代价敏感VBGP在变压器故障诊断中的应用  

Cost-Sensitive Gaussian Process Classification with Variational Bayesian Treatment for Fault Diagnosis of Power Transformers

作　　者：尹金良[1,2] 朱永利[1] 郑晓雨 王国强[1] 

机构地区：[1]华北电力大学电气与电子工程学院,保定071003 [2]天津理工大学自动化学院,天津300384 [3]国家电力调度控制中心,北京100031

出　　处：《电工技术学报》2014年第3期222-227,236,共7页Transactions of China Electrotechnical Society

基　　金：河北省自然科学基金资助项目(E2009001392)

摘　　要：现有变压器诊断方法默认各种误诊代价相同,以全局误诊率最低为目标,而实际问题中误诊代价通常存在差异,不同类型的误诊造成的损失往往不同。针对此提出了代价敏感变分贝叶斯高斯过程(CS-VBGP),并将其应用于变压器故障诊断。该方法将代价敏感机制引入变分贝叶斯高斯过程,以误诊代价最小为目标,按贝叶斯风险理论预测新样本的类别,克服了仅追求低误诊率并不一定会带来符合实际意义的诊断结果的问题。变压器故障诊断实例分析表明,CS-VBGP有较高的诊断正确率,趋于提高高误诊代价类别的诊断正确率,具有代价敏感性,诊断速度足以满足变压器故障诊断的工程需求。The traditional transformer fault diagnosis methods are typical evaluated by estimating their error rate. However this makes sense only if all errors have equal cost. But in practical problems, cost caused by different type of misdiagnosis is usually unequal. In order to overcome the shortcoming that only pursuit of low misdiagnosis may not bring about meaningful diagnosis results, cost-sensitive variational Bayesian treatment for gaussian process classification（CS-VBGP） is proposed and is applied to power transformer fault diagnosis. The method aims to minimize the misdiagnosis cost by introducing cost-sensitive learning mechanism. Class labels of new instances are predicted according to Bayesian risk theory. Experimental results show CS-VBGP is capable of high fault recognition rate and tends to improve the diagnostic accuracy of high misdiagnosis cost class and the diagnosis speed meets the engineering requirements of the transformer fault diagnosis.

关 键 词：高斯过程 误诊代价 代价敏感学习 变压器故障诊断 

分 类 号：TM411[电气工程—电器]