基于最大信息系数的变压器过热故障特征选择

Feature selection of dissolved gases in power transformer based on maximal information coefficient

改良三比值法只关注若干个气体浓度比值,信息利用不充分,而且气体浓度的随机误差对故障诊断结果有影响,因此文中将变压器故障特征气体扩充为单种气体增长率、多种气体比值和相对浓度等62个故障特征,通过具有稳健性的最大信息系数提取与变压器故障状态相关度高的故障特征。同时为了避免筛选特征之间的冗余性,采用相关系数筛选冗余性小的特征组合,并采用距离相关系数和多种分类器进行检验。结果表明与油中溶解气体相比,筛选特征集合与变压器过热故障状态联系更加紧密,且针对过热故障类型的诊断精度,筛选特征集合在样本类别不平衡时性能更优,突破了单一分类器性能上限。

The key gas method and IEC method for dissolved gases analysis only focus several gas index that may be affected by random error and relative percentages,and maybe it is not perfect to be used to conduct preventive maintenance and inspection of transformers.To solve the problem,key gases are mapped into 62 features including traditional combinations of key gases and other new features.The features highly correlated to transformer status are selected by maximum information coefficient,to escape random error of gas concentration.To avoid redundancy between selected features,Pearson coefficient is used to filter features to reduce the redundancy of the selected features.The result showed that correlation coefficient between the selected features and the transformer fault is relatively high.And the result of distance correlation coefficient also shows that the feature set formed by the selected features is more closely related to the transformer overheat fault state than that of the traditional fault characteristic gas.