Decoding Study Of Lower Limb Motion Based On Convolutional Neural Network

As the key technology of human-computer interaction,Brain Computer Interface(BCI)has become a research hotspot in many fields,such as medical care,education,environmental control,entertainment and security guarantee.Among them,BCI equipment based on Electroencephalography(EEG)has been widely studied because of its advantages of portability,non-invasive,high time resolution and low economic cost.However,in the research of exerciserelated EEG,most of them focus on the classification of EEG signals of upper limbs,and few make in-depth exploration on the classification of EEG signals of lower limbs.Based on the above situation,this paper collected the EEG data sets of lower limb movement imagination and actual movement,and used the Convolutional Neural Network(CNN)method to decode the EEG signals of lower limb from three aspects.(1)According to the time-frequency characteristics of EEG signals,a convolution neural network classification method based on time-frequency map was designed.The time-frequency map was generated by Short-time Fourier Transform(STFT),and the time-frequency map was screened from time range and frequency according to the physiological basis of lower limb EEG signals.Because the convolution neural network has a good classification effect on the classified images,the convolution neural network was selected to complete the classification of time-frequency images.In the construction of convolutional neural network,normalization layer and dropout layer were added to improve the gradient update of convolutional neural network and reduce the structure of over fitting.The optimal training state of the model was obtained through parameter training and channel selection experiments,and it indicated that the main activation area of lower limb motor EEG signal was close to the central channel.The classification experimental results of the two data sets showed that the proposed convolution neural network method based on time-frequency map had better classification accuracy than the traditional method,and could effectively classify both motion imagination and actual motion data sets.(2)EEG signals have time-frequency characteristics and spatial domain characteristics,an improved convolutional neural network classification method of Common Spatial Pattern(CSP)was designed.In order to combine time-frequency features with spatial features,the timefrequency-space features of EEG signals were extracted by Wavelet Packet Decomposition(WPD)and CSP.In(1),the two-dimensional convolution layer of convolution neural network was improved to one-dimensional convolution layer,and the time-frequency-space characteristics were used as the input of the new convolution neural network.The CSP feature vector after wavelet packet decomposition was analyzed to judge the difference between the left and right CSP feature vectors,and the effect of model classification was evaluated from the perspectives of frequency band selection and CSP m-value selection.The classification experiment results of the two data sets showed that the improved method of common space mode could improve the classification accuracy to a certain extent.(3)A channel-based EEG data augmentation method was designed and applied to recurrent convolution neural network.Due to the limitations of EEG data acquisition,especially the movement related EEG data,it is impossible to obtain a large EEG data set for the research of neural network.By referring to the data augmentation method in deep learning and according to the characteristics of EEG data,the electrode channel was used to enhance the EEG data set related to lower limb movement.In order to verify the effectiveness and feasibility of the data augmentation method,a recurrent convolution neural network was also applied.Compared with the neural network in(1)and(2),its network structure was more complex,but it increased the ability to integrate context information.The differences of left and right brain data were analyzed,and the conclusion that most of the data of left and right brain had significant differences is obtained.After that,the single subject data augmentation and cross subject data augmentation were studied respectively.There was a significant improvement in the accuracy in the comparison experiment of single subject data augmentation,while the cross subject experiment was relatively poor,but it still indicated that EEG data augmentation could play a positive role in neural network.