Comparison of Support Vector Machine Based Partial Discharge Identification Parameters

Abstract

Partial discharge (PD) may have a significant effect on the insulation performance of power apparatus. Therefore, identification of PD sources is of interest to both power equipment manufacturers and utilities. With the development of PD measurement techniques, data analysis, signal processing and pattern recognition are gaining more interest. Research to date has considered varieties of different identification parameters such as phase resolved information, statistical operators, pulse shape analysis, pulse sequence analysis, frequency spectrum and wavelet analysis, which are also combined with so-called classifiers such as fuzzy logic, neural networks (NN) and learning machines. This paper investigates the performances of PD source identification using a support vector machine (SVM) based on different feature parameters. Due to the unique characteristics of SVM, some feature parameters that are not suitable for other classifiers are applicable. In this paper, comparisons of recognition rate and generalization capability between different feature parameters are discussed. Investigation reveals that recognition rate and generalization capability are influenced by the input PD parameters. Initial results indicate that, by using appropriate kernels and feature parameters, the automatic identification results obtained by using the support vector machine technique are very encouraging.

The role of circumstance monitoring on the diagnostic interpretation of condition monitoring data

“…Secondly, the rules induced by C4.5 are human-readable and can be studied for insight into the classification problem, whereas the support vectors resulting from a trained SVM are a comparatively obscure coding of the knowledge. However, the disparity between accuracy reported in [15] and results given here merits further discussion. Three possible sources of difference are considered below.…”

“…Despite both utilizing PD data, this research attempts to identify harmonic attributes whereas [15] identifies defect type. Harmonic attribute identification is significantly different and perhaps more challenging than defect type identification.…”

“…The sizes of training data sets are different in each case. While [15] uses 30 examples of five defects at three different voltage levels, giving datasets of 90 patterns per class, the research here uses between 999 and 1094 patterns for each test waveform. This order of magnitude difference may result in a more diverse set of patterns, increasing classification difficulty, but potentially more realistic.…”

“…Further research attempted to improve the accuracy of attribute identification by investigating other classifier techniques. In particular, recent research on the use of Support Vector Machines (SVMs) for PD data analysis has suggested accuracies exceeding 99% are possible [15]. On the other hand, comparative research into ship-board fault diagnosis suggests SVMs did not perform better but in some cases more poorly than C4.5 [16].…”

“…For rough grid search, the search space was: C = 2 -5 , 2 -3 , ... , 2 15 ; γ = 2 -15 , 2 -13 , ... , 2 3 Once the SVM with highest accuracy was determined, that <C=2 c , γ=2 g > pair was the source of fine grid search: C = 2 (c-0.5) , 2 (c-0.25) , 2 c , 2 (c+0.25) , 2 (c+0.5) ; γ = 2 (g-0.5) , 2 (g-0.25) , 2 g , 2 (g+0.25) , 2 (g+0.5) The most accurate SVM resulting from fine grid search became the final classifier.…”