The Temporal-frequency-spatial Feature Selection And Classification Methods For Signal Of "Imitating Reading" Brain-computer Interface

Brain-Computer Interface system is a system that can set up a communication and control channel between brain and peripheral directly without the depend of the brain's normal output pathways,and person's intentions can communicate directly with the environment through this channel.One of the important work of Brain-Computer Interface is to analysis the user's intention from the collected EEG signals correctly,and the feature selection and pattern classification methods greatly influence the classification accuracy of Brain-Computer Interface.This paper is a research about pattern recognition method of Imitating-reading Brain-Computer Interface,mainly includes the following aspects.BCI mode can be divided into synchronous model and asynchronous model.Synchronous BCI,in which subjects complete a specific task within a specified period of time,is not completely self-control.However asynchronous BCI is a completely independent way of work,in which subjects can operate BCI system according their wishes.In actual use,most of the time users are in intermittent random working status,therefore,only asynchronous mode can meet the actual demand.There are many researches of feature selection under experimental model of Imitating-reading Brain-Computer Interface,and most of them are based on the single channel signal.These methods can not select feature from time-frequency-spatial three domain at same time even taking into account the multi-channels of information,so some useful information for classification may lose in the process of signal processing.Because of these reasons,this paper studies tensor temporal-frequency-spatial pattern.Thismethods can select feature from multidimension at same time,and diagonalize high-dimensional covariance matrix of EEG signals.It reserves more information of EEG signals,and has higher performance of classification compared to Common Spatial Pattern.Tensor is can be applied not only to feature selection,but also pattern classification.Most recent machine learning algorithm is designed based on vector space,so we need to transform the data of tensor type to vector type if the data to be processed is tensor type,and it will lose space location relationship of feature and generate high dimension vector,so over-fitting occurres in the learning process.In order to overcome these shortcomings we studied the classification method of kernel support tensor machine which can get better classification effect compared to support vector machine in experimental model of Imitating-reading Brain-Computer Interface.According to the characteristics of support tensor machine takes longer time,we studied a algorithm which has faster speed of classification.That is extreme learning machine.The classification accuracy of extreme learning machine is almost the same compared to support vector machine under experimental model of Imitating-reading Brain-Computer Interface,but has simpler network structure and less time consuming.At last,we studied independent component analysis algorithm,and proposed the EEG signal optimal electrode selection method based on independent component analysis and under experimental model of Imitating-reading Brain-Computer Interface,and this method can applied to online BCI.