FUZZY

sets and systems

ELSEVIER Fuzzy Sets and Systems 70 (1995) 135-146

The FLS application of fuzzy logic

L.P. Holmblad, J.-J. ~stergaard*
FLS Automation A/S, 2500 Valby Copenhagen, Denmark
Received June 1994

Abstract

This paper will survey the application of fuzzy logic by F.L. Smidth & Co. A/S (FLS) for control of rotary cement kilns.
The presentation is given in retrospect, starting in 1974 when FLS heard about fuzzy logic for the first time. The most
important milestones in our work with high-level process control are presented, with special emphasis on the role of fuzzy
logic. The present status of the FLS Fuzzy II system is outlined and the development trend for high-level process control
systems as expected by FLS Automation is discussed.

O. Introduction permitting more accurate control of the fuel/air

ratio than was possible simply by observing the
FLS is a major supplier of cement plants and has smoke.
been developing fuzzy logic for controlling cement Central control was introduced in the cement
plants since 1974. industry in the 1960s. Temperature measurements,
Before the 1960s, the operator controlled the pressure measurements, voltage and amps indica-
cement kiln by looking into its hot end, the burning tions, pen recorders, alarms and other indicators
zone, and watching the smoke leaving the chimney. were gathered in a central control room. PID con-
The operator was stationed on a platform just trollers were also installed, mainly for uniform feed-
behind the kiln, with a blue glass to protect his eyes. ing of the raw material and the fuel. The operator
At that time, this was the most important aid for no longer sat next to the hot, noisy and dusty kiln,
controlling the formation of cement clinker. The but controlled the process by looking at instru-
operation was 100% manual and very few meas- ments. However, if there was uncertainty about the
surements were available. In addition to the blue readings, he would still go to the platform and pick
glass for inspection of the burning zone and the up the blue glass to make a visual inspection of the
darkness of the smoke, a temperature measurement burning zone.
was normally available at the cold end of the kiln. Computers for process supervision and control
Later an oxygen analyser was installed, thereby were introduced in the cement industry in the late
1960s. FLS, as one of the world's leading manufac-
turers of cement plants, accepted the challenge of
*Corresponding author. computerized control of cement kiln operation

0165-0114/95/$09.50 © 1995 - Elsevier Science B.V. All rights reserved

SSDI 0 1 6 5 - 0 1 1 4 ( 9 4 ) 0 0 2 1 2 - 6
136 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 13~146

from the very beginning. The cement kiln is the date, more than 100 kilns have been equipped with
main machine at a cement plant and the overall high-level fuzzy control by FLS. Modern control
economy of cement production very much depends systems include comprehensive instrumentation,
on the operation of the kiln. A medium-size cement PLC sequence control, extensive alarm treatment,
kiln producing 5000 tons of cement clinker per day reporting functions, operator colour screens and
consumes about 20 tons of coal per hour. Economi- high-level control strategies. A modern control
cally, it is thus very attractive if even a few per cent room is shown in Fig. 2. But the blue glass has not
of coal can be saved and production increased disappeared. The operator may still want to look
a little through proper control of the kiln. A typical inside the kiln to check that everything is in order.
control room from those days is shown in Fig. 1. However, modern control systems, including high-
After experimental work in the 1970s, FLS in 1980 level fuzzy control, is of invaluable help to the
launched the first commercial computer system for operator in the safe and profitable production of
automatic kiln control based on fuzzy logic. To cement.

Fig. 1. Cement plant control room from the 1960s.

 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 13~ 146 137

Fig. 2. Control r o o m at a m o d e r n cement plant.

