Research And Implementation Of Key Technologies Of Brain-computer Interface Based On Motor Imagination

Brain-computer interface is a novel way of human-computer interaction.It enables the brain to control external devices by parsing the motor intent in EEG signals and converting them into computer commands.However,EEG signals are unstable and susceptible to interference.How to obtain clean EEG signals,extract effective features from EEG signals and complete classification has become the key to EEG recognition research.Therefore,this thesis focuses on the removal of artifacts in motion imagery EEG and the feature recognition and system implementation of EEG signals with important practical application value.First,the current domestic and international research on BCI technology is described.The introduction of EEG signal characteristics and the basic framework of BCI system and its key technologies are completed,and the EEG signal processing methods and the current research difficulties of BCI technology are analyzed.Finally,the framework of the human-computer interaction system based on motion imagery is designed.Second,to address the problem that the traditional EEG artifact removal method needs to add the oculogram artifact signal as a reference and the oculogram artifact removal is incomplete.A convolutional sparse self-coding based method is proposed to remove oculoelectric artifacts from EEG signals.The method adds convolutional operation to the sparse self-coding structure.The convolution is utilized instead of the original inner product.The oculoelectric artifacts in the EEG signal are removed by convolutional pooling in the encoding stage.The pure EEG signal after denoising is restored by reversing the pooling convolution in the decoding stage.Through experimental comparison,it is demonstrated that the convolutional sparse self-coding method improved in this paper can effectively remove the oculoelectric artifacts from EEG signals.Further,address the problem of low recognition rate of four types(left hand,right hand,tongue,and foot)of motor imagery EEG due to insufficient information about EEG signal features.An EEG recognition method based on LFFCNN-GRU is proposed.The method takes interlayer feature fusion convolutional neural network and gated recurrent network as the basic premise.The method is based on the basic premise of interlayer feature fusion convolutional neural network and gated recurrent network.The local features extracted from two adjacent convolutional layers are fused by the feature fusion module,and then the obtained rich features are input to the next layer of the network to enhance the feature information of the boundaries.Afterwards,the information is integrated with the batch normalization algorithm after processing the depth sequence information using the gated recurrent network module,and finally the classification is performed by softmax.Through experimental validation,LFFCNN-GRU can extract EEG feature information more effectively than the traditional convolutional neural network method,and improve the recognition rate of four types of motor imagery EEG.Finally,an intelligent human-computer interaction system based on motion imagery was designed and built.The improved EEG artifact removal algorithm and pattern recognition method were embedded into the system.The algorithm was validated by analyzing the time consumed and the walking trajectory of the "8" trajectory of the intelligent wheelchair controlled by the subject.The experiments were compared with the system controlled by the conventional method.The results show that the trajectory of the system using the improved method is more stable and takes less time to complete one lap.The effectiveness of the improved algorithm and the system built in this thesis verified.