In order to improve the recognition accuracy of motor imagery brain-computer interface, common spatial pattern(CSP) and convolutional neural network (CNN) methods were combined, and an improved filter bank common spatial pattern (FBCSP) and CNN algorithm was proposed for multi-classification motor imagery EEG signal recognition tasks. After the signal was preprocessed, the spatial projection matrix was calculated using the FBCSP containing the overlapping frequency band, and the data was projected to obtain a more discriminative feature sequence. Then the feature sequence was input into the constructed CNN model in a two-dimensional arrangement for classification. The proposed method was verified on the brain-computer interface competition datasets 2a and IIIa, and compared with other literature methods. The results show that the proposed method improves the classification accuracy of motor imagery EEG signals to a certain extent, and provides an effective method for motor imagery research.