Study On Methods Of Feature Extraction In Partial Discharge Pattern Recognition And Analysis Of Feature Correlation

Electric insulation plays a very important role in high voltage electrical equipment. The insulation state directly determines the stability of electrical system. Partial discharge is a very important factor resulting in insulation deterioration. So monitoring partial discharge signals on-line and identifying the types of discharge look like very necessary. By doing so, weakness of insulation and the developing degree can be discovered in time, which can stop accidences. Therefore, a systematic study on pattern recognition of partial discharge has both important academic meaning and engineering value. Based on a great of laboratory investigation carried out in double shielding room, this paper studies systematically the methods of feature extraction in partial discharge pattern recognition and correlation among characteristics. The main works and conclusions in this paper are as follows:Make three kinds of models and corresponding electrode system to simulate corona discharge, surface discharge and cavity discharge. In double shielding room, a large number of experiments were done on these models. Lots of experimental data were obtained.After processing original data, characteristics were extracted using four kinds of methods. These characteristics include tabulated data extracting from pattern, statistics, fractal parameter using gridding methods and moment features. BP artificial network is used to classify different kinds of discharge. The results show that moment features and tabulated data are more effective than statistics and fractal parameters in distinguishing five kinds of discharge. But the former have larger input dimension to BP artificial network. The latter are only two and six.This paper also studies the correlation among the characteristics. As far as moment features, there are four characteristics, which have high relativity. The four ones can be omitted. BP artificial network outputs check it. Thus the number of characteristics formed by moment features is reduced to 22 ,while as far as fractal parameters and statistics, the relativities among them are low. Some characteristics which have low relativities with them can be added to increase the identifying capability of the pattern recognition system.