Mu-Synthesis of robust decentralised PI controllers

E.Gagnon, A. Pomerleau and A. Desbiens

Abstract: A new tuning method for robust decentralised PI controllers is given. The controller
tunings are obtained by minimising a robust performance criterion. The cost function to be
minimised is derived from the standard p-synthesis criterion and it takes into account the process
uncertainty and the desired performance. A constraint on the variations of the controller outputs
can also be added to the performance criterion. The two options, with and without a constraint,
respect the specifications in a minimal way to limit the controller output variations obtained from
the standard p-synthesis criterion. These two options and the standard p-synthesis criterion are
compared on an ill-conditioned two-input two-output (TITO) process example. The generalised
step response for TITO processes is used to analyse the time behaviour of closed-loop systems.

1 Introduction controller parameters. Another criterion, more logical

according to Skogestad and Lundstrm [5], is
Good tuning of decentralised PI controllers for multivari- ir 1 1
able processes is relatively complex. Some authors have
proposed tuning methods that take into account the process
uncertainty. Skogestad and Morari [I] have proposed an The variable z p represents the time constant of the perfor-
independent tuning method for decentralised controllers mance weight. An additional external loop is used in the
based on individual loop conditions. They have derived optimisation of the criterion (eqn. 2 ) to adjust the band-
their conditions from the global robust performance condi- width of the performance weight. Another interesting
tion of the p-synthesis environment. When the individual criterion proposed by Musch and Steiner [6] is
loop conditions are satisfied, the robust performance of the
global system is warranted. The controllers, however, are
tuned independently of each other, which produces conser-
vative tunings. Chiu and Arkun [2], Ito et al. [3] and Hovd
and Skogetad [4] proposed sequential design methods for where k represents the number of frequency points at
decentralised controllers. They also use the robust perfor- which the criterion is evaluated. They have also tested
mance concept defined in the p-synthesis environment. the minimisation of the sum of the squared structured
The sequential design implies that each controller can be singular values, but the deviations from the criterion of
fine-tuned after the fine tuning of another controller. This eqn. 1 are larger. All these criteria aim to obtain the best
operation can converges to simultaneous tuning methods robust performance independently of the variations of the
where all controllers are tuned at the same time. Chiu and actuators. Sonjiao et al. [7] proposed a criterion to tune
Arkun [2] have shown that their sequential method PID on monovariable processes, which could probably be
converges to a classical p-design approach. extended to decentralised PID conpollers. The form of PID
An interesting method for simultaneously tuning decen- used by them is
tralised PI controllers consists in using the standard p-
synthesis criterion [5]. It minimises the superior bound of
the robust performance. The criterion is

and the following criterion is optimised:

