Research On Classification Algorithm Of Motor Imagery EEG Based On Transfer Learning

Brain-computer interface(BCI)can realize the direct control of the external device by the human brain without relying on peripheral nerves and muscle tissues,and it is widely used in medical rehabihtation,education,smart home and other fields.Electroencephalogram(EEG)is a spontaneous brain electrical activity recorded on the scalp surface using electrodes.However,due to the non-stationary characteristics of EEG signals,traditional BCI often require the current user to perform a time-consuming training phase to obtain enough labeled samples to build a reliable classification model.But the long training time increases the user's burden and reduces the practicality of the brain-computer interface system.Motor imagery EEG is a kind of commonly used EEG signal in brain-computer interface,and this thesis introduces the transfer learning idea to reduce the training time required for the motor-imaging brain-computer interface system.In general,the labeled EEG samples of other subjects can be used to assist the classification model training of the current subject through transfer learning.In this learning mode,a good classifier can be obtained without acquiring a large number of labeled EEG samples of the current subject,thereby reducing the training time consumed by the current subject.Currently,most transfer learning algorithms used in braincomputer interfaces still require a small number of labeled EEG samples from the current subject.Unlike those,the algorithms proposed in this thesis only make use of the unlabeled EEG samples of the current subjects for knowledge transfer.The main research contents of this thesis as follows:(1)This thesis proposes a transfer learning algorithm for the manifold embedded distribution alignment under the Riemannian geometric framework.The algorithm takes the spatial covariance matrix of EEG as the initial feature,and then generates discriminative vector features through Riemann tangent plane mapping,and further combines manifold feature transformation and classifier integrated with distribution alignment to align EEG feature distribution of different subjects to generate a prediction model suitable for the target subject's EEG classification task.Experimental results on the public motor imagery dataset verify the effectiveness of the transfer learning algorithm.(2)A transfer learning algorithm based on a convolutional neural network is proposed.The algorithm introduces a CORAL loss for a convolutional neural network used for motor imaging EEG classification tasks.The loss reduces the offset between domains by minimizing the second order statistics(covariance matrix)differences between deep EEG features of different subjects.Through the joint training of classification loss and CORAL loss of the neural network,a prediction model can be generated to classify the target subject's EEG.Experiments show that this transfer learning algorithm can significantly improve the classification accuracy of target subject's EEG.