Feature Extraction And Pattern Recognition Of Partial Discharge In Oil-paper Insulation

Oil-paper insulation is widely used in power equipment. It is very significant tostrengthen the detection of insulation for ensuring power equipment from malfunction. Inaddition to reflect the severity of the insulation defects, partial discharge signals alsocontain a lot of information about partial discharge sources. For finding and eliminatinginsulation defects, it is helpful to ascertain the types of partial discharge sources throughthe information.Firstly, in order to get the data about partial discharge pulses, five typical kinds ofpartial discharge models were tested in the lab. For the subsequent analysis, the paper theamplitude and the phase of each pulse was extracted from the data.Secondly, using morphology to analyze the partial discharge grayscale images, itwas found that the pattern spectrum was not suitable to be used directly for identifyingthe types of partial discharge sources, when the grayscale images not containing enoughdata. Three characteristics were proposed to reflect the pattern spectrum, and theireffectiveness was verified using the experimental data. Then combining the physicalprocess of discharge and the structure of partial discharge models, it can be educed thatthe influence of the models on statistical operators, and this is useful for the choice of thestatistical operators.And then, according to the characteristics of the partial discharge models, themodels were classified reasonably to make the pattern recognition convert into a processof two-category. The characteristic quantities were chosen rationally according to thecontent in each binary classification. In order to improve the effect of the recognition, thefeature vectors would contain the morphological features and the statistical operators inthe paper.Finally, the relevance vector machine was introduced into the recognition of partialdischarge,as a mean of realizing the binary classification process. Each relevance vectormachine was trained and tested using the experimental data. The experimental resultsshow that this method is effective and applicable for the recognition process.