Modern Defense Technology ›› 2022, Vol. 50 ›› Issue (4): 124-131.DOI: 10.3969/j.issn.1009-086x.2022.04.014

• TARGET CHARACTERISTIC, DETECTION AND TRACKING TECHNOLOGY • Previous Articles     Next Articles

A Micro-Motion Feature Importance Evaluation Algorithm Based on Random Forest

Qing-yuan ZHAO, Chun-mao YE, Yao-bing LU   

  1. Beijing Institute of Radio Measurement,Beijing 100854,China
  • Received:2021-12-30 Revised:2022-03-10 Online:2022-08-28 Published:2022-09-02

基于随机森林的微动特征重要性评估研究

赵庆媛, 叶春茂, 鲁耀兵   

  1. 北京无线电测量研究所,北京 100854
  • 作者简介:赵庆媛(1986-),女,河南商丘人。高工,硕士,主要研究方向为雷达智能化应用及目标识别。

Abstract:

In order to reduce the redundancy in the micro-motion feature set of aerodynamic targets and reduce the feature dimensionality, random forest is introduced to evaluate the importance of multi-dimensional micro-motion features. The random forest algorithm and project implementation process are described, and 18 micro-motion features in the time domain, frequency domain and time-frequency domain are extracted. VHF band measured data are used to verify the feature importance evaluation algorithm based on random forest, and the influence of feature selection on the performance of tree classifiers including Fisher, support vector machine and decision tree classifier, as well as the influence of the number of radar coherent accumulation pulses on the feature importance evaluation are analyzed. For Fisher and support vector machine, with the increase of feature dimension, classification accuracy can be improved slightly. For decision trees, classification accuracy is only affected by features with higher importance score.

Key words: random forest, micro-motion feature, feature importance, offline evaluation, aerodynamic target, modulation spectrum interval

摘要:

为了减少气动目标识别微动特征集的冗余,进一步降低特征维度,引入随机森林对多维微动特征进行重要性评估。描述了随机森林算法及其工程实施流程,给出时域、频域及时频域等18维微动特征提取方法,用VHF波段实测数据对基于随机森林的特征重要性评估算法进行验证,并分析了特征选择对Fisher、支持向量机和决策树3种分类器的性能影响,以及雷达相参积累脉冲个数对于特征重要性评估的影响。对于Fisher和支持向量机来说,随着特征维数的增加,能够一定程度地提高分类准确率;而对于决策树来说,只有重要性评分较高的特征对分类准确率有贡献。

关键词: 随机森林, 微动特征, 特征重要性, 离线评估, 气动目标, 调制谱间隔

CLC Number: