The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Real time control of nonlinear dynamic systems using neuro-fuzzy controllers
Author:	Jana, Amitava
View Online:	njit-etd1996-037 (xii, 96 pages ~ 4.3 MB pdf)
Department:	Department of Mechanical Engineering
Degree:	Doctor of Philosophy
Program:	Mechanical Engineering
Document Type:	Dissertation
Advisory Committee:	Dave, Rajesh N. (Committee co-chair) Auslander, David M. (Committee co-chair) Chen, Rong-Yaw (Committee member) Ji, Zhiming (Committee member) Koplik, Bernard (Committee member)
Date:	1996-01
Keywords:	Real-time control. Intelligent control systems. Nonlinear control theory. Fuzzy systems.
Availability:	Unrestricted
Abstract:	The problem of real time control of a nonlinear dynamic system using intelligent control techniques is considered. The current trend is to incorporate neural networks and fuzzy logic into adaptive control strategies. The focus of this work is to investigate the current neuro-fuzzy approaches from literature and adapt them for a specific application. In order to achieve this objective, an experimental nonlinear dynamic system is considered. The motivation for this comes from the desire to solve practical problems and to create a test-bed which can be used to test various control strategies. The nonlinear dynamic system considered here is an unstable balance beam system that contains two fluid tanks, one at each end, and the balance is achieved by pumping the fluid back and forth from the tanks. A popular approach, called ANFIS (Adaptive Networks-based Fuzzy Inference Systems), which combines the structure of fuzzy logic controllers with the learning aspects from neural networks is considered as a basis for developing novel techniques, because it is considered to be one of the most general framework for developing adaptive controllers. However, in the proposed new method, called Generalized Network-based Fuzzy Inferencing Systems (GeNFIS), more conventional fuzzy schemes for the consequent part are used instead of using what is called the Sugeno type rules. Moreover, in contrast to ANFIS which uses a full set of rules, GeNFIS uses only a limited number of rules based on certain expert knowledge. GeNFIS is tested on the balance beam system, both in a real- time actual experiment and the simulation, and is found to perform better than a comparable ANFIS under supervised learning. Based on these results, several modifications of GeNFIS are considered, for example, synchronous defuzzification through triangular as well as bell shaped membership functions. Another modification involves simultaneous use of Sugeno type as well as conventional fuzzy schemes for the consequent part, in an effort to create a more flexible framework. Results of testing different versions of GeNFIS on the balance beam system are presented.