Engng Applic. Artif. lntell. Vol. 8, No. 4, pp.

413-421, 1995
Copyright ~ 1995 Elsevier Science Ltd
Pergamon Printed in Great Britain. All fights reserved
0952-1976(95)00020-8 0952-1976195 $9.50 + 0.00

Contributed Paper

Expert Control Systems I. Concepts, Characteristics

and Issues
D. A. LINKENS
University of Sheffield, U.K.

MINYOU CHEN
Chongqing University, P.R. China
(Received January 1995)

Expert system techniques have been rapidly disseminating into process control engineering, with
many applications being reported within the last decade. Expert control is bringing a new perspective,
as well as new challenges to process control. This paper presents a brief overview of expert control
systems and discusses the impact of expert-system techniques on control engineering. The basic ideas,
concepts, and fundamental characteristics of expert control systems are elucidated. An emphasis is
made on distinctions between conventional expert systems, traditional advanced control systems and
expert control systems. A generic architecture for expert control systems is presented. Some important
issues relating to the realisation of expert control systems in engineering practice are also investigated.

Keywords: Expert systems, intelligent control, engineering applications, characteristics of expert

control systems, implementation issues.

1. INTRODUCTION by control theory alone, but heuristic logic plays an

important role. 2
Although traditional control theory based on mathema-
tical models has dealt successfully with many control Artificial intelligence, in particular expert system
problems, there still exists a large number of practical techniques, provides powerful tools to deal with com-
problems which cannot be solved by existing control plex real-world systems and to aid in the transfer of
system technology in process control engineering. knowledge to engineering practice. Within the last
These problems often arise in situations where plants decade, expert-system techniques have been develop-
are large, complex or uncertain, or there may appear ing rapidly in control engineering. Through the agency
severe changes in operating conditions, such as those of expert-system techniques, knowledge in both docu-
found in chemical processes, metallurgical processes, mented and empirical forms can be utilized to advan-
biomedical systems, etc. Conventional advanced tage. In every area in which problems can be solved
control (AC), e.g. robust control or adaptive control, mainly by knowledge and reasoning instead of analytic
can solve uncertainty only to a limited extent and in computation, expert systems offer effective and avail-
certain cases, and has been considered mainly in the able approaches. Applications of expert-system tech-
aerospace industry. Other industries have mainly incor- niques in control engineering have ranged from
porated PID controllers because they are cheap and control-system design, fault diagnosis, simulation,
reliable, work well, and are easily understood by the modelling and identification, to on-line performance
field engineers and operators. 1 In fact, any practical monitoring, adaptation and auto-tuning, and super-
control problem, even to PID control, cannot be solved visory control. In many cases, expert-system techniques
can provide good closed-loop control where conven-
tional control may be unsuccessful. In chemical pro-
cesses, metallurgical processes and the cement indus-
Correspondence should be sent to: Professor D. A. Linkens, try, expert control can play an important role, and
Department of Automatic Control and Systems Engineering,
University of Sheffield, P.O. Box 600, Mappin Street, Sheffield
many successful applications have been reported in
S1 4DU, England, U.K. recent years. 3-16
413
414 D.A. LINKENSand MINYOUCHEN: EXPERT CONTROLSYSTEMS--I

