GRACY VVC BG ACS 611 DATE......................... DISC USED YES ! NO !

INTERNAIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING

Int. J. Adapt. Control Signal Process. 2000; 14:519}532

Towards &plug and control': self-tuning temperature controller

for PLC

Bernd-Markus Pfei!er*R
Siemens AG, Automation & Drives, Advanced Group Technologies A&D GT 5, D-76181 Karlsruhe, Germany

SUMMARY

Software PID controllers have become a standard component of automation with PLCs (programmable
logic controllers) and work together with classical sequential function charts in the same processing units.
However, traditional methods for the computer-aided design of control systems often involve high math-
ematical e!ort and are usually implemented on additional PCs that are only connected for commissioning
purposes. Therefore a new, very simple concept for self-tuning PID controllers will be presented, which is
implemented together with the controller on the PLC and which enables automatic tuning without prior
knowledge of the process. The concept is designed speci"cally for temperature control, but can also be used
for other control loops in the process industry which show stable time-lag dynamics and allow su$cient
setpoint steps. For users, the most important advantage is the drastic reduction of commissioning time and
e!ort. If, for example, there are 10 temperature loops in one machine, the successive tuning of all 10
controllers used to require several days. With the new concept, there is no longer any need for dedicated
tuning. The self-tuning of all control loops is performed automatically and simultaneously on machine
startup. Control performance is better than it is with the traditionally designed PID controllers, particularly
for slow temperature processes. The concept is based on a systematic startup sequence to explore process
dynamics, a modi"ed in#ection point method for process identi"cation and a separation of setpoint and
disturbance control. Copyright  2000 John Wiley & Sons, Ltd.

KEY WORDS: adaptive control; software PID control; programmable logic controller (PLC); temperature
control

1. INTRODUCTION

Customer demands for modern digital control systems meanwhile include the need for self-tuning
controllers, that automatically adapt to any temperature process without manual parametriz-
ation and without prior knowledge of process dynamics. Such a controller is supposed to learn
enough about an unknown process during the "rst heating phase and to be able to reach even the

* Correspondence to: Bernd-Markus Pfei!er, Siemens AG, Automation & Drives, Advanced Group Technologies A&D
GT 5, D-76181 Karlsruhe, Germany
R E-mail: [email protected]

Received 10 November 1999

Revised 23 November 1999
Copyright  2000 John Wiley & Sons, Ltd. Accepted 20 December 1999
 ACS=611=Gracy=VVC=BG

520 B.-M. PFEIFFER

"rst given setpoint without an overshoot (&plug & control'). This paper describes a new concept to
reach this goal with very low computational e!ort and memory requirements.
The concept can be classi"ed roughly as a model identi"cation adaptive controllers as e.g.
described by Isermann et al. [1, p.11]. Such a system consists of three main parts: process
identi"cation algorithm, controller design algorithm and the controller itself. For process identi-
"cation, parametric discrete-time ARMAX-models and recursive least-squares estimation could
be used. The controller itself could be a standard PID type designed by pole cancellation,
deadbeat or minimum variance considerations. Although this theory is well known since about
1980, it was not yet accepted for a widespread use in industry. Reasons may be robustness issues,
sensitivity to measurement noise, model structure mismatch and other disturbances. Anyway
high e!orts are required for a so-called supervision and co-ordination level, that decides if the
process is su$ciently excited for identi"cation and if estimated models are valid. An industry-
proved implementation of such a fully adaptive controller with discrete square- root "ltering
identi"cation (DSFI [2]) of a third-order ARMAX-model is delivered as a function block for
Simatic S7 by i.p.a.s.-systems, Frankfurt. On the other hand, the lean concept presented here
requires less than 10 per cent of the memory space and is a simpli"cation of these ideas with
respect to:
1. Adaptation mode: in most applications it is neither necessary nor even desired to adapt the
controller online permanently during normal operation. In contrast, only a self-tuning at
commissioning time, and possibly some adaptation on demand during online operation is
needed.
2. Neither the process model, nor the controller is explicitly calculated in time-discrete form.
Customers are used to quasi time-continuous controller implementations and like to
compare estimated model parameters like time-lags with their own experience, whereas they
mostly do not understand the properties of discrete-time models.
3. The process models are restricted to low-order transfer functions suitable for temperature
processes.
Some years ago, even expert systems have been proposed (e.g. by Arzen [5]) for auto-tuning of
controllers. Other suppliers of compact digital controllers e.g. Foxboro, Eurotherm and Philips
are also working on self-tuning schemes, but they usually do not publish their algorithms.
This paper is organized as follows: Section 1 summarizes some basic aspects of modelling and
control of temperature processes. Section 2 gives details on the identi"cation algorithm,
Section 3 on the controller design. In Section 4 some practical aspects of implementation and
industrial application are discussed.

