Intelligent Partial Discharge Diagnosis for Condition Monitoring

This thesis reports on research into intelligent partial discharge (PD) diagnosis for turbogenerator condition monitoring (CM). PD activities occurring in generator stator windings and modern techniques for PD CM are introduced. Research is focused on graphical classification methods, especially for small-size, and incomplete PD database. The research work begins with study on feature extraction methods on different PD patterns, and automated pattern recognition methods involving conventional classifiers and neural networks. Laboratory tests are made to observe PD activities and produce PD database containing typical PD types on industrial model bars. A Hybrid Clustering Method (HCM) and an advanced Self-Organizing Map (SOM) are presented to provide graphical classification results where the relationship of new PD samples and historical samples can be visualized. The work confirms that the graphical classification methods can be used individually or combined with other methods to provide reliable diagnostic information.