However, there still exist some barriers to the imple- time-varying dynamics is thereby achieved. Two new
mentation of real-time expert control systems. The task adaptive schemes, inductive inference and an expert
of investigating the basic ideas, fundamental character- system, are used. The inductive inference system learns
istics and implementation methods of real-time expert the relationship between certain control-system charac-
intelligent control systems is challenging, and is the teristics. This knowledge is used by the expert system to
subject of this paper. detect changes in plant dynamics and to recommend
controller settings that are suited to the new plant
2. OVERVIEW OF EXPERT CONTROL SYSTEMS dynamics. The inductive inference system is a feature
IN PROCESS ENGINEERING of a commercially-available expert-system shell. The
adaptive control scheme is efficient and runs entirely on
Expert control (or knowledge-based control) refers a personal computer. The scheme has been demon-
to methods that utilize expert-system techniques and strated on a single- and two-term controller. Many
control theory to design control systems that can auto- similar researches and applications have been
mate some of the tasks currently performed by human reported) °-23 This type of application could be con-
experts, and which cannot be carried out by traditional sidered as rule-based auto-tuning or adaptive control.
control systems. Since the breakthrough in expert On-line fault diagnosis is a necessary component in
systems (ES) made during the 1970s, interest in apply- intelligent control systems. Within the last decade,
ing expert-system techniques to control engineering has many research results and applications of expert super-
been growing. Early applications of ES in control visory control have appeared in control engineering.
engineering have mainly been focused on off-line PICON is an early expert supervisory control system
systems, such as computer-aided control-system for process fault analysis and diagnosis developed by
design, ~7simulation, ~8and diagnosis, etc. Within these Lisp Machines Inc. 5"6 It could use on-line data to
applications, expert systems play a passive role, per- perform the system diagnoses and alarms. PICON has
forming off-line for consultation or supervision. Such been succeeded by the commercial software package
types of system are supported by conventional ES G2. Dorais et al. use an expert supervisory system to
techniques which are static in both knowledge rep- monitor dynamic behaviour, such that if it deviates
resentation and inference strategies. With the develop- from the normal situation, the supervisory controller
ment of real-time expert control techniques, more switches to the abnormal control mode and hence
research and applications of expert control systems handles the emergency situation. 24
(ECSs) have emerged in process engineering. Within A supervisory scheme aimed at stabilizing adaptive
the last decade, two areas where much work has been controllers was proposed by Liu et al. 2s The scheme
done are expert tuning (or adaptive controllers) and works on the basis of heuristic (expert) rules and
expert supervisory control systems. includes a monitor, a tuner and a structure identifier.
A successful application of an expert controller is the The monitor detects impending instability via the
Foxboro EXACT, 3 which combines heuristics about analysis of the amplitude and trend of the error signal.
control parameter tuning with a conventional PID con- Upon detecting instability, the controller is switched to
troller. EXACT could be considered as a performance a safer algorithm until stability is restored.
adaptive controller embodying the desired damping Simultaneously, new data are gathered and the plant
and overshoot. On the basis of pattern recognition of structure is re-identified. Based on the new structure, a
transients in control error signals, PID parameters are new controller is designed and normal adaptation is
changed to achieve the performance indices by using restored. In similar research, a hierarchical expert
heuristic rules. EXACT has been used widely in indus- system as the supervisory level in adaptive control has
trial process control. Another example is an expert PI been presented by Morant et al. 26
controller proposed by Porter, 4 which is based on step An integrated supervisory control system has been
response analysis. The controlled process is classified presented by Leitch et a1.,27 consisting of generic tasks
according to some qualitative transient response such as adaptive control and fault diagnosis. The super-
features, e.g. monotonic, oscillatory, overshoot, time vision task involves defining a supervisory cost function
delay, etc. The PI controller parameters are adjusted which operates as a boundary-detection mechanism for
by heuristic tuning rules for different process classes. the system behaviour, which is classified into accep-
Introducing rule-based heuristics into adaptive table behaviour, malfunction behaviour and fault
control is a significant advance. An expert adaptive behaviour. The results of simulation and prototype
control scheme is presented by Reyneveld and Dorf.~9 experiments demonstrated that the system could main-
It is useful for automatic control systems that consist of tain good performance under plant parameter changes
a standard, single- or multiple-term controller and a and faults, as well as rejecting the effect of distur-
controlled plant whose dynamic characteristics vary bances.
with time. The scheme works by updating the con- In order to meet the higher requirement of current
troller's settings so that they are continually matched to industrial processes, some knowledge-based control
the plant's dynamics. Good control of a plant with systems that integrate specific knowledge in different
 D. A. LINKENSand MINYOUCHEN: EXPERT CONTROLSYSTEMS--I 415

