Feature Selection Based on Random Forest for Partial Discharges Characteristic Set

Rui Yao,Jun Li,Lin Bai,Qisheng Wu,Meng Hui

doi:10.1109/access.2020.3019377

Abstract

Since the dimension of combined feature set for partial discharge (PD) pattern recognition is higher, the corresponding sample size increases, as does the required amount of storage space and calculation, and there are features with less category-related characteristics in the feature parameters, which may contain redundant information between them. To solve the problem of higher feature dimension and complicated classification model required for the identification of partial discharge insulation defect type in this paper. Random forest sequential forward selection method based on variance analysis (RF-VA) is proposed for the optimal subset selection. This method is improved in two aspects. Firstly, a method based on variance analysis is proposed, which measures feature differences between categories, and obtains a modified arrangement displacement scheme to guide rearrangement of the order of values taken on data sample out of bag. Secondly, the sequence forward search method used to do feature selection could get iteration evaluation results, which solves randomness to determine the size of feature subset and instability of the results existing in the original algorithm. The results show RF-VA can obtain a better subset of features. It is feasible to reduce the dimension of partial discharge characteristic set, and effectively improve the identification rate of partial discharge defect type.

Highlights

Feature space or feature sets required for partial discharge (PD) pattern recognition is obtained through different feature extraction methods
It can be seen from the figure that for the existing Random Forest (RF) feature selection method, when the classification accuracy reaches the highest values of 84.58%, 75%, and 84.58%, the corresponding feature dimensions are 12, 21, and 14, respectively
For the RF-VA feature selection method, when the classification accuracy reaches the highest values of 85.00%, 76.25%, and

Summary

Introduction

Feature space or feature sets required for partial discharge (PD) pattern recognition is obtained through different feature extraction methods. Features that need to be removed are usually divided into two categories: one is redundant features, that is, features are duplicated; the other is irrelevant features This subset transformation from d-dimensional features to d’-dimensional features is called feature selection, and these d’-dimensional features are used for model training. The common feature selection methods are roughly divided into three categories: Filter approach, Wrapper approach, and Embedding approach. These feature selection methods have been applied in various fields and achieved good recognition results.

Methods

Results

Conclusion