where p[o] is the structured singular value, A(s) is the
disturbance matrix which indicates that p[o] depends on
the allowed structure for A@),M(s) is a matrix obtained by
rearranging the uncertain system into the M A-structure
and C(s) indicates that the minimisation is function of the The parameter a is a constant chosen by experience. The
second term reduces the robust performance which can
be obtained from the standard p-synthesis criterion (eqn.
1). Therefore the controller output variations should be
smaller.
0 IEE, 1999 The preceding methods do not take directly into account
IEE Proceedings online no. 19990324 the controller output variations. These variations are an
DOL 10. IO49/ip-cta: 19990324 important point when dealing with industrial processes.
Paper received 4th February 1999 They should be taken into account since large action
The authors are with the GRAIIM, Department of Electrical and Computer variations result in fast deterioration of the actuators,
Engineering, Lava1 University, Quebec, Canada G l K 7P4 therefore generating supplementary maintenance costs.
IEE Proc.-Control Theory Appl., Vol. 146, No. 4, July 1999 289
 Another important point with industrial processes concerns d(s) are calculated. The singular values of the ratio u(s),
the physical constraints of the actuators. These constraints d(s) are
limit the amplitude of the manipulated variables, which
restrict the theoretical performance obtainable. As a conse-
quence, it may not be appropriate to design controllers to
obtain the best theoretical performance if these constraints
fJ
3
[- = a{-C(s)[I + P(s)c(s)J-l}= fJ[-C(s)E(s)]
(8:
are not taken into account.
A new method is proposed to simultaneously tune robust where ~ [ o ]is the singular value. The maximum and
decentralised PI controllers. The method is derived from minimum singular values of u(s)ld(s) delimit an amplifica-
the standard p-synthesis criterion, as the ones developed by tion band. This amplification band is defined by the
Skogestad and Lundstrm [5] and Musch and Steiner [6]. maximum and minimum 2-norm of the controller outpui
However, the criterion proposed limits the controller output vector for each frequency, when the disturbance vector i:
variations while taking into account the process uncertainty unitary 2-norm. Eqn. 8 also gives information on the
and the desired performance. An explicit frequency temporal controller output vector u(t). High frequencie:
constraint on the variations of the controller outputs can correspond to the possible 2-norm of the initial controllei
also be added to the criterion. Using an ill-conditioned output vector, and low frequencies to the possible 2-nom
process, the two options, with and without a constraint, and of the steady-state controller output vector, when a sei
the standard p-synthesis criterion are compared. The point change or an output disturbance step with an unitarq
generalised step response (GSR) is used as a temporal 2-norm is applied. The comparisons of 0 [U( co)/d( 00):
analysis tool to corroborate the frequency domain observa- with 0 [u(O)ld(O)]and 0 [U(C O ) / ~CO)]( with Cr [u(O)ld(O):
tions. permit to establish two possible extreme cases whicl-
depend on the set point change or the disturbance stel:
direction. Note that .[e] and a[.] are 'respectively' the
maximum and minimum singular values. One extreme case
2 Basic tuning method arises when the controllers apply the largest Ilu(t)llz at the
initial time and the smallest Ilu(t)l12 at steady-state. The
2.1 Criterion second extreme case is when the controllers calculate the
smallest Ilu(t)llzat the initial time and the largest Ilu(t)(12a1
The criteria of eqns. 1, 2 and 3 try to reach the best robust
steady-state.
performance regardless of the controller output variations. To facilitate the comparative study between different
An interesting method to limit the variations of the
processes, the u(s)ld(s) singular values are normalised.
manipulated variables would be to use a criterion that
This normalisation consists in multiplying all the u(s)/
satisfies, in a minimal way, the robust performance. The
d(s) singular values of a control system by a constant p
proposed criterion is defined as follows:

Therefore 0 [u(O)ld(O)]= 1 for all control systems. The

where n is a positive integer, which represents a penalty example given in the following Section shows the applica-
factor of the criterion. M(s) is defined as follows for the tion of the new criterion to a simple ill-conditioned
case of multiplicative input uncertainty [SI: process. A comparison with the standard p-synthesis
criterion is also detailed.
M(s) =
[
-P(s) - "(s)P(s)w,(s)
E_(s)P(s)w,(s)
'-
-P(s)- H(s)w,(s)
E(s)rn,(s) I(7)
2.3 Example
The ill-conditioned process (its condition number is seven)
is the following:
where P(s) is the transfer matrix of the nominal process, 4 3
E(s)= [Z+P(s)C(s)] is the sensitivity function,
1+
[ L 1 + 10s ]
10s i
~~

p(s)=
'
H(s)= P(s)C(s)[l+P(s)C(s)] - is the complementary (10)
sensitivity function, oXs) is the relative uncertainty and
1 + 10s 1 +1os
a d s ) is the desired performance weight. Note that C(s)
represents the controller transfer matrix and I the identity Fig. 1 shows the linearised generalised relative dynamic
matrix. The criterion is evaluated over k different fi-equen- gain on a 3D graphic (GRDGL 3D) for the combination of
cies. Eqn. 6 aims to minimise the gap between the robust the GRDGL elements Ibl1(s) and A22 (s) [9] when first-
performance and unity. The desired performance is then order dynamics are assumed in closed-loop for the process
exactly satisfied in the worst case if the robust performance outputs. The frequencies oC1 and wC2represent the band-
is equal to unity for all the frequencies. From a practical widths of the first and the second loop, respectively. A
point of view, it is useless to obtain better robust perfor- small interaction is observed for a small JGRDGLI value.
mance than the design specification because the variations Fig. 1 shows that smaller interactions are obtained when
of the actuators will then be stronger than necessary. both closed-loops have the same bandwidth. To respect this
condition the controllers would be identical. It is possible
that nonidentical controllers can give w,1z w , ~for some
2.2 Controller output variations tunings, but identical controllers warrant wC1= wc2 for all
To obtain information on the variations of controller possible tunings on the nominal process. Also, noniden-
outputs, the singular values of the transfer matrix between tical controllers can improve the robust performance for
the process input u(s) and the process output disturbance the process used in this example [lo], but the possible
290 IEE Proc-Control Theoory AppL, Vol. 146, No. 4, July I999
 50
45