This paper presents how FLS has been working ing in nine actions on the fuel. Today, many fuzzy
the technology of fuzzy logic from the start in 1974 controllers are implemented along exactly the same
to the present version of our fuzzy control system, lines, except that control actions and input condi-
named FUZZY II. tions are expressed in terms of fuzzy membership
functions, as illustrated in Fig. 4.
The idea of implementing a control strategy as
1. The start a decision table, which basically is a rough descrip-
tion of the manual control scheme, was thus born
Before we had even heard of fuzzy logic, some before we learned of the theory of fuzzy logic.
experiments with automatic kiln control were car- The rule-based approach which underlies the
ried out in 1972 at a cement plant in Denmark. The decision tables was also inspired by the instructions
most interesting aspect of these experiments is the we found in a textbook for kiln operators, which
approach that was chosen for the implementation contained the control rules for manual operation of
of the kiln controller. The fuel control strategy, as a cement kiln shown in Fig. 5.
shown in Fig. 3, was programmed as a two-dimen- The first time we heard about fuzzy logic was at
sional decision table with an error signal and the the 4th I F A C / I F I P International Conference on
change in error as inputs. The error and the change Digital Computer Applications to Process Control,
in error were both divided into three levels, result- held in Zfirich, Switzerland, in 1974. As a postscript
138 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146

potentiality of this approach has been demon-

strated by successful implementation of the fuzzy
- +1 +2 +3 logic controller and further research in this area
is strongly recommended."

E 0 -1 0+1 It was immediately clear to us that we had come

upon a theoretical basis for our early attempts to
achieve automatic kiln control. The recommenda-
+ -3 -2 -1 tion was accepted and FLS started to investigate
the feasibility of fuzzy logic control in co-operation
with The Technical University of Denmark.
b, + 0 -
OIL T
2. The first experiments
= OII feed rate changes

E = Error on kiln drive load Before applying the theory of fuzzy logic to con-
= Trend In kiln drive load
trol of a cement kiln, we wanted to investigate the
T
theory under circumstances where experiments
-3/+3 = • OII units
were possible. The experiments allowed in a work-
ing cement kiln are very limited. Experimental
work was therefore carried out at The Technical
University of Denmark. The theoretical under-
Fig. 3. Fuel control decision table.
standing and inspiration in relation to process con-
trol was gained mainly from papers written by
Zadeh and Mamdani [6, 2], and control experi-
velo¢~ ments were performed in laboratory-scale
Positive Negative Zero Positive processes such as, for example, a small heat ex-
dev|atlon "~ changer [4].
Negative Negative Zero
The first experiments using a real cement kiln
Negative
big small were carried out at the beginning of 1978 at a Dan-
ish cement plant, using the control rules for manual
Zero
Negative Zero Positive operation shown in [5]. The ambition of the orig-
small small
inal control strategy was relatively low. We wanted
Positive Positive to start with a strategy that was clear, easy to
Positive Zero discuss and likely to work - at least when the kiln
small big
was in stable operation near the desired operating
point. The idea was to develop a control system
Fig. 4. Fuzzy logic decision table. that could keep the kiln cruising at a given produc-
tion level, like the autopilot for an airplane.
During the experiments, which lasted for about
to a paper on learning controllers [1], fuzzy logic three months, it was possible to change the strategy
was proposed as an alternative approach to hu- by modifying the existing rules or by adding new
man-like controllers. The very last remark of the control rules. Modifications had to be made by the
concluding section says: computer expert on the basis of advice and criti-
cism from the kiln burner specialist.
"Greater dividends can be expected however At this stage of the development work, the atti-
if information is processed in a higher-level tude of the management in FLS towards the new
framework of conceptual representation. The kiln control attempts was sceptical, both because of
 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146 139

Case Condition Action to be taken Reason

10 BZ OK a. Increase I.D. fan speed To raise back-end temperature and increase oxygen percentage for action 'b'
OX low
BE low b. Increase fuel rate To maintain burning zone temperature
11 BZ OK a. Decrease fuel rate slightly To raise percentage of oxygen
OX low
BE OK
12 BZ OK a. Reduce fuel rate To increase percentage of oxygen for action 'b'
OX low b. Reduce I.D. fan speed To lower back-end temperature and maintain burning zone temperature
BE high
13 BZ OK a. Increase I.D. fan speed To raise back-end temperature
OX OK b. Increase fuel rate To maintain burning zone temperature
BE low
14 BZ OK N O N E . However, do not get overconfident,
OX OK and keep all conditions under close observation.
BE OK
OX DK
15 BZ OK When oxygen is in upper part of range
OX OK a. Reduce I.D. fan speed To reduce back-end temperature
BE high When oxygen is in lower part of range
b. Reduce fuel rate To raise oxygen percentage for action 'c'
c. Reduce I.D. fan speed To lower back-end temperature and maintain burning zone temperature
16 BZ O K a. Increase I.D. fan speed To raise back-end temperature
OX high b. Increase fuel rate To maintain burning zone temperature and reduce percentage of oxygen
BE low
17 BZ O K a. Reduce I.D. fan speed To lower percentage of oxygen
OX high slightly
BE O K

