Artificial Intelligence is introduced in process supervision because of its capabilities in dealing with different kinds of information [Gentil, 1996]:
The techniques used in this domain are extensive. Some of them have
been tested in control application with successful results and they have
been introduced in the supervision domain because of their capabilities
in knowledge representation. In fact, the most used AI technologies in
the domain of supervision, are real time ES. They have an extended use
in fault detection and diagnosis. Others are only indirectly used. For
example :
At the same time, the use of these techniques involve knowledge
representation techniques, implicit in the application chosen for the implementation.
Integration problem.
Despite the great number of tools from the AI domain that can be used to represent expert knowledge and to make the procedure of writing rules and describing dependencies between variables, situations and causalities easier, supervisory systems design is still a research field in need of a methodology and frameworks where the complete cycle could be developed. Nowadays, design, test, validation and implementation of supervisory strategies is not possible in a simple way using only one framework with necessary tools. Some commercial packages, such as G2, are as described before but still lack the basic tools, such as data abstraction tools for obtaining significant information (numeric and qualitative) or imprecision management in the rules description. Rule bases describing conditions related to both, numerical and qualitative description of process variables, is a common situation. In such cases, reasoning becomes easier if all possible information about these variables is encapsulated.
When additional capabilities are required, such as qualitative simulation or estimation, other packages must be used and data must be converted to be interpreted. The interface of simulation package and ES is not an easy task, and of course implementation of such systems becomes impossible. Moreover, if only simple relationships between qualitative descriptions of variables are needed, then the use of external complex packages is not justified and makes the option unrealisable.
Imprecision in temporal references.
The advantages of AI managing imprecise data become a serious drawback when translating this imprecision into temporal references. A message alarm about a possible fault "in 2 minutes" is slightly different from the message "is coming soon". Similar problems are derived from the use of qualitative labels for representing process variables state during a period of time. Then, time dependencies are another important factor to be taken into account when dealing with dynamic process. Its importance increases when dealing with states transitions. A change in the behaviour from "normal" to "degraded" can not be suddenly. This kind of transitions must be smoothed in real-time operating conditions in order to detect these situations before they become irreparable.
Benefits of using OOP in supervisory tasks.
The main benefits of using OPP in the implementation of supervisory systems are derived from their characteristics, inheritance, abstraction and encapsulation of data and methods in the same structure. This applies in the construction of numeric to qualitative interfaces where numeric and qualitative data can be encapsulated together with the interface methods for an easier use. If additional methods for data access are provided to other objects, data can be shared and easily accessed.
Tools integration can be solved in this way if any data (numeric, qualitative, symbolic or logic) related to a process variable is encapsulated into objects and all tools are provided with access methods for obtaining desired information related to these variables. This introduces the concept of object-variables explained in chapter five. The use of object-variables with their graphical representation simplifies the conceptual use of them. QR tools can be "connected " to them to access desired information for performing some action or operation. Other numerical tools connected to the same object-variable could operate with numerical attributes.
Conclusions.
The complexity of designing supervisory systems comes from the necessity of dealing with incomplete, imprecise and uncertain data and information related to process behaviour. Expert knowledge, from operators and engineers, must be introduced into computers, despite of the imprecision in the description of data and rules. Knowledge representation tools (logic, production rules, graph, frames and objects) are the interfaces used by AI tools with this purpose. Several AI techniques can help in this task if they are available in the same framework where numerical data is collected from process (measures, simulations or estimations) and specialised tools are provided to interface both numerical data and qualitative methods. The use of QR methods can assist supervisory systems design in the tasks of qualitative estimation (qualitative observer) of variables, more than for defining pure qualitative models for simulation or to be used in model-based approaches.
Numerical to qualitative interfaces are subjected to some drawbacks.
The knowledge representation related to process dynamics needs the conversion
of numerical data into qualitative one. Moreover, multiple qualitative
representation can be necessary and coexist with numerical indices obtained
from these signals. In such cases, encapsulation of methods and data is
needed. Such interfaces must be very close to the process. Its inputs are
numerical row data from the process and its outputs are thought to be connected
to numerical processing and reasoning tools. OOP offers an adequate solution
to this problem.