Deep Learning-based MI-EEG Decoding Method And Its Application In Brain-computer Interface System

Brain-computer interface system constructs a new information transmission channel between the brain and external devices without depending on peripheral nerves and muscles.It can not only enable patients who have lost limb motor ability to complete some life activities by controlling external devices through their brain directly,which is helpful to improve their self-care ability.It can also provide rehabilitation training for patients who have motor dysfunction due to the malfunction of the nerve pathway,and help them rebuild the nerve pathway to restore their motor ability.The key technology in brain computer interface system is EEG decoding.Although the current deep learning-based methods have achieved excellent performance in the MI-EEG decoding tasks,there are still some challenges in the field of deep learning-based MI-EEG decoding.In this study,we have proposed the corresponding solutions to some challenges and have built an upper limb rehabilitation system based on brain-computer interface according the proposed MI-EEG decoding methods.To alleviate the overfitting problem,a data augmentation method based on circular translation strategy is proposed in this study.By adjusting the circular translation step size,a large number of new samples that retain the main temporal and spatial features of original samples and have certain differences can be obtained through this data augmentation method.In addition,this data augmentation method will not introduce additional noise and each new sample contains all the data of the original sample.In order to improve the discrimination of different classes of samples in the feature space,a high discrimination feature learning strategy based on central distance loss,center vector shift process and center vector update process is proposed in this study.This method can reduce the distance of the same class of samples in the feature space through the constraint of central distance loss,and increase the distance of different classes of samples in the feature space through the center vector shift process.Furthermore,this method can also weaken the influence of the initialization and make the central distance loss converge better through the center vector update process.In addition,a MI-EEG decoding method based on the two-branch 3D CNN model is proposed to take full advantage of the spatial features related to electrode distribution and avoid the mutual interference between temporal and spatial features in the feature extraction process in this study.Firstly,a concise 3D representation method of MI-EEG data is adopted to represent each MI-EEG sample as a 3D digital matrix.Then a two-branch 3D CNN model is designed to extract temporal features and spatial features through a time branch and a spatial branch respectively,so as to avoid the mutual interference between temporal features and spatial features.Finally,a combined model for MI-EEG decoding is constructed according to the proposed high discrimination feature learning strategy,feature separation method and the twobranch 3D CNN model.And a brain-computer interface system for upper limb rehabilitation is built based on the combined model in this study.First,the MI-EEG signals of each subject is collected through the data acquisition equipment.Then the collected MI-EEG signal is decoded through the combined model.Next,the decoding results are converted into the control signals and sent to the execution module by the communication module.Finally,the execution module drives the upper limb of experimental subject to carry out the corresponding rehabilitation movement.In order to verify the effectiveness of the proposed MI-EEG decoding methods based on deep learning,a large number of experiments are carried out on the famous public MI-EEG datasets in this study.The experimental results fully prove that these methods can effectively improve the decoding accuracy.In addition,some volunteers were invited to participate in some experiments to verify the effectiveness of the decoding method and the feasibility of the brain-computer interface system for upper limb rehabilitation.The researches of this study are expected to further improve the accuracy and stability of brain-computer interface system,which not only has great scientific and social significance,but also has very broad application prospects.