Identifying defects in SF6:N2 mixtures using external UHF couplers and Neural Network analysis

Abstract

An external measurement system capable of recording Partial Discharge (PD) Ultra High Frequency (UHF) signatures at a measurement frequency of 1 GHz from defects in Gas Insulated Substations (GIS) has been developed. Phase resolved PD patterns for a 15 mm metallic protrusion on the HV bus bar and 15 mm free conducting particle were obtained for SF6:N2 mixtures containing 10:90, 20:80 and 100:0 SF6:N2 content. Simple statistical features such as dis- charge occurrence and mean magnitude were extracted from the phase resolved plots and used as inputs into an Artificial Neural Network (ANN) for the clas- sification of defects. A comparison between defect signatures in the different mixtures is made and an investigation into the feasibility of using data taken in pure SF6 to aid in classifying defects in SF6:N2 mixtures with an ANN is undertaken. The overall classification accuracy of the ANN built and trained on pure SF6 data for the statistical features of discharge occurrence and mean value is 82 % for the two defects in the three different mixtures. However, a 10 % accuracy was obtained for the free particle in the 20:80 SF6:N2 mixture. A second method of extracting statistical features from PD signatures is de- veloped. Statistical features of skewness, kurtosis and variance are computed on the positive and negative half cycles of the phase resolved Maximum En- velope and are used as inputs into the ANN. This ANN achieved an overall classification accuracy of 96.5 % in discriminating between a protrusion on the HV conductor and a free particle in 10:90, 20:80 and 100:0 SF6:N2 mixtures. A 71 % improvement in classification accuracy for a free particle in the 20:80 SF6:N2 mixture is obtained as compared to the orginal ANN using features such as discharge occurrence and mean value of discharges computed on the positive and negative half cycles. The feasibility of using defect signatures taken in pure SF6 for classifying defects in SF6:N2 mixtures is shown. The effect of evolving PD signatures as well as the suitability of statistical features to be used for classification is discussed.