Abstract
Most industrial systems are developed and improved based on experimental data instead of theoretical analysis. In this thesis, the quality and robust design of products and processes are developed based on a generalized artificial neural network response surface. Moreover empirical models derived from plant test are standard in implementation of model predictive control (MPC). The thesis proposed a systematic approach to develop a nonlinear dynamic model for NMPC termed as GAMIA. In the last decade, Artificial neural network has been applied to solve various problems in chemical engineering domain. However, all of the above works did not consider appropriateness of the size of neural network models with or without prior knowledge. The structure and the training policy of the generalized artificial neural network model are determined by the information index of cross validation developed in the dissertation. In case of noisy and few experimental data, this approach proposes a novel concept of smooth training and cross validation development of an artifical neural network model. Polymer composite pultrusion process is studied. Simulation and experimental results of pultrusion process show that this generalized artificial neural model approach is highly effective and promising.GAMIA is a novel systematic approach to acquire good-quality plant data that can be efficiently used to build a complete dynamic empirical model along with the use of partial plant knowledge. A generalized Delaunay triangulation scheme is then implemented to find feasible operating boundaries which may be non-convex based on the existing plant data. Conditional Akaike information index is adopted to assess partial plant knowledge as well as noisy plant data. Information free energy is developed for acquisition of good-quality new plant data that will improve the dynamic model. Information entropy is derived to identify the mutual positions among data points in all feasible regions. In addition, information enthalpy is derived to obtain a system’s dynamic nonlinearity. Hence, the placements of the new data are designed on the basis of a compromise between the information entropy and the information enthalpy - the information free energy. Also included herein are realistic examples such as pH control and high purity separation system. Simulation results demonstrate that the proposed information free energy experiment design approach is highly promising in terms of obtaining a reliable black box model, such as ANN, for model predictive control. The new experimental data suggested by the information analysis, together with the previous data and prior plant knowledge is used to train a generalized dynamic neural model. Multi-variable model predictive control for a high purity separation system using the acquired model based on the proposed approach is also studied. Comparing with PRBS and RAS scheme, the proposed approach outperforms the rest.