Optimal feature extraction for partial discharge recognition

Pattern recognition of partial discharge (PD) is significant for condition judgment of power equipment. To select the optimal features out of the original PD features, this paper introduces a novel feature extraction algorithm, which is based directly upon the decision boundary between the classes, and determines the feature transformation according to the separability as opposed to the more common approach of fitting, such as principal component transformation. Experiment calculations are conducted on PD data obtained through measurements on industrial models of generator stator winding bars. Statistical and moment features, drawn respectively from each PD sample, are used in the experiments. Results show that when used together, the two sets of features are more powerful in discrimination. It is also shown that most discriminant information is concentrated in a subset of features, which can be determined and extracted successfully using the proposed algorithm. With the few selected features, approximate recognition accuracy could be achieved while substantial increases in computation complexity are refrained.