You are on page 1of 5

Automatica, Vo|. 25, No. 2, pp.

273-277, 1989

0005-1098189 $3.00+ 0.O0


Pergamon Pressplc
~) 1989InternationalFederationof AutomaticControl

Printedin GreatBritain.

Brief Paper

Application of Identification-free Algorithms


for Adaptive Control*
J. MAR~IKt and V. STREJCt
Key Words--Automatic control; adaptive control; damping; digital computer applications; digital
systems; computer control; control applications; control system design; heuristic control;
identification-free adaptive control.

simpler algorithms needing less arithmetical operations and


returns the same quality of control.
When applying an adaptive control of any kind, it is always
necessary to take advantage of professional inference
supported by theoretical concepts rather than to rely
exclusively either on experience or on theory.
The heuristic approaches that we have in mind combine
experience and theory. They have a certain theoretical
background but they use it in an unconventional way.
Perhaps, the term "heuristic" is not adequate. It would be
better to speak about unconventional simple control
algorithms rather than heuristic ones. The very simple
algorithms described in the following sections need only a
few arithmetical operations and do not need identification of
process parameters at all, avoiding in this way the most
difficult problem appearing in self-tuning regulator systems.
The main idea was motivated by the following
consideration. When adjusting any controller manually one
does not think about the poles and zeros of the transfer
function and one does not evaluate the quadratic cost
function of the control loop, but the adjuster of the
controller rather interprets the shape of the control process,
mainly its oscillatory behaviour. Since strong oscillating
processes are mostly inadmissible, reasonable damping is the
decision criterion. It was on this experience that the first
adaptive controllers were also based. Unfortunately, they
needed additional exciting test signals in order to make an
evaluation of the damping index.
There are various ways of overcoming this imperfection. In
the following paragraphs, several modifications of the
adapting algorithms are described based on an adequate
damping index or oscillation index, enabling a straightforward updating of controller parameters. The properties of
these algorithms are very similar, even though they are based
on different principles. All adaptation loops act like any
conventional control loop and can be applied, in principle, to
any plant the dynamics of which permits the application of
the usual three-term controllers, i.e. to any stable plant with
monotonous or slightly oscillatory step-response, regardless
of its order and transportation lag.

Abstract--In order to facilitate the utilization of single-chip


microcomputers for adaptive control some extremely simple
algorithms have been developed. The simplicity has been
achieved by using performance criteria based on geometrical
properties of the control error process while the adaptation
acts like an ordinary control loop. No identification, no
mathematical model of the controlled system and no special
test signals are needed.
1. Introduction

THIS PAPER is concerned with simple adaptive digital control


particularly for industrial process control applications. The
simplicity of the algorithms described for this purpose in the
following sections enable one to design single chip
microcomputers representing essentially three-term controllers which may replace in the future conventional
continuously acting and digital PID controllers.
In the last 20 y many contributions to control theory have
been elaborated which may be used for, or at least can
facilitate, the understanding and the design of adaptive
controls. The complete list of publications relevant to this
subject would amount to more than one thousand items.
In spite of all achievements the total number of necessary
numerical operations which must be performed in each
interval of sampling and the total computer memory required
prevent the design of cheap general purpose digital
controllers, which might replace the simple PID controllers
used almost exclusively in all industrial applications.
The question is whether it would not be possible to take
rather more advantage of practical experience for the
construction of control algorithms than has been done until
now. The paper by Marlik and Strejc (1985) shows three
different possibilities of this kind which w e r e called "heuristic
approaches of digital process control". In that paper
input-output orthogonalization, maximum stability margin
of controller parameters and the application of the degree of
the process oscillatory behaviour were discussed.
Actually, the application of the degree of the process
oscillatory behaviour initiated the concept of adaptive
control described in this paper. This concept may be
considered as an extension of the previous one. It yields

2. Basic structure o f algorithms

It is assumed that we can apply the PSD control law in the


incremental form where the abbreviation PSD corresponds
to proportional, summing and differencing terms. Hence

* Received 21 November 1986; revised 1 August 1987;


revised 6 January 1988; revised 25 May 1988; received in
final form 5 July 1988. The original version of this paper was
presented at the 10th IFAC World Congress which was held
in Munich, F.R.G. during July 1987. The Published
Proceedings of this IFAC meeting may be ordered from:
Pergamon Press plc, Headington Hill Hall, Oxford OX3
0BW, England. This paper was recommended for publication
in revised form by Associate Editor H. Unbehauen under the
direction of Editor H. Austin Spang III.
t Institute of Information Theory and Automation,
Czechoslovak Academy of Sciences, 182 08 Prague,
Czechoslovakia.

