Feature Extraction And Classification Of P300 Signal

BCI is a technology of brain computer fusion perception,it does not depend on the information output system of the human body itself,but through the direct connection between the brain and the external machine equipment to realize the communication with the surrounding environment.The P300 signal is an endogenous event-related potential,which has a strict time-locked property,so it is widely used as the control signal of the BCI system.According to the characteristics of P300 signal,such as nonlinear,non-stationary and large individual differences,this paper studied P300 signal processing algorithm from the aspects of preprocessing,feature extraction and classification recognition of EEG signals,combined with the relevant methods in machine learning and deep learning.The main research work of this paper is as follows:(1)A denoising method of P300 signal was proposed,which combined empirical mode decomposition(EMD)algorithm with spectrum analysis algorithm.Firstly,the original EEG signal was decomposed by EMD,and several intrinsic mode components(IMF)were obtained;Then,according to the effective frequency range of P300 signal,we used spectrum analysis to retain useful IMF components and reconstructed them,that is,to get the denoised EEG signal.The experimental results showed that this method had better denoising effect.(2)Wavelet packet decomposition(WPD)algorithm was used to extract the characteristics of P300 EEG signal: the preprocessed P300 signal was decomposed into four layers of wavelet packet,and then the appropriate wavelet packet subband was selected for reconstruction.Experimental results showed that compared with wavelet transform,wavelet packet decomposition algorithm had better performance in feature extraction.At the same time,we used wavelet packet decomposition algorithm and improved convolution neural network model to extract and classify the secondary features of P300 signal,which could get higher classification accuracy.(3)A new classification method of P300 EEG signal based on improved convolutional neural network(CNN)was proposed,that is,T-CNN method: The EEG signal had both time and space features,if we directly carried out the convolution operation would make the space and time information in the extracted features mixed together.Therefore,the convolution kernel in convolution layer was set as vector instead of matrix,so that it could only extract space or time features in convolution operation.At the same time,in order to prevent over fitting in the training process,dropout and batch normalization operations were added to the improved CNN model.In the experiment,machine learning method and T-CNN method were used to classify P300 signals respectively.The results showed that T-CNN method could get higher classification accuracy,which proved the effectiveness of deep learning algorithm in P300 EEG signal recognition.