Research On The Motor Imagery Classification Algorithm Based On Deep Learning

Motor imagery(MI)refers to a thinking activity in which one can imagine performing a specific movement without executing it.Brain-computer interfaces based on motor imagery can assist stroke and epilepsy patients in communication,as well as facilitate the control of external devices,such as wheelchairs and computer mice.However,due to the individual differences in brain anatomy and function,the temporal non-stationarity,and the uneven distribution of information of Electroencephalography(EEG)signals,improving the accuracy of classification tasks across subjects is still a challenging problem.Given the above issues,this thesis carried out two parts of research work on EEG-based MI signals based on deep learning algorithms,mainly including:1)Performing left and right-hand MI tasks demonstrates discrepant neurological patterns between two hemispheres.This thesis incorporates the hemispheric asymmetry neurological mechanism into convolutional neural network design for the first time and proposes a model named dual-branch parallel Hemisphere Discrepancy Network(d HDNet).The model uses a dual-branch structure to extract the spatial difference information of hemispheres.One branch designs the size of the convolution kernel to half the number of electrodes to make it convolve across the hemispheres;the other branch subtracts the left and right brain electrodes to extract the asymmetry information.Finally,the extracted features are fused and input to the fully connected layer for classification.Extensive experiments were conducted on the KU dataset in a leave-one-subject-out fashion.d HDNet achieved an average accuracy of 83.62% across subjects,surpassing the state-of-the-art models with fewer parameters.Furthermore,the performance of the proposed model was analyzed under different training set sizes and number of electrodes,and the proposed model showed strong robustness.The experimental results show that incorporating the hemispheric asymmetry pattern into the model design is beneficial to improve the cross-subject classification accuracy of the model.2)Aiming at the problems of non-stationarity and uneven information distribution of EEG signals,and the lack of generalization of existing static models,this thesis introduces dynamic convolution into EEG-MI classification for the first time,and designs a Dynamic Spatiotemporal Network(Dy TSNet)based on the characteristics of EEG data.The model learns different input-dependent weights through the time dimension and the electrode channel dimension respectively,and performs weighted combination of multiple convolution kernels of different dimensions.Such a design enables the weights of the neural network to have sample adaptive capabilities,thereby improving model capacity and generalization capabilities.Also using the leave-one-subject-out cross-validation strategy,Dy TSNet has an average accuracy of 1.70% higher than that of the comparison model on the BCI-C IV-2a dataset,and an average accuracy of 81.79% has been achieved on the KU dataset with 20 electrodes selected.Experimental results show that the specific design of dynamic convolutions on EEG motor imagery data is effective.The Dy TSNet model based on dynamic convolution can adaptively process complex and changeable EEG signals and has a stronger generalization ability.Based on the neurophysiological characteristics of motor imagery and the characteristics of EEG data,this thesis designed two motor imagery-specific deep learning models.They can effectively mine the transferable EEG pattern information among subjects,improving the classification accuracy of binary classification and multi-classification motor imagery across subjects.The research work of this thesis provides new research ideas and technical support for the design of motor imagery classification algorithm based on deep learning methods.