Research On Class Imbalance Domain Adaptation Method For Eeg Emotion Recognition

With the development of AI and BCIs,there is a growing need for machines to understand the emotional cognition of humans.The non-stationary nature of EEG signals,as well as individual differences and low signal-to-noise ratio,present challenges in utilizing EEG-based artificial intelligence and brain-computer interfaces for extensive training purposes.Moreover,with the introduction of domain adaptation,a large number of cross-subject EEG recognition methods based on domain adaptation have emerged,and there has been a further enhancement in the accuracy of recognition.The issue of imbalanced categories among various subjects is also a practical concern in real-world applications.However,the existing domain adaptation methods are limited to the task of EEG signal recognition under category balance,and category imbalance of cross-subjects results in negative transfer within category space,ultimately leading to decreased recognition accuracy.At present,there are mainly two different category imbalance tasks: the first one is partial domain adaptation,i.e.,the source domain category set contains the target domain categories,and the knowledge migration in this scenario may incorrectly transfer the source domain non-common category knowledge to the target domain,triggering negative transfer.Another one is open-set domain adaptation,which does not restrict the category set of the target domain,and the target domain may have categories outside the category space of the source domain.Therefore,how to separate the unknown classes of the target domain while domain alignment becomes the challenge of this task.For the above two different class imbalance problems,the following models are proposed in this thesis:Regarding the issue of cross-subject partial domain adaptation,this thesis proposes a domain adaptation method based on partial semantic alignment.The model is based on an adversarial learning domain adaptation method,introducing a domain-wise clustering strategy to obtain the target domain class information to generate pseudo-labels,and facilitating the generation of discriminative clustering centers through a prototype alignment loss training model.Then,a partial semantic alignment method is further proposed to reduce the semantic differences and domain differences between the source and target domain,reducing the impact of source domain private classes on domain adaptation and the risk of negative transfer.The experimental validation on the SEED dataset and SEED-Ⅳ dataset in this thesis shows that the method can effectively improve the recognition accuracy of cross-subjects EEG emotion recognition in partial adaptation scenes.Regarding the issue of cross-subject open-set domain adaptation,an open-set domain adaptation method based on unknown sample likelihood evaluation is proposed in this thesis.Firstly,the weights of target samples are generated by the proposed unknown sample likelihood evaluation method,which is used to constrain and adjust the process of adversarial training.Then,a known class compensation strategy based on adaptive feature norms is proposed,in which the uncertainty entropy of the target sample prediction is used to constrain feature norms alignment to encourage the model to learn known class features with larger norms in both domains.Experiments on the SEED-Ⅳ dataset and SEED-Ⅴ dataset demonstrate that the method effectively improves accuracy for cross-subjects EEG emotion recognition in the scenes of open set domain adaptation.