Research On EEG Feature Extraction And Feature Transfer Algorithm

Brain-computer interface(BCI)is a kind of system which can realize the direct control of external devices by the brain without relying on the normal output channels of brain nerve and muscle tissue.The research of BCI technology has brought new changes in many fields,and its application prospect is very broad.However,as a complex multi domain hybrid technology,BCI still has many problems in theory and key technologies,and it needs to study better models and methods to promote its development.Electroencephalogram(EEG)is the focus of BCI research because of its low cost and no operation.EEG signals can effectively analyze and interpret brain physiology,thoughts,emotions and other information.However,each person's EEG signal is very different,and there are many factors affecting the signal,which has great limitations to the recognize the EEG signal.Therefore,this paper aims at the key point of how to improve the accuracy of EEG signal recognition and reduce the calibration time of user verification in BCI research.The main work is as follows:(1)According to the characteristics of multi-channel analysis and bad feature extraction of EEG signals,a feature extraction method based on optimal region common space pattern is proposed and applied to EEG signal classification.Firstly,the region of each channel is obtained by Euclidean distance,and then the region with the highest separability is selected according to the variance ratio.Then,the number of channels in the region is optimized by 5-fold cross validation,and the regional features with the highest degree of differentiation are obtained.Finally,the frequency of the selected regional channels is optimized and classified by support vector machine.The proposed method was applied to the open EEG datasets BCI Competition III dataset IVA and Competition IV dataset 2a.The results of experiment indicate that ORCSP can dislodge a great deal of redundant channels and achieve considerable test accuracy.The results show that the selection of regional channels is necessary,which has positive significance for the further study of EEG feature.(2)According to the characteristics of EEG signal calibration time and great individual difference,a multi-source manifold embedding transfer algorithm based on transfer learning is proposed,which is applied to the study of EEG classification.The algorithm uses the training data of other subjects and achieves good classification results when the target object does not have any training data.Firstly,the average covariance of the source domain features is aligned with the average covariance of the target domain by marginal distribution alignment,and then the vector representation of the features is obtained through tangent space mapping.Finally,the domain invariant classifier with minimum structural risk is constructed by reconstructing manifold features and joint distribution alignment.Finally,the target can be identified by using the data of other objects without the labeled data of the target domain to recognize the target signal.The proposed method has been tested on the public EEG datasets,and the experimental result verifies the effectiveness of the algorithm.It can effectively reduce the long and boring calibration time of subjects,improve the reusability of large amounts of data to solve the key problem in BCI research.