Motion Imagery EEG Feature Classification Algorithm Based On Deep Learning

Brain-Computer Interface(BCI)is a new type of communication system for human-machine interaction,which can directly control external devices through brain consciousness activities.It has established a new bridge between the human brain and devices.Electroencephalogram(EEG)is an important component of the BCI field.It is generated by the electrical activity of neurons in the brain and contains rich physiological information reflecting mental activities.The classification of Motor Imagery(MI)Electroencephalogram is one of the important branches in the BCI field.Based on deep learning,this thesis explores the motor imagery classification algorithm based on Convolutional Neural Networks(CNN)from three perspectives: multi-feature fusion,data augmentation algorithm,and visualization research.Three different CNN models based on motor imagery EEG signals are designed,and the classification performance of the models is verified by publicly available EEG signal datasets.The research contents of the thesis include the following aspects:(1)A fusion feature extraction algorithm based on Sample Entropy(SE)and Common Spatial Pattern(CSP)is proposed to solve the problem that traditional EEG feature extraction algorithms only represent a single feature domain,lacking overall features.After decomposing and reconstructing the EEG signal using wavelet packet transform,the noise interference is removed.Non-linear feature sample entropy and spatial feature CSP are extracted from the reconstructed signals separately,and the two features are combined by serial concatenation to obtain the fusion feature.A novel one-dimensional CNN is designed to classify the fusion feature,and 91.66%classification accuracy is achieved in the 2003 BCI competition dataset III,while an average classification accuracy of 85.29% is achieved in the 2008 BCI Dataset A.(2)To address the overfitting problem of CNN caused by the small sample size of motor imagery EEG signals,a feature extraction method based on wavelet transform and data augmentation is proposed.The noise is removed by wavelet soft thresholding,and the frequency component of 8-30 Hz is extracted.The cmor wavelet is used for wavelet transform,which is then converted into a two-dimensional time-frequency map.The map is used as the input of a novel CNN designed to classify the motor imagery EEG signals.The classification accuracy of 88.89% and Kappa value of0.74 are achieved in the 2003 BCI competition dataset III,and the average classification accuracy of71.12% is achieved in the 2008 BCI Dataset A.Data augmentation algorithm is used to enhance the classification performance of the model.However,when the sample set is replicated too much,noise will be introduced,which may interfere with the model’s classification accuracy.(3)In order to extract the feature information of different frequency components in the EEG signal and avoid the feature loss caused by mixed signals,a multi-channel parallel CNN with frequency division is proposed based on the time-frequency map transformation.The key components of the motor imagery,μ rhythm and β rhythm,and their mixed signals are extracted by wavelet soft threshold algorithm and converted into time-frequency maps,which are input from left to right into the designed novel multi-channel parallel CNN.The CNN contains three input channels and can analyze different frequency feature maps simultaneously through a parallel structure,which speeds up the calculation and improves the efficiency.The classification accuracy of 92.86% is achieved in the 2003 BCI competition dataset III.Then,the t-SNE algorithm is used for CNN visualization to explore the classification process of the CNN.