domains, such as fuzzy logic, expert systems, pattern 3. CONCEPTS AND CHARACTERISTICS OF
recognition and neural networks, have been developed EXPERT CONTROL SYSTEMS
in recent years. Chang and Birdwell proposed a dis-
tributed real-time expert control system based on a Expert systems (ESs) are computerized programs
multiple processing architecture and parallel processing that attempt to imitate the inference process and
techniques. 28 The system is composed of a hierarchical knowledge of human experts in solving special types of
structure with a global knowledge manager at the problems. ES techniques provide a framework for
higher level, and sub-expert systems which contain blending numerical algorithms with the expertise of the
specific domain knowledge in the lower level. A black- control engineer and plant operator. There have been
board technique is used to co-ordinate and control the many attempts at the definition of expert systems, but
different knowledge sources. An on-line learning and only a few on expert control systems. Expert control
fault-tolerant structure is proposed to aid the system (EC) was originally proposed by Astr6m e t al. 3~ One of
robustness. This system has been applied to an indus- the main objectives in applying ES techniques to pro-
trial process. cess control is to: extend the range of conventional
Linkens and Abbod presented a supervisory intelli- control systems, e.g. auto-tuning and adaptation;
gent control system which is constructed using a fuzzy reduce the computing burden of traditional control
logic controller in the direct control level and a fuzzy systems; handle a large amount of sensor data and
knowledge-based supervision in the decision level) 9 uncertain information; and lead to the simplification
This supervisory control system consists of generic and adaptation of conventional systems. A key point of
control tasks for keeping the process on-line within EC is the incorporation of heuristics and logic through
acceptable behaviour boundaries. If a fault occurs, the knowledge-based structures, thus making the control
control task passes first through a primary direct systems more flexible and adaptive than conventional
control regime, then an adaptive control regime, and control systems. 2
finally a fault-diagnosis regime, each attempting to Also, with the growing complexity of process plant
drive the system behaviour into an acceptable zone. and the need for higher efficiency, greater flexibility,
The system has been used effectively in a biomedical better product quality, and lower cost, some new tasks
system and an electro-fluid process. This architecture have been proposed for process control. One of the
could be mapped between medical and industrial appli- challenging tasks is to design intelligent control systems
cations with very little change. which can increase the ability to cope with uncertainty
A real-time expert control system based on a multi- and produce better performance, while guaranteeing
mode intelligent control algorithm, a pattern- system stability. Therefore, a rule-based system which
recognition technique, and knowledge-based on-line pieces together traditional control approaches, such as
fault diagnosis has been investigated by Chen and PID, at a lower level, with loose connection to a rule-
Zhou, 3°'31 and successfully applied in an industrial based controller at a supervisory level, cannot fulfil
rotary kiln. these tasks.
On the other hand, a number of researches on the There exists a misunderstanding that building an
combination of expert-system techniques with neural ECS is simply adding a few rules to a conventional
networks have been reported. Alice and Hulya pre- controller. Actually, designing an ECS is a complex
sented an intelligent composite system which combined task. Multiple knowledge-based and control techniques
statistical process control, artificial neural networks and should be used to build an ECS. As a real-time
an expert system for analysis and control of a plastic knowledge-based control system, an ECS contains
extruder facility. 32 This intelligent approach made the more difficult problems than most existing ESs. In
process analysis and control decisions comprehensive almost all existing expert systems (including design,
and standardised, and thus improved product quality diagnosis, planning, and repairing ESs) the operation is
and reduced the cost. a consultation process. The system usually operates in a
With the growth of system complexity, increases in non-real-time environment, and the human is always in
on-line data and information, and the requirement for the decision-making loop, providing relevant infor-
high performance, simply combining rule-based tech- mation when requested. However, an ECS operates in
niques with conventional control algorithms cannot a continuous real-time environment. It uses real-time
meet the needs of current industrial processes. The information-processing schemes to supervise the dyna-
rapid development of expert-system techniques is mic system performance and give relevant control ac-
bringing about a great impact on control engineering. tion to keep the process under good regulation.
Increasingly, traditional control systems are being Because of the requirement for real-time control and
modified or replaced by intelligent control systems decision-making, an ECS must minimise the time taken
(ICSs) which apply expert systems/artificial intelligence for reaching an appropriate decision in case of system
techniques in computer process-control systems. contingencies.
Expert control has become one of the major Therefore, an ECS is different from an ES in many
approaches to realize intelligent control. aspects. First, the structure and operating mechanism
416 D. A. LINKENS and MINYOU CHEN: EXPERT CONTROL SYSTEMS--I

Table 1. Comparison of conventional expert systems and expert control systems

Dimensions Expert control systems Conventional ES

Execution speed High, real-time operation Lower, mainly consultation

Knowledge base Usually small and simple Usually large and complex
Knowledge source Experts and on-line learning Expertise
Reasoning Symbolic and numeric, but fast Powerful, mostly heuristic, but
time-consuming
Explanation Little Extensive
Implementation Usually conventional languages Usually AI programming