40

35
30
y" 25
20
15
10

5
n
0 20 40 80 80 100
Fig. 1 GRDGL3Dfor combination ofGRDGL elements i.ll(s) and A&) 'I
Fig. 3 Contour of standard p-synthesis criterion as function of PI gain
K, and PI integral TI
improvement is so small that it does not seem necessary to
used nonidentical controllers here.
Assuming an arbitrary uncertainty of 20% on each Using the standard p-synthesis criterion (eqn. 1) with
process input gives identical controllers, the decentralised PI controllers
obtained by optimisation are
0.2 0 +
= [ 0 0.21

and the disturbance matrix A(s) is diagonal. To obtain

closed-loop performance equal or better than a first-order
dynamics with a time constant of five time units, cup(s) has
to be defined as follows:
C(S) =
[
8.11(1 34.74s)
34.74s
0
0
+
8.11(1 34.74s)

Fig. 3 shows the 2D contour graphic of the standard p-

34.74s

synthesis criterion as a function of the PI gain K c and the

I (14)

PI integral TI. This fimction seems also convex on the

WP(S) = [ ; 5s0+
5s+ 1

5s
' (12)
subspace Kc > 0 and TI > 0. Although the criterion is not
very sensitive to the parameters of the PI controllers in a
large part of Fig. 3, the tuning given by eqn. 14 (Kc= 8.1 1
and TI=34.74) seems to be a global minimiser of the
criterion on the subspace Kc > 0 and TI > 0.
Using the criterion developed in eqn. 6 with identical Fig. 4 shows the robust stability (RS), nominal perfor-
controllers and a unitary penalty factor n, the decentralised mance (NP) and robust performance (RP) [8, 11-13]
PI controllers obtained by optimisation are curves for the tuning of eqn. 13. The RP curve follows
unity almost over the entire frequency window except in
+

1
2.63(1 10.53s) the o = 1 frequency area. This small overshoot involves
0 that the desired performance will be slightly violated in the
C(S) = 10.53s
' +
2.63(1 10.53s) (13) worst case for the uncertainty evaluated. The tuning
10.53s obtained is however optimal for the chosen criterion. Fig.
4 also shows that the NP curve is smaller than unity for all
Fig. 2 shows the two-dimensional (2D) contour graphic of frequencies. It indicates that the performance for the
the criterion as a function of the PI gain K c and the PI nominal process will be better than the desired one. The
integral TI. It shows that the function seems convex on the RS curve is largely under unity. It indicates that the system
subspace K c > 0 and TI > 0. Therefore the tuning given by will remain stable in spite of an uncertainty of 20% on each
eqn. 13 (&=2.63 and TI= 10.53) is probably a global
minimiser of the criterion on the subspace KC.> 0 and
T I > 0.

