You are on page 1of 12

Pergamon

ooos1098(95)00105-0

Development

ERIC POULIN,?

ANDRE

Auromatica,Vol. 32, No. 1, pp. 71-82, 1996


Copyright 0 1995 Elsevier Science Ltd
Printed in Great Britain. AU rights reserved
ooo5-m&3/%$15.00 + 0.00

and Evaluation of an Auto-tuning


Adaptive PID Controller*
POMERLEAU,?

ANDRE

DESBIENSt

and DANIEL

and

HODOUINS

An auto-tuning and adaptive PID controller based on the identification of a


second-order discrete-time model with time delay is developed and compared with three commercial adaptive controllers.
Kev Words-PID
control; automatic tunina: adantive control; identification;
filt&ng; industrial control.

Abstract-This
paper describes the design of a practical
auto-tuning and adaptive single-input-single-output
(SISO)
PID controller (AAC). The AAC can control processes with
stable and unstable zeros, processes with an integrator,
unstable processes and standard aperiodic processes. It uses
an explicit identification
with a recursive parameter
estimation of a second-order
with delay model. The
regulator tuning methods are based on an approximate
minimization of the ITAE criterion by applying pole-zero
cancellation, phase margin and maximum peak resonance
specitications, with special considerations for delays, unstable
zeros and poles. The data filtering, the identification, the
tuning mechanism and the supervisory shell are described.
Useful guidelines for PI and PID tuning for SISO processes
are given. The AAC performances are compared using a
benchmark test with commercial adaptive PID controllers:
Foxboro 76OC, Fisher DPR 910 and Leeds & Northrup
Electromax V.
1. INTRODUCIION

Unpredictable
changes in processes dynamics
lead to poor control performances if controller
parameters are not properly adapted. Some of
these changes arise from nonlinearities, process
aging, production strategy changes, raw material
property modifications and changes over equipment maintenance cycles. Adaptive control is
widely discussed in the literature, and reviews of
the evolution of adaptive control algorithms and
products have been given by Astrom (1987),
Keyes and Kaya (1989), Chan (1991) and
Astrom et al. (1993). Comparative studies of
commercial and academic controllers have been
presented
by Nachtigal (1986), Minter and
*Received 26 August 1994; revised 17 March 1995;
received in final form 20 June 1995. This paper was not
presented at any IFAC meeting. This paper was recommended for publication in revised form by Associate Editor
P. J. Gawthrop under the direction of Editor C. C. Hang.
Corresponding author Dr %ic Poulin. Tel. +l 418 656 2131
E-mail
Fax
+l
3159;
4785;
418
656
ext.
Eric.Poulin@gel.ulaval.ca.
t Departement
de Genie filectrique, Universite Laval,
Sainte-Foy, Quebec, Canada GlK 7P4.
$ Departement de Mines et Metallurgic, Universite Laval,
Sainte-Foy, Quebec, Canada GlK 7P4.

parameter

estimation;

Fisher (1988), Dumont et al. (1989), Sorrells


(1989), Cao and McAvoy (1990), Hang and Sin
(1991) and Isermann et al. (1991).
In spite of the increasing number of successful
industrial applications of adaptive control, there
is room for improvement.
Adaptive control
compromises
with adaptation
performances,
stability and robustness in order to cope with
industrial environment
(De Larminat,
1989;
Matko,
1993). In many cases, adaptation
performances are reduced to ensure stability and
robustness in practical applications (Song et al.,
1984; Bartlomiej and Andrzej, 1988). Supervisory shells including safety nets and special
tricks must be used to overcome violations of
theoretical
developments
of the controller
(Albertos et al., 1989; McIntosh et al., 1992).
Finally, the types of processes that adaptive PID
controllers can control are sometimes limited to
standard aperiodic minimum-phase processes.
This paper describes the development of a
practical auto-tuning and adaptive PID controller (AAC). The PID controller structure has
been retained, since it is the most widely used in
industry, and it is able to control a wide range of
processes if properly tuned. The AAC is based
on an explicit identification of a second-order
model with delay and a design step to tune the
controller. Model identification and controller
tuning are under a supervisory shell that ensures
good performances in real conditions. The AAC
controls processes with stable and unstable
zeros, processes with an integrator, overdamped
and unstable processes.
The paper is organized as follows. The data
filtering procedures are presented first. Afterwards, the identification method and its supervision system are described. The rules and the
design procedures
for the PID tuning are
explained, and, finally, comparisons with com-

I?. Poulin et al.

72

mercial adaptive PID controllers are presented.