Au(k) = g ( k ) [ e ( k ) + r o ( k ) A e ( k ) + r ~ ( k ) A e e ( k ) ]

(1)

where u is the output variable of the controller, Au is its


increment, e, Ae and A2e are the control error and its first
and second backward differences, i.e. Ae(k) = e ( k ) - e ( k - 1)
and A2e(k) = e ( k ) - 2 e ( k - 1) + e ( k - 2), respectively, g
is the common gain factor of the controller and r0 and rt are
the parameters of the proportional and difference terms. In
adaptive control all variables and parameters g, ro and r t are
varying with respect to time t. In discrete adaptive control,
the discrete independent time variable k = H A t is intro273

274

Brief Paper

duced, where At is the period of sampling. Taking At as a


basic unit of time (At = 1) and expressing the running time in
multiples of At, then provided e is dimensionless, the
differences Ae and A2e replacing the original derivatives have
physical dimension [l/s] and [l/s2], respectively. In the
following, constants introduced in mathematical relations
must have appropriate correct dimensions.
The parameters ro and r 1 are automatically adjusted in
such a way that the mean absolute values of all terms
constituting the controlling variable Au are equal.
Consequently
le(k)l = r o ( k ) I A e ( k ) l

= rl(k)[A2e(k)l.

le(k)l

and

te(k)l

rl(k) =

IAe(k)l

(3)

le(k)[

IAe(k)l

(7)

The factor L, 1 >> L > 0, influences the speed of convergence


of the iterative procedure. For L approaching zero, the
speed of convergence is slow and vice versa.
The argument (k) denotes the step of iteration and in the
case of adaptive control it coincides with instants of
sampling.
The last relation yields
T~(k) = Z , ( k - 1)[1 - L IAe(k)l] + Z le(k)l
or

T , ( k ) = f [ T , ( k - 1)].

Due to nonlinear relation (8), the


of the iterative procedure can be
principle of contracting mapping,
theorem (cf. for example Vejvoda
(1964)):

I-'A'~'e(k) I

(8)

condition of convergence
verified by applying the
i.e. Banach's fixed point
et al. (1981), and Collatz

IT,(k) - T~(k - 1)1 ~ I f [ T ~ ( k ) ] - f [ T ~ ( k

Condition (2) is in fact a heuristic proposition. It has been


verified by numerous tests and has proved to be adequate. It
can easily be justified when considering sine-wave signals.
Three vectors e, roAe and rlA2e of the same value and
mutually perpendicular yield a resultant of the same
magnitude as e, however, having a positive phase shift of 90
and improving in this way the overall phase characteristic of
the controller. Considering as a basic controller the summing
(integrating) one, then the resultant PSD controller has the
amplitude characteristic of a summing term and the phase
characteristic of a proportional term (Fig. 1). Even though
this type of controller is considerably nonlinear its behaviour
is nearly linear.
From the relations (3) it follows that the physical
dimension of the parameter ro is [s] (time) while that of the
parameter r 1 is [sZ].
Consequently, ro corresponds to a certain variable
time-parameter of the control error
ro(k ) =

T,(k) - T,(k - 1) = L l t e ( k ) l - T , ( k - 1)IAe(kll].

(2)

According to this proposition the controller parameters ro


and r~, correspond approximatively to one half of the maximum
stability margin of the individual controller parameters, i.e.
r i =0.5r,,,~, i = 0 , 1, where r,~, represents the maximum
value on the boundary of stability.
From this proposition it follows that
r0(k) =

by an iterative procedure. We may write

- 1)]1;

~< 1

where Sd is the factor of contraction.


By introducing
T,(k) - T,(k - l) = aT,(k)

and
f[T,(k)] - f [ T , ( k

- 1)] = Af[T~(k)]

we have

~( IAT,(k)l ~ IA/[T,(k)]l.
Hence

~r~> Af[r,(k)]
The continuous version of the last expression requires

af[T,(k)]
.

