| Several areas in chemical engineering exhibit computational and/or representational barriers to the solution of important domain problems. In this dissertation, an artificial neural network (ANN) approach to three such barriers that affect the fault detection and diagnosis, process control, and process modeling problem domains in chemical engineering is discussed. The barriers considered are: the inability of these chemical engineering systems to use large quantities of sensor information, the need for general mapping capabilities, and the need for adaptive or learning systems.; An artificial neural network design and simulation environment called ZNL (Zone Node Link) was built to study practical applications. In the domain of fault detection and diagnosis, it was demonstrated that an ANN can learn useful mappings via the generalized delta learning rule that provides the correct association between fault classes and measurement patterns autonomously. The same network was shown to map the complex decision regions caused by adding sensor noise to deterministic measurement patterns. In the domain of process control, an ANN was used to learn a strategy to control a chemical engineering process, thus obviating the need for a control law to be specified a priori. In the domain of process modeling and simulation, it was demonstrated that an ANN can model a plasma etching system by employing real-valued mappings. A side effect of these studies was the development of a heuristic technique called focused-attention backpropagation (FAB) that results in faster and smoother learning than backpropagation and introduces minimal additional complexity. The results of these studies demonstrate the effectiveness of ANNs to learn and adapt themselves to inputs (which are not constrained in number due to the inherently parallel computational structure of ANNs) from the actual processes, thus allowing representation of complex engineering systems which are difficult to model either with traditional physical engineering relations or knowledge-based systems. |
