$culties as generally
Then
a family of solutions will result from any inversion at-
,) tempt (Sanchez, 1970). Although a general model inverse
y " R ;K . (12)
"G G is still not obtainable, the special structure of the ccFRM
G enables the inversion of the ccFRM at each sample time
The above discussion has concentrated on a model to "nd a required controller manipulation.
for a single output. However, the extension to MIMO
systems is trivial as multi-variable systems are simply
represented by a collection of fuzzy sub-models. 3. Inversion of a ccFRM
It can be seen from the above that exactly the same
predicted output y is yielded with both the full least- The inputs to a ccFRM for any process will be a com-
squares type model, R, and what Postlethwaite and bination of manipulated inputs, x , and other measured
co-workers (Postlethwaite et al., 1997) term the direct- +
variables, x . These measured variables, x , will include
least-squares model, R . However, the direct-least- 4 4
" modelled disturbances and past values of controlled vari-
squares model is a factor of S smaller and yields a crisp ables. Here we will consider a ccFRM with m manipu-
answer immediately without the need for a defuzzi"ca- lated variables and v other input variables. The inversion
tion step. Of course, strictly speaking the direct-least- scheme developed in this work is only capable of hand-
squares model can no longer be referred to as a FRM ling systems with the same number of manipulated and
since it no longer represents a mapping from a fuzzy controlled variables. It is not believed that this is too
input space to a fuzzy output space. However, since it is severe a restriction as most practical systems belong to
exactly equivalent to the least-squares model, which is this category. Therefore, in the case detailed above
a true FRM, it is felt that the direct-least-squares model m fuzzy sub-models will be required to represent the
should retain the nomenclature. It is proposed therefore process. The set of ccFRMs can be represented as
to term the direct-least-squares type of model a crisp-
consequent fuzzy relational model (ccFRM). An interest- y "R k(x ) 2 k(x ) k(x ) 2 k(x ),
" + +K 4 4T
ing feature of this model formulation is that it is also
equivalent to a zeroth-order Takagi}Sugeno model or $ (13)
a neural network with "rst-order B-spline basis y "R k(x ) 2 k(x )k(x ) 2 k(x ).
functions. K "K + +K 4 4T
The development of the smaller ccFRM model com- It should be noted that the inputs for all the models are
bined with an accurate and quick identi"cation method shown to be the same. This is done for simplicity of
was an important step forward in fuzzy relational notation but there is no requirement for this to be the
model-based control (FRMBC). However, the MPC-type case. Now, as before if the fuzzy compositional operator
controller techniques used in FRMBC designs to date used is the summated product then from (12) it can be seen
870 C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877

