Study On Deep-learning Based Classification For P300 Brain-computer Interface

Brain-computer interface(BCI)is a communication and control system that allows interaction between human brain and external devices.The P300 speller is a BCI system based on the oddball stimulus paradigm.It is composed of two classification steps.The first classification is to detect the presence of a P300 in the electroencephalogram(EEG).The second one corresponds to the combination of different P300 responses for determining the target characters.In practical applications,it is important to improve the system's spelling accuracy and information transfer rate(ITR).However,traditional algorithms of P300 detection generally require complex preprocessing and feature extraction before classification,and the character recognition rate(CRR)and ITR obtained by the models are not high enough.In response to these problems,we try to improve the performance of the system from the perspective of deep learning.The research contents of this thesis are as follows:The standard convolutional neural network(CNN)is prone to overfitting in P300 detection,so we use an improved model,termed batch normalized convolutional neural network(BN-CNN),where batch normalization(BN)is introduced in the convolutional layers and Dropout is applied in dense layers to alleviate the problem of gradient dispersion and over-fitting.An improved algorithm is proposed to reduce the interference of non-target characters.Since the interference mainly comes from the adjacent characters of the target,we have designed a three-class model to reduce interference.The three types of samples correspond to the target character,the neighbor characters,and the rest.Use the bi-directional recurrent neural network(RNN)based on the gated recurrent unit(GRU)to keep track of long-term dependencies in the input sequences.RNN is combined with CNN to achieve both spatial filtering and time-domain filtering.So we use the ConvLSTM to solve the long-term dependencies that CNN cannot handle,and also solve the problem that RNN cannot achieve sufficient spatial filtering.The experimental results prove that the above methods of deep learning can achieve better performance in P300 speller,that is,higher CRR and ITR than traditional algorithm and the standard CNNs.