An efficient partial discharge pattern recognition method using texture analysis for transformer defect models

Partial discharge (PD) measurement is one of the best methods for condition monitoring of transformers. In this paper, we use 5 different types of defects as follows: scratch on winding insulation, bubble in oil, moisture in insulation paper, a very small free metal particle in the transformer tank, and a fixed sharp metal point on the transformer tank, for our PD-related studies. Each type of defect is implemented into 1 of the 5 identical transformer models, which had been developed in the authors' recent work. The continuous wavelet transform is applied to each related measured time-domain PD signals. This process results in an image, for each PD pulse in the time-frequency domain. Using these images, a gray-level covariance matrix is constructed. The texture features are extracted from the constructed gray-level covariance matrix of each PD signal. Principal component analysis is applied on the recorded PD data to reduce its dimension, and then support vector machine is used to classify the computed first 6 principal components of those defects' PD signals. The accurate outcome of defects identification in this work verifies efficiency of the proposed method.