0'4 t
- 3 2 1 0 1 2
10 10 10 10 10 10
frequency
Fig. 4 p-plots for tuning ojeqn. 13
....... RS
Fig. 2 Contour ofproposed criterion (eqn. 6) as firnction of PI gain K , ~ _ NP
_
and PI integral TI -RP
IEE Proc.-Control Theory Appl.. Vol. 146, No. 4, July 1999 29 1
 20

Fig. 7 GSR ratio for set point change GSR with tuning of eqn. 13
frequency

Fig. 5 p-plots for tuning of eqn. 14 for the best case and in the worst case, C[u( 00)ld( m)],
....... RS o[u(O)ld(O)]= 56.77. Fig. 6 also indicates, as expected foi
- _ _ NP
-RP both tunings, that .[U( CO )/d(00 )] = g[u( 00 )id( CO )]. Thi:
is because the controllers are identical for each tuning.
To illustrate the previous frequency observations, the
process input. The gap between the RP and the NP curves time domain can be used with the generalised ster
indicates that the performance deterioration will be larger response (GSR) [9]. The GSR is a good tool to obtain i
at low frequencies for the uncertainty chosen. The infor- better feeling of the frequency graphics than a step in on14
mation given in Fig. 4 is of course true only if the one direction. As an example, the 20% uncertainty on each
controlled system is nominally stable [ 131. process input can be applied with a set point change GSR.
Fig. 5 shows the RS, NP and RP curves for the tuning of An interesting case for applying the uncertainty of 20% on
eqn. 14. The high frequencies are almost identical to the each process input is obtained by creating an error of
ones of Fig. 4. The low frequencies are a little worse with
the tuning of the standard p-synthesis criterion. But the
+ 20% on the first process input and an error of - 20% on
the other process input. Figs 7 and 8 show the results for a
major difference is that better performance is obtained in set point change GSR with an unitary 2-norm for the
the middle frequencies with the standard p-synthesis tuning obtained with the proposed criterion. One axis
criterion compared with the one proposed. However, this represents the time t, another the direction of the set
unnecessarily better performance in middle frequencies has point changes Q and the vertical axis represents the GSR
the disadvantage of increasing actuator wear. (Fig. 7) and the 2-norm of the
ratio I(y(t,e)((21((r(t,e)(12
Fig. 6 shows the u(s)ld(s) singular values, which have controller output vector Ilu(t,Q)llz(Fig. 8). Note that y is
been normalised to get g[u(O)ld(O)]= 1. It is observed that the process output vector and r is the set point vector.
g[u(C O ) / ~ CO)]
( is larger than Z[u(O)id(O)] for both These curves are in agreement with the information
criteria. Therefore the 2-norm of the initial controller obtained from the frequency curves of Fig. 4 and 6. Fig.
output vector will always be larger than the 2-norm of 7 shows that the system is stable in all directions of the set
the steady-state controller output vector, when a set point point change. Fig. 8 shows that in the 2.3 radians direction
change or an output disturbance step affects the nominal (best case for the actuator variations), the 2-norm of the
process. From Fig. 6, in the best and worst cases ‘respec- initial controller output vector is 2.47 times larger than the
tively’ for the actuator variations with the criterion 2-norm of the steady-state controller output vector. In the
proposed, o[u(C O ) / ~ ~~)]lCr[u(O)ld(O)]
( = 2.63 and 0.7 radian direction (worst case for the actuator variations),
6[u(C O ) / ~ (0 0 ~ / ~ [ u ( O ) l d ( O=) ]18.41. With the standard the initial 2-norm is 18.45 time lager than the steady-state
@-synthesis criterion, g[u( CO )id( CO )]l~[u(O)/d(O)] = 8.1 1 2-norm. If there were no process input errors, the ratios
would be 2.63 and 18.41, which correspond to the ratios of
Fig. 6. It can also be noted that the 2-norm of the initial
controller output vector is independent of the set point
change direction, as shown in Fig. 6.

frequency
Fig. 6 Normalised singular values of u(s)ld(s) for tunings of eqns. 13
and 14
-eqn. 13 Fig. 8 2-norm of controller output vectorfor set point change GSR with
_ _ _ eqn. 14 unitary 2-norm for tuning of eqn. 13

292 IEE Proc.-Control Theovy Appl., Vol. 146. No. 4, July 1999
 .....
where the subscript p indicates that the maximum singular
values are normalised with the parameter p defined by eqn.
1.o 9. This bound warrants that
-
-
N
0.8
-E-= 0.6
.-
-
- 0.4 Defining ui (s), i = 1, 2, . . ., as the controller outputs and
sx- 0.2 assuming that Ild(s)l12 = 1, then
-
0

0
which constrains the controller output variations. There-
fore the proposed criterion is
Fig. 9 GSR ratio for set point change GSR with tuning of eqn. 14

Different forms of bounds can be used. A simple form, as

the one described by expr. 15, leads to good results.

