Optimal feature selection for defect recognition in varying concentration SF6:N2 mixtures

Using an external UHF coupler the maximum envelope (ME) of PD signatures for defects in SF₆:N₂ mixtures is extracted from data recorded at 1 GHz. Statistical features of skewness, kurtosis and variance are computed on the positive and negative half cycles of the phase resolved ME and is used as inputs into an artificial neural network (ANN). The 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 70% improvement in classification accuracy for a free particle in the 20:80 mixture is obtained as compared to an 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 features such as skewness, kurtosis and variance is discussed.