Creation Of Multi Classification EEG Data Set And Classification Algorithm Research

Brain computer interface(BCI)is a technology that communicates with the outside world through human brain,involving computer science,rehabilitation medicine,information processing and other fields.It is a new way of human-computer interaction.By analyzing and recognizing EEG signals and converting them into control commands for external devices,BCI system provides opportunities for patients with severe paralysis to interact with the outside world.Convolutional neural network(CNN)is a commonly used algorithm to classify EEG signals.Because it has the characteristics of local connection and weight sharing,it can extract the deep-level features of EEG signals,and has excellent performance in the classification of EEG signals.CNN’s classification performance includes classification accuracy,classification speed and generalization ability.Because of the problem of the amount of original data,most of the methods to improve the performance of classification algorithm are to constantly improve the algorithm structure to adapt to the existing labeled data.Although the classification accuracy of some literatures has reached a very high level,the problem that the generalization ability is not strong due to the limitation of the number of database samples and the number of subjects still needs to be solved.The best way to improve the generalization ability of classifiers is to use more labeled training data.However,because the experimental design and data acquisition of motor imagery EEG require higher professional technology and objective environment,most of the current authoritative databases are using limited competition data sets,in which the number of subjects and the effective data collected are less,which seriously restricts the development of related research work.Therefore,in this study,we build our own data set and train the classifier based on the research results of the existing competition data set,and focus on improving the adaptability of the classifier to the data obtained by the subjects under different test modes and more widely.The successful classification of micro motion imagination,such as joints,can make brain computer equipment more flexible and accurate,which is of great significance to the practical application of BCI.However,the existing public data set only includes the classification data of the degree of finger joint curvature.If the relevant research needs to be carried out,it must rely on self built datasets.But the most important point of self built data set is to prove the validity of data,and it is very difficult to collect and verify the correctness of BCI data set.The method used in this paper is to train the classifier by using the competition set data as the training set,and then use the competition set data and the data from the self built data set as the test set of the classifier for classification.When the classification accuracy of the self built data set reaches a preset accuracy rate(> 74%),and the classification accuracy of the competition set data,we think that the subject has become a qualified subject,and its data accuracy has been able to approach the level of the competition set data.In this paper,CNN is used as classifier and improved.Through the continuous screening of qualified test data to the training set,and constantly expand and establish a self built data set with more data and meet the test conditions,lay the foundation for future research work.It is an important research direction of BCI in practical application to accurately recognize more subtle and more joint and muscle group movements.Due to the lack of relevant data,There is no previous study on the classification of wrist motion imagination.This paper proves the accuracy of the subjects’ Sports imagination about the hands,feet and tongue by comparing the data of the competition set.On this basis,the subjects can make more accurate sports imagination than untrained people,so as to establish the basis for expanding the research scope.We increased the imagination of the small joints of the trained subjects,and achieved the classification accuracy of about 82%.