Research On EEG Signal Classification Method Based On Riemann Space And Convolutional Neural Network

Brain-computer interface is a system that can be disconnected from the peripheral nerves and muscles of the human body,allowing signals generated by brain activity to be captured by an acquisition device and transformed into signals that can be recognised by a computer through a series of operations such as pre-processing and signal processing,from which the real intentions of the person can be identified.Brain-computer interface technology allows external devices to read the brain’s neural signals and convert thinking activities into command signals to enable the brain to control external devices;it also has the potential to input computer information directly into the brain,thus greatly shortening the process of learning new knowledge and skills for humans.As one of the paradigms of brain-computer interface,motor imagery EEG signals have the advantage of being spontaneously generated without external stimulation,and have shown a broader application prospect in artificial intelligence,clinical medicine and even military fields.To address the shortcomings in current methods for the analysis and processing of motor imagery EEG signals and the great potential presented by deep learning and Riemannian geometric methods in EEG signal analysis,this thesis is based on signal analysis theory and aims to investigate the methods for motor imagery EEG signal analysis,combining EEG signal analysis,Riemannian geometric features and convolutional neural networks,from Riemannian spatial representation of EEG signals,Riemannian The study aims to combine EEG signal analysis,Riemannian geometric features and convolutional neural networks,and to construct a Riemannian spatial convolutional neural network model for motor imagination EEG signals.The specific research includes the following parts:(1)The current state of research on EEG signals at home and abroad is summarized,focusing on the advantages and challenges in the development of two approaches based on Riemannian space and convolutional neural networks for the analysis and processing of motor imagery EEG signals.(2)In order to obtain more discriminative and informative EEG data,this thesis firstly preprocesses the raw motor imagery EEG data to remove the frequency components of EEG signals that are irrelevant to the target motor imagery task as well as the noise of non-EEG signals such as EMG and ECG.Secondly,the EEG signals described in the spatio-temporal domain are converted to Riemannian space by choosing the covariance matrix as the Riemannian space descriptor of the EEG signals.Finally,a Riemannian space-based data enhancement method is proposed for the EEG signals described in Riemannian space.(3)To further improve the recognition accuracy and robustness of motor imagery EEG signals,the traditional convolutional neural network is improved on the basis of work(2),and a classification model of EEG signals based on Riemannian space convolutional neural network is designed for processing EEG data in Riemannian space.In addition,in order for the model to capture more comprehensive features of the input data,a different scale of convolutional kernel is added to the original convolutional layer,increasing to a convolutional layer with two different sizes of convolutional kernels.(4)In order to verify the validity of this research,a series of experiments were designed on the datasets collected from the International Brain-Machine Interface Competition and the Laboratory of New Human-Machine Collaborative Intelligence Technology and Robot System of Shandong University of Architecture.The results show that the proposed EEG signal analysis method not only works well in terms of Riemannian space description of EEG signals,data enhancement,end-to-end feature extraction and the final classification results,but also has certain robustness and stability.The experimental results verify the feasibility and effectiveness of the algorithm in this thesis.