Research On P300 Brain-computer Interface Classification Algorithm Based On Deep Learning

Brain-computer interface(BCI)is a system that allows information exchange and processing between the brain and the external world,and P300 is a brain-computer interface based on Oddball.The method is divided into two stages,the first stage is to identify the presence or absence of P300 potentials in the EEG,the second stage is to identify the P300 potentials in the previous stage,and the second stage is to identify the target character.Therefore,how to effectively improve the recognition accuracy of P300 images and how to effectively transfer the data is of great importance for the recognition of P300 EEG signal characters.However,the existing P300 detection methods are tedious in the process of pre-processing the target and extracting the features before classifying the target,and the recognition rate and information of the target are relatively low.To address the above problems,this paper takes P300 EEG signal as the research object and takes deep learning as the entry point to optimize it.The following aspects of work are carried out in this thesis.(1)In terms of signal processing,as the original data necessarily contains a large amount of redundant information,which is not convenient for analysis and research;the signal is the original signal collected by the sensor,and the influence of electronic components’ own noise,environment and the subject’s action state will bring certain interference to the collected data.According to the energy concentration of human P300 EEG signal in the frequency band from 0.1Hz to 20 Hz,a 46 th order bandpass Butterworth ⅡR filter is applied to effectively filter the P300 EEG signal to remove the high frequency interference signal,while maintaining the useful information and improving the sensitivity of the EEG signal.(2)The study of feature recognition of the P300 EEG signal is a binary classification problem,in which the excitation functions sigmoid and tanh are both Stype saturated nonlinear functions.In neural networks,data bias occurs when the data traverse multiple layers.Given that many models are trained with gradient descent,however,gradient saturation rarely occurs in the saturation range of the sigmoid and tanh activation functions,and back propagation can be used to eliminate the gradient.To solve this problem,this paper uses batch normalization(BN)to preprocess the experimental data,and this algorithm can effectively solve the gradient disappearance problem.(3)In P300 waveform detection,overfitting problem is likely to occur with a standard neural network(convolutional neural network,CNN),so in this work,we improve the model based on the standard convolutional neural network and use the improved LeNet-5 model for character recognition.Among them,by using small-scale convolutional kernels in each convolutional layer of the network,more features can be extracted to accelerate the training of the model,and in addition,adding the regular term constraint of Triplet Network to the loss function can significantly improve the accuracy of small-sample character recognition.(4)Support vector machines have many unique advantages in solving small-sample,nonlinear and high-dimensional imaging problems,and largely eliminate the problems of "dimensional collapse" and "overfitting".It has important applications in text recognition,handwriting recognition,face recognition,gene classification,and time series prediction.Therefore,in this paper,support vector machine is used to classify and identify P300 EEG signals.First,the training samples are filtered,cleaned,and batch normalized,and these samples are labeled to construct a support vector machine for training and testing samples,and then a classifier for the support vector machine is constructed,and these classifiers are introduced into the support vector machine to finally obtain a support vector machine model.Finally,the P300 EEG signal data to be recognized is similarly preprocessed and its features are extracted to produce a test ensemble,which is then fed into the support vector machine classifier to obtain the character recognition classification of the P300 EEG signal.Experiments show that both of these deep learning methods can obtain good results in the P300 spelling system,which is higher than the conventional and classical convolutional neural networks in terms of Chinese character recognition accuracy and information transfer rate.