Research On Classification Algorithm Of Motion Imagery Signal Based On Brain-computer Interface

Brain-computer interface(BCI)technology has been developed and applied in many fields.As one of the three mainstream brain-computer interface paradigms,brain-computer interface based on Motor Imagery(MI)signals has been a hot spot for research because of its spontaneity and flexibility.With the successful application of Deep Learning(DL)in Computer Vision,natural language processing,etc.,it has received a lot of attention from researchers in Electroencephalogram(EEG)classification.In EEG classification with the application of deep learning,some problems still need to be improved.First,the signals extracted using existing EEG feature extraction algorithms suffer from feature redundancy or information loss.Second,as a class of big data-based methods,the existing methods do not further investigate the impact of data augmentation on model performance when performing EEG data augmentation.In addition,EEG data augmentation is different from image augmentation in Computer Vision,which has led to a lack of clarity among researchers regarding the ability of models to predict new samples learned by training on augmentation EEG data.In this paper,we address these three issues using motor imagery signals from the BCI Competition Set IV dataset 2a:(1)This paper uses a filter bank common spatial pattern to extract spatial and frequency domain feature information of EEG signals.Hilbert transform is used to Downsampling the EEG data after the spectrum shift,which ensures that the EEG data information is not lost,and the data compression is completed.The feature bands of EEG data were investigated by building a Convolutional Neural Network model.Research has shown that using a wider feature frequency band can cause feature redundancy while using a narrower feature frequency band can highlight the frequency domain characteristics of EEG signals.(2)This paper constructs a new feature matrix representation of EEG signals by extracting the spatial-domain features of EEG signals using common spatial patterns and the frequency-domain features of EEG signals using power spectral density analysis.For this feature matrix,a new five-layer convolutional neural network structure is proposed in this paper,and a classification accuracy of 96% on average is obtained after the classification strategy of EEG slicing is introduced.In this paper,two parameters of the window function,namely the window width and step length of the sliding time window,are proposed and introduced in the slicing strategy.The experimental results on adjusting the two parameters of the window function show that the training time cost of the model can be reduced by setting the appropriate window width and step length of the time window.(3)In this paper,the channel-based data enhancement method is introduced to extract time-frequency features of single-channel EEG signals using wavelet packet decomposition.For the newly constructed EEG matrix representation,a new three-layer CNN model is proposed,and the performance of the model is tested in the sense of two strategies.This article clarifies the difference in the classification ability obtained by the model under the two strategic meanings through the analysis of experimental results and reasonable theoretical inference and demonstrates the new significance given to strategy one in the inference through supplementary experiments.Finally,this article found through the study of data augmentation for various numbers of EEG channels with different distributions under the strategy that based on a smaller number of channel augmentations,the model still maintains high performance,which provides a reference for considering device costs in real-time BCI applications.