~t~> ~
= r,(k).

(4)
Consequently

The parameter r~ can be expressed as a product of two


similar relations, i.e.
....

le(k)l _ le(k)l IAe(k)l


rl[K)--~
IAe(k)llA2e(k)l
(5)

= r,(k)T,.(k)

where

~ > I(1 - L IAe(k)l)l.


Evidently, it is possible to find always a value ~ < 1
satisfying the specified condition of convergence. This
condition is equivalent to the condition of stability of the
difference equation (8).
Relation (7) can be written in the simplified form
ATe(k ) = L [ l e ( k ) l - T~(k - 1)IAe(k)l ].

IAe(k)l
T~(k ) = _ _ .2 e ( k )------~l"

(6)

It is possible to simplify considerably the calculation of T,


and To when eliminating the explicit calculation of the mean
absolute valueg of lel, IAet and IA-~-e[ and their reseee~pective
division operations. Namely, the relation T, = I~]/IAel may
be solved in an implicit form as an equation
le(k)l - T~(k) IAe(k)l = 0

ATe(k) = L [ l e ( k ) l -

T~(k - 1)LAe(k)l ].

(10)

As mentioned before, the parameter L influences the


convergence of the iterations. In order to eliminate the
adjusting of this parameter with respect to IAel, let us
introduce a small constant L* and constant increments of
AT, yielding
ATe(k) = L* sign [le(k)t - T,(k - 1) IAe(k)ll

Im

(9)

In addition to this the mean values can be omitted and


only instantaneous values may be applied, because the
iteration itself has an averaging property. Hence

(11)

where 0.05<~ L* ~<0.1.


Using similar reasoning we can write

Re

e Olrl2e_
roAe
FIG. 1. Phase shift of the proposed PSD controller.

ATe(k) = L* sign [IAe(k)l - T,,(k - 1)IA2e(k)l].

(12)

In this way the calculation is simplified and speeded up.


Applying all the proposed recommendations, the controller equation takes the form
A u ( k ) = g ( k ) [ e ( k ) + T~(k ) A e ( k )
+ T~(k)To(k)A2e(k)].

(13)

Brief Paper
Using the former results (cf. Margik and Strejc (1985)) the
optimum value of the ratio F = T~,/T~ is approximatively 0.5,
which yields T, = 2T~.
Hence, the controller equation can be further modified to
the form

275

value of the ratio F = T~IT,, A= ~,, parameter ~ =0.1


verified by experiments, frequency f~, = 1/2~rT,, and index of
oscillation
F(k~
"

fe(k)
"=[~,(k)=

T~,(k)
T,(k)

(16)

Au(k) = g(k )[e(k ) + 2T~(k )Ae(k )


+ 2TZ(k)A2e(k)].

(14)

Thus, in order to obtain the control action Au, we must


measure the control error e, calculate its differences Ae and
A2e and T~ by virtue of (12). These values are also sufficient
for the determination of the value of the gain parameter g.
Of course, it is possible to have a priori estimates of the
controller parameters for the given process to be controlled
and to use them in order to accelerate the tuning procedure.
But it is not necessary. Nevertheless, it is recommended that
the adaptive control be started with the controller
parameters having small positive values in order to start with
a stable control process.

where the possible scope of F is 0 ~ F ~< 1, the approximate


optimum being F = 0.5.
Substituting these relations into (15), we obtain
0 . 1 g ( k - 1) (0.5 T,(k - 1)
Ag(k) = 2~T~(k ~
T~,(k ~ ~

l).

(17)

The gain increment Ag is selected inversely proportional to


T~ ensuring in this way small increments for slow systems and
vice versa. According to the sign of the expression in the
parentheses, Ag can increase or decrease. In particular, for
Te/Tv = 2 the increment of the controller gain Ag = 0.
A simpler relation is given as
1