1.1. Modelling of temperature processes

According to Reference [8], physical models for temperature dynamics can be constructed from
"rst-order models for elementary subsystems. That way it is not necessary to de"ne a system
model with spatially distributed parameters using partial di!erential equations.
Each subsystem i (Figure 1) is characterized by its heat capacity c (Ws/K) and the respective
G
heat transfer coe$cient a [W/K] to the next, colder subsystem. The subsystem receives the
GG>
heat input Q and stores it, i.e. integrating over time, thereby reaching the temperature ¹ . The
G G
heat output to the next subsystem is calculated as a product of temperature di!erence ¹ !¹
G G>
and heat transfer coe$cient. This heat output is taken away from subsystem i, and transferred to

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 521

subsystem i#1. External e.g. electrical energy is an additional input to the subsystem &heating'.
Any measurable temperature of a subsystem can be de"ned as the process output. It is usually
di$cult to determine all the heat transfer coe$cients from construction data, therefore experi-
mental parameter identi"cation is needed.

1.2. Control of Temperature Processes

The resulting transfer functions for temperature dynamics in most cases have time-delay charac-
teristics of rather small order. There are rarely more than three relevant time constants that can
be identi"ed in industrial processes, but these time constants depend in a non-trivial way on the
time constants of the elementary subsystems. Process gains range from 0.5 to 30, depending on
the design of the heating and on measurement normalization factors. Time constants from one to
several thousand seconds are observed. In open loop, there are no oscillations and no extremely
long deadtimes. But anyway there are some properties that make temperature control di$cult:
1. Slow drift time constants. That is, poles of the system transfer function that are close to the
origin of the complex plane, and are causing integrator-like dynamics in the relevant
working range. After applying a step input (&heating on'), it takes a very long time to reach
a steady state. Therefore, identi"cation methods based on evaluation of step responses (e.g.
Reference [7]) are rather time consuming and sometimes not practicable. Moreover, the
relevant time constants time-lag and recovery time di!er from each other by an order of
magnitude or more, posing special problems for identi"cation algorithms.
2. ¹ransfer zeros. That is, numerator terms in the process transfer function. They are found if
the controlled variable is not de"ned at the end of the heat transfer chain, e.g. if an oven is
heated from the inside, but the heat capacity of the outside isolation plays an important role.
3. In many cases, there is no active cooling, and the positive temperature gradients observed are
much bigger than the negative ones. Anyway, this e!ect can be explained with a linear
dynamic model, that does not allow negative values of the manipulated variable.
4. Nonlinearities can be caused by evaporation, condensation or actuators with non-linear
characteristics (e.g. valves). However, in many cases they are considered as a cause for
control problems that actually are bound to one of the "rst three di$culties mentioned
above.
5. The control performance speci"cation dilemma: &no overshoot' is required because even
small overshoots take a very long time to decay without active cooling. On the other hand,
sluggish control is also not acceptable, because the temperature processes are always &too
slow'.
In principle, temperature control is feasible with simple PID-controllers, but the combination of
a process with integrator-like characteristics and a controller with integral action is di$cult.
From a dynamic point of view, PD control would do better, but is not applicable because of
steady-state errors. Many well-established heuristic tuning rules e.g. Reference [3] are not suited
very well for temperature control with the speci"cation &no overshoot'. Others like, e.g. modulus
optimum and symmetric optimum [1,4] are able to "nd su$ciently damped controllers, but for
temperature control, they are much to slow to be accepted by customers. The integral time
constant ¹ of the PID controller cannot be selected according to the dominant, slowest process
G
time constant, as proposed by many tuning formulas (e.g. pole cancellation), because this time
constant usually is in the order of hours. If ¹ is selected much faster, there will be large
G
Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