Fig. 5. Extract from Peray's textbook for kiln operators.

the earlier failures of others who had attempted mill for control of a rotary kiln for the reburning of
computerized kiln control, but perhaps also to lime.
some extent because of the strange name "fuzzy". The programming technique was based on
Other names were suggested like "verbal control- a block-oriented approach. Control rules were pro-
lers", "pragmatic controllers" and "rule-based grammed by interrelating calculation blocks evalu-
controllers" but eventually, with an increasing ating the terms LOW, OK, H I G H , AND, OR, etc.,
understanding by the management of the concept, as shown in the example in Fig. 6. But the possibili-
it was decided to stay with the word fuzzy, a deci- ties were very limited for adding conventional logic
sion that has never been regretted since. schemes and normal calculations to supplement the
fuzzy control rules. It soon became obvious that the
block representation of the strategy was inflexible
3. Fuzzy control language and lost its comprehensibility as more rules and
parameters were added.
The first two industrial installations of fuzzy con- To address these problems, we developed the
trollers occurred towards the end of 1979. One programming language FCL (an abbreviation of
system was installed at the Danish cement plant fuzzy control language). The first version of a FCL-
where the first experiments had been carried out; based fuzzy controller was installed at the cement
the second system was supplied to a Swedish paper plant in Denmark in 1981. To some extent, FCL is
140 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146

..~~ W1C126P
•

) , ~ JIJOlSP
"lUNSCALIEI COAL'
WIA50T1 ~

lr " - '
Fig. 6. Block representation of fuzzy control rules.

similar to Fortran and Basic but it is also different, tion; on the other hand, you do not face the
as it includes evaluation of fuzzy control rules of the demand for very fast execution of control
form: programmes. On this basis, the first version of FCL
was based on an interpreter with repeated transla-
IF (condition) T H E N (control actions) tion of the program lines before each execution.
The minimum time interval between execution of
where the (condition) and the (control actions) programmes was 1 m i n .
are expressed as fuzzy membership functions. Us- However, as applications developed, the execu-
ing F C L for programming of the control rules in tion time did become a problem. The control strat-
Fig. 6 results in the following two program lines, egy increased in complexity through the inclusion
which are easier and more direct to understand of additional process parameters to obtain better
than the block diagram: and more accurate control performance, more than
one kiln was controlled by one computer, and fas-
IF LOW(O2) AND ( O K ( T E M P )
ter-responding processes were also being controlled
OR H I G H ( T E M P ) ) T H E N by the fuzzy. It was therefore found necessary to
modify FCL to include a translator, so that pro-
MNEG(DAMPER), MNEG(COAL)
grammes to be executed appear in a directly execut-
IF OK(O2) AND ( O K ( T E M P ) able form. This second, much faster version of FCL
OR H I G H ( T E M P ) ) T H E N was launched in 1984 and today the minimum time
interval between program executions is 1 s.
SNEG(DAMPER), ZERO(COAL) This, of course, still puts a limit on the extent to
which fast-responding processes can be controlled
where M N E G and SNEG are abbreviations for by our fuzzy system. We are, however, not consid-
medium negative and small negative, respectively. ering that issue because our main application re-
In general, a cement kiln is a process involving mains implementation of control strategies on the
rather long time constants. From the point of view human operator level, rather than underlying basic
of automatic control, this is normally a complica- controllers.
 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135 146 141

