Feature Selection and Classification of Field Leakage Current Waveforms using Genetic Algorithms

摘要

Leakage current monitoring is a commonly employed tool for the investigation of HV insulators' performance. Several techniques have been applied on in order to extract activity indicating information from leakage current waveforms. However, a fully representative value is yet to be defined. A recent approach to cope with this problem is to classify waveforms using a feature set comprising from commonly used values from time and frequency domain, and different feature selection and pattern recognition algorithms for the classification. In this paper, Genetic Algorithms are employed to perform both feature selection and classification. Eleven features were selected (out of the original twenty) and the accuracy achieved ranged from 82.2% to 88.48%, which is a considerable increase compared to earlier GA classification with no feature selection. Results are also compared to other feature selection and classification techniques that were previously applied to the same data set. Results emphasize the importance of feature selection and showed that Genetic Algorithms may give comparable results to nonlinear classifiers, provided that they are also employed for feature selection.