522 B.-M. PFEIFFER

Figure 1. Elementary heat capacity as a subsystem of a temperature process.

overshoots for setpoint step changes due to a closed loop transfer zero, even if closed loop
damping is su$cient and the response to load disturbances is satisfactory. Therefore, the most
important decision even for a simple control concept is a separation of setpoint and load
disturbance design. The PID is optimized for tight disturbance control and is enhanced with
a special structure decomposition to avoid overshooting setpoint responses: proportional and
derivative terms are placed in the feedback path.

2. IDENTIFICATION

Control design is supposed to rely on a rather simple process identi"cation algorithm that easily
"ts into a PLC function block. But the identi"cation must at least deliver a second order model
with two di!erent time constants, because they di!er by an order of magnitude for most
temperature processes. Many identi"cation algorithms for such measurement data deliver only
"rst-order models, that are not su$cient to design good PID control. Therefore, a modi"ed
in#ection point method is applied in a systematic startup sequence (Figure 2) to obtain informa-
tion of the process and successively improve the mathematical model.

2.1. Startup sequence

The startup sequence has the following steps (called PHASE in Figure 2):
Phase 1: Heating is o! and the steady-state c of the temperature is measured. Alternatively, the

self-tuning can be started from another steady state with constant heating power.
Phase 2: As soon as the user demands a setpoint step of su$cient size, constant heating energy
y is applied until the in#ection point of the step response is detected. This requires adaptive

low-pass "ltering of the measured temperature to suppress noise e!ects. As soon as the "ltered
ascent ratio of the step response is decreasing in two successive cycles or the temperature
increases over 70 per cent of the step size, the self-tuner is considering the in#ection point to be

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 523

Figure 2. Self-tuning of controller for unknown process * setpoint SP, process value PV, loop manipulated
variable LMN; constant heating LHLM}TUN for identi"cation and control; bit TUN}ON for activation of
self-tuning (source: Siemens user manual).

found, with time abscissa, ordinate and ascent ratio:

dx
t ,x , 
  dt

From the tangent at the in#ection point, the time delay

x !x
t "t !t !t , t "  
       dx /dt

is calculated, assuming the steady state x "c . For the behaviour observed up to now, an
 
integrating process model with "rst-order time lag (IT )

K dxw/dt
G(s)" G , K" , ¹ "tu
s(¹ s#1) G ymax!y 
 
can be assumed, for which PI-control according to the symmetric optimum [1, p. 263] with
integrating time constant ¹ "4¹ and gain
G 
¹
K" G
 8¹K
 G
is designed. It is slightly detuned to avoid overshoots.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

524 B.-M. PFEIFFER

Phase 3: &Rough' PI-control to reach the setpoint. The estimated IT -model does not have

a steady state, i.e. the process gain is still unknown. Due to the lack of information on process
dynamics, the PI control must be tuned quite conservative. But as soon as this rough control has
achieved a steady state, the process gain

x !x
K"  
y !y
 

can be calculated.

2.2. Modixed inyection point method

At this time at the end of phase 3 it becomes now possible to make additional use of the in#ection
point data from the beginning of phase 3. The open-loop control with "xed heating power (phase
2) had to be interrupted a long time before the respective steady state could be reached. From this
"rst part of the step response, the classical in#ection point method ([9, p. 365], similar ideas might
be found in English standard textbooks on control engineering) only delivers the time-lag, but not
the recovery time. But with the additional information on process gain, an enhancement of the
classical in#ection point method becomes feasible: the intersection of the in#ection tangent and
a virtual (not measurable) steady state at the end of the step response (that was interrupted!) is
de"ned, from which the recovery time is obtained:

x #K(y !y )!x
t "t #t , t "    U
      dx /dt


In this formula, x #K(y !y ) is the virtual steady state at the end of the step response. The
  
classical in#ection point method for a complete step response delivers only P¹ !¹ ("rst-order
 R
plus dead time) models. If the real process does not have a physical dead time, the approximation
by a non-minimum-phase model will result in quite conservative controllers. Therefore a second-
order Laplace transfer function model of the form

