Because of the non-stationary and individual differences of EEG,it is difficult to construct a common EEG emotion recogni-tion model.To solve this problem,Multi-Source Subdomain Adaptation Network is proposed to realize EEG emotion recognition.First-ly,each subject is regarded as an different source domain,and a domain-specific feature extractor is designed to make full use of data samples from multiple source domains to improve the generalization ability of the model.Secondly,in view of the distribution discrep-ancy of EEG signals from each source domain in the process of multi-source domain transfer,the multi-source domain discrepancy loss is proposed to reduce the variance and improve the stability of the model.Finally,training a domain-invariant classifier,the source do-main and target domain are divided into different subdomains.The local maximum mean discrepancy loss is used to realize subdomain adaptation and improve the accuracy of emotion recognition.Experiments are carried out on two emotional EEG datasets SEED and SEED-Ⅳ.The results indicate that the method effectively increases the accuracy of cross-subject and cross-session EEG emotion recog-nition,and the model is more stable and has strong generalization ability.
万方数据
维普
