Copyright © IFAC Adaptive Systems in Control

and Signal Processing, Grenoble, France, 1992

AUTOMATIC TUNING AND ADAPTATION FOR PID

CONTROLLERS-A SURVEY
K.J. Astrom·, T. Hagglund·, C.C. Hang·· and W.K. Ho"
*Department of Automatic Control, Lund Institute of Technology, Lund, Sweden
**Department of Electrical Engifll!ering, National University of Singapore, Singapore

Ab.dract Adaptive techniques such as gain scheduling, automatic tuning and continuous
adaptation have been used in industrial single-loop controllers for about ten years. This
paper gives a survey of the different adaptive techniques, the underlying process models
and control designs. The paper ends with an overview of industrial adaptive single-loop
controllers.

Ke1lword6 Adaptive control, automatic tuning, gain scheduling, PID control.

1. Introduction to look at the future. To make such an evaluation is

the purpose of this paper.
Single-loop controllers have been used in industry for
a long time. For example, it was about fifty years
ago that derivative action was introduced in pneu- 2. Adaptive Techniques
matic controllers by Taylor Instruments. Industrial
single-loop controllers have gone through an interest- The first general purpose adaptive controller for in-
ing change in technology, from pneumatic via ana- dustrial process control was introduced around 1983.
log electronics to microprocessor implementation . In There are now lots of adaptive controllers for indus-
spite of these changes in technology, the functional- trial process control and about 100 000 loops being
ity of the controllers has remained the same, namely controlled by such controllers. As a result of this there
essentially an implementation of the standard PID al- is also a growing experience of using adaptive tech-
gorithm. The controllers are currently going through niques. Unfortunately there is also some confusion in
an interesting phase of development where features terminology. Therefore we will introduce an appropri-
like auto-tuning, gain scheduling and adaptation are ate terminology at the same time as we describe some
added. There is hardly any new product announced of the adaptive techniques that have been useful.
that does not incorporate some of these features . The
single-loop controller is therefore rapidly becoming a Automatic tuning
test bench for control theory. Since the controllers are
By automatic tuning (or auto-tuning) we mean a
made and used in large number, much industrial ex-
method where a controller is tuned automatically on
perience of using sophisticated control is also accumu-
demand from a user. Typically the user will either
lated. One reason for this development is the advances
push a button or send a command to the controller.
in microelectronics which give cheap microprocessors
Industrial experience has clearly indicated that this
whose computational power is continuously increasing.
is a highly desirable and useful feature. Automatic
Another reason is the pressure from users and appli-
tuning is sometimes called one-shot tuning.
cations and a third is the growing experience of using
Automatic tuning can also be performed using
advanced control.
external equipment. These products are connected to
It may be asked why a similar development is the control loop only during the tuning phase. When
not taking place in distributed control systems. One the tuning experiment is finished, the products sug-
reason is that the distributed control systems are gest controller parameters. Since these products are
quite complicated and that their development cycle supposed to work together with controllers from differ-
is quite long. Therefore it takes a long time for new ent manufacturers, they must be provided with quite a
features to be introduced in such systems. There are lot of information about the controller structure in or-
however indications that the experiences from single- der to give an appropriate parameter suggestion. Such
loop controllers are migrating into larger systems. information is e.g. controller structure (series or par-
It is therefore appropriate to take stock of the allel form), sampling rate, filter time constants, and
development, to assess the current state of the art and units of the different controller parameters (gain or

371
proportional band, minutes or seconds, time or re-
peats/time).

Gain scheduling
By gain scheduling we mean a system where controller
parameters are changed depending on measured oper- y
ating conditions. The scheduling variable can e.g. be
measurement signal, controller output or an external
signal. For historical reasons the word gain scheduling
is used even if other parameters like derivative time or
integral time are changed. Gain scheduling is a very
effective way of controlling systems whose dynamics Figure 1. Block diagram of an indirect adap-
change with the operating conditions. Gain schedul- tive system
ing has however not been used much because of the ef-
fort required to implement and use it. When combined
ate disturbance. Adaptation is therefore particularly
with auto-tuning, gain scheduling is however very easy
useful for feedforward.
to use.

