Research On Multi-Subject Motor Imagery Brain-Computer Interface Algorithm Based On Temporal Attention Model

A Brain Computer Interface(BCI)system is a communication and control system that creates non-muscular output channels for the brain.Common BCI paradigms include Motor Imagery(MI),Steady-state Visual Evoked Potential,Event-related Potential,and Emotion Recognition,etc.Among them,Motor Imagery Brain-Computer Interface(MI-BCI)is a technique that decodes the subject’s motor imagery intention by recording neuronal activity in the brain and converts it into control signals to control external devices.It has been playing an increasingly important role in the rehabilitation of patients with stroke and muscular dystrophy because of its ability to activate partially damaged motor areas in the brain and achieve motor recovery.In addition,it can help elderly or disabled people control wheelchairs or other motor equipment,which is one of the classical paradigms of BCI.Therefore,the study of EEG recognition algorithms for motor imagery has important practical significance and broad application prospects.EEG signals are characterized by difficult acquisition and subjects are easily fatigued,so the trainable samples of motor imagery EEG are small,which is a typical small sample problem.This feature leads to models that are easily overfitted to the training set,and also makes it difficult to apply complex models to the motor imagery EEG task.In addition,influenced by human anatomy,behavioral habits,response behaviors and other factors,current motor imagery EEG recognition algorithms face the problem of subject differences,which are mainly reflected in the time domain as the phenomenon of Event-related Desynchronization/Event-related Synchronization appear and persist at different times,resulting in different effective signal segments of motor imagery EEG between subjects,and thus the model has poor generalization ability between subjects and poor experimental results on new subjects.How to make the model adaptively recognize and utilize the effective signal segments of motor imagery from different subjects is the key to improve the performance of motor imagery EEG recognition algorithm.Based on the above questions,this paper will carry out research in the following two areas:(1)A mirror convolutional neural network model for motor imagery EEG recognition is proposed to address the problem of poor model training due to the small sample size of motor imagery EEG training.Specifically,in the training phase,the model constructs mirror EEG data by exchanging the corresponding EEG channels of the left and right brain hemispheres,which effectively expands the motor imagery EEG data set,and in the testing phase,the mirror convolutional network structure is constructed for probability ensemble,which further improves the performance of the network model without increasing the training cost.Experiments using two motor imagery EEG benchmark datasets on three benchmark networks,Shallow ConvNet,Deep ConvNet and EEGNet,demonstrate the generality of the mirror convolutional neural network model that can be built based on a variety of motor imagery EEG recognition benchmark network.The excellent performance of the model lays the foundation for research on multisubject motor imagery brain-computer interface based on the time-domain attention model.(2)To address the problem of time-domain effective signal segment differences in motor imagery EEG subject differences,a shallow mirror Transformer model for motor imagery EEG recognition is proposed,which can capture EEG global time-domain information in the shallow layer of the network by using a multi-headed self-attention mechanism in the shallow layer of the network,and adaptively identify and utilize motor imagery effective signal segments from different subjects,effectively compensating for the Convolutional neural network has the disadvantage of small receptive field and cannot capture global information.In addition,the performance of the network is further enhanced by the mirror network structure.Experimental results on three motor imagery EEG benchmark datasets demonstrate the effectiveness of the shallow mirror Transformer model and show the ability of the model to detect and utilize motor imagery EEG valid signal segments through visualization of time-domain attention weights.In summary,the research conducted in this paper on multi-subject motor imagery braincomputer interface based on time-domain attention model provides a feasible solution to the problem of time-domain subject differences in multi-subject motor imagery EEG and promotes the rapid application of motor imagery brain-computer interface among different subjects,which has important research significance and practical application value.