| Accurately modeling the process data is an important part of process control, quality control, process monitoring, and sensor development. This research was focused on demonstrating that artificial neural networks (ANNs) are good tools for modeling linear and nonlinear process data. The problems and issues that this research had to deal with were related to a single objective--the correct modeling of the available process data. To reach such a goal, this research studied: (1) how to construct a proper ANN model structure, (2) how to handle limited calibration data sets, (3) how to reduce the calibration or training time, and (4) how to correctly validate the resulting model. It is well understood though that solutions for these problems are problem-dependent and that there is no one heuristic which can be applied to all problems. This research identified the advantageous modeling and analytical properties of the direct linear feedthrough (DLF) network structure which incorporated the linear mapping abilities to the nonlinear network. When the DLF network was compared with a wide spectrum of linear and nonlinear regression tools, the DLF network always produced one of the best models. A powerful analytical method was developed using a four-layer DLF network where one can determine the severity of the nonlinearity of the process data. For handling the limited data set, different data treatment methods were tested and a vast amount of knowledge has been obtained on the relationships between data structures and model structures. Also, the training time was significantly reduced by the sequential quadratic programming (SQP) method. For the model validation, statistical validation methods were used to improve the confidence and robustness of the final model. |
