Background for knowledge-based control: Holistic approaches in knowledge-based process control; introduction to knowledge-based systems for process control; basic theory and algorithms for fuzzy sets and logic; knowledge engineering and process control. Artificial intelligence issues: Cognitive models from subcognitive skills; a review of the approaches to the qualitative modelling of complex systems; solving process engineering problems using artificial neural networks; parallel processing architecture for real-time control.
Inspec keywords: process control; multivariable control systems; artificial intelligence; knowledge based systems
Other keywords: industrial control; artificial intelligence; knowledge based process control; deductive control
Subjects: Control technology and theory (production); Industrial processes; Control in industrial production systems; Industrial applications of IT; Multivariable control systems; Control engineering computing; Expert systems and other AI software and techniques
A consideration of the nature and scope of knowledge based process control (KBPC), leads to the proposal that Artificial Intelligence, Systems Engineering and Information Technology are three core elements of the discipline. Some aspects of all three are described. Systems Engineering methods are based upon whole life whole system considerations which are argued to be fundamental to KBPC. Applying the principles of taxonomy, or classification science, leads to a hierarchical perspective of control instrumentation systems which helps to gain breadth of comprehension of information machines. By first introducing the Systems Engineering approach and the methods of taxonomy, the chapter allows further development of the core theoretical elements in automatic control systems in general but especially in process control. Highlighting of the cardinal elements in the technology of knowledge based process control systems is also allowed using these methods.
This chapter gives an introduction to expert systems as they have been and are likely to be applied to the domain of process control. The evolutionary context of expert systems is discussed, so as to pick out those features of process control expert systems which make them different from expert systems in other domains and artificial intelligence in general. This chapter on expert systems for process control has taken an evolutionary view of the field, examining how expert systems have evolved from their roots in artificial intelligence to become something quite distinct from the medical expert systems from which they sprang. However, they are not so separate that there is no more to be learnt from the applications in the medical domain. While the experts in process control and medicine may not discuss the intricacies of their knowledge-based systems, the designers and academics remain in communication.
Fuzzy sets were first introduced by Zadeh as a method of handling 'real-world' classes of objects. Ambiguities abound in these real-world sets, examples given by Zadeh include the 'class of all real numbers which are much greater than 1, and the 'class of tall men'. Examples of these ambiguous sets are easily found in the process control field, where operators may talk about 'very high temperatures' or a 'slight increase in flowrate'. Conventional set theory is clearly inadequate to handle these ambiguous concepts since set members either do, or do not, belong to a set. For example, consider the set 'tall men' a man who is seven feet tall will clearly belong to the set and one who is four feet tall will not, but what about someone who measures five feet ten inches? Zadeh's solution to this problem was to create the fuzzy set, in which members could have a continuous range of membership ranging from zero, or not belonging, to one indicating definite belonging.
The lack of well-tried and reliable knowledge engineering techniques has been a limiting factor in the development of expert systems. The initial successes of early systems resulted in too much emphasis being placed upon the bottom-up strategy for knowledge acquisition. At the present time the technology used for implementing knowledge based systems tends to exert too great an influence on the acquisition process itself. There is an urgent need for better tools to aid the knowledge acquisition process. In the process control area we need better tools for creating models of the process and relating these to operator actions and cognitive processes. There is also a need for tools to assist in validating expert systems since progress will almost certainly be limited by safety considerations.
This chapter considers the acquisition of skill in dynamical control tasks for which the 'recognise-act' cycle is relatively fast, as in piloting a helicopter. Human pilots commonly receive their initial training on computer simulations. From such trial-and-error learning they acquire cognitive capabilities which they cannot articulate.
Much work has been done in developing basic representational paradigms for the qualitative representation of complex systems. This work, however, is just the beginning. Much more needs to be done in determining the relationship between the various approaches so that the systems modeller can select the most appropriate method for the particular problem under study. Also, further work on developing other quantity spaces that reduce the qualitative ambiguity and hence result in less spurious behaviors needs to be done. Finally, realistic applications need to be tack led to determined whether these techniques will 'scale-up' to real size practical applications.
Artificial neural networks are made up of highly inter-connected layers of simple 'neuron' like nodes. The neurons act as nonlinear processing elements within the network. An attractive property of artificial neural networks is that given the appropriate network topology, they are capable of characterising nonlinear functional relationships. Furthermore, the structure of the resulting neural network based process model may be considered generic, in the sense that little prior process knowledge is required in its determination. The methodology therefore provides a cost efficient and reliable process modelling technique.
In this chapter we study architectures, algorithms, and applications for parallel processing in real-time control. The nature of advances in VLSI technology have resulted in increased computing power generally being made more available through parallel processing architectures of different types rather than increased clock rate in uniprocessor systems. Despite the development of faster processors, the real attraction of parallel processing to system designers is its scalability to meet increasing demands. There is a plethora of control engineering application areas.
