摘要

This paper proposes a novel partial discharge (PD) pattern recognition method based on extension neural network (ENN) using fractal features. Five types of defect models are well-designed on the base of investigation of power apparatus failures. A PD detector is used to measure the raw three-dimension (3D) PD patterns, from which the fractal dimension, the lacunarity, and the mean discharges of phase windows are extracted as PD features. These critical features form the cluster domains of defect types. An ENN is then developed to recognize the pattern of partial discharge, which utilizes an extension distance (ED) instead of Euclidean distance to measure the similarities among the recognized data and the cluster domains. The ENN with simpler structure than traditional neural networks is capable of processing the clustering problems which have a range of feature values, supervised learning, continuous input, and descriptive output. Moreover, the ENN has the advantages of higher accuracy, shorter learning times, and noise tolerance, which are useful in recognizing the PD patterns of electrical apparatus. To demonstrate the effectiveness of the proposed method, comparative studies among multilayer neural network (MNN), extension theory, and K-means are conducted on 200 sets of field-test PD patterns with rather encouraging results.