Research On Classification Algorithms For EEG Signals Based On Motor Imagery

With the continuous exploration of brain by humans,brain science has developed rapidly.As a branch of neuroscience,brain computer interfaces aims to establish a pathway for information exchange between the brain and external devices,enabling control of external devices.The brain computer interface based on motor imagery EEG signal is considered one of the most promising brain computer interface technologies due to its low cost,convenient use,safety,and no need to present stimulus signals.Due to the nonlinear,non-stationary,and low signal-to-noise ratio characteristics of EEG signal,the processing of EEG signal is particularly crucial.The accuracy of signal classification determines the performance of brain computer interface technology.At present,the accuracy of classification algorithms based on motor imagery EEG signal is relatively low.This article designs two classification algorithms based on the latest network structure of deep learning,and the specific work is as follows:(1)In response to the current shortage of motor imagery EEG signal datasets,this article designs an experimental paradigm for motor imagery signal collection based on the publicly available dataset BCIC IV 2a.Participants are gathered to collect highdensity motor imagery EEG signal containing 64 electrodes using EGI devices,and a set of preprocessing methods for motor imagery EEG signal is summarized.The preprocessing of motor imagery EEG signal mainly includes filtering and selecting frequency bands of interest,re-referencing,segmenting useful EEG signals,baseline correction,interpolating bad channels to reconstruct electrode signals with high noise,and deleting data segments with poor quality.(2)To solve the problem of poor classification performance of EEG signal based on traditional machine learning algorithms,this paper extracts time-domain features of EEG signal from BCIC IV 2a public dataset and self-collected dataset,including variance,detrended fluctuation analysis,fractal dimension,etc.,frequency domain features: power spectral density and spatial domain features: common spatial pattern,uses feature selection algorithm based on chi square test to reduce dimensions,and finally uses K nearest neighbor algorithm,support vector machine and random forest machine learning algorithm to classify the extracted features within and between subjects.The average of the four classification accuracy rates of all subjects was calculated as the evaluation item of the model.The results showed that all subjects basically achieved the best classification performance using support vector machines after feature selection.The four categories classification accuracy rates within and between subjects in BCIC IV 2a reached 0.642 and 0.569,which were 11.4% and 4.4%higher than those without feature selection.The four categories classification accuracy within and between subjects in the self-collected dataset reached 0.593 and 0.35,respectively.(3)In response to the problem of weak feature extraction ability and the need to improve classification accuracy of current neural network-based classification algorithms for motor imagery EEG signal,this paper uses filter banks to perform bandpass filtering on EEG signal with minimal preprocessing,while fully utilizing the time-domain,frequency-domain,and spatial features of EEG signals,in order to obtain a multi-view representation of EEG signals,Using convolutional neural networks as spatial filters for EEG signals to extract spatial features,using attention mechanism based long-short term memory networks and multi-layer Transformer encoder based networks to extract temporal information of EEG signals,two EEG signal classification models FBLSTM and FBTransformer are proposed.We conducted intra-subject and inter-subject experiments on the BCIC IV 2a dataset using our proposed model and comparison method Deep Conv Net,EEGNet and FBCNet,calculated the average classification accuracy of all subjects as the evaluation item of the models.The results showed that the intra-subject and inter-subject four classes accuracy of FBLSTM reached 0.742 and 0.536,which were improved by 1.9% and 3.2% compared to the mainstream method FBCNet,respectively;The intra-subject four categories classification accuracy of FBTransformer reached 0.729,which is superior to the comparison method.The accuracy of inter-subject four categories classification reached0.517.