Classification Of Multi-Channel EEG Signal Based On Brain Networks And Deep Learning

Modeling and analysis of complex systems based on multivariate time series can better describe the dynamic behavior of the system.However,multivariate time series generated by observing complex systems are characterized by high dimensionality and complexity,and it is difficult to analyze the change pattern of the data by traditional methods.As a typical complex system,the human brain generates multi-channel EEG signals with low signal-to-noise ratio,low spatial resolution,and complex interchannel relationships.To address these major challenges in analyzing multichannel EEG signals,this project combines deep learning and brain networks to explore the role of multi-domain information fusion in multichannel EEG signal analysis.The subject designs deep learning models from three perspectives of time domain,frequency domain and channels,and takes emotional EEG signals and motor imagery EEG signals as the research objects,and mainly accomplishes the following work:1.Two EEG experiments were designed,in which the emotional EEG signals of8 subjects and the motor imagery EEG signals of 10 subjects were acquired.For the internal and external interference during the acquisition process,band-pass filtering and Independent Component Analysis were used to remove the noise from the signals.The two datasets obtained from experiments are used as the performance test of the model together with two public datasets.2.A multi-input Convolutional Neural Network model is designed for accurate emotion classification.The proposed method is capable of mining EEG information in the time domain,frequency domain and inter-channel.First,a time-frequency decomposition is performed by Continuous Wavelet Transform to segment the original signal under critical frequency bands of the human brain.Further,brain network that can portray the correlation between individual channels is constructed from EEG signal using a complex network approach and saved in the form of an adjacency matrix.Meanwhile,nodes in brain network that contribute more to the emotion classification are selected based on differential entropy feature,and then core brain network is constructed.The differential entropy and the core brain network under different frequency bands were simultaneously used as the input of the multi-input Convolutional Neural Network.The average classification accuracy of the proposed model reached 91.45%for emotion classification on the SEED dataset and 90.26% on the autonomously collected emotion dataset.3.A Residual Network based on Complementary Ensemble Empirical Mode Decomposition and attention mechanism is designed and applied to motor imagery EEG classification.The proposed model is designed to address the channel selection problem,and the attention mechanism enables the model to achieve dynamic weighting of each channel of the EEG multivariate time series during training.First,the original signal is decomposed into different frequency components by Complementary Ensemble Empirical Mode Decomposition; then the attention mechanism is used to weight the different components as well as the channels of the sequence; and Residual Network is used for further feature analysis.Classification accuracy results of 76.45% and 84.97% were obtained on the publicly available dataset BCI IV 2a and the autonomously collected dataset,respectively.