Study On Brain-computer Interface Classification Based On Transfer Learning

Brain-Computer Interface(BCI)is a technology that used to establish a new channel between human brain and external devices and not depend on the brain's normal output pathways of peripheral nerves and muscles in order to realize the communication between human thoughts and other electric instruments.This new type of human-computer interaction technology can provide patients with normal thinking but severe disabilities a way to communicate with outside world and control special devices.BCI technology has great application value in medical rehabilitation field and also has potential value in entertainment,industry and military fields.There are two basic assumptions in the SVM machine learning for the accuracy and reliability of the classification model after training.First,the training samples and the new test samples meet the independent distribution conditions.And second assumption is that there must be enough available training sample to learn a exact classification model.However,the nonstationarity of EEG signal causes differences of statistical distribution of signals between different sessions of one subject in BCI system,which limits the reusability of large amount of training data.Therefore,the SVM BCI system needs to be re-trained for a long time before it is officially used.Consequently,reducing the number of training samples required to shorten the training time is one of the main difficulties in the problem of brain-computer interface classification.Around the main challenges of classification,this thesis proposes two kind of brain-computer interface classification methods based on transfer learning.The main research results include:1.A transfer learning method based on self-training support vector machine is proposed.In this thesis,with the basic frame based on a semi-supervised learning method of self-training support vector machine algorithm,the labels of unlabeled samples in the target subject test set are predicted during the self-training process and the confidence level of classification label of unlabeled sample is calculated by utilizing the target subject training set and the auxiliary training set.Then we select the unlabeled sample with the largest target confidence level and its prediction label to join the target subject training set,in order to expand scale of the target subject training set and solve the problem of a small number of training samples when training target subject classification model.2.A transfer learning method based on common spatial pattern is proposed.Based on the feature extraction method of common space model,this method calculates the similarity between the target subject training set and the auxiliary training set and reconstructs the characteristic data of the target subject training set according to the similarity.This method build a new classification model to and improve the accuracy of classifier.The experimental results show that the two methods can achieve the goal of training a reliable and accurate classification model in the case of a small number of training samples,and finally achieve better classification accuracy compared to the situation of the SVM training and testing process.