Research On Classification Of Motor Imagery Eeg Signals Based On Ensemble Learning

Brain-Computer Interface(BCI),as a newly emerged cross-discipline,has been a hot topic in many fields such as neuroscience,artificial intelligence,pattern recognition and so on.Brain-Computer Interface has received wide attention for its external information communication and control technology which enables communication between the brain and external devices without relying on muscle tissue and peripheral neural circuits.In brain activity,Motor Imagery only requires participants to imagine simple actions to generate corresponding signals in the cerebral cortex as input to BCI,so it has received special attention.BCI system mainly include signal acquisition,preprocessing,feature extraction,classification and recognition,and output equipment and so on.The prerequisite for any BCI device is identify the classification features well.Therefore,this thesis mainly focuses on the study of EEG preprocessing,feature extraction,classification and recognition in Brain-Computer Interface.Due to the characteristics of the low signal-to-noise ratio,high dimension and non-stationary of EEG signals,single time-domain or frequency-domain analysis has several limitations.Although wavelet transform can combine time-domain and frequency-domain analysis,some artifacts and EEG signals still have similar time-frequency characteristics.Therefore,a single time-frequency analysis cannot completely distinguish artifacts and EEG signals,which would reduce the classification recognition rate.In view of this,this thesis proposes a method based on the random subspace ensemble learning of multi-domain features,aiming at EEG classification of motor imagery This thesis analyzes the ERD/ERS characteristics of MI signals and extracts the multi-domain features of the best effective time and frequency bands as the feature vectors.And then,this method adaptively chooses the scale of the random subspace ensemble with cross-validation.Finally,it realizes EEG classification using linear discriminant analysis(LDA)classifiers ensemble.The accuracy of the multi-domain features extracted in this thesis is about 3% higher than that for single feature,and the accuracy of random subspace ensemble is at least 2.14% higher than that for traditional ensemble learning algorithm.In addition,the classification accuracy of the suggested method can reach 90.71% on BCI competition 2003 dataset III,and the Kappa coefficient can reach 0.63 on BCI competition 2008 dataset 2b,which both are better than the first prize in the BCI competition.In the BCI system that incorporates features of time-domain,frequency-domain,spatial domain,and time-frequency,which can improve the reliability of the system and obtain a higher classification accuracy rate.The study found that ensemble learning has obvious advantages over traditional methods in motor imaging tasks,and can well implement classification.This thesis proposes an XGBoost-based Stacking ensemble model for the study of EEG classification of motor imagery.The first layer is composed of XGBOOST,Random Forest(RF)and Logistic Regression(LR)models,and the second layer is finally output by XGBOOST model.The experimental results show that the accuracy of the Stacking model proposed in this thesis in the dataset BCI 2003(Dataset III)can reach 89.29%,the precision,recall rate and F1 value can reach 89.86%,88.57%,and 89.21% respectively.Compared with other classification algorithms,such as KNN,SVM,RF,GBDT,XGBoost,the classification model used in this thesis can improve the accuracy rate of at least 3.58% compared with a single model,which can provide a certain reference value for subsequent studies of EEG classification of motor imagery.