Kernel entropy component analysis（KECA）method have shown unique advantages for removing redundantinformation and adjusting nonlinear structure of input data. In view of the difficulty of effective feature extraction inpower transformer fault diagnosis

the tedious selection of core parameters in KECA and the neglect of sample type in-formation

etc

a novel feature extraction method base on the optimization learning of feature extraction is proposed inthis paper

it is used to extract sensitive feature of power transformer. Firstly

various state samples are collected

andthen a data-dependent kernel function is build by considering sample class information. Secondly

based on cosine simi-larity

the parameters representing the class information in kernel function are optimized to minimize the similarity be-tween classes and maximize the similarity within classes;Then the optimized kernel function and kernel entropy com-ponent analysis are combined to form a supervised adaptive kernel entropy component analysis（SAKECA）feature ex-traction method

and the sensitive features which can effectively represent the state of power transformer are extractedfrom the original feature quantity by using this method;Finally

the extracted features are taken as the input of KELMto set up transformer fault diagnosis model. Compared with common feature extraction methods

the results show thatthe proposed method has good feature extraction effect and can effectively improve the accuracy of fault diagnosis.