With the advent of artificial intelligence (AI) techniques and with the increased interest in applying the new technology to a wide variety of problems, there is a proliferation of software tools marketed for developing knowledge based systems. There are many factors that influence the selection of a tool for a particular project. For example, machine availability, supplier credibility, etc. These factors, though important, are not considered in this chapter. The primary concern of this chapter is to look at the AI aspects of tools and see how they influence tool selection. The objective is, therefore, threefold. First, it describes the AI features that are found in KBS tools. Second, it considers the problem of mapping application characteristics to these tool features. Third, it describes three representative tools that implement some of these features.
This chapter concentrates on application studies carried out within the Chemical Engineering Department at Strathclyde University. The second section of this chapter describes application studies using a fermenter problem. Both rule-based and relational models has been investigated and it has been found that both could produce very good results. The third section of the chapter describes the early stages of work on a project to apply relational modelling techniques to the control of a real industrial process. It has been demonstrated that rule based and relational system can be made to represent the sort of processes occurring in fermentation, and that relational modelling is easily applied to real industrial problems.
In this chapter the issues involved with the design and implementation of real-time knowledge based systems (RTKBS) are reviewed. These will be demonstrated with reference to a RTKBS approach for the improved supervision and control of fermentation plant. The system is composed of three elements; a supervisory knowledge base, a scheduling knowledge base and an on-line relational database. The supervisor is responsible for monitoring and controlling individual fermenters and is instructed when to initiate or terminate a fermentation by commands issued by the scheduler, which aims to maximise plant-wide productivity. The system also performs sensor validation, fault detection/ fault diagnosis and incorporates relevant expertise and experience drawn from both the process engineering domain and the control engineering domain. The database provides a centralised store of information which can be retrieved on line by both knowledge bases.
Machine-learned rule-based control differs from more typical approaches to the engineering of controllers for physical systems in the following respect. In traditional control theory, a mathematical model of the system is constructed and then analysed in order to synthesise a control method. This approach is clearly deductive. A machine-learning approach to the synthesis of controllers aims to inductively acquire control knowledge, thereby avoiding the necessity of constructing a mathematical model of the system. In applications where systems are very complex, or insufficient knowledge is available, the construction of such a model may be impossible, and traditional methods therefore inappropriate. It is for these applications that an inductive approach promises solutions.
The application of self-tuning control can in certain process control problems lead to improvements in economic, safety and control performance. However, the components of a self-tuning controller are more complex than conventional loop controllers and require considerable knowledge of a variety of advanced algorithmic techniques. In addition, as much detailed process specific knowledge as possible must be built into the self-tuner in order to select good design parameters for a given application. We identify a two-way knowledge threshold which must be crossed before self tuning control can be used and fully exploited on industrial processes. This chapter examines the components of the knowledge threshold present for self tuning controllers and considers some possible solutions. The use of expert systems to lower the knowledge threshold is discussed with reference to a prototype system.
In this chapter, a number of systems developed by intelligent applications, primarily in heavy industry are summarised. This chapter discusses some of the applications developed in the mechanical and process area. In this talk, a number of practical expert system applications both in the process industry, where they have process data, or PLC data and also in the rotating machinery area both on-line and off-line have been illustrated.
As an environment for building expert systems, COGSYS provides the framework and functions for the configuration of realtime reasoning systems for a wide variety of applications. COGSYS can be employed in monitoring realtime process systems, interpreting financial data, managing critical resource situations, controlling systems at a supervisory level, and regulating complex communication networks.
This chapter describes the application of COGSYS to a small gas processing plant at the British Gas Midlands Research Station, Solihull. COGSYS (for COGnitive SYStem), is a new real-time expert system which is particularly suited to industrial process control. It was specified and developed within a collaborative group of some 35 major companies and has been made commercially available through the formation of a new company, COGSYS Ltd.
In this chapter some of the issues related to the off-line design of single-input single-output (SISO) and multiple-input multiple-output (MIMO) control systems are discussed. Particular emphasis has been placed on the issues concerning the transfer of the expertise implicit in control theoretical knowledge into the design environment for general use by industrial design engineers. It is noted that considerable insight is obtained by the analysis of the procedures (both implicit and explicit) of classical SISO design.
The computation of H∞ and LQG optimal controllers is considered for process control applications. There are many process control problems where significant uncertanties exist in the system models which therefore require robust control designs. A simple solution for the optimal H∞ robust design problem is considered and the relationship to super-optimal solutions is discussed. For special types of weighted plant model the main H∞ equations to be solved are shown to be decoupled so that the calculations are similar to the scalar case. This situation is shown to arise when a mixed-sensitivity cost-function is selected and the plant has an interaction structure typical of many hot or cold rolling mill gauge control applications. A simplified design procedure is also introduced which further simplifies the calculations of the optimal controller and enables standard eigenvector/eigenvalue algorithms to be employed in solving the equations. The procedures are illustrated using a multivariable metal processing control design example.