Abstract：According to the signal processing of EEG in brain-computer interface (BCI), a generalized kernel linear discriminant analysis (GKLDA) method based on kernel method and generalized singular value decomposition (GSVD) was proposed to extract feature of EEG with two classes. First, the samples were mapped using the linear discriminant analysis in the feature space defined by a nonlinear mapping through kernel functions. Secondly, a nonlinear spatial filtering was solved through the GSVD which can solve the small sample size problem. In the experiment, the GKLDA was contrasted with kernel common spatial pattern (KCSP), kernel linear discriminant analysis (KDA) and generalized linear discriminant analysis (GDA) for three public data which are dataset of BCI Competition Ⅰ, dataset Ⅳ of BCI Competition Ⅱ and S4b in dataset ⅢIB of BCI Competition Ⅲ. The same method was used on the dataset from ourselves with the fisher linear discriminant analysis classifier. The accuracy of the propose GKLDA feature of the three data are 93%,77%,80%, and 97% on the dataset from ourselves, better than the other kernel method. Experiment results indicate that, the GKLDA method can be well a new effective feature extraction method in brain computer interface.