that (13) can be rewritten as

,)
y " R ;K ,
" G
G
$ (14)
,)
y " R ;K ,
K "K G
G
where in this case

K "kron(k(x ),2, k(x ), k(x ),2, k(x )). (15) Fig. 2. A variable range de"ned by fully overlapping triangular refer-
G + +K 4 4T ence sets.
Now, at any particular sample time desired values for the
controlled variables, i.e. the setpoints, will be known. The
ccFRM sub-models, R , are also known as will be An illustrative example of the inversion procedure
"K
the fuzzi"ed measured variables, k (x ). If the pre- applied to the multi-variable pH process is given in
4T
dicted output terms, y , are equated with the setpoints Appendix A.
K
then by inspection of (14) and (15) it can be seen that the Two important issues in obtaining the manipulation
only the fuzzi"ed manipulated variables, k (x ), are using the above technique are the existence of a model
+K
unknown. If S is the number of fuzzy reference sets used inverse and the possibility of multiple inverses. Clearly, if
to de"ne the variable range of a manipulated variable a model inverse does not exist then the above algorithm
then the set of m equations, (14), will have (m; K S ) will fail to obtain a suitable manipulation. Recalling
H H
unknown terms. It is clear therefore that this set of that during the inversion procedure y is equated to
equations alone does not provide enough information to G
y ∀i3[1, m] then it can be seen that if the required
"nd k(x ), more information is required and can be 1.G
+K setpoint is not achievable within the next sample time,
obtained by looking at the structural properties of the and the model correctly represents this fact, then no
ccFRM. model inverse will exist. In the FIMC controller this
It was mentioned earlier that a ccFRM uses fully problem is dealt with by reducing the required change in
overlapping triangular reference sets to describe the vari- setpoint until the required setpoint is achievable by the
able ranges of the model inputs. This is illustrated in next sample time. This ensures that the model can always
Fig. 2, from which it can be seen that any crisp variable be inverted and that a manipulation can always be ob-
value will belong to a maximum of two fuzzy sets and tained. The other main issue related to the calculation of
that these sets must be adjoining. Furthermore, the use of the model inverse is the possibility of the existence of
fully overlapping triangular reference sets means that multiple model inverses. Multiple model inverses will
only occur if the model is multi-modal with respect to
1H
k (x )"1 for j"1,2, m. (16) one or more of the manipulations. Consequently, to
G +H
G ensure that multiple inverses do not occur it is necessary
Consider the two fuzzy sets to which the crisp value of to ensure that the FRM used is uni-modal with respect to
x belongs to be A and B . Therefore since the degree of all the manipulated variables.
+H H H
belonging to all the other sets must be zero then from (16)

k (x )#k (x )"1. (17) 4. The fuzzy internal model controller (FIMC)

H +H H +H
Now bearing in mind that A and B must be adjacent to
H H The inversion procedure described above allows
one another, then there are (S !1) di!erent locations in
H a ccFRM to be incorporated into a true IMC design. The
k(x ) where A and B may be located. Extending the
+H H H FIMC controller structure is shown in Fig. 3 and it can
analysis to the set of m equations, (14), there will be be seen that it is based upon the standard IMC structure.
a total of S possible &solutions' for k(x ). However, However, for a FIMC the process model, PI , is the FRM
L +K
a constraint placed on membership to any fuzzy set is and the FIMC departs slightly from the IMC format in
k (x )3[0,1]. (18) that the controller, C, is not a true general model inverse
G +H but rather a block which inverts the FRM at each sample
From the many possible &solutions' the correct solution is time to produce a manipulation using the inversion pro-
the one where k (x ) and k (x ), for j"1,2, m, obey cedure detailed above. The FIMC uses simple low-pass
+H +H
Eq. (18). "lters for both F and F which are currently set by a trial

Once k(x ) has been found crisp manipulation and error method. It is through the tuning of these "lters
+K
values are found using the fuzzy mean method of defuz- that any inaccuracy in the FRM can be compensated for
zi"cation given in (4). and controller robustness increased.
C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877 871

Fig. 3. FIMC controller structure.