The main contribution to the paper is to present
an adaptive and auto-tuning PID controller that
can handle most types of industrial processes.
The paper describes the explicit design of the
data
acquisition
and filtering
system
for
recursive
identification.
It also summarizes
efficient PID tuning methods based on frequency
responses.
2. DATA

FILTERING

Data filtering is necessary to obtain satisfactory estimation


results. Noise and external
disturbance effects must be attenuated as much
as possible in order to use reliable data for
identification and model updating. Even when
noises and disturbances are not present, filtering
is required to mask the effect of unmodeled
dynamics, since one is attempting to model
high-order
systems with a low-order model
(Mohtadi, 1988). Filtering the input and output
data acts on the bias distribution
of the
estimated
transfer
function
(Wahlberg
and
Ljung, 1986; Ljung, 1987). It is thus possible to
emphasize and depress different frequency bands
by a proper choice of filters.
Figure 1 presents the data filtering system for
identification.
First, an anti-aliasing
analog
lowpass filter is used and is considered as a part
of the process. Afterwards, a fast sampling is
performed (w,& and both input and output data
are bandpass-filtered. Identical digital filters are
applied in order to avoid modifications of the
identified transfer function. The bandpass filter is
realized by combining digital lowpass and
highpass filters. The digital lowpass filter
attenuates high-frequency noise and reduces the
effect of unmodeled
dynamics. The digital
highpass filter allows identification of a process
around a set point and attenuates the effect of
nonzero mean disturbances. Finally, following
the digital filtering, some samples can be

IDENTIFICATION

DIGITAL HP

DIGITAL LP

dropped to go down to a lower sampling


frequency ( w,~) appropriate
for identification
purposes.
A detailed procedure for the design of the
data acquisition system for recursive identification is given by Poulin et al. (1994). The analog
lowpass filter is chosen according to the desired
attenuation at the sampling frequency osl, the
allowable phase margin reduction for control
applications and the acceptable system modification by addition
of supplementary
poles
(Pomerleau and Hodouin, 1994). The digital
lowpass filter is selected to match the desired
closed-loop bandwidth o,, (Shook et al., 1991)
and the choice of the digital highpass filter is a
compromise between the elimination time of the
effect of nonzero mean disturbances and the
allowable information
attenuation
that slows
estimation convergence
(Astrbm and Wittenmark, 1989).
In the AAC the analog lowpass filter and the
fast sampling frequency w,, are not managed by
the auto-tuning
and adaptation
mechanisms.
This part of the data acquisition system does not
need to be adapted, since the controlled systems
are oversampled at this step. A second-order
analog lowpass filter with a cutoff frequency
1s
.
used. This gives a 40 dB
%3 = +)Q
attenuation at w,~.
The digital filter parameters and the second
sampling frequency ws2 are determined automatically at the auto-tuning (Section 5). The
controller
roughly
evaluates
the open-loop
dynamics of the system. This information is used
to determine the initial closed-loop bandwidth
w,,. The cutoff frequencies of the first-order
lowpass filter w,~, and the first-order highpass
filter ucdh are set to o,~, = w,~,, = w,,. The second
sampling frequency
w,~ (identification
and
control frequency) is chosen according to the
relation w,* = (10-20)~~ (Isermann, 1980). After
auto-tuning, the digital filters and w,~ are not
modified.
The data acquisition
system is
sufficiently flexible to cope with process dynamics variations. It is worth noting that the initial
closed-loop bandwidth o,, is only used for the
data acquisition system setting. The information
about the process dynamics required for the PID
tuning is continuously
extracted
from the
identified model and recalculated.
3. IDENTIFICATION

A second-order

model with delay is identified:


b,z

- + bzzm2 -d
z
+ a,Z-2

Gil(Z)= 1 +a,z-
Fig. 1. Data filtering

for identification.

where

d is the delay in number

(1)

of sampling

Auto-tuning

and adaptive PID

periods
estimated
at the auto-tuning
step
(Section 5). The identification is based on the
damped least-squares (DLS) algorithm, which is
an extended version of the recursive simple
least-squares (RLS) algorithm. The DLS algorithm is more appropriate for adaptive control,
since it weights increments of the estimated
parameter vector. This gives more control on the
adaptation rate. The DLS criterion is

J(6)= i

rk[y(k)

CpT(k)ii(t)]2

k=r-N

II&(0@(0
- &t - 1))11.

(2)

The weighting matrix Ad(t) is diagonal and


weights the parameters
variations.
For an
n-parameter model,
A,(t) = diag [al(t)

a&)

. . . %(f)l*

(3)

A standard form of the DLS algorithm is given


by Lambert (1987):
6(t) = @(t - 1) + K(t)[y(t) - ii=@ - l)cp(t)]
+ P(t - l)h(t)A,(t)[@t

- 1) - i& - 2)],
(4)

K(t) =

P(t) = $j

P(t - l)cp(t)
A(?) + cp(t - l)P(t - l)cp(t - 1)
P(t) -

P(t)cp(t - 1)&t

- l)P(t)

(5)

h(t) + cp(t - l)P(t)cp(t - 1)

(6)
P(t) = P(t - 1)
_ 2 P[_& - l)r$:P;_*(t - l)(Y@)
1 + rTPi_,(t - l)ricYi(t)

i=l

(7)

PJ(t - 1) = P[_l(f - 1)
_ Pi_,(t - l)rirFP:_I(t - l)al(t)
1 + rTPf-I(t - l)riaI

PA(t- 1) = P(t - l),

(8)

(9)

where ri are the successive basis vectors, e.g.


rl = [l

...

O]=,

(10)

and
a:(t) =

a&) - A(+#
A(t)

normalization

constant
factor;

trace through

a variable

information measurement
tation on and off;

stability/instability

parameter

forgetting

for turning adap-

check;

variations and maximum limits.