of an ECS cannot be as complex as a conventional ES possess all the major features of expert systems, such as
because of its real-time application. Also, the knowl- transparency, but they must possess the main features
edge sources of an ES come mainly from expertise or of advanced control systems, such as real-time control
prior knowledge, but in an ECS on-line information and supervision, reliability, and adaptation. More pre-
processing and learning play important roles. cisely, ECSs are referred to as knowledge-based intelli-
Furthermore, ESs usually have a large knowledge base, gent control systems. The primary distinctions between
a powerful symbolic inference engine and an explana- an ECS and an ES are shown in Table 1. In comparison
tion mechanism in order to meet the need for complex with traditional advanced control systems, the funda-
problem-solving. However, ECSs usually have a small mental characteristics of an ECS are its knowledge-
knowledge base, and effective but simple temporal based structure and the ability to deal with uncertainty.
reasoning mechanisms which use both symbolic reason- Although many new approaches have been developed
ing and numerical computation to realize real-time to improve the ability to deal with uncertainty for
operation. The response time should be fast enough to traditional advanced control systems, for instance
control and supervise the process. In addition, ESs are robust control, adaptive control, etc., they are difficult
often implemented via AI programming techniques and to use in industrial processes because of their purely
languages, such as LISP and PROLOG, while ECSs analytic structure, linear and time-invariant con-
are usually implemented via conventional languages, straints, and poor user understanding. By introducing
such as C, BASIC, FORTRAN, Pascal, etc., so as to ES techniques, an ECS can provide some benefits to
increase the execution speed. Sometimes, however, AI control systems, such as flexibility, reliability, and the
languages may be used in the high levels of ECSs for ability to deal with incomplete or uncertain infor-
system management and coordination. Finally, ESs mation, and make the control system repeatedly inter-
usually have good explanation mechanisms to clarify pret the current situation, predict the future, diagnose
their reasoning, recommendations, or other actions. In the causes of anticipated problems, formulate a
contrast, ECSs pay more attention to real-time ope- remedial plan, and monitor its execution to ensure
ration, reasoning and decision-making, and less atten- control performance. The primary distinctions between
tion to explanation. an ECS and traditional AC are described in Table 2.
Thus, ECSs cannot be considered simply as conven- Functionally, the distinction between expert control
tional expert systems with real-time application to a and traditional control lies in the degree to which the
dynamic control environment. In concept, the defini- system can deal with uncertainty and complexity. The
tion of an ECS is wider than that of a control expert fundamental functions of ECSs are to:
system. The ECS may not be in the form of a typical (1) take over the skilled operators' routine tasks
expert system, but could be in versatile forms including and give effective controls for processes which
control expert system, expert tuning controller, expert are time-varying, non-linear, and subjective to
supervisory control system, expert adaptive control various disturbances.
system, expert fuzzy control system, etc. ECSs may not (2) take advantage of all the available prior

Table 2. Comparison of expert control and traditional advanced control

Dimensions Expert control Advanced control

System structure Knowledge-based Model-based

Information processing Symbolicor numeric Numeric
Knowledge s o u r c e Document and/or empirical Document
Nature of input May be incomplete Must be complete
Search Heuristic and algorithmic Algorithmic
Process model May be incomplete or qualitative Mustbe precise
Maintenance and update Relatively easy Usually difficult
Explanation Can be provided Usually not
Execution Heuristic, logic and algorithmic Purely algorithmic
User understanding Easy Difficult
 D. A. LINKENSand MINYOUCHEN: EXPERT CONTROL SYSTEMS--I 417

I Learning I_
subsystem r

@ aquisition H "-'°*
°°"°* processor H
base H '---- Sensors

[ interface
,.,se, H mechanism
",,,,'ationH 'nengine
torfaceH Icontroller
ntoi"gentH Actuators
Fig. 1. Structure of an expert control system.

knowledge and on-line information; Existing expert control systems only cover some of the
(3) perform fault diagnosis on the control system above functions to a certain degree.
operation and components, including the In contrast with traditional control systems,
detection of actuator and sensor problems; knowledge-based expert control systems are particu-
(4) operate reliably and conveniently; larly suitable to deal with complex, uncertain or ill-
(5) increase the amount of process knowledge, defined plants. Generally speaking, suitable application
and accordingly improve the control system's areas for ECSs are:
performance; (1) ill-structured processes for which mathemati-
(6) represent control knowledge in an effective cal models do not exist or are inadequate;
way which easily allows for modification and (2) complex problems which require answers
extension; within a limited time interval, such as fault
(7) maintain dialogue with the user and give diagnosis and emergency handling;
explanation of reasoning results, and also (3) situations where expertise is required for
obtain information from the user; problem-solving but where there are not
(8) require a minimal amount of prior knowledge; enough experts for the task;
(9) have a capability for real-time reasoning and (4) situations where qualitative or uncertain infor-
decision making. mation must be processed, and symbolic logic
Hence, a general architecture for expert control is required for problem-solving;
systems is proposed, as shown in Fig. 1. (5) complicated problems where a heavy comput-
In this architecture, the knowledge base is indepen- ing burden and high cost would be involved
dent of the knowledge-processing mechanism. The when using conventional algorithmic methods;
system is highly flexible and easily extended and modi- (6) cases where operating conditions change fre-
fied, due to its modular structure. The knowledge base, quently and/or severely.
which stores the specific domain knowledge about a Finally, as a summary of the above discussion, tenta-
process, control engineering principles, and experience tive definitions are given as follows:
of control experts and operators, etc., is the basis for • Expert control (or knowledge-based control) is
the system's operation and is large-scale. The task of one of the intelligent control methods, which
the inference engine is to achieve problem-solving combines control theory and expert-system
according to a certain reasoning strategy, using knowl- techniques to design and realise in the auton-
edge and on-line information. The functions of the omous operation of complex, uncertain or ill-
characteristic information processor are: processing on- defined physical processes.
line information, and extracting and recognizing the • An ECS is an intelligent control system which
characteristic information which can show the current uses expert-system techniques on difficult
status and tendency of a dynamic process, and provid- control problems where analytic models do not
ing useful information for the decision-making pro- exist or are inadequate, and require expert
cedure of the knowledge base and inference engine. knowledge for their problem-solving.
The learning subsystem refines the knowledge base
according to on-line characteristic information, and
4. CLASSIFICATION OF EXPERT CONTROL
increases the ability to solve problems. The knowledge-
SYSTEMS
acquisition subsystem encodes the expert's knowledge
automatically or semi-automatically. The explanation Although the development of expert control has
mechanism explains the procedures and results of reas- taken place over only about 10 years, various types of
oning to operators. Plant operators can input instruc- expert control systems have appeared in control engin-
tions or index through the knowledge base via the user- eering. According to the systems' structural and func-
interface, and also monitor the system's operational tional implementation methods, expert control systems
status. These represent the desired features of an ECS. could broadly be classified into four types.
418 D.A. LINKENS and MINYOU CHEN: EXPERT CONTROL SYSTEMS--I

