Study On Transfer Learning Classification Based On Motor Imagery

Brain-Computer Interface(BCI)is a computer-based system that converts EEG signals into control signals for external devices.BCI technology is unique in that it does not rely on the normal output channels of the nerves and muscles around the brain.The main application areas of BCI technology are medical rehabilitation,such as amyotrophic lateral sclerosis,brain stem stroke or spinal cord injury.BCI technology allows people with severe motor impairments to communicate with computers or other external devices.At the same time,this new technology has great potential application value in the fields of artificial intelligence,new entertainment,military and aerospace.This thesis studies the classification of EEG signals in motor imagery.When classifying EEG signals,traditional machine learning algorithms often require sufficient training samples to obtain higher classification accuracy.However,when the number of training samples is small,it is difficult to construct a classifier with good performance.In the BCI system,due to the large individual differences in EEG signals,the reusability of experimental data for different experimental subjects is limited.Therefore,how to use the experimental data of different experimental objects to obtain good classification accuracy under the condition of insufficient training samples is a major difficulty in the BCI system.Based on the BCI system of motor imagery,this thesis proposes two kinds of transfer learning classification algorithms based on motor imagery,focusing on the classification of small training samples in the BCI system mentioned above,the correlation between the auxiliary samples and the target training samples,and the influence of the auxiliary samples on the construction of the target classifier.The main research results include:1.A transfer learning method for removing unrelated auxiliary samples is proposed.The method considers the correlation between the auxiliary sample and the target training sample,The auxiliary sample of the misclassified is regarded as the auxiliary sample not related to the target training set,and is removed from the auxiliary sample set.Then perform instance-based transfer learning.2.A transfer learning method with different proportions of auxiliary samples is proposed.This thesis proposes a new evaluation standard for evaluating the impact of auxiliary samples on target training samples.And based on this evaluation standard,the most suitable proportion of auxiliary samples is selected.At the same time,for the multi-source system studied in this thesis,based on this evaluation standard,different weights are given to different auxiliary objects.The experimental results show that the two methods proposed in this thesis can build a reliable classifier with only a small number of training samples.Compared with the traditional machine learning classification algorithm,the classification accuracy rate is significantly improved.