K
G(s)" , ¹ '¹
(¹ s#1)(¹ s#1)  
 

is considered "rstly. A very important characteristic of this model is the ratio

¹
f" 
¹


of the two time constants. The in#ection point of the step response of such a model is known in
analytic form

ln f
t "f ¹
 f!1 

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 525

The intersection of the in#ection tangent with the lines x"x and x "x #Ky yields time
   
lag t and recovery time t :



¹ 22\2
t "¹ 
 ¹



¹¹ ¹
t "   ln  !t #¹ #¹
 ¹ !¹ ¹  
  
Introducing f , the following non-linear equation system is obtained:

t 1
¹ "t f\DD\, "
 t f\DD\(1#f#( f ln f )/( f!1))!1

To solve this non-linear system for ¹ and f, the exponential functions must be replaced with

linear approximations for the relevant parameter space 2(f(20:

1
¹ +t , p "1.0722, p "2.0982
 p f#p  
 
t
+p f#p , p "1.1919, p "8.0633)
t    

The second of these equations is solved for f:


1 t
f" !p
p t 
 
then the "rst equation delivers ¹ , and "nally ¹ "f ¹ . This model form is only valid if
  
t /t '9.64. Otherwise a P¹n-model of the form

K
G(s)" , n'2
(¹ s#1)L

is assumed. There are tables in literature (e.g. Reference [4]) to "nd time-lag and recovery time of
this type of models. The table entries are now approximated by hyberbols to obtain the inverse
functions. Given t and t , the system order can be estimated to be

7.9826
n" #1.1099
(t /t )!0.3954

and the time constant to be

t
¹" 
 0.0165n#0.5078n#0.8387

This approach is feasible from a practical point of view up to the order of n)4.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

526 B.-M. PFEIFFER

Figure 3. Heating and cooling identi"cation for split range control, heating and cooling
in#ection tangents are marked as arrows.

2.3. Heating and cooling

In some temperature control applications, there are two manipulated variables for one controlled
variable. Considering e.g. a plastic pro"le extruder, the cylinder under temperature control must
be heated at machine startup, but must be cooled during normal production due to friction
energy of the extruder screws. There are two actuators with di!erent physical principles: direct
electical heating using puls-width-modulation, and air cooling using a blast engine. In these cases,
a split-range controller is needed, and care must be taken for di!erent heating and cooling
dynamics. Therefore, a cooling experiment (cf. Figure 3) with constant cooling power y is

performed after self-tuning of the heating is "nished (with a steady state y ). The ratio of heating

and cooling gain is deduced from a comparison of heating and cooling in#ection tangents:

K (dx /dt)#K y
r " G , K "  G 
 K G y
G 
It must be noted here, that the cooling descent ratio is not only due to the active cooling, but also
to the simultanous shutdown of the heating.

3. CONTROL CONCEPT

The control concept is based on classical PID control, but allows a separation of setpoint and
load disturbance design: the PID is optimized for tight disturbance control and is enhanced with

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 527

a special structure decomposition to avoid overshooting setpoint responses: proportional and

derivative terms are placed in the feedback path.

3.1. Structure decomposition control

Structure decomposition as shown in Figure 4 is well known in literature for integrating
processes, because there is a fundamental dilemma between good disturbance control and good
setpoint control: a PID controller designed for optimal disturbance control leads to overshoots of
10}20 per cent of the step size on setpoint changes, even if all closed-loop poles are su$ciently
damped. This is due to a dominant transfer zero in the closed-loop reference transfer function.
The same dilemma is also found for temperature control, where usually the requirement is &no
overshoot', because even small overshoots take a very long time to decay without active cooling.
On the other hand, sluggish disturbance control is also not acceptable, because the temperature
processes are always &too slow'. One approach to solve this dilemma is a controller with two
degrees of freedom (RST-control), actually including a dynamic "lter for the reference input. But
this approach is not feasible with many industial controllers, because they do not have two
degrees of freedom.
The new discovery is: structure decomposition originally intended for integrating processes is also
helpful for temperature control with integrator-like dynamics in the relevant working range.
Optional structure decomposition is already available in many software PID controllers.

