The Detection Of Idle State Of Motor-Imagery BCI By Multimodality Features

Act as a translator,Brain Computer Interface(BCI)can decode human brain's intention directly and send the parsed commands to external devices to implement the control of external devices.Motor Imagery(MI)BCI can be divided into synchronous and asynchronous system according to the used control strategies.Among them,the asynchronous BCI system gives subjects enough control right,which can realize the free will control and improve comfort of the system.The asynchronous system based on MI has potential practical value in the motion assistance of the handicapped,the rehabilitation of motor function for stroke and cerebral palsy because of its spontaneous,harmlessness,convenience,and comfort.The dissertation mainly focuses on MI-BCI asynchronous system,and carried out the corresponding studies from below two aspects.First,the main difficulty in implementing the asynchronous MI-BCI system is how to distinguish the task state and the idle state of subject reliably.As a high level cognitive process,MI needs to involve multiple brain regions to perform the information exchange.In this paper,we combined EEG functional brain network information with traditional CSP method to realize the detection of idle state of asynchronous MI-BCI using F-score method to screen the features of the two states.This paper compares classification accuracy and information transfer rate(ITR)performance among three methods(i.e.,CSP,network properties and the proposed approach)at four time intervals windows of 0.5s,1s,2.5s and 2.5s.The results show that the proposed method has significantly improved performances compared to the other two methods in all the considered conditions.Besides,the further analysis actually demonstrates that the performance improvement of the proposed approach is due to the additional utilization of the network properties that can provide the complementary information to CSP features for the discrimination of idle state and task.Moreover,the method proposed in this paper also demonstrates a greater performance improvement for the subjects with poor BCI performance,which is meaningful for the practical MI BCI.All results consistently demonstrate the effectiveness and practicability of the proposed method for the idle state detection.In addition,this dissertation further probed the possibility of using deep learning to detect idle state across subjects.Considering the high-dimensional time domain features and the complicated structure of raw EEG data,we extract the CSP features,power spectrum features of multiple subjects as the input of deep belief network(DBN)and proposes a deep structure containing two hidden layers to detect the subject's idle state.The results show that even without additional training and just use the existing information of other subjects,the accuracy rate of identification of idle state can reach 0.74.The results prove the effectiveness of idle state identification across subjects using learned complex network structure of DBN,which may provide an alternative solution for online MI-BCI.