4. The second generation Control objectives with different priorities are

thus among the basic elements of a Fuzzy II high-
The mid-1980s were very busy, with production level control strategy.
and installation of fuzzy systems at cement kilns all
over the world. It was a very exciting period too,
with growth and a lot of encouraging and success- 5. The Fuzzy II structure
ful projects, resulting in increasingly advanced and
complicated control strategies. Problems, however, The elements of a Fuzzy II process control strat-
started to emerge like small clouds coming over the egy are the following:
horizon, as it turned out that some of our fuzzy
systems were not performing satisfactorily after State indices: Calculations concerning the actual
a period of time. Although they were claimed to be process conditions.
easy to comprehend and adapt because of their Control groups: Arrangement of the overall con-
inherent pragmatic and human modelling ap- trol strategy into groups of control objectives.
proach, it was found in practice that the control Priority management: Determines the extent to
strategies became rather complex and not always which the control actions to fulfil the individual
easy for the plant engineers to maintain. objective should be executed.
Against this background, development work for Control objectives: Specifications of what should
the second generation of our fuzzy system was be attained by the fuzzy control strategy.
initiated in 1988, with the main design objectives
focusing on user maintainability. The new fuzzy Normally, a state index combines various
system was launched in early 1990 under the name measurements into a single figure. The degree of
of Fuzzy II. The objectives relating to comprehensi- process stability, the product quality, the produc-
bility and maintainability have been met through tion level, etc., are all typical examples of state
a simple and well-defined structure for implementa- indices for a kiln control strategy. The state indices
tion of high-level control strategies. are important to the structure of a Fuzzy II control
For an operator, control of the process consists strategy, as they form the basis for dividing the
in achieving various goals, which are more or less overall strategy into control groups which can be
precisely defined: maximum product output, min- treated independently. The state indices are used to
imum consumption of raw materials and energy, co-ordinate control actions from the various con-
high product quality, safe process operation, etc. trol groups. Fig. 7 shows a diagram of the control
Different processes have different control objectives strategy structure for which the Fuzzy II system has
but, in general, good process control may be de- been prepared.
fined through a list of control objectives which Design of a control strategy consists in filling in
should be fulfilled as much as possible. the structure with specific state index calculations,
The concept of control objectives is a key ele- control groups and control objectives, whereas the
ment in the structure of a Fuzzy II control strategy. priority management system normally is a fixed
For a cement kiln, typical control objectives are module in the system. For a new process, it is
• stable operation, perhaps not so clear which are the most appropri-
• good cement clinker quality, ate indices and control groups. This is not a
• complete combustion, major problem, however, as the structure facilitates
• low fuel consumption. a stepwise implementation of control objectives
As control objectives are frequently in conflict, co- and index calculations concurrent with more and
ordination means approaching optimal conditions more process control knowledge being available.
in succession of importance. Priorities, in other Every control objective is implemented as what
words, have to be assigned to the various objec- is known as an objective module made up of the
tives, specifying which objectives are considered the following four basic elements:
most important to fulfill. 1. deviation calculation,
 142 L.P. Holmblad, J,J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146

uniny
programs i_, 1--nu I---,onl... PriorityManagement

Control
groups
I Burning
zone contr. I
I Icombustlon
contro """
I
Deviation
I I calculation

I PriorityManagement
Timing
Weight factor
Control
objectives

Rule ~ ) Output
, Stable I block , I block
operation[ control
Fig. 8. Control objective module diagram.

