一种改进的无监督学习SVM及其在故障识别中的应用  

DECISION IMPROVING OF UNSUPERVISED SVM FOR FAULT IDENTIFICATION

作　　者：柳新民[1] 刘冠军[1] 邱静[1] 胡茑庆[1] 

机构地区：[1]国防科技大学机电工程与自动化学院,长沙410073

出　　处：《机械工程学报》2006年第4期107-111,共5页Journal of Mechanical Engineering

基　　金：国家自然科学基金(50375153)资助项目

摘　　要：提出一种改进决策1-SVM方法(1-DISVM),并由此构建了基于单类样本训练的1-DISVM多分类模型。 1-DISVM是1-SVM方法的改进,通过对决策算法的修正,解决了1-SVM分类精度低的不足,并将其应用于直升机减速器故障识别中。结果表明该方法能够在训练样本数量少、不准确的情况下,自动排除错误样本的干扰,获得很好的分类结果,且具有无监督学习、分类精度高、易于扩展和代价小等优点。One-class support vector machine （1-SVM） has the ability to find outliers from a dataset without any class information, but it has been rarely applied to classification for it＇s low classification precision resulted from the algorithm limits. By modifying the decision-function of 1-SVM, a decisionimproved 1-SVM （1-DISVM） is presented to adjust the classification precision. Based on it, multi-classes classification models trained by single-class samples are designed. The 1-DISVM models are applied to a helicopter＇s gearbox faultidentification. The experimental results show that this method can get rid of the influence of wrong samples to achieve precise classification with small fault samples, and this method has the merits of unsupervised learning, precise classification, easy to expand and low cost.

关 键 词：支持矢量机 无监督学习 分类 故障识别 

分 类 号：TH17[机械工程—机械制造及自动化]