One of the principle advantages cited earlier for the use 5. Control of a multivariable pH neutralisation process
of an advanced control scheme such as IMC is its ability
to cope with deadtimes. This is straightforward for The control of pH is extremely important industrially
a SISO system but extending the controller formulation for many di!erent systems including boiler water treat-
into MIMO systems produces complications. Garcia and ment, chemical and biological reactions, municipal waste
Morari (1985) investigated this problem for linear IMC digestion, acid pickling and etching, cooling tower water
controllers and noted that for MIMO systems with mul- treatment, electrohydrolysis and coagulation/precipita-
tiple di!erent time delays that selection of an optimum tion (Williams, Rhinehart & Riggs, 1990). Because of this
inverse model for use in the controller was not immedi- industrial importance the control of pH has been a popu-
ately obvious. They go on to suggest that the trivially lar subject for research and there are many examples of
optimal model inverse is that which yields the least pH controllers presented in the literature (e.g. Nahas,
possible delay in each output response. For a MIMO Henson & Seborg, 1992; Qin & Borders, 1994; Sing &
FIMC the optimal control response in this sense, derived Postlethwaite, 1997). However, most of the controllers
from a model inverse, can be found using the following implemented so far have been purely single output pH
procedure. controllers and little work has been done on multi-vari-
Consider again the general form ccFRM represented able pH systems. The modelling of pH processes for
in (13) but this time with delays in the input variables simulation purposes has also received attention and
a general methodology for determining dynamic equa-
y "R k(x (k!q )) 2 k(x (k!q )) tions for a pH process was given by McAvoy and co-
" + + +K +K
k(x (k!q )) 2 k(x (k!q )), workers (McAvoy, Hsu & Lowenthal, 1972).
4 4 4T 4GT In this work a simulation of a multivariable pH pro-
$ (19)
cess was used to test the FIMC controller presented in
y "R k(x (k!q )) 2 k(x (k!q )) this paper. The process chosen was a neutralisation of
K "K + +K +K +KK
k(x (k!q )) 2 k(x (k!q )), a weak acid, acetic acid, and a strong base, sodium
4 4K 4T 4KT hydroxide, in a stirred tank. The aim of the controller
where k is the current sample and the q terms represent was to regulate both the liquid level in the tank and the
the lags for each model input. pH of the exit stream using the inlet acid and base
To obtain a controller manipulation from this model #owrates as manipulations. The system is shown sche-
the ccFRM represented in (19) would be inverted as matically in Fig. 4 where F represents the acid #owrate;

detailed earlier to yield crisp values for the suggested F represents the base #owrate and h represents the
manipulations, M , ∀j3[1,2, m]. However, this inver- liquid level in the tank. A dynamic model for the simula-
H
sion does not take account of the lags in the model and tion can be derived using the modelling methodology of
compensation must now be made. McAvoy (McAvoy et al., 1972).
Consider The reactions taking place in the neutralisation of
acetic acid and sodium hydroxide are
¸ "S (q H G ) for i"1,2, m, (20)
H +
H O0H>#OH\, (22)
i.e. ¸ will be the maximum time delay for any particular
H
fuzzy sub-model. The appropriate lag, j , for each of the
G HAC0H>#AC\, (23)
suggested manipulations, M , will be
G
j "R (¸ !q ) ∀j3[1, m]. (21) NaOHPNa>#OH\. (24)
G H HG
This procedure allows control of systems incorporating Now if [X] denotes the concentration of ion X,
di!erent deadtimes for the di!erent manipulations. then equations can be written for both the acetic acid
872 C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877

Table 1
Parameters for pH simulation

Parameter Value

Cross-sectional area of tank, A 1 m

Friction coe$cient for tank out#ow, C 0.01405
4
Concentration of acid, s 0.05005 M

Concentration of base, s 0.32 M
Flowrate of acid, F 0}160 l/s

Flowrate of base, F 0}1000 l/s
Controller sample time, ¹ 5s
1

dw F s !w (F #F )
@" @ , (34)
Fig. 4. Schematic representation of pH neutralisation system. dt Ah
where s and s are the molar concentrations of the acid

