Research On Human Motion Intention Recognition Based On Graph Convolution Neural Network

Brain-computer interface has enabled communication between the injured patient and the external environment.Spinal cord injury patients can be effectively rehabilitated and partially restored motor functions by a brain-computer interface rehabilitation system based on motor imagery.The main function of this system is to control external devices by pattern recognition of EEG signals from spinal cord injury patients.In order to study brain-computer interface rehabilitation systems in functional analysis of EEG signals based on motor imagery,it is necessary to improve the quality of EEG signal features so that the patient’s motor intent is more obvious and can be accurately and effectively identified by classification through model training.This thesis focuses on signal processing and classification recognition based on brain-computer interface systems,characterizing EEG signals through different feature extraction.In this thesis,graph convolutional neural networks are fully investigated to achieve classification and recognition of EEG signals in motor imagery tasks using spectral and spatial domain graph convolutional neural networks.1.The study proposes a feature representation of EEG signals based on timefrequency spatial domain feature fusion.It is a feature fusion of traditional timefrequency and spatial domain features,where the time-frequency features mainly use the wavelet transform and modified S-transform,and the spatial domain features mainly use modified common spatial pattern.It extracts the features of EEG signals by multiple methods,retaining the advantages of the features extracted by different methods.There is not only more concentrated information of time-frequency features,but also spatial electrode position relationship,which has more obvious distinction for classification recognition of left and right hand motor imagery tasks.2.The proposed algorithm of EEG signals classification and recognition based on graph convolutional neural network.The algorithm contains spectral domain graph convolution and spatial domain graph convolution,while improving the two different graph convolutions.The improvement of the spectral domain graph convolution is mainly reflected in the decomposition of the Laplace matrix,which reduces the parameters that the model can learn and makes the model structure simpler and does not affect the performance of classification recognition.The spatial domain graph convolution improvements mainly improve the sampling and weight functions.They let the channels of the data share the convolution kernel parameters and then use the normalized adjacency matrix as the transfer matrix to highlight the relationships between channels.This algorithm can reduce the complexity of the algorithm to a large extent and can avoid the overfitting phenomenon.Finally,the thesis introduces multiple models for comparison and demonstrates the effectiveness of graph convolutional neural networks by comparing diverse classification results.3.The study uses a novel approach to comprehensively analyze and evaluate the brain-computer interface system and EEG signal characteristics.The fact that the experimental paradigm is tailored to patients with spinal cord injury makes the assessment criteria also need to be analyzed and evaluated on a case-by-case basis.For the fusion characteristics of EEG signals,the functional analysis of brain regions is selected,which mainly contains: the analysis of the influence of brain regions on the fusion characteristics in the time-frequency spatial domain,the ERD/ERS analysis of EEG signals and the brain network analysis.The study appropriate statistical analysis methods were selected to perform differential analysis and correlation analysis on the experimental results to demonstrate the significance of the data.The model complexity is also analytically evaluated to highlight the efficiency and speed of the graph convolution model.The research content of this thesis can further advance the development of the brain-computer interface technology based on motor imagery tasks to determine the motor intention of patients with spinal cord injury by classifying and recognizing the EEG signals of the patients.It provides effective rehabilitation training for patients to partially restore motor functions.