4.1. Rule-based expert tuning or adaptive controllers

Rule-based auto-tuning controllers are becoming
common in process control. The knowledge required
for the efficient tuning of conventional controller par-
ameters, e.g. PID, is acquired from both control engin-
~ Knowledge-based system I
System management [ ~
Fault detection & diagnosis ~ Identification [ ~
Performance monitoring ] I
1
•
T
I [

eers and plant operators, and is stored in a knowledge

base in the form of I F . . . T H E N . . . ELSE rules.
During system operation, the characteristic behaviour
of the process is captured by a pattern classifier or Fig. 3. Block diagram of an expert supervisory control system.
identifier. An inference engine uses the tuning rules
and classified patterns to tune the controller parameters The decision-making in the supervisory control
automatically, such that the system's performance is system is related to situations involving major distur-
improved. A general architecture of a rule-based auto- bances, technical faults, inappropriate human actions,
tuning controller is shown in Fig. 2. and a combination of such events. More tasks can be
Expert tuning controllers provide an architecture for handled by a supervisory control scheme, e.g. start-up
the combination of real-time algorithms (from simple and shut-down procedures, process optimization, fault
PID to adaptive control) and logic. /~strom proposed diagnosis, response to malfunctioning behaviour, start
an Intelligent PID controller, 34 in which the different and stop parameter estimation procedures, and alarm-
control algorithms, such as P, PI, PD, and PID, and the handling procedures.
control parameters could be selected and adjusted Supervisory expert control systems are an important
according to the knowledge base and classification of type of knowledge-based control systems. They are
processes, such as the range of normalized dead-time usually used in process and/or manufacturing systems in
and gain of the processes. order to obtain higher quality, lower cost, fault diagno-
Rule-based expert tuning or adaptive controllers sis, emergency handling, and hazard prediction.
have been widely applied in process-control engineer-
4.3. Hybrid expert control systems
ing because they are easy to implement in microproces-
sors and are readily commercialized. The combination A hybrid ECS is a composite intelligent control
of algorithms and heuristics allows the controller signi- system which utilizes a multi-layer hierarchical struc-
ficant simplification in programming and modification. ture and the incorporation of various techniques,
It is an effective and simple way of implementing including expert systems, pattern recognition, fuzzy
intelligent process control. logic, neural networks, and computer process control.
Because of the requirement for multiple knowledge
4.2. Expert supervisory control systems sources, a blackboard architecture is commonly used in
This is a popular approach for the implementation of hybrid expert control systems. A "blackboard" is a
expert control systems, by introducing ES techniques working memory that maximizes independence among
into the supervisory levels of control systems. Expert knowledge sources by appropriately dividing the prob-
supervisory control systems are concerned primarily lem domain. This architecture can accept many types of
with on-line identification, process monitoring, fault knowledge. The user may store or retrieve information
detection and diagnosis. The structure of an expert from any of the knowledge sources. The blackboard
supervisory control system usually comprises a direct provides a centralized way of recording and tabulating
control level which involves signal processing and a intermediate decisions about a problem.
conventional control algorithm, and a supervisory level A hybrid ECS can effectively address the com-
which consists of a knowledge base and an inference promise between completeness and simplicity, improve
engine for on-line performance monitoring, fault detec- the system's performance, and increase the system
tion and diagnosis. Supervisory control pays more functions by cross-fertilization between different disci-
attention to the system supervision, objective optimiza- plines. For instance, a rule-based approach can only
tion, fault analysis and diagnosis, emergency handling, deal with the problems in specific domain, but a neural
and decision-making. Figure 3 shows a typical architec- network has the capability of parallel processing and
ture for supervisory expert control systems. on-line learning, and can be used in generic situations.
The combination of ES techniques and neural networks
would bring great benefit to control systems.
Tuning ~ Pattern A hybrid ECS possesses some outstanding features,
rules I ~ " ~ classifier
such as non-monotonic reasoning, complete pattern
matching, object-oriented programming, meta-rules,
Plant _~ I
Control [ _ _ ~ Controlled and real-time capabilities. Although hybrid expert
operator algorithm [ [ process
control systems are not as mature as rule-based con-
trollers or supervisory expert control systems, and
Fig. 2. Block diagram of a rule-based auto-tuning controller. there are some problems for further investigation, it is
 D. A. LINKENSand MINYOUCHEN: EXPERT CONTROL SYSTEMS--I 419

