改进投票策略的多类SVM及在故障诊断中应用  

Multi-class SVM based on improved voting strategy and its application in fault diagnosis

作　　者：吴德会[1,2] 

机构地区：[1]九江大学数字控制技术与应用江西省重点实验室,江西九江332005 [2]清华大学电机工程与应用电子技术系,北京100084

出　　处：《系统工程与电子技术》2009年第4期982-987,共6页Systems Engineering and Electronics

基　　金：国家自然科学基金资助课题(50705039)

摘　　要：针对一对一（OVO）分解法,提出了一种改进的投票（MWV）策略,解决了传统策略中的不可分区域问题。首先,由训练iω类和ωj（j≠i,j=1,…,n）类而得到的SVM决策函数;再对iω类定义了一个取值在0~1之间的调节函数,并使改进的得票值等于传统得票值加上调节函数。最后,根据改进的得票值进行分类决策。对于可分区域的样本,改进MWV策略的分类结果与传统策略完全相同;对于不可分区域的数据,由调节函数的值决定。将所提法应用于齿轮传动箱故障诊断实例并与传统得票策略诊断进行了对比,实验结果验证了所提方法的上述优越性。An improved max-wins-voting （MWV） strategy for one-versus-one （OVO） classification is developed and the unclassifiable regions existing in conventional one are resolved. Firstly, using the decision functions obtained by training the SVM for classes wi and ωj（j≠i,j=1,…,n）, for class wi, a novel tuning function is defined in the range of 0-1. Secondly, the improved voting value for class wi equals to the traditional voting value plus the tuning function. Finally, a classification decision is made according to the improved voting value. For the data in the classifiable regions, the classification results using improved MWV strategy are the same as that using the traditional one. Whereas, the data in the unclassifiable region are determined by the tuning func- tion. The comparison is done with experimental data in the application of fault diagnosis for gearbox. Experimental results demonstrate the superiority of the presented strategy.

关 键 词：模式识别 多类支持向量机 投票法 故障诊断 一对一分解 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]