Research On Classification Algorithm Based On Motor Imagery EEG

In recent years,stroke has become the leading cause of death and disability in China.Once a person has suffered a stroke,they will have different degrees of disability,which brings a huge burden to society and family.Brain-computer interface(BCI),as an emerging technology integrating multidisciplinary knowledge,provides the possibility for patients with motor dysfunction to carry out active rehabilitation training.The brain-computer interface system based on motor imagery(MI)is to obtain the EEG data of the subjects' motion intentions,adopt effective feature extraction methods,design accurate classification algorithms,and then issue commands to control external devices to help patients undergo rehabilitation treatment.It has theoretical research value and medical application value to improve the recognition accuracy of EEG signals based on motor imagery.This paper analyzes the current research status of MI EEG recognition,and introduces data preprocessing,feature extraction and classification methods in the study of MI EEG recognition.Based on the open MI EEG data provided by the BCI competition,an MSC-DMLP framework is constructed to recognize the EEG signal.The main content of the thesis includes: 1)Firstly,the open motor imagery EEG data of BCI competition is preprocessed byband-pass filtering,then the frequency band is divided into several scales.Thecovariance feature extraction based on Riemannian geometry is carried out bycombining the multi-scale frequency sub-band EEG data.2)In this paper,a new decreasing multi-layer perceptron based on the MI EEGclassifier is designed.By introducing norm to improve focal loss function,theloss function is improved from simple empirical risk loss to structural risk loss,and the overfitting of the model is reduced to a certain extent.3)The performance of the model was evaluated by multiple sets of experiments.Inthis paper,the recognition effect of MI EEG data is tested from different scalespectral division,different feature extraction methods,various common classifiermodels,and improved loss function.A recognition frame with high accuracy is obtained.In the end,the mean accuracy of 75% is achieved in the BCI competition IV dataset 2A.This method is applied to BCI competition II set III binary classification motor imagery EEG data and compared with common classifiers,and still achieves high accuracy.It is proved that the improved decreasing multi-layer perceptron model(DMLP)is effective for MI EEG signal recognition.