3.2 Example
In this example the same process as the one defined by eqn.
10 is used. The same uncertainty and performance are
specified with the same structure for A(s), but a constraint
on the controller output variations is added. From Fig. 6,
Cr[u(O)ld(O)]is equal to seven and is fixed by the process.
The constraint in this frequency region cannot be smaller
Fig. 10 2-norm ojcontroller output vector for set point change GSR than seven. Then, suppose that the desired constraint is
with unitary 2-norm for tuning of eqn. 14 defined as

Fig. 9 and 10 show the results for a set point change

GSR with an unitary 2-norm for the tuning using the
standard y-synthesis criterion. From Fig. 9, it is clear This constraint is the dashed line plotted in Fig. 12. It
that the tuning obtained with the standard p-synthesis warrants that the 2-norm of the initial controller output
criterion produces better performance in some directions vector will be smaller than the 2-norm of the steady-state
than the tuning obtained with the criterion proposed (Fig. controller output vector for set point change or output
7). However, in other directions, it seems that the integral disturbance steps on the nominal process. Using the
constant TI of the tuning based on the standard y-synthesis criterion of eqn. 18 with identical controllers, a unitary
criterion is too large. Fig. IO shows that the tuning of the penalty factor n and the constraint of expr. 19, the decen-
standard p-synthesis criterion produces higher 2-norms for tralised PI controllers obtained by optimisation are
the initial controller output vector than the tuning obtained
+

1
with the criterion proposed (Fig. 8). 0.1414(1 2.748s)
0
2.748s
C(s) =
0
+
0.1414(1 2.748s)
3 Tuning with constraint 2.748s
(20)
3.1 Criterion
Fig. 11 shows the RS, NP and RP curves for that tuning.
The criterion of eqn. 6 can decrease the variations of the The NP and RP curves are largely over unity at low
actuators, but it cannot. warrant that the variations will be frequencies. This means that the desired performance has
less than a fixed threshold. A solution to limit the varia- not been respected owing to the optimisation constraint on
tions of the actuators consists in using a constrained the controller output variations. The system is slower than
criterion. To be able to fix a constraint on the variations the desired one. The RS curve is however well under unity.
of the actuators, it is necessary to obtain a measure of these The system will then remain stable in spite of an uncey-
possible variations. As previously explained, a good tainty of 20% on each process input. Fig. 12 shows the
measure is obtained from the normalised singular. values u(s)/d(s) normalised singular values (solid line) for the
of u(s)/d(s). Therefore the constraint on the controller tuning given by eqn. 20. The dashed line-is the constraint
output variations can be defined by setting a maximum defined by expr. 19. It is clear that the controllers have been
bound on the normalised maximum singular values of U @ ) / optimised to reach the desired performance without violat-
d(s). The following equation shows a simple form of the
ing the constraint.
constraint that can be used: Fig. 13 and 14 show the results of a set point change
GSR with an unitary 2-norm. The 20% uncertainty on each
(15) process input has, been applied by creating an error of
+ 20% on the first process input and an error of - 20% on

IEE Proc -Control Theory Appl, Vol 146, No 4, JUG 1999 293
 4.5

3.5 .
3.0 .
m 2.5
U
2 .
.-
c

:::I
0.5
......
0

410
.,,,,,,,....,,.

10
-2
I....., ..,.

10
.
1...1-.-.,..

