A Study On Deep Transfer Learning For Motor Imagery Electroencephalography Decoding

Brain-computer interface(BCI)system can establish a communication pathway between the brain and external devices directly,and enables the brain to control the external devices by decoding the EEG signals into a series of executable machine instructions.As a kind of typical spontaneous BCI system,motor imagery(MI)based BCI does not require the external stimulation,and records the EEG signals during imagining movements,so it has received great attention from many researchers.Because of the difficulty of EEG acquisition,brain computer interface system has the problem of lack of training data.Therefore,how to learn domain-irrelevant knowledge from different subjects and apply the knowledge learned from the existing subject to the new subject is a problem to be solved in the research of brain computer interface system.To solve the problems and challenges in the existing MI decoding methods,in this thesis we study the motor imagery coding method based on deep transfer learning,focusing on the single-source unsupervised domain adaptation problem.The main contents are as follows:1.The Wasserstein distance can be used as a criterion for the initial selection of source domain from target domain.Based on this,a training strategy including three stages of pre training,source domain selection and transfer learning is designed.This strategy can reduce the training time of transfer learning and effectively use the existing data.2.This paper proposes a metric-based deep transfer learning method which is called Wasserstein-guided domain adaptation.In order to achieve better effect of domain alignment,this method uses Wasserstein distance as the loss,and the influence of different Wasserstein distance calculation methods on the results is studied.Experimental results on open motor imagery data sets demonstrate the effectiveness of the proposed transfer learning method.3.This paper proposes a subdomain adversarial-based deep transfer learning method which is called Wasserstein-guided multi-adversarial domain adaptation.In order to obtain fine-grained effect of distribution alignment,Wasserstein distance is used to segment sub domains,and multi domain discriminators are used to align different sub domains.At the same time,domain shrinkage loss is used to improve the effect of domain alignment.Experimental results on open motor imagery data sets demonstrate the effectiveness of the proposed transfer learning method.