The complexity of designing supervisory systems has been introduced in the last chapter. The necessity of incorporating expert knowledge in such design is present in all tasks involved in supervisory systems. Fault detection can be performed by using analytical models, but these models are not always available and final resolution about residual generated, is always submitted to the human expert decision. On the other hand, the use of knowledge-based representations is needed by numeric to qualitative interfaces to perform reasoning about process variables. The important role of human knowledge in process supervision is concentrating the attention on these techniques that permit the use of expert knowledge to automate these tasks. From this perspective, the increasing use of Artificial Intelligence (AI) techniques in control and process engineering must improve supervisory designs, and coexist with numerical methods.

The most extended AI tools in the control domain, are overviewed in this chapter, and benefits and drawbacks of using some of the AI techniques for knowledge representation and qualitative reasoning, are discussed from a supervisory systems design point of view. The aim is to take advantage of control experience in using AI resources, to choose a set of adequate tools to be integrated in a framework for assisting supervisory tasks avoiding.

In the design of supervisory systems, the information related to process variables is available in several ways: numerical data (from sensors, analytical models, numerical estimations and so on), qualitative data (from human perception of process variable trends, qualitative models and qualitative estimations) and relationship or dependencies among these data. The complexity in managing together such different kinds of data is in the scope of AI techniques.

Nowadays, the main contribution for designing supervisory systems comes from operators and expert engineers experience. In fact, they are also present in the majority of applications for final decision making. Sometimes, the description and translation into computers of this expert knowledge, related to process variables, becomes very difficult or impossible due to the different nature of human descriptions and data obtained from process.

Usually, experts describe situations, while data are instantaneous samples of measures, or estimations of these situations. In the procedure of matching process variables evolution and those situations, humans use an imprecise description of magnitudes. An example can clarify these difficulties : The following sentence, "when temperature in the reactor increases, opens the input valve slowly", describes an action (opens valve slowly) to be performed when a process variable (temperature) experiments certain behaviour (increasing). This expert description are easily interpreted by humans, but difficult to interface with numerical magnitudes coming from the process (temperature) or actuators to perform this action (open valve). They are imprecise descriptions of numerical magnitudes available in the process. This imprecise description must be processed before to be used in the control structure. The representation of these kinds of information and the capability of dealing with the relationship among imprecise variables is in the scope of AI. The use of qualitative reasoning and modelling techniques can be useful for these purposes.

Another important feature of AI techniques is the capability of dealing with uncertain data or information. The description of an intermittent fault is an example of this situation. In spite of some conditions matched, the fault is not sure. Certainty can be introduced as an index of confidence in the description of situations or data. A similar situation is given when contrasting information among several sources. Sometimes it leads to incongruous descriptions and a decreasing confidence of a source of information. In such cases, the use of different certainty indices for each source can be useful to merge incoming information to work in co-operation.

An additional inconvenience of expert knowledge for process supervision refers to temporal references. Usually it describes process behaviour in an uncertain period of time, or changes in the evolution of process variables without dating these events. In the example of the previous paragraph, the label increasing is related to a characteristic of a process variable during an imprecise period of time. This consideration must be taken into account for building numeric to qualitative interfaces in order to take benefit of this kind of descriptions of variables evolution. This also applies more general descriptions of process behaviour such as transient or steady state, for instance. In this case, both possibilities are exclusive, but real transition between both states is gradual. Then difficulties exist in determining the limits between both, because it implies all process variables analysed.

Actual AI techniques, as fuzzy logic or some qualitative reasoning formalisms, are applied to deal with this imprecision in the description magnitudes and the relationship between variables. On the other hand, temporal imprecision is inherent to qualitative methods.