The purpose of this research is to analyze the basic principles of motor fault detection using neural/fuzzy architectures and to develop a framework for incorporating a priori heuristic information into the process of fault detection with neural/fuzzy architectures. First, two fault detectors based on prevalent neural/fuzzy architectures are analyzed and compared with regard to several attributes, including learning algorithms, initial knowledge requirements, and extracted knowledge types. Comparative experimental results are presented for a three-phase induction motor fault detection problem. Then, a framework is developed for incorporating a priori information into the training of neural/fuzzy architectures using set theoretic concepts. The method, called heuristic constraint enforcement, is integrated into the training of one of the analyzed neural/fuzzy architectures to obtain accurate fault detection along with fuzzy sets that agree with a priori heuristic knowledge. Finally, an alternative training scheme is developed for the neural/fuzzy architecture based on line search methods. |