Partial discharge is an important characteristics of GIS insulation deterioration, and a comprehensive and accurate analysis of Partial discharge signal can identify the type of insulation defects, thus effectively preventing insulation accidents and improving equipment operating life. Accurate pattern recognition depends on effective feature extraction, however, there are many feature parameters, and combining all of them for signal analysis will lead to dimensional disasters, and the fused features after dimensionality reduction may not be ideal. Fractal features contain a large amount of data information and are concise, which are extremely ideal feature parameters, but they are insensitive to the information of the distribution characteristics of the data, and are prone to misjudgment in the face of different discharge distributions and similar changes in the number of discharges. This paper proposes a multi-scale fractal statistical feature extraction method using the primary and secondary feature vector model. From the half-cycle PRPD patterns, the box dimension is extracted as the primary feature vector by the differential box counting method; the statistical features of the probability density curve of PRPD patterns are extracted as the secondary feature vector. The statistical features are used to correct the fractal feature misclassification recognition results, and the 110kV GIS experimental platform is built to verify the feature extraction effect by collecting Partial discharge data through embedded UHF. The final results show that fractal statistical feature extraction can effectively improve the correct rate of pattern recognition. It has important guidance value for the actual GIS detection system.

gis,partial discharge,measurement,pattern recognition