| Brain-computer interface technology can enable some paralyzed patients with neurological diseases to control external devices directly through brain activity without the control of any nerves and muscles,and this technology provides a new communication channel for patients to communicate with the outside world.However,EEG signals have a low signal-to-noise ratio,and they are susceptible to interference from acquisition devices or other physiological signals of subjects which brings some challenges to the feature extraction and classification of EEG signals.In addition,for different types of perceptual tasks,the stimulated areas in the cerebral cortex are not the same.For example,when performing visual stimulation,the visual cortex will produce EEG signals induced by specific frequencies,resulting in EEG signals collected by some electrodes unrelated to stimulation,and the use of too many electrodes will cause information redundancy.Based on the above problems,the project is based on the research of the P300 character recognition system and is studied from two angles: brain-computer interface recognition algorithm and channel selection,aiming to improve the target character recognition rate and information transmission rate.The main work of this article is as follows:(1)Aiming at the characteristics of EEG signals and the problem of low character recognition rate,this thesis constructs an MTS-CNN network for feature extraction and character recognition of EEG signals.MTS-CNN has two parallel convolutional layers,and the time sample point of the EEG signal is an integer multiple of the length of the convolutional kernel,and the convolutional kernel moves in steps for the respective lengths of time of the two convolutional nuclei.Spatio-temporal features extracted by the two convolutions are superimposed on the time dimension.The fully connected layer randomly inactivates neurons with a dropout probability of 0.2,which alleviates the occurrence of overfitting and can extract diversified temporal information in the same spatial dimension.Experimental results show that MTS-CNN has a higher character recognition rate and information transmission efficiency compared with other classical networks.(2)A correlation method for channel selection of EEG signals is proposed,which is a biased hard processing method.The degree of correlation between channels is expl ored through the correlation analysis between channels.The corresponding correlation threshold is set according to the characteristics of different subjects,and the channels with strong correlation are removed to retain the channels that are valuable for classification in the EEG signal.This method reduces data redundancy and improves character recognition and information transmission rate.In 15 replicated experiments,MTS-CNN-CSC achieved a character recognition ratio of 98.0%,which was significantly better than 95.5%of MTS-CNN.In addition,the information transmission rate of MTS-CNN-CSC is higher than that of MTS-CNN under most repeated experiments.(3)A channel attention-weighting algorithm is proposed to process the EEG signal.The spatial dimension of the EEG signal is channel attention weighted.The different channels extract different features,and the channel attention mechanism can measure the importance of these channels,which is a biased soft processing method.In this thesis,the channel attention module is embedded in the MTS-CNN model and then continuously trained.Finally,the channel attention module will produce a weight matrix,which weights the characteristics of the weight matrix and the EEG signal by channel dimension,and adaptively correct the valuable channel characteristics.Experimental results show that the total average character recognition rate of MTS-CNN-CAM in one to fifteen repeated experiments is 1.4% higher than that of MTS-CNN.In the previous repeated experiments,the character recognition rate and information transmission rate obtained by MTS-CNNCAM have been significantly improved compared with MTS-CNN. |
PDF Full Text Request