Research On Feature Extraction And Visualization Of Illiteracy Subjects In Motion Imaging Based On CNN

Brain-computer interface is a system that does not rely on peripheral muscles and nerves.Directly reading human brain electrical signals can enable people to control peripheral equipment.Its appearance not only brings hope to disabled people,but also provides social security for the elderly.Among the many brain-computer interface systems,motor imagination has become the focus of current brain-computer interface research with less external stimuli.CSP is currently the most commonly used method for extracting electroencephalographic features of motor imagination,which uses the ERD / ERS phenomenon of motor imagination.However,some studies have shown that about 20%-30% of the general population are "blind" in motor imagination.They have no ERD / ERS phenomenon when imagining limb movements,so the CSP algorithm compares the effects of "illiteracy" on these motor imaginations.difference.The existence of "illiteracy" in motor imagination has restricted the use of brain-computer interfaces and made people question the technology of brain-computer interfaces.In order to better study the "illiteracy" of motor imagination,some studies have defined the CSP-LDA two-category motor imagination EEG signal below 70% as "impairment" of motor imagination,and used the "entropy" value of resting state EEG Motion imagination is "blind".None of these can solve the problem fundamentally.By consulting the literature,there is currently no algorithm research to solve the problem of "blindness" in motor imagination.Therefore,combined with the CSP extraction of EEG defects and the characteristics of "blind" EEG signals of motor imaging,we propose to use CNN to study this problem.Through end-to-end training,a CNN for motor imaging "illiteracy" is established.model.Traditional EEG algorithm research generally consists of three aspects: EEG preprocessing,feature extraction and classifier.Most of the researches study each module separately,and then couple them together to get a better classification effect.This method has the advantage that Interpretable is better.CNN algorithm has made major breakthroughs in machine vision,natural language processing and other fields.With its powerful feature extraction and network representation capabilities,it has begun to be applied in motor imaging EEG classification and achieved good results.Due to the particularity and flexibility of convolutional neural networks,most of the research focuses on structural design combinations such as the number of convolutional layers,convolution kernels,activation functions,etc.The interpretation of the results is still not intuitive.In view of the above,our research work is as follows:In this study,an experimental paradigm of motor imagination was designed.Data from 20 subjects were collected and experimental data were preprocessed.CSP + LDA and FBCSP + LDA were used to classify the EEG data of 15 subjects,including 9 “blind” subjects and 6 normal subjects.The results obtained by CSP + LDA classification serve as our benchmark for CNN research.A "illiteracy" CNN model for motor imaging was trained,and the influence of the sampling rate of the EEG signal and the activation function in the CNN model on the classification of the EEG signal was investigated,and a good classification effect was obtained.When the EEG signal sampling rate is 1000 Hz,the average classification accuracy rate of the 15 participants is the highest.The classification effect is best when a linear activation function is used in the CNN structure.Based on comprehensive research results,the average classification accuracy rate of 9 “blind” CNNs who were tried was 91.57 %%,which was an increase of 33.35% compared to 58.22% of CSP + LDA.83.38% of CSP + LDA increased by 3.73%.I put forward my own visualization method for the results we got,and try to explain the results.We extract the weights obtained after training the spatial filter in CNN,and correlate them with the first filter and the sixth filter obtained by CSP,and get the two sets of weights with the largest correlation value in CNN weights for the topographic map.It is found that the spatial distribution of EEG features extracted by CNN is very similar to CSP.From this we conclude that the obvious EEG features of "blind" motor imaging may be distributed in other frequency bands.