Feature extraction of partial discharge gray-scale images for identification based on multifractal spectrum using fluorescence fiber sensor

This document proposes the utilization of partial discharge (PD) optical signals from a fluorescence optical sensor system for constructing phi-u-n charts and gray-scale images. The fractal characteristics cannot fully characterize the PD gray-scale images, which reduce the PD pattern recognition rates. Multifractal spectrum is used for analyzing the characteristics of gray-scale images, and a new probability calculation method is proposed for computing the multifractal spectrogram of these images. By conducting a series of experiments, the multifractal spectrum is proven effective in describing the variations in the geometric characteristics of the gray-scale images. The main physical features of the multifractal spectrum are extracted and used as pattern recognition features. The backpropagation neural network with an improved conjugate gradient algorithm is used as a classifier in identifying the different PD types, which achieves recognition rates that are >87%. The multifractal spectrum improves the accuracy of the PD pattern recognition unlike other pattern recognition features, such as the box-counting dimension and the information dimension. (C) 2014 Society of Photo-Optical Instrumentation Engineers (SPIE)