1
”. ...,...,,.

frequency
10
0
10
1
10
I2
0 0

Fig. 13 GSR ratio for set point change GSR with tuning of eqn. 20

Fig. 11 p-plots for tuning of eqn. 20

....... RS
--_ NP
-RP

6 ,,
,,
5 , I
c I
-34
0, ,,
I

,,
3
I
,
2

1
\

.
>
’----

-2 1 0 1
I2
10 10 10 10 10 10 Fig. 14 2-norm of controller output vector for set point change GSR
frequency with unitary 2-norm for tuning of eqn. 20
Fig. 12 Constraint of eqn. 19 and normalised singular values of
u(s)ld(s)for tuning ofeqn. 20
-__ eqn. 19
~ eqn. 20 5 References
1 SKOGESTAD, S., and MORARI, M.: ‘Robust performance of decen-
tralized control systems by independent designs’, Automatica, 1989,25,
the other process input. Fig. 13 shows that the system does (I), pp. 119-125
not respect the desired performance. However, Fig. 14 2 CHIU, M.-S., and ARKUN, Y.: ‘A methodology for sequential design of
robust decentralized control systems’, Automatica, 1992, 28, (5), pp.
shows that (lu(t=Of)l12 5 Ilu(t= 0 0 ) 1 1 ~ in all directions 997-1001
as illustrated in Fig. 12. This part of the constraint is then 3 ITO, H., OHMORI, H., and SANO, A.: ‘Robust performance of
respected, even with .the errors of +20% and -20% decentralized control systems by expanding sequential designs’, Int. J;
Control, 1995, 61, (6), pp. 1297-1311
created on the process inputs. ~
4 HOVD, M., and SKOGESTAD, S.: ‘Sequential design of decentralized
controllers’, Automatica, 1994, 30 pp. 1601-1607
5 SKOGESTAD, S., and LUNDSTROM, P.: ‘Mu-optimal LV-control of
4 Conclusion distillation columns’, Comput. Chem. Eng., 1990,14, (4/5), pp. 4 0 1 4 1 3
6 MUSCH, H.E., and STEINER, M.: ‘Robust PID control for an industrial
distillation column’, IEEE Control Syst., 1995, 15, (4), pp. 46-55
Since actuator deterioration is an important factor for 7 SONJIAO. S.. SHAN, C.. and ZHONGJUN, Z.: ‘A design method for
industrial maintenance costs, actuator variations are very robust PID controller ‘based on structured singular value’, . IShanghai
important. To satisfy this requirement, a tuning method for Jiaotong Univ., 1995, 29, (3), pp. 1-5
8 MORARI, M., and ZAFIRIOU, E.: ‘Robust process control’ (Prentice-
robust decentralised PI controllers has been proposed. The Hall, Englewood Cliffs, NJ, 1989)
method takes into account the process uncertainty and the 9 GAGNON, E., POMERLEAU, A., and DESBIENS, A.: ‘On new
desired performance. It tries to satisfy the latter in a representations of two input-two output (TITO) process characteristics’.
GRAIIM document 98-10, 1998, Dept. of mining and metallurgy, Lava1
minimal way as opposed to the standard y-synthesis University, Quebec, Canada
criterion where a better robust performance can be reached 10 LUNDSTROM, P., SKOGESTAD, S., HOVD, M., and WANG, 2.:
‘Non-uniqueness of robust H , decentralized PI-control’. Proceedings
at the cost of stronger variations of the controller outputs. of the 1991 American Control conference, IEEE cat. 91CH2939-7,2 pp.
The addition of an optimisation constraint on the varia- 1830-1835
tions of the controller outputs permits to tune the control- 11 DOYLE, J.C., WALL, J.E., and STEIN, G.: ‘Performance and robustness
analysis for structured uncertainty’. Proceedings of IEEE conference on
lers to approach the desired performance in the worst case, Decision and Control, Orlando, FL. 1982
while respecting the desired variations of the actuators. 12 DOYLE, J.C., and STEIN, G.: ‘Multivariable feedback design: concepts
The criterion proposed has been applied with PI control- for a classicallmodem synthesis’, IEEE Trans., 1981, AC-26 (1) pp. 4-
16
lers but it could also be used with other parametric 13 SKOGESTAD, S., and POSTLETH WAITE, I.: ‘Multivariable feedback
controllers. control’ (John Wiley, UK, 1996)

294 LEE Proc.-Control Theoly Appl., Vol. 146, No. 4, July 1999