| Emotion is a unique psychological activity of human beings.It not only reflects our physiological and psychological states,but also has significant effects on our cognition,communication,and decision-making.Electroencephalogram(EEG)signals,which represent the collective response of neural cells in the cerebral cortex,can objectively and accurately reflect human emotions.However,the field of EEG-based emotion recognition has long faced the challenge of improving the recognition accuracy and generalization ability of emotion classification models,particularly when dealing with samples from different individuals or time periods.In this regard,domain adaptation(DA)methods provide an effective solution that allows training and testing across different domains through knowledge transfer,thus relaxing the requirement that training and testing samples must satisfy the same probability distribution.Based on the characteristics of electroencephalogram emotion recognition technology,this paper proposes a multi-branch multi-source domain adaptation based on attention mechanism(MBMDA)electroencephalogram emotion model,building upon the traditional method of multi-source domain adaptation(MSDA)to address the generalization problem of emotion classification models.This model can achieve multi-source domain adaptation transfer learning across datasets and subjects,thereby improving the accuracy of electroencephalogram emotion recognition.The main contributions of this paper are as follows.In this article,we first introduce different types of common artifacts found in EEG signals,as well as a range of methods for preprocessing such signals.We also conduct preprocessing on three EEG datasets used in this study,namely DEAP,SEED,and SEED-Ⅳ.The goal of preprocessing is to eliminate noise from EEG signals,which is a critical first step in improving classification accuracy.Subsequently,in response to the common issue of traditional domain adaptation methods neglecting the temporal and spatial information of EEG signals in public feature extraction,this paper proposes a brain EEG public feature extraction method based on attention mechanism,and conducts research on the mixed optimization model(FFT_CLA).On the public dataset DEAP,FFT_CLA demonstrates exceptional feature extraction capabilities,with an average accuracy of 92.38%for EEG emotion recognition.This outcome indicates that utilizing the attentionbased deep model FFT CLA for brain EEG public feature extraction is a feasible approach.In the final stage of the research,the brain EEG public feature extractor FFT CLA is utilized to extract the corresponding brain EEG public features.These features are then sequentially input into the domainspecific feature extractor(DSFE)and domain specific feature classifier(DSFC)for the purpose of feature extraction and classification of brain EEG signals,respectively.Furthermore,the paper employs Maximum Mean Discrepancy(MMD)to minimize the distribution difference between the source and target domains,and utilizes Minimize Classification Loss(MCL)to obtain more accurate classifiers.Additionally,Minimize Discrepancy Loss(MDL)is used to obtain more convergent classifiers.Finally,the proposed method is evaluated on the public datasets SEED and SEED-Ⅳ.According to the experimental results,the accuracy of cross-session and cross-subject scenarios are 88.88%and 90.98%,respectively,on the SEED dataset.On the SEED-Ⅳ dataset,the accuracy of cross-session and cross-subject scenarios are 69.28%and 68.04%,respectively.Furthermore,the paper explores the impact of different normalization methods,combinations of loss functions,and various hyperparameters used in the experiments,and the results indicate that they have a significant impact on this research.In summary,the use of an attention-based deep learning hybrid model for brain EEG public feature extraction,and the utilization of a multi-branch multi-source domain strategy for EEG emotion recognition are feasible approaches in the domain adaptation field of transfer learning. |
PDF Full Text Request