3. Gain adaptation
Several algorithms were developed for gain adaptation
which may also be mutually combined. They are based on
different principles but finally they give very similar results.
Gain adaptation corresponds always to the conventional
control action of maintaining a certain variable or index at a
desired value according to a given criterion. For this aim
extremely simple control algorithms were developed. They
are based on new performance criteria utilizing certain
geometrical properties of the control error process. The
algorithms evaluate either the damping or the degree of the
oscillatory behaviour of the control process. It is to be
stressed that no special test signals or identification
procedures are required. The current control process is quite
sufficient for self-tuning.
Since the quadratic optimum yields, as a rule, oscillatory
control processes, the overall closed loop behaviour can be
approximated by a second order model, irrespective of the
plant order. Among all roots of the characteristic equation
there is a dominant pair of complex conjugate roots which
determines the shape of the transient response. Therefore,
optimal step responses of different control loops are fairly
similar, consisting of exponentials and a single damped sine
wave, with additional sine waves being negligible. Thus
adequate damping is a sufficient quality criterion to enable us
to design simple and robust adaptive controllers.
The performance indices presented in this paper serve as
detectors of the sine wave level in the control process. These
criteria are rather intuitive, nevertheless they can be
calibrated by means of sine waves with various damping
ratios. The common approach to derivation of the
identification-free algorithms for adaptive control can be
characterized by the following steps:
--finding a suitable formula for the index of the damping or
of the control process oscillatory behaviour by educated
reasoning;
----calculation of boundary values of the selected index using,
for example, sine wave and by exponential functions,
thereby defining the interval in which it is possible to select
the index value corresponding to the desired shape of the
control process;
---find out the index value according to control requirements
by simulation and experiments, verifying in this way the
applicability and usefulness of the proposed index of
control.
3.1. Application o f the index o f the control process
oscillatory behaviour. The approach established using the
index of the control process oscillatory behaviour was
described in detail in the publication by Margik and Strejc
(1985). The increment of gain was derived in the form
Ag(k) = ),g(k - 1 )/~ - ( ~~'~f
_ 1)

1)_

(15)

where Frcf=0.5 corresponds to approximately an optimum

Ag(k)

To(k - 1) cg(k - 1)

(18)

where 0.025 ~<c <~0.05. In this case Ag can increase only.


Therefore, if sign e(k) ~ sign e(k - 1), we set g o p t = 0.75g.
The speed of gain increase is inversely proportional to T~, but
it decreases to 0.75 of its current value whenever the control
error changes sign. The mean interval between two
subsequent changes of sign is equal to the half period of
control error oscillation, i.e. to ~tT,. In the case of a slow
control process T~>> Tv; on the boundary of stability T~
approaches T~.
Therefore the constant c was determined so that the gain
increases if Te > T, and, on the other hand, decreases if T~
approaches T~. In view of this fact, the value c may be
denoted as an index of the oscillatory behaviour of the
system.
There is a large variety of adaptive control algorithms
possible. Some of them are briefly mentioned in the
following sections. For example, the next algorithm, based
on professional experience, shows another index of
oscillation differing from the previous one. It is expressed
by the mean value of the following product:
W (t) = sign e(t) sign [d2e(t)/dt 2]

(19)

or approximately by
W ( k ) = sign e(k) sign [A2e(k)].

(20)

For the sake of straightforward derivation of marginal values


of this kind of criterion, let us apply for a while the
continuous version (19). For e(t) = E sin tot, the value of the
criterion in accordance with (19) yields W = - 1 .
On the
contrary, if e is not oscillating, for example, if e ( t ) =
exp ( - a t ) , then in conformity with (19), W = + 1. Hence, W
expressed by relation (19) can be considered as a measure of
oscillatory behaviour of the system and can take values in the
interval - 1 <~ W <~ + 1.
It is recommendable to use the criterion (20), with central
differences Ae(k) = (e(k + 1) - e(k - 1))/2. The gain adjusting algorithm seeks the equilibrium W + W,~f = 0 and may be
expressed in terms of the following formula:
Ag(k) = ~

g(k - 1). {sign e(k - 2) sign [e(k)

- 2e(k - 2) + e(k - 4)] + Wry,}.

(21)

Due to increment of gain the mean value relating to W can


be omitted.
It may be assumed that the optimum value of Wr~f is in the
middle of the interval - 1 , +1. In that case the control
process is over-damped and consequently too slow.
However, a very good adaptive control process can be
obtained with reference value W, c f = 0 . 3 corresponding to
a slightly oscillating process. In conformity with (21)
Ag(k)<>0

if

( W ( k ) + W , cO~O.