equilibrium and water equilibrium and base, respectively.
Using the titration function de"ned by (30) and (31)
[AC\][H>] and the three di!erential equations (32)}(34) the system
K " , (25)
? [HAC] was simulated using Simulink, part of the Matlab package.
The parameters used in the simulation are de"ned in
K "[H>][OH\], (26)
Table 1.
where K is the acid ionization constant for the acetic The ccFRM used in the FIMC was constructed by
? gathering 2500 samples of past process data at 5 s inter-
acid and K is the ion-product constant for water. Now if
vals. This data was generated by supplying band-limited
w "[HAC]#[AC\], (27) white noise signals as the acid and base #owrates over
?
w "[Na>], (28) a period of 12,500 s. The ccFRM was identi"ed from this
@ data using the least-squares identi"cation technique and
where w represents the total acetate concentration and consisted of the two fuzzy sub-models de"ned by (35)
?
w represents the sodium ion concentration. The solution and (36), using three fuzzy sets to partition all the input
@
in the tank must be electrically neutral and thus it can be variable ranges.
seen that
h(k)"R k(h(k!1)) k(F (k!1)) k(F (k!1)), (35)
"
[Na>]#[H>]"[AC\]#[OH\]. (29)
pH(k)"R k(pH(k!1)) k(F (k!1))
"
Combining the above equations yields
k(F (k!1)). (36)
[H>]#(K #w ) . [H>]#(K (w !w )
? @ ? @ ? Three tests were carried out to evaluate the controller
!K ) . [H>]! K .K "0. (30) performance. The "rst test was to perform servo action
?
on the pH across the entire pH operating range whilst
The pH of a solution is de"ned as maintaining the level at a constant value. The required
pH"!log [H>] (31) setpoint changes are given in Table 2 and the control
results for this test are presented in Figs. 5a and b. The
and when combined (30) and (31) represent a relationship second test was to perform servo action on the level
between w , w and pH termed the titration function. whilst holding the pH constant. In many real-world ap-
? @
An overall mass balance on the tank, assuming con- plications, such as waste water treatment, the pH of
stant density, yields a solution must be kept as close to neutral as possible.
This is a di$cult task as in this region the pH is very
dh
A "F #F !C (h, (32) sensitive to small #uctuations in either acid or base #ow.
dt 4
Consequently, for this second test the pH was held con-
where A is the cross-sectional area of the stirred tank stant at 7 to re#ect possible applications and to provide
and C is the friction coe$cient for the exit #ow. Molar the most challenging task for the controller. The required
4 servo action is given in Table 3 and the control results are
balances can be carried out for the acetate and sodium
ion concentrations and this gives presented in Figs. 6a and b. The "nal test conducted was
carried out to discover how robust the FIMC was in
dw F s !w (F #F ) coping with unexpected changes in process conditions. In
?" ? , (33)
dt Ah this test the concentration of the base was subjected to
C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877 873

Table 2 Table 3
Required setpoints for test 1 * servo action Required setpoints for test 2 * servo action on
on pH level

Time (s) pH Setpoint Time (s) Level setpoint (m)

0 5.5 0 0.5
500 7 1000 0.4
1000 10 2000 0.5
1500 11.5 3000 0.6
2000 10 4000 0.5
2500 7
3500 5.5

Fig. 5. (a) Controlled variables for test 1. (b) Manipulated variables for Fig. 6. (a) Controlled variables for test 2. (b) Manipulated variables for
test 1. test 2.
874 C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877

Table 4
Test 3 * step changes in base concentration

Time (s) Base concentration (mol/l)

0 0.05005
500 0.04750
1000 0.05005

a step change of approximately 10% as shown in Table 4.

The controller's task was to maintain the level at 0.5
m and the pH at 7 despite this unmeasured disturbance.
The results for this test are shown in Figs. 7a and b.
The controller "lter constants were set to the values
given in Table 5 in order to obtain good control for all of
the three tests.

6. Discussion

The performance of the controller for the "rst test is

shown in Fig. 5a and from this it can be seen that good
servo control of the pH was obtained whilst the level was
regulated very well. Control of pH necessitates coping
with the extreme non-linearity of the titration function
and the performance of the FIMC demonstrates the
power of the methodology in dealing with very non-
linear systems.
A particular problem in multivariable control is inter-
action between the process outputs. Such interaction is
caused by a manipulation having an a!ect on more than
one output and consequently any manipulation change
selected to achieve setpoint on a particular output can
disturb other outputs. In the sample system considered
pH can be adjusted with relatively little disturbance in
the level as only very small #owrate changes are required
to alter the pH. However, the servo action on the level
requires large changes in the manipulated #owrates and
consequently the pH is severely a!ected. In order to cope
with this severe interaction the setpoint "lter on the level
was set to a high value, see Table 5, in order to slow down
the required changes in acid and base #owrates. This
minimised the disturbance to the pH of the solution and Fig. 7. (a) Controlled variables for test 3. (b) Manipulated Variables for
as can be seen from Fig. 6a reasonably good regulation of test 3.
the pH was achieved during the servo action on the level.
In many commercial systems tight control of one par- Table 5
ticular output is more important than the control of the Filter constants for FIMC controller
other outputs. For example if pH control in waste water
Output variable Error "lter Setpoint "lter
treatment is considered it can be recognised that a vari-
ation in #owrates or level is much more tolerable than h 0.5 0.9
signi"cant disturbances in the pH. The design of the pH 0.9 0.5
FIMC permits prioritisation of certain outputs by careful
tuning of the "lter constants. As was demonstrated in
test 2 this prioritisation of controlled outputs can also The "nal test was conducted to check the robustness of
be used to improve the control of inherently sensitive the controller to unmeasured disturbances entering the
outputs. system. This is a particular problem in pH control as,
C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877 875