Direct and indirect methods

Adaptive control
The techniques used for auto-tuning and adaptation
By adaptive control we mean a controller whose pa-
are very similar. In broad terms we can distinguish be-
rameters are continuously adjusted to accommodate
tween direct and indirect methods. In a direct method
changes in process dynamics and disturbances. Adap-
controller parameters are adjusted directly from data
tation can be applied both to feedback and feedfor-
in closed loop operation. In indirect methods a model
ward parameters. It has proven particularly useful for
of the process is first developed from on-line data. The
feedforward control. The reason for this is that feed-
controller parameters are then determined from this
forward requires good models. Adaptation is therefore
model.
almost a prerequisite for using feedforward.
There is a large number of methods available both
for direct and indirect methods. They can conveniently
Use of adaptive techniques be described in terms of the methods used for model-
The notation adaptive techniques will be used to ing and control design.
cover auto-tuning, gain scheduling and adaptation. In the direct methods the key issues are to find
Although research on adaptive techniques has almost suitable features that characterize relevant properties
exclusively focused on adaptation, experience has of the closed loop system and appropriate ways of
shown that auto-tuning and gain scheduling have a changing the controller parameters so that the desired
much wider industrial interest . properties are obtained.
The first issue to consider is controller perfor- The indirect systems can all be represented by
mance. If the requirements are modest, a controller the block diagram in Figure 1. There is a parameter
with constant parameters and conservative tuning can estimator that determines the parameters of the model
be used. With higher demands on performance other based on observations of process inputs and outputs.
solutions should be considered . If process dynamics There is also a design block that computes controller
are constant, a controller with constant parameters parameters from the model parameters. If the system
should be used. The parameters of the controller can is operated as a tuner the process is excited by an
be obtained using auto-tuning. If process dynamics or input signal. The parameters can either be estimated
the nature of the disturbances are changing it is use- recursively or in batch mode. Controller parameters
ful to compensate for these changes by changing the are computed and the controller is run. If the system
controller. If the variations can be predicted from mea- is operated as an adaptive controller parameters are
sured signals, gain scheduling should be used. Typical computed recursively and controller parameters are
examples are variations caused by nonlinearities in the updated when new parameter values are obtained.
control loop. Auto-tuning can be used to build up the
gain schedules. There are also cases where the varia-
tions in process dynamics are not predictable. Typical
3. Modeling
examples are changes due to nonmeasurable variations
in raw material, wear, fouling etc. These variations
A model of a system can be any type of abstract
cannot be handled by gain scheduling but must be
description that captures useful relevant features of
dealt with by adaptation. An auto-tuning procedure
a process. Modeling can mean many different things,
is often used to initialize the adaptive controller. It is
from extraction of some simple features of a transient
then sometimes called pre-tuning or initial tuning.
response to development of a traditional control model
Feedforward control deserves special mentioning.
in terms of a transfer function. We will now describe
It is a very powerful method for dealing with measur-
some models that are used in adaptive PID controllers.
able disturbances. Use of feed forward requires how-
ever good models of process dynamics. It is difficult
Time domain models
to tune feedforward loops automatically on demand,
since the operator often cannot manipulate the distur- Typical time domain features are static gain, dom-
bance used for the feedforward. To tune the feed for- inant time constant and dominant dead time. They
ward it is therefore necessary to wait for an appropri- can all be determined from the step response of the

372
a)
r- ...IT
Kpr---------------~~__--------- -- ~ Process r--r-
"'-+ PlO U
a{ ,
-1
W "V
L T
Figure S. The relay autotuner. In the tuning
mode the process is connected to relay feedback
b)
There are many methods available to determine
Kp the parameters in equation (4), e.g. the recursive least
squares method.

Frequency domain models

Typical frequency domain characteristics are ampli-
tude margin, phase margin, Mp-value, ultimate gain,
ultimate period. These quantities are all related to
Figure 2. Determining a first order plus dead- simple properties of the Nyquist curves of the open
time model from a step response loop transfer function or of the loop transfer function.

process. See Figure 2a. Static gain K p , dominant time Relay ezcitation
constant T and dominant deadtime L can be used to Frequency domain characteristics can be deter-
obtain a first order plus dead time model for the pro- mined from experiments with relay feedback in the
cess as given in equation (1). following way. When the controller is to be tuned, a
relay with hysteresis is introduced in the loop, and the
PID controller is temporarily disconnected. See figure
(1)
3. For large classes of processes, relay feedback gives
an oscillation with period close to the ultimate fre-
The construction in Figure 2a is sensitive because it quency Wl80. The gain of the transfer function at this
relies on drawing a tangent to the step response. An- frequency is also easy to obtain from amplitude mea-
other method is to calculate L as the time interval surements. Details and conditions are given in [Astrom
between the step change and the moment the mea- and Hiigglund, 1984, Astrom and Hiigglund, 1988).
surement signal crosses a certain small level. The time
constant T is then calculated as the time to reach 63% COf'f'elation Method
of the final value. This method is also sensitive, since In this method, a small pseudo random binary
the calculations are based on a few values of the step sequence (PRBS) test signal u(t) is injected and the
response only. resultant process output y(t) is logged. The cross
Another method, which is based on determination correlation, t/> ... (r), between u(t) and y(t) is then used
of areas can be used. See Figure 2b. In this method, to compute the impulse response g( r) of the process
gain Kp is first determined from the steady-state value [Hang and Sin, 1991) as follows
of the step response. Area Ao is then determined. The
average residence time of the system is then