276

Brief Paper

Instead of using the fraction 1/T~,, it is worth while to


apply again the procedure explained in connection with
relations (11) and (12). Starting with equation (12) and
denoting 1/T~, = m we can derive

intervals corresponding to control error decrease must be in


balance with time T~ corresponding to all error increases
according to the following equation:

Ta = T, + q,dT, + Td).

(27)

Am(k) = 0 . 1 m ( k - 1)
x sign [IA2e(k)[ - m(k - I) IAe(k)[].

(22)

3.2. Gain adaptation based on the damping index. In


contrast to formula (20), the following criterion applies to
sign e and sign A2e in additive form, i.e.

d(k) = Isign e(k) + sign A2e(k)l.

(23)

For e(t) = E sin tot the value of Isign e + sign A2e] is always
zero and for e(t)= exp (-at) it is equal to two. In view of
this fact, d was denoted as damping index which may vary in
the interval 0 ~<d ~< 2. The optimum value of d found by
experiments is d~f = 0.6 yielding a slightly oscillating control
process as in the case of the criterion (20). Equilibrium is
reached for d - dr~f = 0. In that case Ag = 0.
For the gain tuning we may write the following relation:
Ag(k) = 0.1m(k - 1)g(k - 1)
x [Isign e(k) + sign A2e(k)l - d,ef].

(24)

The next relation, belonging to the same category of


heuristic strategies of control, but working on a different
principle, has the form

ag(k) = - 0 . 1 m ( k -

1 ) g ( k - 1)
x [sign e(k) sign Ae(k) + qref].

(25)

For illustration let us assume the error signal indicated in


Fig. 2. In the time-interval t 1 - to, q = sign e sign Ae = 1. On
the other hand, for t > q , q = - 1 . Consequently, the index q
of error signal damping can vary in the interval + 1 >~ q ~> - 1.
The reference value q,ef = 0.3, ensuring as in the foregoing
cases a slightly oscillating control process, was obtained by
experiments. Equilibrium is reached for q + q,ef = 0. It holds
that
Ag(k)X0

if

q(k)+q,ef<~O.

The same results can be obtained for the criterion based on


the first difference of the absolute value of the control error

Ag(k) = - 0 . 1 m ( k - 1)g(k - 1)[sign A le(k))+ h,ed.

(26)

Applying again the error signal of Fig. 2, it can be verified


that in the time-interval t l - t o , l e l > 0 , A l e [ > 0 and
s i g n A l e l = + l . However, for t > t 1, l e l > 0 , A l e l < 0 and
sign A lel = - 1 . The same is valid for a negative error signal.
As in the case (24) where dret was denoted as damping index,
q,~f in (25) and h,~f in (26) characterizes the damping of the
process, too.
The last two relations are suitable particularly in the case
of step function disturbances. Even for these events it is easy
to explain the action of the given algorithms, by taking into
account the fact that the rise of the control error corresponds
to a rather shorter interval of time than its decay, i.e. it
grows faster in comparison with the decrease caused by
control action.
The increase or the decrease of the control error is
indicated by sign A lel or by sign e sign Ae. The increase of
the control error is accompanied by a reduction of gain g
and, contrary, an error decrease raises the gain. In order to
reach an equilibrium state, the entire sum Td of time

If for example, qref = 0 , then Ta = T,. This relates to regular


oscillations on the boundary of stability.

4. Applications
Applications are concerned with control as well as
compensation of disturbances. Transport lag in the
controlled system does not present any difficulties.
Numerous examples have been simulated with good
results. Two examples of simulated processes are shown in
Figs 3 and 4. In both examples the controlled system was
described by the discrete transfer function

S(z)

,~o

FtG. 2. Control

Ti "

Itl

Td

I,t

process relating to step change of a


disturbing variable.

