Brain-Computer Interface (BCI) based on motor imagery (MI) has been applied in the rehabilitation exoskeleton widely. In the practical use, the low signal-noise ratio of electroencephalogram (EEG) signal results in the low classification accuracy in BCI. Therefore, many studies have focused on the improvement of feature extraction and classification algorithms. In this paper, we proposed an original method based on the deep convolutional neural network (CNN) to perform feature extraction and classification for single-trial MI EEG signal. Firstly, according to the EEG signal's characteristic that combining time and space information, we constructed a 5-layer CNN model to classify the MI; secondly, MI experimental paradigm was designed based on imagining left hand movement and foot movement, and the experimental data of MI were collected; thirdly, the proposed method was used in the public data set and experimental data set to build classification model, compared with the other three methods (power+SVM, CSP+SVM and MRA+LDA); finally, the classification model which achieved the best classification performance was applied in real-time control of upper-limb exoskeleton to verify the effectiveness of our proposed method. The results demonstrate that CNN can further improve classification performance: the average accuracies of public data set (90.75%±2.47%) and experimental data set (89.51%±2.95%) using CNN are both higher than that using the other three methods. Furthermore, in real-time control of upper-limb exoskeleton, the average accuracy of all subjects reaches to 88.75%±3.42%, which verifies the effectiveness of the CNN method. The proposed method can recognize MI, and provides theoretical basis and technical support for BCI applications in the field of rehabilitation exoskeleton. © 2017, Science Press. All right reserved.