(2)
g(r) = A!h (N: J (t/>",(T) + ~ t/>".. (k») (5)

Area A l , which is the area under the step response up where A is the amplitude of the PRBS signal, h is
to time L + T, is then determined and T is given by the sampling interval and Nh is the period of the
PRBS signal. The impulse response computed is then
(3) numerically transformed into its frequency response
from which the ultimate gain, ultimate period, static
This method is less sensitive to high frequency dis- gain and normalized dead time of the process can be
turbances because it is based on area determination. determined.
More details can be found in [Astrom and Hiigglund,
1988) and [Nishikawa et al., 1984).
A discrete time model as given in equation (4)
can also be used to describe the process. 4. Control Design
The PID control law can be implemented in either the
(4)
parallel or the series form. The parallel form is given

373
below. drawn. The intersections between this tangent and the
coordinate axes give the two parameters a and L. A
simple model of the process which can be derived from
these parameters is given by the transfer function

e = Yr - Y (6) a -.L
Gplf
() = - e (8)
ilL

Ziegler and Nichols have given PID parameters di-

rectly as functions of a and L, see [Ziegler and Nichols,
The controller output, process output and set-point
1942]. The formulas will roughly give quarter ampli-
are u, Y and Yr, respectively. A weighting factor /3
tude damping ratio for load disturbance response for
can be placed on the set-point. The factor /3 when
processes that can be modeled by equation (8). In
set to less than one can reduce the set-point response
practice some form of fine tuning has to be done.
overshoot without affecting the load disturbance re-
There are several design methods which are simi-
sponse. In a conventional PID controller with /3 = 1,
lar to the Ziegler-Nichols step response method in the
tuning to give good load disturbances responses will
sense that they are based on a step response experi-
often give set-point responses that are too oscillatory.
ment combined with a table that relates the controller
On the other hand, tuning to give good set point re-
parameters to the characteristics of the step response.
sponse often gives sluggish load disturbance response.
The most common method is the Cohen-Coon method
The factor /3 when set to less than one can reduce
[Cohen and Coon, 1953].
the set-point response overshoot for tuning that op-
timizes the load disturbance response. In some lit- Optimization Techniquelf
erature, this form of the PID control law is known There are other tuning formulas derived to op-
as a two-degree-of-freedom PID controller [Shigemasa timize certain criteria such as the integrated abso-
et al., 1987], [Takatsu et al., 1991]. Several industrial lute error (IAE), the integrated time absolute error
controllers have /3 = o. (ITAE), the integrated square error (ISE) etc. They
The series form of the PID control law is given by are mostly derived for the first-order plus dead-time
model as given in equation (1). A word of caution is in
order - formulas derived for the first order plus dead-
time model may not give good results for higher order
systems.
It is not possible to obtain complex zeros from this
form of the controller. However, it is often claimed Pole Placement
that the series form is easier to tune by trial and error If the process is described by a low-order transfer
procedures. function, a complete pole-placement design can be
This section describes some methods for deter- performed. Consider the process described by the
mining the parameters of a PID controller. They can second-order model
be classified broadly into direct and indirect meth-
ods. The direct control design techniques are simply
(9)
prescriptions that tell how the controller parameters
should be changed in order to obtain the desired fea-
tures. The indirect control design methods give con- This model has three parameters. By using a PID con-
troller parameters in terms of model parameters. troller, which also has three parameters, it is possible
to arbitrarily place the three poles of the closed-loop
Direct Methods system. A suitable closed-loop characteristic equation
of a third-order system is
A vast majority of the PID controllers in the indus-
tries are tuned manually by control engineers and op-
erators . The tuning is done based on past experiences
and heuristics. By observing the pattern of the closed-
loop response to a set-point change, the operators uses which contains two dominant poles with relative
heuristics to directly adjust the controller parameters. damping (() and frequency (w), and a real pole located
Much of these heuristics have been captured and put in -o:w . Straight-forward calculations show that this
into commercial products . Most products wait for set characteristic equation is obtained with the following
point changes or major load disturbances . When these controller parameters;
occur, properties like damping, overshoot, period of
oscillation and static gain are estimated . Based on K = TS2w2(1 + 2(0:) - 1
these properties, rules for changing the controller pa- Kp
rameters to meet desired specifications are executed.
Ti = TiT2W2(1 + 2(0:) - 1 (11)
T i T2 0:w 3
Indirect Methods
Td = Ti T2 w(0: + 2() - Ti - T2
Zieglef'-NichoI6 6tep f'e6pOn6e method TiT2(1 + 2(0:)w 2 - 1
This method is based on a registration of the
open-loop step response of the system, which is char- Notice that pure PI control is obtained for
acterized by two parameters (see Figure 3a). The point
where the slope of the step response has its maximum
(12)
is first determined, and the tangent at this point is

374
 Table 1. Examples of industrial adaptive process controllers.

Manufacturer Controller Automatic Gain Adaptive Adaptive

tuning scheduling feedback feedforward

Bailey Controls CLC04 Step Yes Model based

Control Techniques Expert controller Ramps Model based
Fisher Controls DPR900 Relay Yes
DPR910 Relay Yes Model based Model based
Foxboro Exact Step Rule based
Fuji CC-S:PNA 3 Steps Yes
Hartmann le Braun Protroruc P Step
Digitric P Step
Honeywell UDC 6000 Step Yes Rule based
SattControl ECA40 Relay Yes
ECA400 Relay Yes Model based Model based
Siemens SIPART DR22 Step Yes
Toaruba TOSDIC-215D PRBS Yea Model based
EC300 PRBS Yes Model based
Thrnbull Control Systems TCS 6355 Steps Model based
Yokogawa SLPC-171,271 Step Yes Rule based
SLPC-181,281 Step Yes Model based

The choice of W may be critical. The derivative time 1S known from classical control that the damping of
is negative for W < wc. The frequency Wc thus gives the system is related to the phase margin. The phase
a lower bound to the bandwidth. Also notice that the margin is given by
gain increases rapidly with w . The upper bound to the
bandwidth is given by the validity of the simplified (13)
model.
where IGp(iwg)Gc(iwg)1 1, and the amplitude
Cancellation of proceu polu margin is equal to the ultimate gain .
The PID controller has two zeros. A popular
design method is to choose these zeros so that they
cancel the two dominant process poles . The method
is popular since it is simple and gives good response 5. Overview of industrial products
to set point changes . The method will, however, often
give poor response to load disturbances. See [Astrom Commercial PID controllers with adaptive techniques
and Hiigglund, 1988]. have been available since the beginning of the eighties.
In this section, we will classify these products with
Ziegler-Nichol, Frequency Re,pon,e Method respect to their applications and adaptive techniques
This method is based on a very simple character- used.
ization of the process dynamics. The design is based We will distinguish between temperature con-
on knowledge of the point on the Nyquist curve of the trollers and process controllers. Temperature con-
process transfer function G where the Nyquist curve trollers are primarily designed for temperature control,
intersects the negative real axis. For historical reasons whereas process controllers are supposed to work in
this point is characterized by the parameters K,. and not only temperature control loops, but also in other
T. which are called the ultimate gain and the ultimate loops in the process industry such as flow, pressure,
period. These two quantities can be obtained for many level and pH control loops. For some reason, tempera-
systems by using proportional feedback and choosing ture controllers have mostly a quadratic front (96x96
sufficiently high gain until continuous cycling occurs. mm), whereas process controllers normally have a rect-
The ultimate gain K,. is then given by the propor- angular front (72x144 mm) .
tional gain and the ultimate period T ... is given by the
period of the oscillations. The parameters can also be Temperature Controllers
obtained using the relay feedback experiment or the
correlation method presented in Section 3. Most PID controllers are designed for temperature
Ziegler and Nichols derived PID parameters di- control only. These controllers are normally cheaper
rectly as functions of K... and T ... , see [Ziegler and than process controllers. Automatic tuning and adap-
Nichols, 1942]. These formulas were derived to give tation is easier to implement in temperature con-
quarter amplitude damping for load disturbance re- trollers, since most temperature control loops have
sponse. Refined Ziegler-Nichols formulas are given in several common features. This is the main reason why
[Hang et al., 1991]. automatic tuning has been introduced more rapidly in
temperature controllers.
Pha,e and Amplitude margin, In temperature control loops, a serious nonlinear-
Phase and amplitude margins design methods ity may result if the heating and cooling dynamics are
give a measure of the robustness of the system . It different. This nonlinearity should be treated by gain

