| This doctoral thesis presents the advances in three major research fronts of artificial neural network (ANN): a graphical mapping technique to interpret the internal activities of ANN model, an ANN hierarchical modeling protocol enhanced with a pattern recognition technique, and the actual industrial application and implementation of ANN technology. The advances in these three aspects of ANN technology interact and benefit each other.;The ability to interpret and manipulate internal workings of neural network is a major breakthrough in the ANN theoretical research. The author proves that most of the feed-forward neural networks are functions. Thus, many properties of functions can be applied to analyze neural network behaviors. Based on this proof, a graphical mapping technique was proposed to interpret the internal activities of ANN model. With this technique, it is possible to study the impact of noisy data on ANN modeling, and several key features of ANN models such as memorization, robustness, and sensitivity from the perspective of artificial neurons and their connection weights.;Based on the new knowledge of ANN models, a systematic modeling approach is proposed. Pattern recognition is integrated into the ANN modeling approach to provide additional capacities to analyze the source data in a noisy or complex study domain. The pattern recognition technique classifies the source data into many small modular domains to allow much more accurate modeling within smaller domains. This modeling approach creates an ANN modeling system with a hierarchical structure, modular components, and an important built-in ability to automatically detect new features of a study domain, and thus prevent meaningless prediction on the out-of-bound data and maintain the integrity of the model prediction.;Finally, the improved ANN modeling approach is used to build two industrial applications: one in water treatment industry and the other one in the oil sand treatment process. The success of these applications not only complements the theoretical research, but also provides precious experience on how to implement ANN modeling system on-line for the industrial applications. |
