Research On The Classification Method Of Motor Imagery Base On Transfer Learning

In recent decades,the research on brain-computer interfaces(BCI)technology has attracted worldwide attention and developed rapidly.BCI is a communication technique that aims to identify a user's brain intents and translate them into a special set of commands to control the equipment in real world.Electroencephalography(EEG)is one of the most widely used noninvasive imaging technique in BCI.The EEG-based brain computer interface system can replace the damaged nerve and provide a new communication method.Traditional brain-computer interface techniques typically require a large amount of labeled training data from a target subject to construct a classifier model,which assume that training data and test data are in the same feature space and subject to the same statistical distribution.However,due to the nonstationarity of EEG signals,there may be different probability distributions between training data and test data.In addition,the BCI system is mainly used in the field of medical rehabilitation,it is often difficult to obtain a large amount of training data from the same subject.Moreover,there are individual differences between different subjects,which makes the data from other subjects cannot be directly utilized.These problems not only affect the performance of the system,but also limit its application.BCI-based subject specific motor imagery has been widely investigated.Based on the traditional processing methods of motor imagery EEG signal,we applies the idea of transfer learning to the classification of motor imagery EEG signals.In order to address the limitations of traditional brain-computer interface systems,two transfer learning algorithms are proposed in this thesis.The specific research work of this thesis is as follows.In this thesis,based on the subspace alignment,we proposed the weighted subspace alignment,in which the source domain samples are given different weights according to the similarity of the sample features,and the weighted source domain subspaces are aligned with the target domain subspaces.Considering the influence of the nonlinearity of EEG signals,we proposed a method to convert the original data into high-dimensional space by nonlinear mapping and then make subspace alignment,and derived the feasibility of the algorithm.Finally,we designed two experiments,which are cross-sessions EEG data transfer learning experiment and cross-subjects EEG data transfer learning experiment,verifying the effectiveness of the algorithm.In this thesis,we proposed a transfer learning algorithm based on common spatial pattern,which combined with the composite common spatial pattern and adaptive method.The method updates the covariance matrix for both classes simultaneously by calculating the similarity among the samples,and reconstructs the spatial filter to extract features for classification,to improve the performance of the classifier.Finally,we designed the experiment to verify the effectiveness of the algorithm on the public EEG dataset.