The parameter setting for these modifications


and the supervisory shell is performed at the
auto-tuning step (Section 5)
Normalization of input-output
data is useful
to prevent the effect of unbounded modelling
errors and to keep values in the same magnitude
range. The output y(t) and the regressor vector
q(t) are divided by
n(t) = p~(t - 1) + (1 - p) max [max IcpWL77d1

(12)
where
Os/J.l,

To>O.

(13)

The trace of the covariance matrix P(t) is a


measure of the magnitude of P(t) and hence of
the magnitude of the adaptation gain K(t).
Setting the trace of P(t) constant ensures a
constant adaptation gain and prevents covariante blowup. The algorithm is then able, by a
proper choice of the trace, to track process
parameter variations.
A measurement of information is given by
mi(t) = cp(t - l)P(t - l)cp(t - 1).

(14)

It gives the possibility to switch the estimation


off when the excitation is poor and on when the
excitation is rich. This prevents model degradation during good control periods.
Some verifications are made on the new
model before it is transferred to the tuning
module. A stability check ensures that a stable
model is used to control a stable open-loop
process, and vice versa. Gain limits constrain the
gain of the model between certain limits.
Maximum values of the parameters and the
parameter variations can be added to ensure a
smoother adaptation.

- 1)

(11)

When the weighting matrix Ad(t) is taken as


diag [0 0 . . . 01, the DLS algorithm is equivalent to the RLS algorithm.
Some modifications that improve the algorithm performances in practical situations have
been implemented (Sripada and Fisher, 1987;
Shah and Cluett, 1991; Desbiens et al., 1994).
They are the same as in the RLS case:
l

73

of the regression vector;

4. CONTROLLER

TUNING

The PID has an interacting structure. This


structure is equivalent, with proper parameter
conversions
(Besharati
Rad and Gawthrop,
1991) to the noninteracting one, but it has the
advantage
that the poles can be directly
cancelled. Its equation in the s domain is given
by
U(s)

G&)=-=E(s)

K, (1 + 7$)(l+

TTi

1+&s

Tds)

(15)

E. Poulin et al.

74
where
E(s) = R,(s) - Y(s),

R,(s) = $$)

(16)
f

and, in its discrete form,

G,(z) =

1 -p
I-pz-

Po+p,Z-'p*Z-2

l-z-

(17)

in the controller to avoid discrete-to-continuous


and continuous-to-discrete
conversions. Limited
calculations in the s domain are used for
nomininmum-phase
and unstable processes to
overcome problems due to poles and zeros
outside the unit circle.
Case I

with

Kpeeas
= (1 + T,s)(l + T2s) J

q4s)

T 55%

(22)

(18)
PI =K,

(19)
p2= K$
p

(20)

e-7;7;.

The process is stable, and has two significant


time constants. A PID with filter compensation
is used.
Speci,fications.

The desired closed-loop


transfer function
G,,(s) should behave as a first-order system
with delay, (23), in the ideal case (Seborg et
al., 1989). The specification can be achieved
when no derivative filter is used:

(21)

It should be noted that the set-point changes


R(s) are filtered (R,(s)) to avoid the derivative
kick. This gives the possibility of having a
two-degree-of-freedom
controller. By a proper
choice of the gain of the controller and a proper
set-point filtering, it is possible to have different
dynamics for load disturbances and set-point
response.
The tuning rules used by AAC are classified
into seven different cases that include all types of
process controlled by the AAC. The rules are
gathered in Table 1. The user has first to specify
the class of the process, i.e. if it is stable,
unstable or has an integrator. Afterwards, the
characterization
is perfromed automatically by
the controller at the auto-tuning level. The
process is allowed to change from one case to an
other during adaptation while staying in the
same process class. A complete discussion of the
tuning rules is given in Pomerleau (1994) and
Poulin and Pomerleau (1995).
According to the certainly equivalent principle, the model transfer function G,(s) is
assumed to be identical to the process transfer
function G,,(s) for the tuning step. A sufficient
phase margin A@ is preserved for modelling
errors. The tuning methods are presented in the
s domain, but are implemented in the z domain

G,,(s) =

&.
c

In order to obtain similar dynamics to the


open-loop response, the desired closed-loop
time constant T, is set to T, = T,.
For real applications, a derivative filter must
be used to prevent high-frequency
noise
amplification. The time constant of this filter is
chosen as Tf = 0.2T,.
The phase margin for overdamped
.
given by
AQ,

180

180
?r

It is minimum (A~)mi) when T1= T2, and


decreases with increasing delay 8. For processes
without delay, the minimum phase margin is
AR,,, = 78, and for processes with delay 6 = T,,
it decreases to Audi, = 55.
Case 2
KpemS
G,(s) = t1 + T,sjtl + T2s) J

Table

1. AAC

PID tuning

for the 7 different

cases

CCW

7;

1
2
3
4

T,lK,(T + 6)

r,

0.2Tz

T,,IK,V, + 6)
T,IK,(T, + 6)
T,IK,(T, + 27;,+ 6)

T,
TI
T,

T?
r,

T, 21(0.27, + 8)

7;

5
6

-s,T,lK,
-3.53/K,
2.5 wm..lKp

7;,
?;~T,l(r, +2U

4m+
6)
l.?T,-4(T,+8)
0.27,

processes is

i? 5Tz.

(25)

The process is stable, and has one significant


time constant Tl. A PI compensation is used.
Switching the derivative Td off when T,/ T2 > 5 is
quasi-optimal
for a second-order
process
(Pomerleau,
1994) according to the ITAE
criterion (integral time average error):
ITAE =

,E(T), z dr.

