Multiple SVM-RFE for Feature Subset Selection in Partial Discharge Pattern Recognition

作者：Tang Ju^*; Tao Jiagui; Zhang Xiaoxing; Liu Fan

来源：International Review of Electrical Engineering, 2012, 7(4): 5240-5246.

• PD
• Feature Extraction
• Ultra High Frequency
• SVM-RFE
• Recognition

On study of the PRPD spectrum samples produced by typical insulation defects in Gas Insulated Switch gear (GIS), a group of fifteen statistical features are calculated to describe the PD characteristics. All the input features are reordered through the support vector machine recursive feature elimination (SVM-RFE) method, and the optimal feature subset is selected, with cross-validation and grid search introduced to improve SVM adaptive and classification accuracy The results indicate the validity on majority types of defects recognition except for a little mistake on viod in spacer. In addition, we also find that the SVM classifier input with only nine optimum features can obtain a better diagnosis performance than choosing all the calculated features as the input vectors.

• 出版日期2012-8
• 单位重庆大学