Research On Key Technologies Of EEG Motion Imagination Recognition Based On Graph Neural Network

Motion imagery EEG is a popular research direction in the field of brain-computer interface,and its recognition technology has great potential in applications such as humancomputer interaction,rehabilitation therapy and virtual reality.Currently,scholars have proposed several methods for motor imagery EEG recognition,however,most of the existing methods isolate EEG signals in a fixed channel order as input,while ignoring the complex correlation among topological structural properties and EEG channels due to EEG channel distribution,making the decoding of EEG signals with non-Euclidean structures ineffective.On the other hand,subjects’ EEG signals do not produce significant specific motion information throughout,and only several segments may have motion-related information.Existing methods tend to calculate statistical values for EEG sequences as time-domain features or capture temporal context dependence through temporal relationships only,ignoring the global dependence of EEG signal sequences.Finally,there is a general lack of practical application of existing work.To address the above issues,this thesis proposes a motion imagery recognition method based on graph neural networks and applies it to a UAV brain-machine interface system.The details include:(1)A motor imagery recognition method based on Adaptive Spatio-Temporal Graph Neural Network for Motor Imagery Classification(ASTGNN)is proposed.Firstly,the raw EEG signal is band-pass filtered and normalized;secondly,the EEG signal is fed into the brain graph topology learning module,which applies the spatial location relationship to define the common hierarchical topology;and the single-layer neural network is trained to quantify the degree of association among EEG channels to obtain the personality hierarchical topology.The common and personality EEG channel association weights are fused and summed to obtain the optimal brain map topology;again,the original EEG signal and the optimal brain map topology are used as the input of the graph convolution layer,and the spatial features are extracted through the graph convolution network.Then the spatially encoded EEG features are sliced in the temporal dimension and input to the temporal encoding module.The influence weights among EEG segments at different locations are calculated using a gated position-aware selfattentive mechanism,which supplements each EEG signal segment with learnable relative position information and captures the global dependencies of EEG segments that significantly contain motion information.Then the decoding module applies the pooling layer and the fully connected layer to obtain recognition results;finally,ASTGNN is designed and implemented,and experiments are conducted on a publicly available dataset,and the experimental results show that the method in this thesis can effectively recognize four types of motion imagery tasks and outperforms existing methods.(2)A brain-computer interface control system for UAV with motion imagery control is designed and implemented based on the above method.Firstly,the Emotiv Epoc EEG helmet is used to collect EEG and send it to the upper computer by wireless communication;secondly,the model program in the upper computer preprocesses the EEG signal and recognizes the motion imagination,and the recognition result is transmitted to the lower computer by wireless communication;thirdly,the lower computer converts it into control commands to the UAV and controls the UAV to make corresponding actions;finally,the UAV brain-machine interface control system is designed and implemented.Finally,the UAV brain-machine interface control system was designed and implemented,and flight experiments were conducted in the simulation environment and the real environment to verify the effectiveness and reliability of the system.Finally,the full text of the work was summarized and the next step was looked forward to.