I0

(26)

Auto-tuning
Specifications.
l

The desired closed-loop


response
should
behave as a first-order system with delay (see
(23)), with similar dynamics to the open-loop
response
T, = Tl in the ideal case. The
specification
can be achieved for purely
first-order processes Tz = 0.

The phase margin for overdamped


is given by

processes

and adaptive PID

75

open loop response


T, = Tl (Morari
and
Zahriou, 1989) is used. The IMC specification
is
GIMc(s)G,,,(s) =

(- Tos)e-
+ T,s)
(31)

(1 + T$)(l

The phase margin for overdamped


given by

processes is

180
A@ = 180 - 7r

AQ,= 180 - y[;+tani(T,+;+T,)


X{E+tan-l[

s
T,+a+T,

+tar-

(27)

It is minimum (A@mi) when T2 = f T,, and


decreases with increasing delay 6. For processes
without delay, the minimum phase margin is
Alvin = 80, and decreases to A@,i = 58 for
processes with delay S = &.

(& + 2T,)(T, + 2To + 6)


To

(t,+2To+S)

1
(32)

For processes with To = & and no delay the


minimum
phase margin is A<P,i = 65. It
decreases to A@min= 56 for processes with
T,=T,andS=T,.

Case 3
K,(l + Td)e-&
G,(s) = (1 + T,s)(l + T,s)

Case 5
To > O.

t2*)

The process is stable, with a stable zero. A PID


with filter compensation is used.
Specijications.
l

The desired closed-loop


response
should
behave as a first-order system with delay (see
(23)), and with similar dynamics to the
open-loop response, T, = Tl.

The phase margin for overdamped


given by
60 = 180 - F

processes is

(;+&).

(29)

It decreases
with increasing delay S. For
processes without delay, the minimum phase
margin is A@i = 90, and decreases to A@min=
60 for processes with delay S = Tl.

- lK,epS
G(s) = (1 + T,s)(l + T2s)
O.l6T,i(T,+

S)10.3T,.

(33)

The process is unstable, with a dominant


unstable time constant. A PI compensation is
used. The tuning method is based on the
calculation of the frequency where the phase of
G,(s)G,,,(s) is maximum and on a phase-margin
specification AQ, = 25 (Poulin and Pomerleau,
1995). The frequency
where the phase of
is
maximum
is chosen to be
G&)G&),
Q(w),
the open-loop crossover frequency w,. Knowing
the expressions for (P(o), o,, and the phase
margin specification, K can be calculated. K, is
then adjusted to have the desired w,. It should
be noted that if T2 + 6 > 0.3Tl, the phase-margin
specification cannot be respected,
since the
phase of G,(s) is always smaller than -155. A
PID must then be used, and this case is not
discussed here. The expression for w,, is

Case 4

K,(l - Tos)ewSs
GP(s) = (1 + T,s)(l + T2s)

(34)
To *

(30)

The process is stable, with an unstable zero. A


PID with filter compensation is used.

Case 6

Specifications.
l

The adaptation of the internal model controller (IMC) design with dynamics similar to

The process is unstable with a largely dominant


time constant & and a delay small compared

E. Poulin et al.

76

with T,. In order to limit the controller gain K,,


which reaches, very large values when T2 + 6 <<
T,, the specification A@ = 25 is dropped when
Tz + 6 10*16T,.
The values of T = T, and
KcKp = -3.53
calculated from Case 5 when
Tz + S = 0.16T, are fixed. The phase margin
increases from 25 to 90 as T, + 6 decreases.
Case 7
K,e

G,(s)= ~(1

+ T,s)

(36)

The process has an integrator, a dominant time


constant and a delay. A PID with filter
compensation
is used. The tuning method is
similar to that used for unstable processes. The
derivative time constant Td is used to cancel T,
and the filter time constant is set to 7; = 0.2T,.
The frequency
w,,,
where the phase of
G,(s)G,,(s) is maximum is calculated, and 7; is
adjusted to give a phase margin of 65. The gain
K, is then adjusted to give a maximum peak
resonance Mp of 1 dB (Poulin and Pomerleau,
1995). The expression for wmilXis
%l,X =

7;(6 + 0.2T,)

initial closed-loop bandwidth wb (used at the


auto-tuning step only). The sampling frequency
wS2, the digital filter cutoff frequencies and the
step length (in the case of open-loop autotuning) are selected according to wb. The delay
is provided to the identification
algorithm.
During the second step of the auto-tuning, the
process is excited by step changes (in both types
of auto-tuning). The identification algorithm is
initialized with a high trace value and no
constraint. As the auto-tuning progresses, the
trace is slowly decreased to facilitate convergence. In the last step of auto-tuning,
the
parameters of the supervisory shell are adjusted
according to the identified model, the process
dynamics and the noise level.

