Subdomain Adaptive Network Algorithm Cross-subject EEG Emotion Recognition

‍ ‍In order to solve the problem of low generalization performance of emotion recognition model among different subjects caused by the non-stationarity and individual differences of EEG， this paper proposes a model， the Series System of Global Domain Adaptation Network and Related Subdomain Adaptation Network （SS_GDAN_RSAN）， to realize emotion recognition across subjects. In this paper， the whole model is composed of feature extractor， global domain classifier， subdomain classifier. Firstly， in the shallow neural network， the feature extractor and the global domain classifier generate domain invariant expression， and make global domain adaptation by minimizing the source domain emotion classification error and the loss of edge distribution similarity between the source domain and the target domain data， Secondly， in the deep neural network， the transmission network is learned based on the weighted maximum mean discrepancy， and the subdomain classifier is trained by minimizing the error in subdomain classification and the loss of subdomain adaptation. So as to capture the fine-grained information of each category to achieve subdomain adaptation. The experimental results show that the proposed method SS_GDAN_RSAN is simple but effective， and the recognition rate is 84.05%±5.91% in many-to-one cross subject migration experiment， and 91.66%±7.32% in single subject cross time migration experiment. Compared with the traditional classifier model， SS_GDAN_RSAN performs well in improving the generalization ability of cross subject and cross time emotion classification tasks.