CAJ | 학술논문

网络故障诊断中大量无关或冗余的特征会降低诊断的精度，需要对初始特征进行选择。Wrapper模式特征选择方法分类算法计算量大，为了降低计算量，本文提出了基于支持向量的二进制粒子群(SVB-BPSO)的故障特征选择方法。该算法以SVM为分类器，首先通过对所有样本的SVM训练选出SV集，在封装的分类训练中仅使用SV集，然后采用异类支持向量之间的平均距离作为SVM的参数进行训练，最后根据分类结果，利用BPSO在特征空间中进行全局搜索选出最优特征集。在DARPA数据集上的实验表明本文提出的方法能够降低封装模式特征选择的计算量且获得了较高的分类精度以及较明显的降维效果。
망락고장진단중대량무관혹용여적특정회강저진단적정도，수요대초시특정진행선택。Wrapper모식특정선택방법분류산법계산량대，위료강저계산량，본문제출료기우지지향량적이진제입자군(SVB-BPSO)적고장특정선택방법。해산법이SVM위분류기，수선통과대소유양본적SVM훈련선출SV집，재봉장적분류훈련중부사용SV집，연후채용이류지지향량지간적평균거리작위SVM적삼수진행훈련，최후근거분류결과，이용BPSO재특정공간중진행전국수색선출최우특정집。재DARPA수거집상적실험표명본문제출적방법능구강저봉장모식특정선택적계산량차획득료교고적분류정도이급교명현적강유효과。
In network fault diagnosis, many irrelevant and redundant features lessen the performance of diagnosis, feature selection is introduced on this condition. The wrapper feature selection algorithms get large calculation cost, a support vector based binary particle swarm optimization(SVB-BPSO) feature selection algorithm was proposed in this paper. The support vectors(SVs) are selected from the whole datasets by SVM training, the following wrapper classification focus only on these SVs. The training parameter is decided by average distance between different class SVs. Based on the SVM classifiers, the BPSO is used for searching the whole feature space to find the best feature subset. Experiments on DARPA datasets show the proposed method can reduce the wrapper feature selection's calculation cost while gets good performance on diagnosis accuracy and dimensional decrease.