6. TESTS

AND

COMPARISONS

The AAC auto-tuning and adaptive control


performances are evaluated and compared with
three commercial
adaptive PID controllers:
Fisher DPR 910, Foxboro 760C (Exact) and
Leeds & Northrup Electromax V. The adaptive
control and auto-tuning
principles of these
controllers are described briefly.
Fisher DPR 910

5. AAC AUTO-TUNING

The AAC software offers two types of


auto-tuning. The first is performed in open loop
for stable processes, and the second in closed
loop for processes with an integrator
and
unstable processes. In open loop, the process is
excited by manipulated variable steps around the
operating point. In closed loop, the process
output is kept in a narrow band around the
operating point by the controller (the controller
does not need initial values of PID parameters).
The user has to preclassify the process in one
of the three classes: stable, unstable, or with an
integrator.
Whether
the process is stable
(open-loop auto-tuning), unstable or with an
integrator (closed-loop auto-tuning), the magnitude of the step changes or the width of the
narrow band within which the process must be
kept must be specified. The other parameters are
adjusted by the controller. It should be noted
that the user has the possibility to modify these
parameters.
The first step of the auto-tuning consists in
measurement of the noise band and estimation
of the delay and the open-loop dynamics. The
noise band is used to detect the effective delay
and to adjust some parameters of the supervisory
shell in the last step of the auto-tuning. The
open-loop dynamics is used to determine the

In the adaptive mode, the controller tracks the


point on the Nyquist curve where the phase of
the process is - 180 by bandpass-filtering the
process input-output.
A combination of phase
and gain margin is used to tune the controller
(Hagglund and Astrbm, 1991). The auto-tuning
function is based on the relay feedback principle
that can be seen as ultimate cycle with limited
amplitude.
The ultimate gain K, and the
ultimate frequency w, are used to tune the
controller according to modified Ziegler-Nichols
rules.
Foxboro 760C

The controller adaptation mechanism is based


on the analysis of the transient response of the
closed-loop system to set point changes or load
disturbances (Kraus and Myron, 1984). Parameters are tuned by an expert system according
to overshoot specifications and damping of the
transient response. The auto-tuning evaluates
the effective dead time, the process gain and
sensitivity from a step response, and initially
tunes the controller.
Leeds & Northrup Electromax

The controller
recursively
identifies
the
parameters of a second-order model with the
instrumental variable algorithm. The PID para-

Auto-tuning

77

and adaptive PID


.(

I.

::
70 -

..
:

ti5-

t
:

..

_ . .._
. .._.______..~~

____._. . . . . --?

__,___,_____
,._. _...~-.,-..~-,------,::

20

40

60

SO

TIME

100

120

140

MO

20

40

M)

80
TTME

CW3JNDS)

,@I

,u)

140

,a

180

(SECONDS)

Fig. 2. Set-point change response for Process A: AAC (-);


Fisher (..); Foxboro (--); Leeds & Northrup (-.).

Fig. 3. Set-point response for Process B: AAC (-); Fisher


(..); Foxboro (--); Leeds & Northrup (-.).

meters are adjusted to produce the desired


closed loop set point change response (Hoopes et
al., 1983). The user provides the desired 90%
time response of the closed-loop system. The
auto-tuning is performed in closed loop (the
controller needs initial PID parameters),
and
consists in a series of set-point changes to excite
the process.

Process D:

Test 1: evaluation of the auto-tuning functions

An auto-tuning phase is performed,


followed by a set-point change (4060%)
analog processes:

and is
for five

-2s

+ 4s) ;

(38)

Process B:
(1 + 20s)eW2
Gr(s) = (1 + lOs)(l + 4s) ;

(39)

(1 - lOs)e-a
Gp(s) = (1 + lOs)(l + 4s) ;

(40)

Process C:

Table 2. PID settings and ITAE criterion


Test 1

(41)

Process E:
G,(s) =

(42)

s(1 + 4s).

To compare the PID tunings on the same basis,


the PID parameters from the different structures
are converted to the noninteracting PID form
(43)

Process A:

Gp(s) = (1+ l&l

-1
Gp(s) = (1 - lOs)(l + 2s) ;

for Process A,

and the ITAE, criterion (26) is calculated for


each response.
Figure 2 presents the set point change
response for Process A. The PID parameters for
each controller
and the ITAE criteria are
presented in Table 2. The AAC response is
nearly optimum according to the ITAE criterion
(57% overshoot). The responses of the other
controllers are also satisfactory.
Set-point change responses for Process B are
shown in Fig. 3, and the PID parameters and
ITAE criterion are given in Table 3. The AAC
has a good response, since it uses a filter to
cancel the zero. Fisher and Foxboro have to
reduce their gain to cope with the zero, because
they do not have a filter. The result is a sluggish
Table 3. PID settings and ITAE criterion
Test 1

Controller

K,

7;

Td

Tf

ITAE

Controller

K,

?;

AAC
Fisher
Foxboro
L&N

1.12
1.67
1.47
0.89

11.23
18.8
12.0
9.2

1.90
3.03
1.8
3.6

0.49
1.00

1063
2571
1459
1560

AAC
Fisher
Foxboro
L&N

1.07
0.31
0.10
0.51

11.95
8.20
5.4
5.53

Td

2.21
1.29
0.53
12.6

for Process B,

Tf

ITAE

28.03
20.00

1853
8 763
24 275
5 531

6.. Poulin et al.

78

:./
30
15
0

50

loo
TIME

Fig. 4. Set-point

Table

4. PID

150

200

(SECONDS)

change response for Process


Fisher (..): Foxboro (--).

settings

and

iI

ITAE
Test 1

criterion

C: AAC

for

(-):

Process

C.

Controller

K,

7;

Td

7;

ITAE

AAC
Fisher
Foxboro
L&N

0.23
0.30
0.70
0.95

10.61
42.40
30.60
13.90

2.2 I
6.74
3.62
6.00

2.88
-3.00

8 403
89517
23 954
392 150

Figure 4 and Table 4 give the results for


Process C. The AAC response is satisfactory.
The Fisher and Foxboro controllers are not able
to cope with the zero. The Leeds & Northrup
response is not presented, since it is unstable.
For process D, only the AAC and Fisher
controllers have been able to succeed in the
auto-tuning phase. The results are presented in
Fig. 5 and Table 5. The response has a large
overshoot.
It is typical of unstable process
response (Poulin and Pomerleau, 1995), but a
phase margin A@ = 25 (AAC) does not lead to
an oscillatory response.
Figure 6 and Table 6 gives the results of the
AAC and Fisher controllers for Process E. The
other controllers failed the auto-tuning phase.
The AAC and Fisher controllers have approximately the same time responses (&5% of the final
value), but the Fisher response is much more
oscillatory. The AAC response has a small
overshoot that is typical of this type of process
(Poulin and Pomerleau, 1995).
Test 2: dominant time delay and high-order
system

response. The Leeds & Northrup controller has


a filter, but it is not managed by the controller.
In spite of the good initial tuning (K, = 1.
T= 14s Td= 3s and T,= 20s) provided to
initiate the auto-tuning, the final settings gives an
unsatisfactory response.

This test shows the results of the AAC


auto-tuning for a process with a dominant time
delay, (44), and a high-order process, (45).
Figure 7 presents the response for the dominant
time-delay process. Even if PID control is not
suitable for a process with large time delay, it is

70

65

20

40

&ONDS)

80

100

120

TIME
TIME

Fig. 5. Set-point

Table

5. PID

change

settings

(SECONDS)

response for Process


Fisher (.).

and

D: AAC

ITAE
Test 1

criterion

7;

~
-

Controller

K,

7;

AAC
Fisher
Foxboro
L&N

3.17
3.37

16.20
8.80

for Process

(-);

D.

Fig. 6. Set-point

Table

6. PID

change

settings

response for Process


Fisher (..).

and

ITAE
Test 1

criterion

E: AAC

for Process

ITAE

Controller

K,

7;

Td

Tr

ITAE

1591
2147

AAC
Fisher
Foxboro

0.45
0.49

14.74
8.70

3.51
2.2

0.7
-

2484
3544
-

L&N

(-):

E.

Auto-tuning

and adaptive PID

79

lOs+l]

Fig. 9. Benchmark for Tests 3 and 4.

I
0

20

40

a0
TIME

80

:&OF& 140

I
160

180

200

Fig. 7. AAC set-point change response for the process with a


dominant time delay: process output (-); controller output
(..).

possible to have a satisfactory response. In the


present case the delay is equal to the dominant
time constant, and the phase margin A@ = 45.
The response for the fourth-order
process is
shown in Fig. 8. It is satisfactory, even if the
model structure does not match the system. The
effect of the small time constants is approximated by a time delay by the controller:
-10s

Gp(S) = (1+ &I

(4.4)

+ 4s)

Gp(S) = (1+ lOS)(l + %)(l + 4s)(l+

2s).

(45)

Test 3: evaluation of the adaptation performances

Test 3 combines disturbance rejections, parametric variations (gain) and set-point tracking to
accelerate
controller
returning.
The analog
process is presented in Fig. 9. There are step
disturbances
acting
at different
locations.
Disturbance
A is completely filtered by the

process dynamics, and Disturbance B is only


filtered by the 4s lag. The gain variations are
also filtered by the 4 s lag. Table 7 summarizes
the events of Test 3. It should be noted that
disturbance A is amplified by the process gain.
The test is first done with a PID with fixed
parameters to show that if the controller is not
adaptive, the system will progressively be less
stable, producing an unsatisfactory
response.
The response is presented
in Fig. 10. An
adaptive controller has to be used to improve
performances. Figure 11 shows the results of the
AAC. The system is stable, and the responses to
the different events are quite constant, even with
increasing gain. At each of the second set-point
changes (t = 200 s, 440 s and 680 s), the controller is completely adapted and accurately
tuned. The load-disturbance response is a little
slow compared with the set-point-change
response. This can be improved by the two-degreeof-freedom possibility of the AAC. But decreasing the phase margin to produce
faster
load-disturbance
response reduces the robustness and stability of the system. This also leads
to large actions of the manipulated variable. The
response to slow disturbances would be satisfactory, but the response to fast disturbances would
be oscillatory.
Figure 12 presents the Fisher results. The
controller had some problems during the gain
changes, notably at t = 80 s. The first set-pointchange response (t = 140 s) is unsatisfactory,
since the controller is not yet adapted. Foxboro
results are given in Fig. 13. Foxboro performed
well, but has an oscillatory response that is
Table 7. Events for Test 3
Time (s)
20
80
140
200

0I

20

&ONDS)
80

40

100

I
120

TIME

Fig. 8. AAC set-point change response for the fourth-order


process: process output (-) controller output (..).

z
380
440
500
560
620
680
740

Event
Auto-tuning phase
Disturbance B: -10%
Gain variation: l-2
Set-point change: 40-50%
Set-point change: 50-40%
Disturbance B: +lO%
Gain variation: 2-3
Set-point change: 40-50%
Set-point change: 50-4096
Disturbance A: -10%
Gain variation: 3 to 4
Set-point change: 40-50%
Set-point change: 50-40%
Disturbance A: +lO%

l2. Poulin et al.

