Research On Brain-computer Interface Technology Based On Spatial Hearing P300

Without the central nervous system or peripheral muscles,the Brain Computer Interface(BCI)can use brain activity as a control signal to achieve direct communication with the computer.At present,the visually evoked BCI paradigm has made great progress in domestic and foreign research.In contrast,the development of BCI of auditory evoked has lagged behind.The BCI of auditory evoked is mainly for patients with amyotrophic lateral sclerosis(ALS)and those with visual pathway obstruction.There are many types of electroencephalography(EEG)signals based on auditory stimuli.P300,as a kind of event-related potential(ERP),has the characteristics of constant latency and constant waveform,which is characterized by phase locking between potentials.The topic of research in this paper is based on the research of spatial auditory P300 BCI technology.The relevant research is shown as follows:(1)The design of the spatial auditory P300 BCI system.Firstly,the experimental paradigm of P300 signal evoked by spatial auditory attention was designed: using the three-tone oddball paradigm(the stimulus sequence contains two target stimuli,one non-target stimulus),each subject needs to pay attention to a target stimuli sound,and ignore non-target stimuli.Secondly,EEG signals of 8 electrodes are collected.Then,based on Microsoft Foundation Classes(MFC),a signal acquisition system platform is built,and a software system with user graphical interface,parameter setting module,sound stimulation module,timing module,data acquisition display module and data saving module are designed.(2)EEG signal classification based on the features of P300 in time domain.The wavelet signal interpolation,the empirical mode decomposition(EMD)based on the noise statistics,the autocorrelation denoising and the low-order averaging are used to improve the signal-to-noise ratio of the P300 signal,and then through independent component analysis(ICA)related algorithm selects the electrode combination to select the obvious electrodes with P300 component.Finally,the time-domain features of P300 at selected electrodes were extracted and classified by stepwise linear discriminant analysis(SWLDA)and Fisher linear discriminant analysis.The classification accuracy rate can reach 81.03%,and the average recognition rate of 9 participants reaches 77.43%.(3)Classification of EEG signals based on deep learning.The powerful capability of nonlinear feature representation of convolutional neural networks makes it one of the most widely used.This paper describes a model based on CNN and an EEGNet that uses depthwise convolution and the separable convolution to identify EEG-specific models.Experimental result shows that EEGNet can better promote the paradigm and achieve relatively high performance in the reference algorithm when limited training data is available in all test cases.The classification accuracy of CNN is up to 71.33%,and the average accuracy is 65.67%.The classification accuracy of EEGNet is up to 72.67%,and the average accuracy is 68.78%,which is 3.11% higher than that of CNN.In terms of runtime,EEGNet saves an average of 25.44 seconds over CNN.