evident that the development of hybrid expert control information, missing sensor data, etc. Expert control
systems is significant and important for intelligent pro- systems should find an effective way to collect, cata-
cess control. logue, extract, dispense and organize the on-line infor-
mation so as to support the system's reasoning and
4.4. Real-time control expert system decision-making in these areas.
Another way to realize expert control is to build a
real-time control expert system using a knowledge (4) Multiple knowledge representation
engineering methodology, with the design principles Most existing expert control systems used rule-based
and implementation approaches of ES. Hence, a techniques. This approach has proved effective for
control expert system is a typical real-time expert representing shallow knowledge, e.g. empirical rules,
system with all the characteristics of an expert system, heuristics, etc. However, some problem-solving
such as modularity (flexibility), heuristics and trans- requires deep knowledge regarding the internal and
parency, as well as the features of a control system, e.g. causal structure of the system, and fundamental princi-
real-time operation, reliability, and adaptation, etc. ples of the domain behaviour, etc. Therefore, develop-
Real-time expert systems usually have a complex struc- ing better methods for representing both shallow and
ture, a powerful inference capability, and relatively deep knowledge effectively is a current task in building
complete functions. A general structure of an expert ECS. One possible way is to combine rule-based tech-
control system is shown in Fig. 1. Developing a real- niques with frame-based or object-oriented techniques.
time expert control system is a very difficult task
because of the tough requirements in closed-loop (5) Real-time intelligent control algorithms
control, on-line information processing, dynamic reas- The majority of existing expert control systems intro-
oning, on-line self-learning and knowledge refinement, duce intelligence into the higher level of control
process supervision, user-interfaces and explanation. systems, such as tuning or adaptation, monitoring, and
supervisory levels, and pay less attention to intelligent
5. IMPLEMENTATION ISSUES FOR ECSs control at the real-time level. Actually, developing
various real-time intelligent control algorithms with a
In contrast to the majority of expert system appli-
simple structure, high execution speed and good per-
cations, expert control is an on-line real-time appli-
formance would improve the system's performance and
cation. As such it contains several difficult problems,
increase the degree of intelligence greatly. This is a
relating to its real-time nature.
significant task.
(1) Real-time operating knowledge base
(6) Non-monotonic reasoning
In expert control systems, real-time operation and
system reliability require that the knowledge base In a dynamic system, data as well as inferred facts are
should be small in scale and simple in structure, and the either not durable, and decay with time, or they cease
knowledge and on-line information should be well to be valid due to events which have changed the state
organized. On the other hand, the complexity of plant, of the system. In order to maintain a consistent view of
variation of environment, and completeness of system the environment, the reasoning system must be able to
functions require multiple knowledge sources and a automatically retract inferred facts.
large amount of knowledge, thus resulting in a compli-
(7) Temporal reasoning
cated knowledge base and more time taken in search
procedures. Simplicity and completeness are two con- Time is a very important variable in real-time super-
tradictory requirements. A compromise between com- vision and control systems. The best possible solution
pleteness in functions and simplicity in structure is within a given deadline is desired. It is necessary to
necessary. Therefore, the building of an appropriate reason about a number of past, present and future
real-time operating knowledge base should be investi- events. This topic also includes the problem of estimat-
gated further. ing the time needed for internal reasoning activities.