80

,,--.
0

TIME

Fig.

10. Fixed-mode
output

I
loo

200

(SECONDS)

ml

400
TIME

controller
response-Test
(-): controller
output (..).

3: process

Fig.

12. Fisher

Fig.

13. Foxhoro

500

600

700

800

(SECONDS)

response-Test
3: process
controller
output (..).

output

(-):

output

(--):

required by its adaptation


mechanism. The
controller gain often reaches high values. This
gives large and noisy actions of the manipulated
variable,
but fast response
to slow load
disturbances. Figure 14 presents the Leeds &
Northrup results. The controller has satisfactory
responses in the first part of the test (t = O300 s), but the controller
slowly becomes
mistuned, and the system is unstable at the end
of the test.
Test 4: adaptation during a sequence of load
disturbances

The adaptation mechanism is much slower


than the auto-tuning mechanism. In fact, during
the auto-tuning phase, the process is submitted
to a deterministic excitation. It is thus easy for
the tuner to extract the required characteristics
of the process. In the process industry many
controllers are used in regulation. The closed
loop is mostly excited by noise, load disturbances
and parametric variations. The adaptation must
take place only when good information is
present, i.e. during parametric variation and

80,

0'
0

I-

100

ZM)

300
TIME

Fig.

Il.

AAC

400

500

._

Mx)

700

after the transients of load disturbances. Without


deterministic excitation (dither signal, set-point
changes etc.), the adaptation must be slow to
prevent model degradation.
Test 4 shows the adaptation performances of
the AAC during a load-disturbance
sequence.
Sensor noise is also present. The same analog

,._.

800

I
100

(SECONDS)

response-Test
3: process
controller
output (..).

response-Test
3: process
controller
output (.).

200

1W

(-):

Fig. 14. Leeds

5W

703

8Cnl

& Northrup
response-Test
3: process
(-); controller
output (..).

output

TIME

output

400

60l

(SECONDS)

Auto-tuning

and adaptive PID

Table 8. Events for Test 4


Time (s)

iz

120
180
240
300
360
440

500
580
640

01
0

Fig.

100

15. AAC

200

81
REFERENCES

Event
Auto-tuning phase
Disturbance B: -10%
Gain variation: 1-2
Disturbance B: +lO%
Disturbance B: -10%
Gain variation: 2-3
Disturbance B: +lO%
Disturbance A: -10%
Gain variation: 3-4
Disturbance A: +lO%
Disturbance B: -10%
Disturbance B: +10%

TIE(SECo%)
so0

response-Test
4: process
controller output (..).

600

output

I
700

(-);

process as that used in Test 3 is used (see Fig. 9).


Table 8 summarizes the events of Test 4, and
Fig. 15 presents the result of the AAC. The
response is quite constant, and the action is
smooth and clean. It is important to note that
Disturbance A (360s and 500s) is amplified by
the process gain (3 and 4).
7. CONCLUSIONS

The development
and evaluation
of an
auto-tuning and adaptive controller (AAC) has
been presented. The performance of the AAC
relies heavily on three elements: efficient data
filtering, supervision of the identification procedure and proper tuning rules. The data filtering
is the basis for good identification.
The
supervisory shell for the identification procedure
is essential to provide a suitable model to the
tuning system. Proper
tuning rules permit
control of a wide range of processes with a PID
controller. The AAC auto-tuning function and
adaptive performances have been compared with
commercial PID controllers (Fisher DPR 910,
Foxboro 760C and Leeds & Northrup Electromax V), and have generally given better results.

Albertos, P., F. Morant, J. Salt and M. Martinez (1989).


