The development of electronic and computer technology makes non-invasive acquisition systems of EEG more universal adoption. Therefore this paper focus on the BCI based on motor imagery, which is a popular studied spontaneous EEG, producing without external stimulus. 6 healthy people participate in the examination, 5 of which are saved while one of which is thrown due to disturbing. Independent component correlation algorithm (ICA) was used to extract motor related component, meanwhile other filters also designed for better signal quality. Taking the motor related components PSD distribution after ICA as features, signal modal decoding based on reinforcement learning is processed in two algorithms. Both of them have good performance. Finally, on-line decoding model is descripted on the basis of the off-line decoding model. The result shows the accuracy of training model has significant difference with random level. Furthermore, the conclusion can be developed into human-computer integration based on EEG signals. © 2019 IEEE.