Weighted Multi-Source Domain Adversarial Transfer Based Learning Motor Imagery EEG Recognition
Motor imagery EEG signals have large individual differences and high data collection costs, and building a small sample, cross-subject motor imagery EEG recognition model is a key problem in Brain-Computer Interfaces. A weighted multi-source domain adaptive method is proposed based on pre-alignment strategy and adversarial transfer learning for small sample,cross-domain learning, which combines transfer learning and adversarial training,extends the domain adversarial neural network to multiple source domains, weights the pearson correlation coefficient for each source domain to achieve weighted alignment of multiple source domains and target domains in features,and uses the pre-alignment strategy to improve the consistency of data distribution between domains. On the BCI competition motor imagery dataset,the recognition accuracy of cross-subject motor imagery task reaches 84.43％,which is 9.17％ higher than that of the non-transfer method and 5.0％ higher than that of the domain adversarial neural network. The experimental results show that the proposed method can effectively reduce the differences in EEG data distribution and feature distribution between different sub-jects,and achieve double alignment of data and features,thus improving the cross-subject motor imagery EEG classification performance.
motor imageryadversarial transfer learningweighted multi-source domaincross-subjects
万方数据
维普