Practical Issues in the Implementation
of Adaptive
Controllers. In Proc. ZFAC workshop on Eualuatio~ of
AdarNive Control Strateaies. Tbilisi. DD. 69-74.
Astrom, K. J. (1987). Adaptive Feedback Control. Proc.
IEEE, 75,185-217.
Astrom, K. J. and B. Wittenmark (1989). Adaptive Control.
Addison-Wesley, Reading, MA.
Astrom, K. J., T. I-RQglund, C. C. Hang and W. K. Ho
(1993). Automatic
tuning and adaptation
for PID
controllers-a
survey. Control Engng Practice, 1,699-714.
Bartlomiej, B. and L. Andrzej (1988). The adaptive robust
PID controller. In Proc. 8th IFAC Symp. on Identijcation
and System Parameter Estimhtion, Beijing, Vol. 1, pp.
227-232.
Besharati Rad, A. and P. J. Gawthrop (1991). Explicit self
tuning control for systems with unknown time delay. In
Proc. IFAC Symp. on Intelligent Tuning and Adaptive
Control, Singapore, pp. 251-257.
Cao. R. and T. McAvov (1990). Evaluation of a oattern
recognition adaptive PID controller. Automatiia, 26,
797-801.
Chan, C. W. (1991). Self-tuning controllers-a
review. Hong
Kong Engng. 19,26-29.
Ph. (1989). Adaptive
PID regulators.
De Larminat,
Ambitions and limitations. In Proc. IFAC Workshop on
Robust Adaptive Control, Newcastle, pp. 105-110.
Desbiens, A., D. Hodouin, K. Najim and F. Flament (1994).
Long-range predictive control of a rougher flotation unit.
Minerals Engng, 7, 21-37.
Dumont, G. A., J. M. Martin-Sanchez and C. C. Zervos
(1989). Comparison of an auto-tuned PID regulator and an
adaptive predictive control system on an industrial bleach
plant. Automatica, 25, 33-40.
I-B&hind, T. and K. J. Astrom (1991). Industrial adaptive
controllers
based on frequency response techniques.
Automatica, 27, 599-609.
Hang, C. C. and K. K. Sin (1991). A comparative
performance study of PID auto-tuners. IEEE Control
Syst. Msg., 11,41-47.
Hoopes, H. S., W. M. Hawk and R. C. Lewis (1983). A
self-tuning controller. ISA Trans., 22,49-58.
Isermann, R. (1980). Practical aspects of process identification. Automatica, 16, 575-587.
Isermann, R., Th. Knapp and K.-H. Peter (1991). Adaptive
control with continuous
and discrete time process
models-a
comparison. In Proc. 9th IFAC Symp. on
Identification and System Parameter Estimation, Budapest,
Vol. 1, pp. 423-428.
Keyes, M. A. and A. Kaya (1989). Evolution of adaptive
control algorithms and products: a critical review and
evaluation. In Proc. IFAC Workshop on Evaluation of
Adaptive Control Strategies in Industrial Applications,
Tbilisi, pp. l-8.
Kraus, T. W. and T. J. Myron (1984). Self-tuning controller
uses pattern recognition approach. Control Engng, June,
pp. 106-111.
Lambert, E. P. (1987). Process control applications of
long-range- prediction.
DPhil thesis, University of Oxford.
_
Ljung, L. (1987). System Identification-Theory-for the User.
Prentice-Hall. Englewood Cliffs. NJ.
McIntosh, A. R., b. G. Fisher and S. L. Shah (1992).
Performance adaptive control-general
structure and a
case study. J. Proc. Cont., 2,213-221.
Matko, D. (1993). Adaptive or robust control? Robust
adaptive control! Elekt&tehnirki Vesmik, 60,158-165.
Minter. B. J. and D. G. Fisher (1988). A comoarison of
adaptive controllers: academic . vs industrial. - In Proc.
American Control Conf, Atlanta, GA, Vol. 2, pp.
1653-1658.
Mohtadi, C. (1988). On the role of prefiltering in parameter
estimation and control. In Proc. _Workshop on Adaptive
Control Strategies for Industrial Use, Lodge Kananaskis,
Alberta, Canada, pp. 261-282.

82

fi. Poulin et al.

Morari, M. and E. Zatiriou (1989). Robust Process Control.


Prentice-Hall. Enelewood Cliffs. NJ.
Nachtigal, C. L. $986). Adaptive controller performance
evaluation: Foxboro EXACT and ASEA Novatune. In
Proc.
American
Control
Conj,
Seattle, WA, pp.
1428-1433.
Pomerleau, A. (1994). Reglage des regulateurs PID par une
mtthode Mquentielle-Partie
1: rtgles get&ales; Partie
2: applications. Report GRAIIM 94-31, prepared for
Mineral Research Center, Quebec Government, pp. l-64.
Pomerleau, A. and D. Hodouin (1994). Analog I/O design
for computer applications.
Report GRAIIM
94-06.
prepared for Mineral Research Center, Quebec Government, pp. l-22.
Poulin, E. and A. Pomerleau (1995). PID tuning for
integrating and unstable processes. Int. J. Control,
submitted.
Poulin, I?., A. Pomerleau, A. Desbiens and D. Hodouin
(1994). Analog and digital filtering for recursive identification. Report GRAIIM 94-08, prepared for Mineral

Research Center, Quebec Government, pp. l-43.


Sebore. D. E.. T. F. Edear and D. A. Mellichamo
Pro&s Dyrkmics and ?ontrol. Wiley, New York:
Shah, S. I. and W. R. Cluett (1991). Recursive least
based estimation schemes for self-tuning control.

0989).

squares
Can. J.

Chem. Engng, 69,89-96.

Shook. D. S., C. Mohtadi and S. L. Shah (1991).


Identification for long-range predictive control. IEE Proc.
Pt D,

138,74-84.

Song, H. K., D. G. Fisher and S. L. Shah (1984).


Experimental evaluation of a robust self-tuning PID
controller. Gun. J. Chem. Engng, 62,755-763.
Sorrels, J. (1989). Comparison of contemporary adaptive
control design techniques. In Proc. 21st Southeastern
Symp. on System Theory, Tallahassee, FI, pp. 636-640.
Sripada, N. R. and D. G. Fisher (1987). Improved least
squares identification. Int. J. Control, 46, 1889-1913.
Wahlberg, B. and L. Ljung (1986). Design variables for bias
distribution in transfer function estimation. IEEE Trans.
Aurom. Conrrol, AC-31, 134-144.