Abstract: In our report， a novel filter bank-based preprocessing structure and feature extraction model was designed to alleviate the effects of the variance of subject on BCI system. During the filter bank-based stage， EEG signals were filtered by multi-band pass filter to alleviate the effects of the difference of related sub-band frequency corresponding to different subjects. During feature extraction stage， common spatial pattern （CSP） was used to extract features library， then， group lasso （GL）， extreme learning machine autoencoder （ELM_AE） and mutual information （MI） were used for feature selection to improve feature separability. The results showed that the proposed method can highlight the sub-band frequency of the alpha and beta wave， which has high separable features and classification performance.