| Reports on the application of neural networks in chemical engineering started in late 1980's. Since then, the research in this field has been growing rapidly. This work deals with the particular application of neural network models for Fault Detection.; This project was initially motivated by the observations of misclassification due to similar symptoms in input patterns, and the overlapping among the basis functions of a Neural Network. The motivation to solve these problems led to a comparison of three neural network models: Haar Wave-net, Projection Pursuit Regression, and Backpropagation Network. It was found that, although the orthonormal local basis functions of a Haar Wave-Net avoid the overlapping among the basis functions and can prevent the propagation of noise effect, they possess poor generalization ability. This work shows that both the crisp classification on the class boundary and reasonable generalization accuracy inside the classes are desired in a pattern recognition problem. Backpropagation Network has the potential drawback of propagating the noise effect throughout the input space, though it is reliable and easy to implement. Projection Pursuit Regression seems to have a good balance between sensitivity to noise and generalization accuracy.; One of the major achievements in this project is the development of two innovative experimental design methods for the pattern recognition task. The first design method, the Gaussian Probability Design explores the sparseness of training data in the process input space, by calculating the Gaussian probability with respect to the original training data. The second design method, the Fuzzy Boundary Design is derived from the bootstrapping technique. When compared to the conventional Factorial Design, these two new design methods show a significant advantage in improving pattern recognition model. On the other hand, it is realized that the Optimum Experimental Design for pattern recognition depends on both the design problem to be solved and the modeling toot to be used. When the class boundary between two classes is linear, and the modeling tool, such as Projection Pursuit Regression, can generalize well, one conjugate pair of training data, with one on each side of the class boundary, is usually enough. But, if the class boundary is curvilinear, many conjugate pairs may be needed, along the class boundary. It is always desirable to have equal density of training data on two sides of a class boundary. Otherwise, the identified class boundary will be biased towards the side with less training data. If the fault detection is based on Haar Wave-Net models, the training data need to be distributed evenly throughout the data domain, due to the poor generalization ability of these networks.; This project has investigated the application of neural network models to fault diagnosis in a CSTR process and a copolymerization process of STY/MMA. It is found that neural network models can handle fault diagnosis reasonably well in both steady state and dynamic situations. Analyses of system observability have been conducted on the CSTR and the copolymerization processes. The analyses reveal that the similar symptoms in input patterns could be due to a lack of observability, which may result in misclassification of faults. (Abstract shortened by UMI.)... |