particularly around neutral pH, the system is extremely Expanding the above expressions yields
sensitive to very small changes in the concentration of the
h(k#1)"R ;k (h(k));k (F (k));k (F (k))
acetate and sodium. In the third test the concentration of "
the base was reduced by approximately 10% and it can R ;k (h(k));k (F (k));k (F (k))
"
be seen, from Fig. 7a, that the controller was robust
#2
enough to limit the impact of this disturbance and return
to steady state operation reasonably quickly. #R ;k (h(k));k (F (k));k (F (k)),
"
Although the test results presented look promising the
(A.3)
work is still in an early stage of development and a num-
ber of issues remain to be looked at. At the present time
the controller can cope with time-varying systems
pH(k#1)"R ;k (pH(k));k (F (k));k (F (k))
through careful tuning of the "lter constants to ensure "
robustness. However, a much better solution would be to R ;k (pH(k));k (F (k));k (F (k))
"
use some form of on-line model adaptation and this is
#2
certainly an area worthy of investigation. One area where
problems are envisaged is if control is required for pro- #R ;k (pH(k));k (F (k));k (F (k))
"
cesses which are not input}output stable or which are
(A.4)
inverse phase. At the moment processes which are not
input}output stable require stabilisation by a secondary where R represents element j in relational matrix i.
"
controller before the FIMC can be applied. Inverse phase Now, the use of fully overlapping triangular reference
systems can be handled somewhat better by selection of sets to de"ne the variable ranges means that
appropriate pseudo-deadtimes but obviously the control
k (F (k))#k (F (k))#k (F (k))"1, (A.5)
achieved is sub-optimal and the selection of a suitable
time lag for the model is also di$cult. Certainly these are k (F (k))#k (F (k))#k (F (k))"1. (A.6)

not trivial problems but a number of systems fall into
Furthermore, since any particular crisp value can belong
these categories and a solution to these di$culties would
to a maximum of two fuzzy sets it can be said that
be a signi"cant step forward in the general applicability
of the methodology. (F (k) , set1 OR F (k) , set3) AND

;(F (k) , set1 OR F (k) , set3),
7. Conclusions i.e.
(k (F (k))"0 AND k (F (k))"0)
The novel FIMC methodology has been introduced
and explained. The controller uses a ccFRM to represent OR
the process and controller action is determined by on- (k (F (k))"0 AND k (F (k))"0)
line inversion of this ccFRM. A multivariable pH simula-
tion has been presented in some detail and results have OR
been given showing the FIMC's performance in control- (k (F (k))"0 AND k (F (k))"0)
ling this system. These results demonstrate that the
FIMC is capable of good regulation of a di$cult non- OR
linear multivariable system. (k (F (k))"0 AND k (F (k))"0) (A.7)

