Abstract: In order to handle the invalidation problem of fault detection when Kernel Principal Component Analysis（KPCA） modeling sample is impure, a new KPCA monitoring model for power transformer fault detection based on Modified Feature Sample（MFS） method is proposed. A new purification algorithm is proposed to eliminate abnormal samples from original database by using eigenvalue variation. Then a Feature Sample（FS） method is adopted to extract modeling samples of KPCA; compound statistics is used to verify the state of power transformer. Experimental results show effectiveness of the modified feature sample algorithm and the proposed method has high fault sensitivity and diagnosis accuracy.

Key words: power transformer, fault detection, Kernel Principal Component Analysis（KPCA）, feature sample

摘要： 针对核主元分析（KPCA）监控模型由于建模样本不纯而导致故障检测失效问题，提出基于改进特征样本方法的KPCA故障检测模型并应用于变压器故障检测中。利用特征值变化信息，设计出异常样本剔除算法以避免异常样本被选入特征样本集；采用特征样本方法提取建模样本集，建立KPCA监控模型，采用复合统计量对变压器运行状态进行检测，实验结果验证了改进特征样本算法的有效性，表明提出的方法具有较高的故障敏感性和检测效率。

关键词: 电力变压器, 故障检测, 核主元分析, 特征样本