Self-learning predictive control using relational-based fuzzy logic

This thesis documents the development of a Model-based Self-Learning Predictive Fuzzy Logic (MSPF) Controller for use in applications where the inherent uncertainty in the process model and/or data precludes the use of conventional discrete control algorithms. This work required not only a translation of the concepts of discrete model-based control systems into the fuzzy domain but also significant extensions to fuzzy logic theory.;The extensions to fuzzy logic theory in this thesis pertain mostly to the max-product composition, which several authors have shown to produce better results than the widely used max-min composition. The superiority of the max-product composition was also confined in this thesis for a variety of process oriented applications. The new theory developed for the max-product composition includes eigen fuzzy stability, powers of ;Since the max-product composition has not been used extensively, there was very little existing literature on effective identification algorithms for this composition. This thesis therefore reviews and compares several important fuzzy identification strategies for the max-min composition and then applies them using the max-product composition. Based on this work, a new identification algorithm was developed that is better, from a least squares perspective, than the existing algorithms when applies to the Box-Jenkins gas furnace data. The new identification algorithm also includes a new procedure that permits an identification aigorithm to adapt quickly to process changes while maintaining a complete solution.;Most of the rule-based fuzzy logic controller designs in the literature are based on a ;The MSPF controller gave very good closed-loop performance in simulation using underdamped, overdamped and non-linear processes plus processes with large time delays and/or disturbances. A direct comparison of the MSPF controller versus a conventional discrete: it PI controller using a very (smoothly) non-linear process showed that (based on minimization of the discrete control error) the MSPF controller gave better performances over the full operating domain than PI control even when three-level gain scheduling was used.;The development and evaluating of the Self-Learning Predictive Fuzzy Logic Controller described above is complemented by a extensive fuzzy logic tutorial which includes a literature survey and examples for each aspect of the controller development.