3.2. Controller design

A general PID controller of the form


1 ¹s
G (s)"K 1# #  , v"5
0  ¹ s (1/v)¹ s#1
G 

is considered and then decomposed. The whole con"guration in Figure 4 can be considered as
a cascade control: in an inner loop, the slow-temperature dynamics are accelerated by PD
control; in the outer loop, the whole inner loop is directed to steady-state precision by I-control.

Figure 4. Structure decomposition * PD in feedback path.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

528 B.-M. PFEIFFER

The design procedure follows the usual strategy for cascade control: "rstly, the inner loop is
tuned.
Each second-order process has two states: the process value x and its time derivative xR "dx/dt.
The transfer function can be transformed to an equivalent state space representation

  
x xR "Ax#by
x" ,
xR x"cx#dy

 
0 1
A"
!1/¹ ¹ !(¹ #¹ )/¹ ¹
     

 
0
b" , c"[1 0], d"0
K/¹ ¹
 
A state feedback for this process is de"ned by

y"!kx"!(k x#k xR )
 
and can be implemented using

 
1
k"K
 ¹

as conventional PD controller with gain K and derivative time ¹ ! This allows to design PD
 
control for a second-order process with the rigorous mathematical methods of state space
optimization, minimizing a quadratic performance index:

 
 1 0


J" (xQx#yry) dt, Q" \ , r"0.002
0 1


The time-consuming optimizations are performed o%ine for the relevant parameter space
1.1(f"¹ /¹ (40 and the results are approximated with hyperbols for fast online computa-
 
tion on the PLC:


21.4 8.417
K " 1.5 , ¹ " 0.985! ¹
 K  f#10.66 

The same strategy applies to the design of the I-controller: it is optimized in o%ine simulations for
the relevant parameter space and the resulting tables for the integral time constant are approxi-
mated with a "rst-order polynomial, depending on the already known derivative time constant:

¹ "(0.1236 f#3.322)¹
G 
Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 529

Figure 5. Setpoint step with structure decomposition control, process

parameters K"3, ¹1"200, ¹2"30.

A simulation example for structure decomposition control is shown in Figure 5. Obviously, after
a setpoint step, the manipulated variable does not increase stepwise, but is increasing smoothly
via the integral action of the controller. The setpoint is reached without overshoot.
For processes not allowing a second-order model, controller design is based on the PTn-model
and the modulus optimum, in the form that is already used in many other contoller commission-
ing tools by Siemens [7].

3.3. Control zone

There is one more alternative way to get away from the dilemma of setpoint and disturbance
control: the control zone, as shown in Figure 6. Only inside this narrow band around the setpoint,
closed-loop PID control is performed. Outside the control zone, the controller is switched to
manual mode with full power (feedforward control): e.g. if temperature is too low, maximal
heating energy is applied. This way, the closed control loop with the transfer zero mentioned
above is cut temporarily for setpoint steps, avoiding the overshoot. The size c } of the control
 
zone is chosen according to the proportional gain of the controller: c } "100 per cent/K .
  
For di!erent heating and cooling dynamics, the zone must be selected non-symmetrical to the
setpoint. A simulation example with control zone is shown in Figure 7. The control zone is only
recommended for f'10 to deliver zero overshoot. For processes with bigger relative time-lag
( f(10), a full speed temperature increase cannot be arrested fast enough by switching back to
closed-loop control.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

530 B.-M. PFEIFFER

Figure 6. Control Zone with closed-loop control only in a narrow band around the setpoint SP.
LMN is the loop manipulated variable.

Figure 7. Setpoint step with control zone, process parameters K"3, ¹1"200, ¹2"10 ( f"20).

4. PRACTICAL ASPECTS

4.1. Implementation

The new methods described in this paper are implemented as a PLC function block for SIMATIC
S7. Programming language is structured text according to IEC 1131-3. The product name is &PID

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=Venkatachala=BG

SELF-TUNING TEMPERATURE CONTROL 531

Figure 8. Control commissioning for plastic extruder. Setpoint, process value and manipulated variable
for startup procedure to 1803C. (scales: 0}100 per cent manipulated variable for heating only, !100..#100
per cent for heating and cooling, 0}2003C setpoint and process value).

