Towards EEG Decoding For Motor Imagery Through Graph Convolution Neural Network

Motor imagery is one of the most widely used paradigms in the application of braincomputer interface technology.The brain-computer interface based on motor imagery maps the identified patterns to the control instructions by detecting the EEG signals of the subjects imagining specific motor tasks,and then realizes the communication between the human brain and the external device.Therefore,the number of identifiable classifications and accuracy of EEG signals under the motor imagery paradigm is directly related to the application of brain-computer interface systems.This paper takes the motor imagery EEG recognition task as the background,and explores the feature extraction and classification methods suitable for this task.The work mainly includes the following three aspects:Modeling of spatio-temporal graph for motor imagery EEG recognition task.This work introduces a functional connection analysis method based on EEG signals,calculates connection coefficients for different brain regions,and constructs a spatiotemporal graph in combination with time series.A spatio-temporal graph can be regarded as a spatially-graph time series,that is,at each time sampling point,the observation value is a spatially dependent graph.Therefore,this paper converts the recognition problem of motor imagery EEG signal into a classification problem of spatio-temporal graph.The graph convolutional neural network models for motor imagery based on EEG recognition task.With the development of deep learning,many scholars have introduced deep neural network models,such as convolutional neural network and recurrent neural network,into motor imagery EEG recognition tasks.The structure of deep neural network makes end-to-end classification possible,but since EEG signals are not regularly arranged in space,this paper abstracts them into spatio-temporal graphs.In terms of feature extraction and classification of spatio-temporal graphs,this paper constructs a neural network model based on graph convolution to realize end-to-end classification algorithms.In order to improve the feature expression ability of the model,this paper further introduces the residual connection and attention mechanism to construct a deep attention graph convolutional neural network model.In the course of the experiment,this paper takes into account the offline feature extraction capability and the feasibility of deployment in a real-time brain-computer interface system,and then proposes two dataset-partitioning protocols.The models are tested separately under two protocols,verifying the performance of the proposed model on multiple scales such as offline and online,single subject and multiple subjects,binary classification and multiclassification,which will bring to the motor imaging EEG recognition task a significant performance improvement.A training framework for noisy motor imagery EEG datasets.In order to avoid the over-fitting of deep network on noise samples due to the high noise of EEG signals,this paper first presents a noise reduction learning framework based on Co-Teaching,and then learns a pre-training model with general motor imagery knowledge from multiple subjects.The pre-training model can be used as a priori knowledge for new subjects.It can effectively reduce the cost of algorithm deployment in the case of long collection cycle of EEG signals and limited computing power of terminal devices,and has practical application value.