Research On Classification Method Based On Feature Learning Of Eeg And Fnirs Brain Topographic Map

Brain-computer interface is a technology that converts signals from the human brain into instructions that a computer can recognize.Compared with the traditional single brain computer interface,hybrid brain computer interface has improved the control accuracy and reliability,and has attracted wide attention.The brain topographic map can be used to monitor the brain activity state,which contains the feature information and corresponding spatial information used to generate pictures.This paper takes the hybrid brain-computer interface of EEG and functional near-infrared spectral signals as the research object,and proposes a classification algorithm combining EEG and f NIRS brain topographic map for the classification of single subjects.It was verified in two experimental paradigms of motor imagery and mental arithmetic from multiple angles.Aiming at the problem of large individual differences among different subjects and the loss of temporal characteristics of brain topography,a cross-subject brain topography classification method based on style transfer and domain confrontation was proposed.The main work of this paper is summarized as follows:(1)The EEG-f NIRS public dataset used in this paper was preprocessed and characterized,and an automatic EOG artifact removal method was proposed based on the data structure.The method could make full use of the collected EOG data,avoid the chance of manual removal by experience,and reduce the time of data preprocessing.(2)The features of EEG and f NIRS were extracted to generate brain topographic map data of different structures,and the advanced features of the two types of brain topographic map signals were extracted and fused in the feature layer in the design of a single subject experiment.The fusion model was superior to the single modal model in the four classification tasks by combining the two paradigm data of motion imagination and mental calculation.The classification accuracy is higher than other methods in the same data set,and the average classification accuracy of the four classifications is 78.18%,which verifies the feasibility of the classification of brain topographic map and hybrid brain-computer interface in this paper.(3)A deep transfer learning method based on style transfer and domain rivalry is proposed,which uses image style transfer to achieve multimodal fusion of brain topography,and designs a multi-layer brain topography structure to solve the problem of temporal feature loss of brain topography to a certain extent.Combined with the channel attention module and domain adversarity deep neural network,the gap in the distribution of different subjects’ data features was reduced,and the brain topographic map channels that were conducive to transfer were highlighted.Aiming at the negative transfer problem,the information entropy of the output result of the domain discriminator was used as the measure of sample migration,and the alignment effect of sample feature distribution was optimized.The ablation experiment was conducted on the mental arithmetic paradigm of the public data set,and the average classification accuracy reached 79.62% after the final migration,which was 9.84% higher than that without migration,and the standard deviation was 3.48% lower.