A self-adaptive classifier which combines kernel principal component analysis（KPCA） with real-value negative selection algorithm is proposed to prevent missed diagnosis and insufficiently detailed detection in fault diagnosis. Real value encoding is employed

and Euclidean distance measure is utilized to measure the affinity of sample and detector. Thus

the classifier achieves the ability of learning and classification for the purpose of detailed classification of faults. Kernel principal component analysis technique is applied to extract the characteristic quantities of sample according to the data of dissolved gases in transformer oil

and the normalized characteristic quantities are used for training the classifying algorithm. Test results show that compared with BP neural network method and least squares support vector machine method

the proposed self-adaptive classifier obtains higher fault detection rate

and its fault detection rates for four types of power transformer faults all reach more than 90%.