Combining (A.5) and (A.6) with (A.7) yields four alterna-
Appendix A. Worked example of inversion procedure tive pairs of equations
k (F (k))#k (F (k))"1 AND
In order to clarify the inversion procedure a worked
example is presented here for the pH simulation pres- k (F (k))#k (F (k))"1, (A.8a)

ented in this paper. k (F (k))#k (F (k))"1 AND
The FRM for the pH system
k (F (k))#k (F (k))"1, (A.8b)
h(k#1)"R k(h(k)) k(F (k)) k(F (k)), (A.1)
" k (F (k))#k (F (k))"1 AND

pH(k#1)"R k(pH(k)) k(F (k)) k(F (k)), (A.2)
" k (F (k))#k (F (k))"1, (A.8c)

where all the symbols have the same meanings as before
k (F (k))#k (F (k))"1 AND
and three fuzzy reference sets have been used to partition
each variable range. k (F (k))#k (F (k))"1. (A.8d)

876 C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877

Combining equations (A.3), (A.4) and (A.8a)}(A.8d) and C"(c !c );pH !(c !c );h
1. 1.
equating the future values of the model outputs with the
!(c !c );c #c ;(c !c ),
desired setpoints yields a set of equations of the form
k (F (k))"1!k (F (k)), (A.11)
h "C ;k (F (k));k (F (k)) @> @
1. ? @
pH !(c !c );k (F (k))!c
# C ;k (F (k));k (F (k)) k (F (k))" 1. @ ,
? @> ? (c !c !c #c );k (F (k))#(c !c )
@
# C ;k (F (k));k (F (k))
?> @ (A.12)
# C ;k (F (k));k (F (k)),
?> @> k (F (k))"1!k (F (k)). (A.13)
?> ?
pH "C ;k (F (k));k (F (k))
1. ? @ Solving Eqs. (A.10)}(A.13) yields values for
# C ;k (F (k));k k (F (k)),k (F (k)), k (F (k)) and k (F (k)). This
? @>
(F (k)) ? ?> @ @>
means that we now have four possible &solutions' for
# C ;k (F (k));k (F (k)) k(F (k)) and k(F (k)):
?> @
# C ;k (F (k));k (F (k)), (1) k(F (k))"[k (F (k)) k (F (k)) 0],
?> @> ? ?>
k (F (k))#k (F (k))"1, k(F (k))"[k (F (k)) k (F (k)) 0],
? ?> (A.9) @ @>
k (F (k))#k (F (k))"1, (2) k(F (k))"[k (F (k)) k (F (k)) 0],
@ @> ? ?>
There are four di!erent sets of Eqs. (A.9) each represent- k(F (k))"[0 k (F (k)) k (F (k))],
@ @>
ing one possible &solution' for the manipulations F and
(3) k(F (k))"[0 k (F (k)) k (F (k))],
F . The C terms in the equations are the multiplication ? ?>
co-factors for each possible manipulation set combina- k(F (k))"[k (F (k)) k (F (k)) 0],
tion and are found by summating the product of all the @ @>
appropriate R terms and the fuzzi"ed past output values. (4) k(F (k))"[0 k (F (k)) k (F (k))],
? ?>
Consequently, the C values will be di!erent for each of
k(F (k))"[0 k (F (k)) k (F (k))].
the four &solutions'. In (A.9) the a and b terms have the @ @>
following values Provided that the model is uni-modal with respect to
F (k) and F (k) and that the required setpoints are
a"1, b"1, (A.9a)
achievable by the next sample then only one of the four
a"1, b"2, (A.9b) possible &solutions' above will obey the constraint:
k (F (k)) AND k (F (k)) AND k (F (k)))
a"2, b"1, (A.9c) ? ?> @
AND k (F (k))3[0,1].
a"2, b"2. (A.9d) @>
The correct &solution' for k(F (k)) and k(F (k)) is the
The set of Eqs. (A.9) can be re-arranged to yield
one which obeys the above constraint and crisp values
!B$(4AB for the manipulation can be found using fuzzy mean
k (F (k))" , (A.10) defuzzi"cation.
2C

where
References
A"! (c !c !c #c );(c !c )

