Statistical error tolerances of partial discharge recognition rates

Abstract

This paper compares the statistical error tolerances of the single neural network (SNN) and the ensemble neural network (ENN) recognition efficiencies, when both the SNN and ENN are applied to recognize partial discharge (PD) patterns. Statistical fingerprints from the phased and amplitude resolved patterns of PDs, have been applied for training and testing the SNN and the ENN. Statistical mean and variances of the SNN and ENN recognition rates were compared and evaluated over several iterations in order to obtain an acceptable value. The results show that the ENN is generally more robust and often provides an improved recognition rate with higher mean value and lower variance when compared with the SNN. The result implies that it is possible to determine the accurate statistical error tolerances for the SNN and ENN recognition probability for correct diagnosis of PD fault.