Knowledge representation for expert systems in chemical process control design

The subject of this dissertation is knowledge representation for expert systems applied to chemical process control. For the purposes of the dissertation, control system design is defined as establishment of the necessary single input-single output controllers for regulatory and constrained control. One of the most difficult tasks in creating an expert system to solve any problem is organization of the knowledge such that creation, modification, and extension of the expert system can progress smoothly.; An introduction to expert systems and the way knowledge is represented within them is given. Included in the introduction is a description of the requirements of an expert system shell to be used to solve the control design problem. Idiomatic control and the goal tree success tree model for knowledge are presented as organizational tools to represent the problem solution prior to expert system construction.; A methodology for constructing the knowledge base of an expert system for control system design is proposed. That method is used to create the expert system DICODE (DIstillation COntrol Design Expert) using a commercially available expert system shell. The expert system DICODE is presented in detail, and generalizations are drawn concerning use of the method for creating expert systems for control system design. Customization of the shell in the form of a partial knowledge base specific to control system design but not specific to distillation is discussed.; Most of the control design problem solution is cast into an organized structure which facilitates creation of the expert system. The only exception to this is the specific knowledge used to determine variable pairings for regulatory control. The knowledge associated with this variable pairing is expert dependent, and therefore generalizations concerning structure are difficult to make.; A simulated neural net is applied to the regulatory control variable pairing problem in the hopes of providing a "learn by example" solution. A neural net has the ability to learn an input/output pattern mapping through repeated presentation of a learning set. The net used for regulatory control design is taught 100 pairs of input/output data vectors and tested on data not in the learning set. The net is found to adequately provide a regulatory control design for input data not in the learning set.