基于两阶段学习的半监督SVM故障检测方法  

Fault detection based on two-stage learning and semi-supervised SVM

作　　者：陶新民[1] 曹盼东[1] 宋少宇[1] 付丹丹[1] 

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《振动与冲击》2012年第23期39-43,56,共6页Journal of Vibration and Shock

基　　金：国家自然科学基金面上项目(61074076);中国博士后科学基金(20090450119);中国博士点新教师基金(20092304120017);黑龙江省博士后基金(LBH-Z08227);黑龙江省教育厅项目(11555009)

摘　　要：提出一种基于两阶段学习的半监督SVM故障检测方法。该方法首先使用标识传递算法给未标识样本赋予初始伪标识,并通过k近邻图对比样本点标识值,将可能是噪声的样本点识别并剔除;然后将去噪处理后的样本集输入到SVM中,使得SVM在训练时能兼顾整个样本集的信息,从而提高SVM的故障检测性能。实验中将该方法同支持向量机(SVM)、模糊支持向量机(FSVM)、直推式支持向量机(TSVM)及拉普拉斯支持向量机算法(LapSVM)进行比较,结果表明该方法在不同数目标识样本集合的情况下,检测精度较其他算法有较大幅度提高,同时该方法还比较了不含测试样本和含测试样本训练条件下的故障检测性能,结果表明结合测试样本可进一步提高算法的故障检测性能。A novel semisupervised support vector machine fault detection model based on twostage learning was presented here. First of all, propagation algorithm was used to provide pseudo labels for unlabeled samples. Secondly, k nearneighbor graph was utilized to distinguish and delete the noisy samples. Then, the denoised samples were input into a support vector machine（SVM） so that the global information of the whole sample set could be considered by the SVM to enhance its fault detection ability. The comparisons among support vector machine, fuzzy support vector machine, transductive support vector machine and Laplacian support vector machine fault detection algorithms were performed. The results showed that the proposed approach has a more substantial detection accuracy than other algorithms with different labeled sample sets. The behaviors of the proposed fault detection method with test samples and without those were compared. The results illustrated that the proposed method with test samples can further raise its ability to detect faults.

关 键 词：故障检测 半监督 两阶段 伪标识 

分 类 号：TH17[机械工程—机械制造及自动化] TP306[自动化与计算机技术—计算机系统结构]