| Summary: | 碩士 === 中原大學 === 電機工程研究所 === 86 === To keep the power system in operation, detection of the incipient faults of the equipment in the systems play an important role, especially of the power transformers which have a wide-ranged effects on the power supply. If the incipient fault of the power transformer cannot be detected earlier, it will lead to serious consequences due to the evolution of the fault. As a result, protectin the power transformers and reducing the customers' loss and inconvenience arising from the equipment faults, periodic examination of the transformers must be rigorously conducted to find the fault as early as possible and to prevent it from further deterioration. To investigate the operating situations and performance of the insulation oil inside the power transformers, regular testing of the isolation oil is a widely-adopted technique. Through chromatographic analysis of the insulation oil, we can have the concentrations of diverse dissolved gases created from partial discharge or over heating of the transformers. The concentrations of the dissolved gases, past operating records of the transformers and detailed synthetic fault causes. This process is call the dissolved gas analysis (DGA). To enhance the fault diagnosis abilities for the dissolved gas analysis of the power transformers, this thesis proposes a novel adaptive fuzzy system for the incipient fault recognition through natural construction rules and evolution-enhanced rule design approach. First, the IEC/IEEE D assessment codes are relied on to construct the preliminary fuzzy diagnosis system. Complying with the practical gas records and associated fault causes as much as possible, a fuzzy reasoning algorithm is then presented to enhance dagnosis system. In the systems, an evolutionary optimization algorithm is further used to fine-tune the membership functions of the if-then inference rules. Hence, the condition parts of the rules can be automatically adjusted to reach minimal diagnostic error and thus massively improve situations accuracy. Moreover, to make the diagnosis system intensively compact and the inference process more understandable, a pruning scheme using genetic algorithm is developed to filter out the insignificant or redundant rules on the condition that the diagnosis accuracy is not degraded. The pruned compact system not only needs less reasoning time, but also it makes the reasoning process simpler and more explainable. At the same time, users can generalize the DGA diagnosis rules which maybe still unknown before through the support of the developed diagnosis system. The capabilities of the proposed diagnosis system have been extensively verified through the practical DGA test data collected from Taiwan Power Company (TPC). The general performance of the proposed system has also been compared to that of the previously established fuzzy system and the artificial neural network.
|
|---|