Garcia, G. E., & Morari, M. (1982). Internal model control * 1.
# (c !c );(c !c !c #c ), A unifying review and some new results. Industrial Engineering

Chemical Process Design Development, 21, 308}323.
B"(c !c !c #c );pH Garcia, C. E., & Morari, M. (1985). Internal model control. 2. Design
1.
procedure for multivariable systems. Industrial Engineering
! (c !c !c #c );h Chemical Process Design and Development, 24, 472}484.
1.
Hunt, K. J., & Sbabaro, D. (1991). Neural networks for nonlinear
# (c !c );(c !c ) internal model control. IEE Proceedings * D, 138(5), 431}438.

McAvoy, T. J., Hsu, E., & Lowenthal, S. (1972). Dynamics of pH in
! (c !c );(c !c ) controlled stirred tank reactor. Industrial Engineering Chemical
Process Design and Development, 11, 68}70.
# c ;(c !c !c #c ) Nahas, E. P., Henson, M. A., & Seborg, D. E. (1992). Non-linear
internal model control strategy for neural network models.
! (c !c !c #c );c . Computers and Chemical Engineering, 16, 1039}1057.

C.R. Edgar, B.E. Postlethwaite / Automatica 36 (2000) 867}877 877

Lowe, D., & Webb, A. R. (1991). Optimised feature extraction and Bruce Postlethwaite graduated from the
the Bayes decision in a feedforward classi"er network. IEEE Trans University of Strathclyde in 1978 with
Pattern Analysis and Machine Intelligence, 13(4), 355}364. a BSc in Chemical Engineering. He then
Postlethwaite, B. E., Brown, M., & Sing, C. H. (1997). A new identi"ca- spent several years in the process indus-
tion algorithm for fuzzy relational models and its application in tries before returning to Strathclyde as
a lecturer in 1986. He completed his PhD
model-based control. Transactions of Institute of Chemical Engineers, in Fuzzy Modelling and Control while
75(A), 453}458. a member of sta!. He currently lectures in
Qin, S. J., & Borders, G. (1994). A Multiregion Fuzzy Logic Controller Process Control, Biochemical Engineering
for Nonlinear Process Control. IEEE Transactions on Fuzzy Systems, and Computer Programming. He has been
2, 74}81. working in the area of Fuzzy Control since
Ridley, J. N., Shaw, I. S., & Kruger, J. J. (1988). Probabilistic fuzzy returning to Strathclyde. He is a Director
model for dynamic systems. Electronics Letters, 24(14), 890}892. in the spin-o! company Process Cybernetics Ltd.
Sanchez, E. (1970). Resolution of composite fuzzy relational equations.
Information and Control, 30, 30}38. Craig Edgar studied Chemical Engineering
Sing, C. H., & Postlethwaite, B. (1996). Fuzzy relational model-based at Strathclyde University in Glasgow
control applying stochastic and iterative methods for model identi- graduating with at BEng (Hons) in
"cation. Transactions of Institute of Chemical Engineers, 74(A), 7}76. 1995. He then went on to study for a
Sing, C. H., & Postlethwaite, B. E. (1997). pH control: handling non- PhD, also in the Chemical Engineering
linearity and deadtime with fuzzy relational model based control. Department at Strathclyde University,
IEE Proceedings Control Theory Applications, 144, 263}268. which he obtained in 1988. For the last
18 months he has been acting as a
Williams, G. L., Rhinehart, R. R., & Riggs, J. B. (1990). In-line process-
technical consultant for Process Cyber-
model-based control of wastewater pH using dual base injection. netics Ltd in addition to his duties as
Industrial Engineering Chemical Research, 29, 1254}1259. a teaching assistant at Strathclyde Univer-
Zadeh, L. A. (1977). Fuzzy sets. In: J. Belzer, A. Nolzman, & A. Kent sity. His research interests include chem-
(Eds.), Encyclopedia of computer science and technology. New York: ical process control and in particular fuzzy modelling and model-based
Marcel Decker. control.