Auditory Attention Research Based On EEG And Convolutional Neural Network

With the rapid development of deep learning,a large number of researchers try to apply deep learning methods to various EEG decoding tasks,and deeply mine the features of EEG data,which greatly improves the recognition accuracy.Attention,as an indispensable cognitive ability in human life,is closely related to EEG signals.In the face of different degrees of sound stimulation,human brain waves fluctuate.Meanwhile,the fluctuation changes can reflect people’s attention level when stimulated by external sound.At present,the traditional methods still have a lot of room for improvement in dealing with EEG auditory decoding problems,while deep learning methods such as convolution neural network are more and more widely used in EEG signal-related fields,showing great advantages.Therefore,in order to fully extract the EEG information and reduce the error value of auditory attention recognition task as much as possible,this thesis takes auditory attention as the research goal,explores a variety of attention mechanisms,combines attention mechanisms with convolution neural networks,and designs two different EEG decoding models,namely 3D-AttCNN and C-DeepConvNet.The main research contents are as follows:(1)In order to explore the role of deep learning in analyzing EEG auditory recognition tasks and utilize the ability of attention mechanism to deal with EEG channel differences and frequency band differences,a three-dimensional convolution neural network 3D-AttCNN based on power spectral density and attention mechanism is designed.Firstly,this thesis compares several methods for data filtering,preprocesses the EEG signals,removes common artifact signals,and avoids the interference of irrelevant signals to the experiment.Then,the EEG signal is decomposed into several sub-frequency bands,the corresponding power spectral density features are calculated,and the features are mapped into the 3D input matrix according to the distribution position of electrodes.Finally,the mapping features are input into the 3DCNN,and the appropriate spatial information and frequency band information is adaptively selected using the channel spatial attention mechanism and frequency band attention mechanism,and the input feature vectors are learned to predict auditory attention level.The auditory attention recognition prediction task is carried out on 25 subjects in the PhyAAt dataset.The experimental results show that the 3D-AttCNN model utilizes two attention mechanisms to reduce the error of EEG auditory attention recognition tasks and achieve EEG auditory attention recognition across individuals.(2)In order to solve the problem that the feature extraction method in support vector machine and 3D-AttCNN model relies too much on theoretical accumulation,this thesis also designs an end-to-end prediction model,which improves the deep convolution neural network DeepConvNet.The C-DeepConvNet model stacks multi-layer convolution neural networks,and uses the channel attention mechanism to learn channel diversity.The experimental results obtained on the PhyAAt dataset show that,compared with the traditional support vector machine(SVM)method,the average absolute error of this model is reduced by about 8%,and the experimental performance is better than that of the 3DAttCNN model.The C-DeepConvNet model can directly extract effective features from the original EEG signals,and predict the level of auditory attention more accurately.In general,this model has better generalization and effectively improves the decoding efficiency.