(2) On-line learning and knowledge refinement (8) Parallel reasoning

The domain knowledge in an ECS is not always It is often natural to view problem-solving activities
readily available, especially in changing situations. as composed of a set of parallel reasoning activities.
Therefore, on-line learning and rule modification or Parallel reasoning activities impose requirements on
generation are important in expert control systems, but synchronization between different activities, and possi-
that is a very difficult task. bilities for suspending an activity for a certain time or
until a certain event has occurred.
(3) Effective on-line information extraction and organ-
ization (9) Intelligent interface
Industrial process control has to attempt to deal with In expert control systems, interfaces must be
various aspects of information, including uncertain provided to both conventional software and the user
420 D . A . LINKENS and MINYOU CHEN: EXPERT CONTROL SYSTEMS--I

environment. Thus, an effective multimedia interface 5. Moore R. L. et al. A real-time expert control system for process
control. Proceedings of the First Conference on Artificial
with a user-friendly style is needed. Intelligence Applications, pp. 569-576 (1984).
Many of the problems described above are extremely 6. Knickerbocker C. G. and Moore R. L. The PICON: expert
difficult, and far from a solution. In many cases, exist- system for process control. Expert Systems and Their
Applications 5th International Workshop, pp. 59-66 (1985).
ing expert control systems only deal with some of the 7. Francis J. G. and Leitch R. Intelligent knowledge-based process
problems to some degree. The way to solve these control, lEE International Conference on Control '85. Vol. 2, pp.
problems is to adopt a multi-disciplinary approach 483-488 (1985).
8. Arzen K. E. Use of expert systems in closed loop feedback
which utilizes the research results from different fields, control. Proceedings of ACC (1986).
including expert systems, pattern recognition, fuzzy 9. Wright M. L. et al. An expert system for real-time control. IEEE
logic, and neural networks. Software 16-24 (1986).
10. Moore R. L., Hawkinson L. B., Levin M., Hoffman A. G.,
Matthews B. L. and David M. H. Expert systems methodology
for real-time process control. Proceedings of the lOth IFAC
6. CONCLUSIONS World Congress on Automatic Control (1987).
11. Chen Minyou and Zhou Qijian. Intelligent control for pH-
As a relatively new technique, expert control has process. Inform. Control 16, No. 2, 1-6 (1987).
become attractive to the process-control community. 12. Chen Minyou and Zhou Qijian. An expert controller for indus-
The rapid development of expert control systems has trial processes. J. Chongqing Univ. II, No. 3, 24-32 (1988).
opened a new door towards the realization of intelli- 13. Reath P. G. Expert systems in process observation and control.
AI Expert 5, No. 9, 54-59 (1990).
gent automation. Although ECSs are significantly 14. Lou An, Jia Xueyan and Qiusi Xiong. An intelligent control
different from traditional control systems, both in system for industrial furnace. Proceedings of IECON'90. 16th
design philosophy and in operating procedures, they Annual Conference of lEEE Industrial Electronics Society Vol. 1,
pp. 62-65 (1990).
are not a replacement for but an extension and modifi- 15. Rosenof H. P. Expert systems come to batch process automa-
cation of conventional control systems. Rule-based tion. 1SA Trans. 31, No. 2, 67-76 (1992).
expert tuning and adaptive control, and expert super- 16. Otsuka K., Matoba Y., Kojima M., Yoshida M. and Matsuoda
S. Expert system for blast furnace operation. Sumitomo Search
visory control systems, are two major implementation No. 50, 43-50 (1992).
forms in ECSs. They have been used successfully in 17. Taylor J. H. and Frederick D. K. An expert system architecture
process engineering. Hybrid expert control systems for computer-aided control engineering. Proc. IEEE 72, 1795-
1805 (1984).
comprise different disciplines in artificial intelligence, 18. Shannon R. E. and Mayer R. Expert systems and simulation.
fuzzy logic, pattern recognition, and neural networks, Simulation 44, 275-284 (1985).
and are becoming an important tendency in the devel- 19. Reyneveld K. E. and Doff R. C. An expert adaptive control
system. Twenty-third Asilomar Conference on Signals, Systems
opment of intelligent control systems. Expert-system and Computers, Vol. 2, pp. 1013-1016, 1989.
techniques are changing the face of current process- 20. Lieslehto J., Tanttu J. T. and Koivo H. N. An expert system for
control systems: from purely analytic model-based tuning PID controllers. 4th IFAC Symposium on Computer
Aided Design in Control Systems (CADCS'88), Beijing, P.R.C.,
systems to knowledge-based synergetic information pp. 524-529 (1988).
systems; from simple feedback control systems to com- 21. Davanathan R. et al. An expert PID controller. Proceedings of
posite intelligent systems involving real-time control, IEEE International Workshop on Artificial Intelligence for
Industrial Applications (1988).
fault diagnosis and system management. Although 22. Ollero A., Jimenez J., Cerezo A. and Silva C. F. Expert tuning
there exist some major issues for further investigation, of automatic controllers. 4th 1EEE Symposium on Intelligent
expert control has a vast potential in intelligent process Control, pp. 207-216 (1989).
23. Neat G. W., Kaufman H. and Roy R. J. Expert adaptive control
control systems. Via the integration of ES techniques, for drug delivery systems. 1EEE Control Syst. Mag. 9, No. 4, 20-
pattern recognition, neural networks and fuzzy logic, 24 (1989).
expert control will be developed significantly, and will 24. Dorais R. and Jiang J. Performance monitoring in expert control
systems. Automatica 25, 799-811 (1989).
play an increasingly important role in process control 25. Liu K., Gertler J., Chizek H. and Katona P. A supervisory
engineering. (expert) adaptive control scheme. Proceedings of the lOth World
Part 2 of this paper will focus on the design methods Congress of IFAC, Munich, Germany. Vol. 6, pp. 381-386
(1987).
for real-time ECSs. Some important issues, such as 26. Morant F., Albertos P., Martinez M. and Serrano J. J.
design principles, generic architectures, and develop- Hierarchical expert system as supervisory level in an adaptive
ment methods for ECSs, as well as some important control. 4th IEEE Symposium on Intelligent Control, pp. 18-27
(1989).
factors for the engineering implementation of ECSs, 27. Leitch R. and Quek C. Architecture for integrated process
are also discussed in Part 2. supervision. IEE Proc. D 139, 317-327 (1992).
28. Chang G. and Birdwell J. D. Distributed real-time expert control
system. Proceedings of 29th IEEE Control & Decision. Vol. 3,
REFERENCES pp. 1525-1529 (1990).
29. Linkens D. A. and Abbod M. F. Supervisory intelligent control
1. Rahmat S. Intelligent control: are they for real? J. Dynam. Syst. using a fuzzy logic hierarchy. Trans. Inst. MC, 15, 112-132
Meas. Control 115, 392--401 (1993). (1993).
2. ,~strom K. J., Anton J. J. and Arzen K. E. Expert control. 30. Chert Minyou. An intelligent control approach based on pattern
Automatica 22, 227-286 (1986). recognition techniques. Pattern Recogn. Artif. Intell. 5,229-234
3. Kravs T. W. Myron. Self-tuning PID controller use pattern (1992).
recognition approach. Control Engng. No. 6, 106-111 (1984). 31. Chen Minyou and Zhou Qijian. A real-time expert controller for
4. Porter B. and Jones A. H. Real-time expert tuner for PI industrial rotary kiln. Proceedings of IFAC and Symposium
controller. IEE Proc. D 34, 260-263 (1987). MMM'92, Beijing, P.R. China, pp. 391-395 (1992).
 D. A. LINKENS and MINYOU CHEN: EXPERT CONTROL SYSTEMS--I 421

