Application of an ensemble neural network for classifying partial discharge patterns

This paper proposes a technique for classifying partial discharge (PD) patterns based on ensemble neural network (ENN) learning. The ENN technique is based on training a number of neural network (NN) models with statistical parameters from PD patterns and combining their predictions. In this paper, six constituent NN models form the ensemble. Combining the outputs of the constituent NNs through an aggregating unit using dynamically weighted averaging strategy gives a final evaluation of PD patterns in relation to a range of PD fault types. Using the data sets of measured PD patterns as the system input fingerprints, the classification performance of the ENN has been compared statistically and quantitatively with a single neural network (SNN). This is achieved through evaluating the average, variance and standard error of the means of ENN and SNN recognition performances over 100 different initial states of the NNs thus providing an effective comparison to be made. The result shows that the ENN appears to be more robust with statistically improved performance in recognizing untrained PD patterns for a number of PD fault geometries.