-t)3"

The adaptive controller was the simplest one according to


equation (18). Variables and parameters for the first
simulation run have been selected as follows:
controlled and controlling variable y(0) = 0 and u(0) = 0
controller gain g(0) = 0.3
parameter of time To(0) = 1
command variable w(k) = 0 for k < 1, w(k) = 1 for k >/1
disturbing variable v(k) = 0 for k < 15 and for k > 30
acting on the output of
the process v(k) = 1 for 15 <~ k ~< 30
simulation time 1 ~<k ~<40
sampling period T = 1.
Figure 3 shows the result of the first simulation run. Figure
4 represents the result of the second run starting with values
u(0), g(0) and T~(0) obtained at the end of the previous run.
The command variable and disturbing variable have been
changed identically in each run. Comparing Fig. 3 with Fig.
4, the improvement due to adaptive action is clear. At the
end of the first run To(40) = 1.67 and at the end of the second
run T~(40) = 2.16.
It is well known that changes of the command variable and
disturbing variable in a control loop equipped with just one
controller yield only a compromise of adapted variables as
was shown in the examples of Figs 3 and 4, respectively. If
only the command variable is varied (v = 0 for all k) and all
other conditions remain unchanged, then in the second run
zero error is reached already in four steps of control. On the
other hand, if only compensation of the disturbing variable v
is undertaken and the command variable w = 0, then, after
an optimum adaptive setting of the controller, the maximum
error is only 13%.
Figure 5 shows the first run of the simulated adaptive
control process of a control loop with a system to be
controlled described by the indicated transfer function and
with controller according to equation (21). By inspection it
can be seen that the command variable w was changed at
0.5-

1.5

0.4--

0.3--

0.5

gi

0.125
(1 - 0 . 5 z

yiu
5

10

15

20

25

30

35

40

FIG. 3. First run of a simulated adaptive control loop;


discrete controlled system.

Brief Paper

0.5-

1.5

0.4

0.3g

There are two situations in real applications where the


adaptive control must be switched off, first when the error
remains zero for two or three control intervals and secondly
when the actuator saturates.
Several successful in-line tests and applications have been
realized on industrial plants using process computers as well
as microcomputers. As an example the steam temperature
control of a steam generator of 200 and 500 MW should be
mentioned (Krasnick~, 1985).
Another example of industrial application concerns
adaptive control of wood drying kiln of a charge type
(Zentko and Alexik, 1986).

0.5
y,u

T1

o.2

10

15

~ ~

20

25

30

277

35
=,k

40

FIG. 4. Second run of a simulated adaptive control loop;


discrete controlled system.
time instant t = 0 from w = 0 to 1, then at t = 1 6 to w - = - 1
and in the interval from t = 3 0 . 5 to 40.7 the command
variable was changed by a constant speed and finally reached
a constant value of 0.083. The following starting values have
been applied:

5. Conclusion
The presented contribution is devoted to adaptive control
algorithms not needing a mathematical model of the system
to be controlled and, consequently not needing special test
signals and the identification of the system parameters
evaluated by the data measured on the input and output of
the system.
The simple algorithms make it possible to design
single-chip microcomputers for three-term controllers. The
algorithms are rather robust and tested in numerous
simulations and real time experiments.

controlled and controlling variable y(0) = 0 and u(0) = 0

controller gain g(0) = 0.1


final value of gain g(90) = 0.35
sampling period T = 1.
w

References
L. (1964). Funktionalanalysis
Mathematik, 371 pp. Springer, Berlin.

Collatz,

(p+l) 4

20
I

Flo.

5. Adaptive

MUTO 2 ~ / 2 - H

30
I

"

control process; continuously


controlled system,

acting

und nurnerische

Kl~n, P., J. Mar~ik and J. Drozen (1986). Adaptive PID


controllers on a single-chip microcomputer. IFAC Symp.
LCA 86, Valencia, Spain.
Krasnick~, S. (1985). Steam superheaters control of power
station block 200MW (in Czech). Automatizace, 28,
271-275.
Mar~lk, J. (1983). A new conception of digital adaptive PSD
control. In Problems of Control and Information Theory,
Vol. 12, pp. 267-277.
Mar~lk, J. and V. Strejc (1985). Heuristic adaptive process
computer control. Preprints IFAC/IFIP/IMACS 7th Conf.
on Digital Computer Applications to Process Control,
Vienna, pp. 387-392.
Vejvoda, O. et al. (1981). Partial Differential Equations:
Time Periodic Functions, 358 pp. Sijthoff Noordhoff,
Rockville, U.S.A.
Zentko P. and M. Alex~ (1986). Drying kiln parameters
adaptive control. Preprints of IFAC Symp. Lignoautomatica '86, Bratislava, (~SSR, pp. 107-110.

You might also like