32. Alice E. S. and Hulya Y. An intelligent composite system for 33. ,~strom K. J. and Anton J. J. Expert control. Proceedings of 9th
statistical process control. Engng Applic. Artif. lntell. 5,519--526 IFAC World Congress, Budapest, Hungary (1984).
(1992). 34. Astrom K. J. Towards intelligent PID control. Automatica 28,
1-6 (1992).

AUTHORS' BIOGRAPHIES
Derek Linkens has held a Personal Chair at the Department of Automatic Control and Systems Engineering, University of
Sheffield since 1984, and is currently the Dean of Engineering. He has received the degrees of B.Sc.(Eng.) from Imperial
College, M.Sc. from the University of Surrey, Ph.D. from the University of Sheffield and D.Sc.(Eng.) from the University of
London. His main research areas are in intelligent and adaptive control and modelling in biomedicine and the process
industries. He has published 300 papers in these fields, and edited and co-authored five research monographs. In 1993 he was
President of the Institute of Measurement and Control. Prior to entering academia he held industrial posts in underwater and
aerospace technology.
Minyou Chen received the M.S. degree in Control Engineering from the Department of Automation, Chongqing University,
China, in 1987. From December 1988 to August 1993, he served as a lecturer in the Department of Automation, Chongqing
University, China. Since September 1993, he has been engaged in research work in the Department of Automatic Control &
Systems Engineering, University of Sheffield, U.K. His current research interests include knowledge-based systems, expert
control, fuzzy logic systems, integrated intelligent control system design and implementation.