I ]*[ I upon the actual index values. The logic for selecting
alternative adjustments may be fuzzy or non-fuzzy
depending on whether a gradual or an either/or
switch is the most appropriate. In most cases, the
fuzzy logic approach gives the best control perfor-
Fig. 7. Fuzzy II control strategy structure.
mance, simply because no process operates
with sudden changes between alternative control
2. rule block, actions.
3. output block, Timin9 calculation
4. timing calculation. The timing calculation determines when control
Fig. 8 shows the objective module diagram, which actions are to be executed. It is just as important as
is used for implementation of all types of control the rule block for determining the proper function
objectives. of the control strategy. Also the timing calculation
Deviation calculation is normally fuzzy in the sense that the time interval
The deviation calculation evaluates the degree to between control actions changes gradually as
which the objective is fulfilled. Normally, the calcu- a function of the deviation value. The larger the
lation results in a "fuzzy value" between - 1 and deviation, the more frequent the control actions.
+ 1, which expresses how far the actual process Priority management
situation is from fulfilment of the objective. The rule block for each objective results in con-
Rule block trol actions which are multiplied by a weight factor
The rule block holds the set of control rules for between 0 and 1. The smaller the weight factor, the
fulfilment of the specific objective. Normally, the more the control actions are suppressed. The
rule block is formulated as a set of fuzzy control weight factor for a given objective is a function of
rules using FCL, but other techniques may also be the deviation values for the higher priority objec-
used, like PID, neural networks, mathematical tives, which ensures that control actions to fulfill
models, etc. a given objective are reduced if higher priority
Output block objectives are not fulfilled.
The output block evaluates process constraints The priority management system reflects our
and selects among alternative control actions based "fuzzy way of thinking", as it is designed to give
 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146 143

a smooth and gradual transition between the objec- As shown in Fig. 11, the timing program calcu-
tives, as opposed to the non-fuzzy approach, lates the time interval between control actions, ap-
in which only one objective at a time would be plying a linear function of the absolute value of
active. DEVI. The closer the oxygen is to either the low
Years ago, oxygen control was used to illustrate limit or the high limit, the more frequent the con-
the way FLS have applied the theory of fuzzy logic trol actions.
to process control [5]. At that time, the purpose The output block, shown in Fig. 12, implements
was to explain membership functions, the combina- the evaluated increase or decrease or the draught
tion of rules and the fuzzy control language. Now by adjustment of the I.D. fan speed or the damper
oxygen control will be used to illustrate the struc- position, or by another action if the I.D. fan speed
ture of a control objective module. or damper position has reached a high or low limit.
The deviation calculation, shown in Fig. 9, con- The alternative actions will be decided on the basis
sists of a transformation of the actual oxygen per- of the actual values of the state indices.
centage into a deviation value (DEVI) between - 1
and + 1, applying the user-specified low limit, set-
point and high limit. 6. The next generation
The rule block, shown in Fig. 10, specifies more or
less draught as a function of the deviation value. The fuzzy system has been used for rule-based
For instance, a negative DEVI value indicates that implementation of single input/single output
oxygen is too low, and the control rules specify an (SI/SO) control loops. Even though an increasing
increase in the draught. trend has been registered for beneficial application
The weight factor W, calculated by the priority of more advanced control schemes to solve SI/SO
management, determines the actual degree of influ- problems, our main application of the fuzzy system
ence of the oxygen control objective. W is a func- is still high-level process control, which may be
tion of the deviation values for the higher priority defined as co-ordination of single loop process con-
objectives. The larger these higher priority devi- troller setpoints, normally done by a human
ation values, the more the oxygen control actions operator.
will be suppressed, whereas low values of the higher To a certain extent, the Fuzzy II system acts as
priority deviations will allow oxygen control ac- a human operator by taking over most of his job of
tions to be executed. controlling the process, including co-ordination of

+1

) 02

-1

Fig. 9. Deviation calculation.

144 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135 146

Weight factor (W)

Rule block
if upos (oev~) the. L.NeG (/1DRAF)
If MPO6 (Dlvf) thin MNEG (A DRAF)
If SPO6 (Devl) then SNEG (A DRAF) A t~t~ A w,A am~r
If ZF~IO (De,d) then (ADRAF) V
If SNEG (Devl) then SPOS (A DRAF)
If MNEG (Devl) then MPOS (ADRAF)
If LNEG (De~ the. L.nOS (ADRAF)

Fig. 10. Rule block for oxygen control.

I~vl Process constraints

& State indices

Timing calculation
I,D. Fan
Output b l o c k
Coal
Implementationof control
action considering: Speed
w . A m~F • Procelm constraints
MIn • Plro~mll indices

Fig. 12. Output co-ordination for oxygen control.

