Research On The Classification Of Motor Imagery Based On The Combination Of CSP And Deep Learning

Brain-Computer Interface(BCI)technology enables the human brain to communicate with the outside world and to control external devices a communication system that does not rely on peripheral nerves and muscles,mainly by creating a transmission channel between the brain and external devices.BCI technology has promising applications in the fields of medical rehabilitation,transport and the military.The core of BCI technology is the rational analysis of electroencephalogram(EEG)signals and the extraction and classification of features.Traditional methods for classifying EEG signals in Motor Imagery(MI)usually require a lot of data pre-processing and tedious feature extraction and classification.Compared to traditional methods,deep neural networks can automatically extract and classify features.However,the low signal-to-noise ratio of EEG signals makes them unsuitable as input to neural network models.In this paper,different deep neural networks are used to classify MIEEG,and a Common Spatial Pattern(CSP)spatial filtering is used to improve the signal-to-noise ratio of the data and increase the classification accuracy before classification.The detailed work is as follows.(1)Three basic neural network models,MLP,CNN and LSTM,were constructed and later combined to obtain the CNN-LSTM and CNN-Bi LSTM models.The results of the experiments conducted on the BCI competition dataset 2b showed that the CNN-Bi LSTM classification outperformed the MLP,CNN,LSTM and CNN-LSTM models,and the average Kappa values of the classification results for different subjects were calculated based on the CNN-Bi LSTM model and compared with the top three competition results.The average Kappa value was 0.1 higher than that of the top contest result,and the Kappa of each subject was higher than that of the contest winner.(2)In addition the five models were applied to the laboratory dataset and the results showed that the CNN-Bi LSTM model also had good classification accuracy.The training accuracy of the models was also recorded and the losses of the training models were plotted,and the CNN-Bi LSTM was found to have better generalisation.And the effect of band-pass filtering below 7Hz and above 40 Hz on the classification results of the neural network was also explored for the BCI competition dataset 2b and the laboratory dataset,and the results showed that there was some improvement in the classification results of each model after band-pass filtering of the raw data.(3)In order to improve the signal-to-noise ratio of the input data,this thesis also combines CSP with a deep learning model,that is,the raw EEG signal that has been band-pass filtered is followed by CSP spatial filtering,after which a neural network model is used to learn and classify the features.On the competition dataset,experimental results showed that the average results of CNN,LSTM,CNN-LSTM and CNN-Bi LSTM models combined with CSP spatial filtering were 0.83,0.41,1.86 and 0.19 higher than the average results of band-pass filtering only;on the laboratory dataset,experimental results showed that MLP,CNN,LSTM,CNN-LSTM and CNN-Bi LSTM models combined with CSP spatial filtering gave average results that were 1.67,1.84,1.67,1 and 2.5higher than the average results with bandpass filtering only.The band-pass filtering followed by CSP spatial filtering on the raw data proposed in this thesis can effectively help the deep learning model to extract effective features and improve the classification results.