Research On Classification Of EEG Signals Of Motor Imagery Based On Multi-Scale Feature Attention

The Brain Computer Interface(BCI)is a kind of technology which connects the brain with the external equipment,and uses the EEG signal to communicate with the outside world.Motor imagery is one of the important applications of brain computer interface technology,which is widely used in medical rehabilitation,military and entertainment fields.Therefore,it is of great significance to improve the correct classification of motor imagery tasks.In the research of motor imagery EEG signal classification,using traditional machine learning methods requires manual extraction of EEG signal features,which may lead to loss of important features and feature coverage by noise;In addition,the single scale convolution using the depth learning method is difficult to adapt to the differences between subjects and the non-stationary characteristics of EEG signals,and the multi-scale convolution has the problems of feature importance differences within the scale branches and feature redundancy between the scale branches when extracting time-frequency features.To solve the above problems,this thesis uses the deep learning method to design two models,namely,based the Multi-Scale Feature Attention(MSCA)network model and based the Multi-Scale Feature Attention Gate Recurrent Unit(MSCA-GRU)network model,to classify the motor imagery EEG signals.The specific research contents of this thesis include:(1)Design of attention network model based on multi-scale features: The model is composed of improved Inception network and improved residual network.First,attention mechanism and residuals are introduced into each scale convolution branch of Inception network to construct residual attention Inception(RS-Inception)module;Then,five scale progressive RS-Inception modules are used to extract multi-scale time-frequency characteristics of motor imagery EEG signals in parallel;Secondly,the improved residual module,which introduces attention through residual jump,optimizes the multi-scale time-frequency features extracted to suppress feature redundancy between scale branches;Finally,the EEG signals of motor imagery were classified through the full connectivity layer.The original data is enhanced as the input data of the network model.In the model,the activation function selects the scaling exponential linear unit(Selu)activation function with negative transmission characteristics.The Selu activation function can enable this part of neurons to be activated when the input is negative,so as to better train the network.The experimental results on the 2a four classification and 2b two classification data sets of BCI Competition IV show that the average classification accuracy of MSCA network model is 83.46% and 88.98% respectively,which is superior to other advanced motor imagery EEG signal classification methods.(2)Based on the multi-scale feature attention network model Gate Recurrent Unit(MSCA-GRU)network model design: this model removes the flattening layer from the MSCA network model.After feature fusion,SE attention mechanism is added to remove the information redundancy in the fused features.Then,through the flattening layer and the Dropout layer,the Dropout layer is used to reduce the impact of over fitting on the classification results.Finally,the GRU structure is added to better mine the information in the time domain.It was compared with MSCA model and other advanced methods in the same subject experiment,and cross subject comparison experiment and ablation experiment were conducted on MSCA-GRU model.The results showed that the average classification accuracy of MSCA-GRU for the same subject was 83.69% and 89.23% respectively on the 2a four classification and 2b two classification data set of BCI competition IV,and the average classification accuracy of cross-subjects was 65.80% and 77.52%respectively,with improved accuracy.At the same time,it verified the effectiveness of adding Dropout layer and SE attention mechanism into the model,and proved the classification performance of MSCA-GRU model.