| Most chemical processes are nonlinear in nature, yet the overwhelming majority of the ones being computer controlled are using linear techniques. The main reason for not using nonlinear controller techniques is that detailed nonlinear models of processes are generally not available. Furthermore, nonlinear controller design usually require a mathematical sophistication above the level of the typical control engineer. Based on these observations, there is a need for nonlinear controllers which, even though based on sound process control principles, are intuitive in their design. The nonlinear controllers should also make use of the one piece of information readily available in most manufacturing sites: process data. The control algorithms should not require detailed fundamental models of the process for their implementation.It is thus the objective of this dissertation to investigate in a comprehensive fashion the control of nonlinear systems using available plant data (input-output) information. To this end, the dissertation is composed of three major parts: identification, analysis and control of nonlinear input-output models.The first part of the dissertation provides an overview of the identification problem. It concentrates on techniques available for the identification of nonlinear systems. An algorithm for the identification of nonlinear systems using input-output information is constructed. The underlying structure of the model identified is of a polynomial ARX. This choice is justified over others from both theoretical and practical viewpoints.Part II of the dissertation focuses on the analysis of discrete time input-output models. The analysis concentrates on the stability, invertibility and stability of the inverse model. These properties are first analyzed at a local level and then in an "extended neighborhood". The extended neighborhood analysis uses linear robust control theory and classical nonlinear system theory to extend and improve analysis results obtained earlier.The last part of the dissertation considers the problem of controller design as relaxations to the inverse model constructed in Part II. Several control algorithms available in the literature are studied as relaxations to the model inverse. In addition, a new control algorithm is obtained which relaxes the inverse model from an optimal viewpoint. The algorithm belongs to the general class of model predictive controllers, but it is shown to generalize other controllers presented earlier. |
