| Process industries are abundant in complex nonlinear systems for which it is difficult and costly to build relevant mathematical models from first principles. Fuzzy logic systems (FLSs) can be employed to learn process behavior using operating data. FLSs describe process behavior in terms of linguistic rules that allow straightforward incorporation of a priori information about the process and allow verification of the acquired knowledge. In an effort to bring this advantageous modeling technique to wider use, this dissertation presents systematic design methodology for FLSs as well as an analysis of fundamental properties of FLS models and FLS based control schemes.; In FLS design, the following three parts are to be determined: the number of roles, the structure of each rule, and the membership function parameters. Most techniques treat these parts separately, which may result in a suboptimal solution due to the high dependence of the design parts on each other. In this study, two new FLS design methods were developed. Both are based on genetic algorithms that simultaneously optimize all three parts. The capabilities of the new methods and the properties of the resulting FLS models were tested on several major industrial problems: dynamic process modeling, material property prediction, and process fault diagnosis. FLS models developed by the new design methods compare well with other fuzzy models. They also compare well with state-of-the-art nonlinear modeling techniques.; The performance of FLS based model predictive control (FLS-MPC) was evaluated on two highly nonlinear processes. Results show that FLS based MPC possesses all the important qualities of a reliable control scheme, including good set point tracking and unmodeled disturbance rejection. It was also shown that FLS models can be expressed in a state space realization and local asymptotic stability of their equilibrium points can be proven. In addition, FLS models can be parameterized as a linear combination of fuzzy basis functions, providing ground for adaptive control applications. The results show that on-line adaptation of FLSs using recursive least squares provides rapid and reliable convergence to correct parameter values even in situations when a sudden change in process parameters occurs. |