375
scheduling. The temperature controllers have some- Gain scheduling effectively compensates for non-
times a facility to shift between different controller pa- linearities and other predictable variations in the pro-
rameters when the control action shifts between heat- cess dynamics. Since it is much easier to implement
ing and cooling. than other adaptive techniques, it is surprising that
several adaptive controllers lack this feature.
Process Controllers The adaptive controllers can be divided into
Since the processes that are controlled with process two categories, namely the model based and the
controllers may have large differences in their dynam- rule based. It is interesting to note that there are
ics, tuning and adaptation becomes more difficult com- many successful applications of rule based adaptive
pared to the pure temperature control loops. controllers, although almost all university research has
If a temperature controller is applied to e.g. a been focused on model based adaptation.
pressure loop, the tuning procedure may be very poor, Adaptive feedforward is rarely used in single-
since pressure control loops normally are much faster loop controllers. This is also surprising, since adaptive
and more sensitive than temperature control loops. In feed forward has proven to be very useful in other
the following presentation, we will focus on process applications of adaptive control. Adaptive feedforward
controllers. is also easier to implement than adaptive feedback
In Section 2, different adaptive techniques were control.
discussed. The techniques are: Automatic tuning, gain
scheduling and adaptive feedback and feedforward
control. In Table 1, a collection of process controllers
is presented together with information about their
adaptive techniques.
7. References
Automatic tuning is the most common adaptive
technique in the industrial products. The usefulness ASTR()M, K. J. and T. HXGGLUND (1984): "Auto-
of this technique is also obvious from Section 2. The matic tuning of simple regulators with specifica-
auto-tuning procedures are necessary to obtain a rea- tions on phase and amplitude margins." Auto-
sonably comfortable operation of the other adaptive matica, 20, pp. 645-651.
techniques, i.e. to build gain schedules and to initial- ASTR()M, K. J. and T. HXGGLUND (1988): Automatic
ize adaptive controllers. Most auto-tuning procedures TImings of PID Controllers. Instrument Society
are based on step response analysis. of America, Research Triangle Park, North Car-
The controllers listed in Table 1 belong to the olina.
most advanced category of PID controllers on the BENGTSSON, G. and B. EGARDT (1984): "Experi-
market . It is therefore surprising that gain scheduling ences with self-tuning control in the process in-
is not available in all of them, since the technique is dustry." In Preprints 9th IFAC World Congress,
much simpler to implement than automatic tuning or Budapest, Hungary.
adaptation. Gain scheduling is also found to be more
COHEN, G. H . and G. A . COON (1953): "Theoretical
useful than adaptive control in most situations .
consideration of retarded controL" Trans. ASME,
It is interesting to see that many adaptive con-
75, pp. 827-834.
trollers are rule based instead of model based. The
research on adaptive control at universities has been HANG, C. C., K. J . ASTR()M, and W . K . Ho (1991):
almost exclusively focused on model based adaptive "Refinements of the Ziegler-Nichols tuning for-
control. mula." lEE Proceedings-D, 138:2, pp. 111-118.
Adaptive feedforward control is seldom provided HANG, C. C. and K. S. SIN (1991): "On-line auto-
in the industrial controllers. This is surprising, since tuning of PID controllers based on the cross-
adaptive feed forward control is known to be of great correlation technique." IEEE Transactions on In-
value. See e.g. [Bengtsson and Egardt, 1984]. Further- dustrial Electronics, December. To be published.
more, it is easier to develop robust adaptive feedfor- NISHIKAWA, Y., N. SANNOMIYA, T. OHTA, and
ward control than adaptive feedback control. H. TANAKA (1984): "A method for auto-tuning
of PID control parameters." Automatica, 20,
pp. 321-332.
6. Conclusions
SHIGEMASA , T., Y. lINO , and M. KANDA (1987):
"Two degrees of freedom PID auto-tuning con-
Various adaptive techniques have been implemented
troller." In Proceedings of ISA Annual Confer-
in single-loop PID controllers during the last decade.
ence, pp. 703-711.
In this paper, a summary of these different techniques
has been given together with an overview of industrial TAKATSU, H., T. KAWANO, and K. KITANO
products. (1991): "Intelligent self-tuning PID controller."
Almost all controllers that use adaptive tech- In Preprints IFAC International Symposium on
niques have some form of automatic tuning function. Intelligent Tuning and Adaptive Control, Singa-
The automatic tuning function is not only useful to pore.
help the operator in finding suitable controller param- ZIEGLER, J. G . and N. B. NICHOLS (1942): "Opti-
eters. It is also useful to build gain schedules and to mum settings for automatic controllers." Trans.
initialize adaptive controllers. ASME, 64, pp. 759-768.

376