W*A DPaI~

Fig. l 1. Timing calculation for oxygen control. fuzzy controller is generally well accepted by the
operators, but there is no doubt that the system will
be even better if conflicts can be avoided and real
setpoints and decisions on production level. co-operation between the system and the operator
A Fuzzy II system may thus be used to reduce the can be established.
number of operators in the control room. However, It is even possible that the conflict is of a psycho-
an unmanned control room at a cement plant is not logical nature. The operator may feel that the com-
yet possible, nor is it a goal that one should aim for. puter has been appointed to do the j o b better than
If the operator does not agree with the way the he can do it. Perhaps he is of the opinion that the
fuzzy system controls the process, he can switch off fuzzy controller has taken away the most interest-
the fuzzy controller and control the process man- ing part of his daily work. The operator, in other
ually. If the operator judges rightly that the fuzzy words, does not consider the fuzzy his tool for
controller is not producing optimal operation, then producing optimal process control but a competi-
the need is simply to adjust the control strategy. tor that he is very much inclined to switch off at the
If, however, the operator does not judge the first opportunity.
situation correctly, a conflict situation may arise The fuzzy logic development work within FLS
between the operator and the fuzzy system. The Automation is thus focusing on:
 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146 145

Priority Management

I Deviation
-I Process calculation

WelgMfactor I Tlmlng I
,,,t
Output
block

Slate
Index

r
Fig. 13. Fuzzy II system applying operator's information.

"How to get the operator back into the driver's in a structured way as an active part of the auto-
seat". matic control system.
The goal is still to approach optimal operation Fig. 13 shows a diagram of an operator-activated
through automatic control of the process. But the control objective. The neural net has been trained
high-level control system should be placed in the to reveal key information about the actual process
hands of the operator to enable him to use the state, the so-called state indices in Fuzzy II termin-
system as a tool instead of the either fuzzy or ology. The operator may judge the same index
operator control situation. The fuzzy control sys- variables and a Fuzzy II control objective has been
tem should be able to accept and process inputs defined to cope with the situation in which the
from the operator, who very often possesses useful neural net and the operator do not agree.
information for the achievement of optimal opera- In this way, the operator has the possibility of
tion. A control system that works with inputs from interfering with the automatic control strategy. His
a human operator who expresses himself in every- interference is not a necessity, but when he inter-
day terms is another step in the direction of ap- feres, it is done in a well-defined way that has been
plying the basic ideas of fuzzy logic. decided through the definition of the special oper-
For future high-level control systems, FLS Auto- ator-activated control objective.
mation is working with integration of various con-
trol and prediction techniques. In our opinion, the
optimal overall control system is most likely to be
a combination of the rule-based fuzzy logic References
approach, neural network techniques and model-
[1] S. Assilian and E.H. Mamdani, Learning control algorithms
based systems, with the possibility of operator in- in real dynamic systems, Proc. 1FAC/IFIP Int. Conf. on
terference. In particular, neural nets seem to be Digital Computer Applications to Process Control, Zfirich
a promising technique for integrating the operator (Springer, Berlin, 1974).
146 L.P. Holmblad, J.J. Ostergaard / Fuzzy Sets and Systems 70 (1995) 135-146

[2] F. Eshragh and E.H. Mamdani, A general approach to [5] K.E. Peray and J.J. Wadell, The Rotary Cement Kiln
linguistic approximation, lnternat. J. Man-Machine Studies (Chemical, New York, 1972).
11 (1979) 501-519. [6] L.A. Zadeh, Outline of a new approach to the analysis of
[3] L.P. Hoimblad and J.-J. Ostergaard, Control of a cement complex systems and decision processes, IEEE Trans. Sys-
kiln by fuzzy logic, FLS-review No. 67, FLS Automation, tems Man Cybernet. SMC-3 (1) (1973) 28-44.
Ramsingsvej 32, DK-2500 Valby, Copenhagen, Denmark.
1-4] J.-J. l~stergaard, Fuzzy logic control of a heat exchanger
process, in: Gupta, M. Madan et al., Eds., Fuzzy Automata
and Decision Processes (North-Holland, New York, 1977).