A Study On Classification Methods For EEG Based On Deep Learning

Brain-Computer Interface(BCI)is a communication system independent of the human internal neural path,which can be used for the purpose of the direct interaction between the human brains with the outer world.By decoding the recorded P300 into control message of external device,the BCI based on P300 potentials has been realized in this way.The offline analysis of the P300 has attracted lots of scholars home and aboard owing to the potential significant in Psychology and Neuroscience,and a variety of analysis methods have been proposed.However,two prominent problems also arose in existing methods:1 The diffusion gradient would be caused by back propagation applied in Convolutional Neural Network(CNN)training and Deep Brief Net(DBN)has no capacity to capture the local P300 spatial information.2 Sharing temporal filters has not considered the completion representation of features owing to one filter only can obtain one feature.Considering the excellent result in image detection and artificial intelligent field achieved by Deep learning,this thesis has applied it into P300 analysis for combining the defeats of the existing methods.And based on that,our novel frameworks have been proposed to analyze the P300 potential and the classification performance has been boosted effectively.The major research contents in this thesis include the following parts:1 The background of BCI and EEG has been discussed in detail and the past analysis methods used for P300 has also been demonstrated in a systematical fashion.Besides,the two methods in deep learning-CNN and DBN have been discussed for their structure realization and the analysis of the new P300 data.2 Based on the diffusion gradient caused by the application of MSE in CNN training and without the capacity to capture the local P300 spatial information for DBN,this thesis has proposed to adopt layer-wise energy model-Convolutional Deep Brief Net(CDBN)to analyze the P300 for the first time.Being different from the original structure,we have considered the nature of data and attached the structure at new spatio-temporal signification.The adaption of the spatial and temporal extraction layer have been realized to capture the spatial and temporal domain information respectively.Finally,our results have distinctly outperformed the CNN one.3 Based on the linear spatial mixture model and the incompletion representation obtained by the sharing kernels,this thesis has proposed a novel spatio-temporal structure and the unshared temporal filters initialized by the Denoising Auto-Encoder(DAE),respectively.And the DAE abandons the MSE cost function and replaces it with the cross-entropy cost function to fit the distribution of the output,which avoids some problems like insensitive updating process and other problem caused by MSE.Besides,in order to obtain a robust and compact representation that have a powerful capacity to capture the nature features,some constrains have been imposed on hidden layer for limiting the activation.A better result,naturally,would be obtained comparing to the existing methods.In order to verify the performance for our proposed methods,this thesis has conduct a detail analysis for P300 of total 10 subjects in our experiments.And comparing to the existing methods,both of our proposed methods have a better result.Based on this thesis,our proposed methods also can be applied to other data with a spatio-temporal information.And the further application like online experiments also can be conducted.