Study On Improved Convolutional Neural Network Based Classification For P300 Brain-Computer Interface

A Brain Computer Interface(BCI)character speller based on P300 allows human beings to directly spell characters through brain responses induced by visual stimuli,thereby building direct communication between the human brain and a computer.Detecting and analyzing P300 signals from electroencephalography(EEG)is significant to establish the BCI spelling system.Convolutional Neural Networks(CNNs)is an approach that have shown better P300 detection performance than traditional machine learning methods.The P300 dataset has a limited number of available samples,in order to achieve competitive accuracy on under limited model complexity,this paper improves the spatial and temporal processing parts of the classic CNN according to the characteristics of the P300 signal itself,and proposes two improved CNN models to classify the P300 signal.The main work of this paper is as follows:(1)The purpose of channel selection is to select a few electrode channels that are most helpful for classification,thereby reducing the requirements for model complexity.In order to realize the automatic selection of electrode channels,this paper proposes a convolutional neural network based on the attention mechanism is proposed for detecting P300 signals,where SE Block is introduced as the attention module before the spatial filter of the classical CNN to obtain a weight vector of the electrode channel of the input EEG signal,and the automatic selection of the electrode channel is realized through the weight vector.At the same time,the L1 norm of the weight vector is added to the loss function of the network,so that the learned weight vector is sparser and the channel selection ability is stronger.(2)Convolution with a fixed-size kernel can only learn single pattern,it is not flexible to adapt to the P300 waveform changes caused by the psycho-physiological condition of the subjects.To address the problem of fixed size of convolution kernel,a multi-scale dilated convolutional neural network is proposed to detect the P300 signal,where dilated convolutions with different dilation rates are introduced to extract and aggregation multi-scale time domain features,thereby enhancing the feature extraction capabilities without increasing the complexity of the model.In addition,according to the multi-task learning method,an auxiliary task is employed in the reconstruction of the input signal from the temporal features extracted,which plays a regularizing role to alleviate overfitting.This paper performs experiments on datasets of BCI Competition ? and results show that the above two models outperforms classical CNN with the higher character recognition rate and information transfer rate.