Research On Brain-computer Interface Of Motor Imagery Based On Deep Learning

Motion imagery the brain machine interface is one of the three brain electrical experimental paradigms whose accuracy rate of classification directly relates to whether brain machine interface systems can be applied.Poor classification performance at the motion imagination brain machine interface resulted from large individual variability,high randomness,and the need for fixed paradigms for cueing experiments.With the development of deep learning,researchers began to use deep learning models on motor imagery EEG classification.The advent of convolutional networks has changed the dilemma facing EEG signal research,and the structural properties of convolutional networks make it easier to implement end-to-end classification methods.But convolutional networks need to have a certain size dataset to train on them to achieve a desired effect,and there are relatively few published standard datasets of motor imagery EEG signal.Based on the problem of convolutional networks on classification of EEG signals and its convenience on classification problems,the following methods are proposed for analysis.First,the four types of motor imaginary EEG signals of 11 healthy subjects were collected,and the sampling frequency was 250 Hz.After preprocessing and screening,the data of one subject with poor EEG signal quality was removed,and 10 subjects were obtained.The EEG data of the author is composed of the self-made EEG data set sc of motor imagination,which solves the problem of small sample size.Then the public data set 2a and self-made data set sc of the 2008 brain-computer interface competition are preprocessed by band-pass filtering and artifact removal.To validate the classification performance of multi feature fused convolutional network models,multi kernel filter sets,and long short memory combined with convolutional network models.Proposed an algorithm for classifying motor imagery EEG signals by convolutional networks with multi feature fusion,which was applied to traditional classification algorithms.Convolutional networks are applied to traditional classification algorithms.Use short-time Fourier transform and continuous wavelet transform to extract time-domain features of EEG data and generate time-domain maps;use wavelet packet decomposition to extract the energy feature matrix of EEG signals;apply OVR-FBCSP to perform spatial features of EEG signals Extract and generate feature matrix.The time-frequency graph,energy feature matrix,and spatial feature matrix are used for feature fusion learning and classification through the convolutional network.The experimental results show that the classification accuracy of the convolutional network model of multi-feature fusion is 95.7%,which is higher than the existing model classification accuracy.A convolutional network model combining multi-kernel filter banks with long short term memory is proposed(MKFB-LSTMNet).This model can describe the temporal characteristics of EEG signals through time convolution combined with long and short-term memory network modules,and then extract spatial features through spatial convolution to achieve the effect of spatiotemporal convolution.This model was verified using the public data set 2a.According to the results,the classification accuracy of this network model is 77.9%,which is higher than that of the existing models.The self-made data set sc was verified by MKFB-LSTMNet and the results showed the classification result of this data set is 86.8%,demonstrating the validity of this network model.