Motor Imagery EEG Signal Accurate Analysis Based On Deep Learning

Brain Computer Interface(BCI)system has the advantage of bypassing peripheral nerves and muscles to establish direct connections between the brain and external devices.The BCI system based on motor-imagining electroencephalograph(EEG)signal is a spontaneous braincomputer interface system in which the subjects can generate EEG signals through their own imagined movements without using external stimulation.However,the motor-imaging EEG signal is a quite weak sequence signal with the characteristics of non-stationary and low Signal to Noise Ratio(SNR).When selecting the signal channel,there are some traditional methods that use manual selection,as well as other methods is to input all channel signals into network.However,the irrelevant or poor-relevant channel information also will be input into the signal channels without any focus,resulting in the loss or redundancy of motor-imaging EEG information,and low classification accuracy of motor-imaging EEG signals in four classifications——left hand,right hand,feet,tongue or rest tasks.Moreover,different subjects have different time from the prompt of the motor-imaging task to the execution of it,and the amplitude of the motor-imaging EEG signal generated by each subject for the same motorimaging task is different as well.That means there are differences in the reflection time and amplitude of the EEG signals of different subjects' motor imagery,which causes low accuracy in the classification among subjects.First,in response to the above problems,this thesis proposes an end-to-end multi-scale spatiotemporal self-attention network model to improve the classification accuracy of the four classification tasks of motor-imaging EEG signals.This model consists of two modules: feature extraction module and feature classification module.The feature extraction module includes two layers,namely the spatial self-attention layer and the parallel multi-scale temporal convolution layer,which extract the original motor-imaging EEG signal feature information from space domain and time domain respectively.The spatial self-attention module calculates the similarity degree between each signal channel,automatically learns the similar weight value between the channels,adaptively aggregates the signal data of all channels,and uses weighted summation to update the information of each channel.This module eliminates the information loss caused by manual selection of signal channels in traditional machine learning.At the same time,it can automatically select signal channels related to motor imagery to extract the discriminant features of the subjects and enhance spatial information.In the time domain,multiscale temporal convolution module is used to extract the time domain feature information under different time scales and merge them to eliminate the interference of noise in the motor-imaging EEG signal,then obtain enhanced time domain feature information.After that,the feature fusion module is used to combine the output of the two small modules to get the enhanced spatiotemporal features in the space and time domains,and the final classification results are output through the feature classification modules.Experiments have proved that this model has good classification performance and robustness in the classification results of single subject,as well as a certain transfer learning ability.Next,purpose of improving the accuracy of the model's cross-subject classification,enrich the frequency domain information of the time-frequency domain and enhance its feature representation,and solve the problem of low cross-subject classification accuracy due to the storage of a large number of other subjects' historical information in the multi-scale temporal convolution layer,a time-frequency-spatial self-attention network model is proposed.Firstly,the original motor-imaging EEG signal is converted into the time-frequency feature map of each channel through morlet wavelet transform,and the spatial self-attention network module is also used in the spatial domain.In the time domain,the time self-attention network module is introduced to replace the multi-scale temporal convolution module.This module does not store a large amount of other historical information of the subjects and can extract the dependency between sampling points at different times,and each time point contains global time information,thereby enhancing the ability to represent features in the time domain.In the frequency domain,the frequency self-attention network module is introduced.This module can extract the dependency between different frequency sampling points,and each frequency point contains global frequency information,which can enhance the feature representation ability in the frequency domain and improve the cross-subject classification accuracy of the model.The experimental results show that the network model has higher cross-subject classification accuracy than the previous model and is more suitable for cross-subject classification.