Self-Tuner'. It does not contain a PID controller itself, but is working together with all software
controllers of SIMATIC S7, i.e. &standard PID control', &modular PID control' and the integrated
&step 7 PID control'. The self-tuner takes actual setpoint, process value and manipulated variable
as input signals. During the startup sequence the self-tuner sets the controller in manual mode,
and makes use of the bumpless switching from manual to automatic mode. As a result of
automatic commissioning, the parameters gain, integral time and derivative time of the PID are
de"ned. In a PLC program, controller and self-tuner are called one after each other in the same
cycle. The self-tuner is a very lean function block, needing only 5 kB of code plus 0.3 kB of data
per control channel. It actually requires less resources than the PID controller.

4.2. Industry applications

The new concept was veri"ed in simulation and laboratory experiments before it was tested with
selected pilot customers. As an example, some results from the practical application to a plastic
pro"le extruder &Weber CE7 272K' for PVC-window frames, located at KoK mmerling (Pirmasens,
Germany), will be given.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532
 ACS=611=Gracy=VVC=BG

532 B.-M. PFEIFFER

At zone 1 plastic powder is "lled into the machine, and is pressed with extruder screws through
the cylinder zones 2}6, before it leaves the machine with high pressure through a heated tool
(zones 7}10), giving the material the desired pro"le shape. There are 10 temperature zones to be
controlled, all of them measured with thermocouples. Actuators are electrical heating inductors
controlled by solid-state relays. Zones 2}4 are additionally equipped with cooling blast engines
controlled by mechanical relays.
Automatic self-tuning is performed for all 10 channels at the same time as shown in Figure 8.
The control loops are just switched on and supplied with a setpoint, before all self-tuners step
through their startup sequence, identify the process model for the respective control channel, and
take their controllers in automatic operation. After self-tuning has "nished, regular production
can be started in controlled mode. This example does not show the additional cooling test to
identify the di!erence between heating and cooling dynamics.

5. CONCLUSIONS

The &&PID self-tuner'' V1.0 with variable structure control instead of structure decomposition is
delivered since January 1998 to the customers. Patents are pending under Appl. No. 19548909.8,
19722431.8, 19851827.7 and 19851826.9. With the self-tuner, customers are enabled to perform
control commissioning in the sense of a &plug & play'-philosophy: the controller is connected to
input and output signals, afterwards switched on, and is immediately ready to drive the process to
the "rst given setpoint. There are still some control parameters that can be handled by the user,
e.g. to select operating modes, but they only need attention if there are special problems or desires.
The self tuner is applicable to continuous output control, to puls-width modulation control for
binary actuators (relays) and to step control for motor valves with or without position feedback
(repeated manipulated variable). Version 5.0 with structure decomposition is capable to handle
three-point control for two actuators per loop (heating and active cooling) and is available since
September 1999.

REFERENCES

1. Isermann R, Lachmann K-H, Matko D. Adaptive Control Systems. Prentice-Hall:UK, 1992.

2. Biermann GJ. Factorization Methods for Discrete Sequential Estimation. Academic Press:New York, 1977.
3. Ziegler JG, Nichols NB. Optimum settings for automatic controllers. Transactions of the ASME 1942; 64.
4. FroK hr F, Orttenburger F. Introduction to Electronic Control Engineering. Siemens AG, Heyden & Son: London, 1982.
5. Arzen K-E. An architecture for expert system based feedback control. Automatica 1989; 25(6).
6. FoK llinger O. Regelungstechnik, vol. 6. Au#age, HuK thig: Heidelberg, 1990.
7. Preu{ H-P, Linzenkirchner E, Kirchberg K-H. SIEPID * ein InbetriebsetzungsgeraK t zur automatischen Reglerop-
timierung. atp 29, 9/1987.
8. Preu{ H-P. Modellbildung und Regelung von Temperaturstrecken, vorlaK u"ger Bericht, Siemens AG, A&D GT 34,
6/94.
9. Unbehauen H. Regelungstechnik 1, Vieweg: Braunschweig, 1992.

Copyright  2000 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2000; 14:519}532