EEG Emotion Classification Based On Fusion Features And Joint Model

EEG emotion recognition has always had important research significance in the fields of medical health,psychological medicine,and human-computer interaction.However,EEG signals are not stable enough,difficult to analyze,strong noise artifacts,nonlinear,and unbalanced,which add many difficulties to the emotion recognition task.At the same time,it is difficult for the previous emotion recognition methods to simultaneously fit the characteristic information of the EEG signal in the time domain,frequency domain,wavelet domain,and spatial domain,which makes it difficult to improve the accuracy of emotion recognition.To address the above research status,this paper starts from improving the accuracy of emotion recognition,then performs multidomain feature fusion based on generalized features learned from deep models,and finally the fused features are input to the joint model constructed to handle the emotion classification task.The specific research contents and main achievements are as follows:(1)In view of the research foundation and current situation of EEG emotion recognition based on feature fusion,The fusion of traditional computing features is difficult to improve the accuracy of emotion recognition.Innovative Generalization feature extraction using deep hybrid network(LF).To select the best combined features and verify the recognition performance of generalized features.In the DEAP dataset,a nine-category emotion classification experiment based on the SVM model is carried out.The experimental results show that the combined features of Hjorth parameter,difference asymmetry,wavelet feature and generalization feature achieve the highest accuracy and score of 74.73%,0.63,and 0.61 in the three evaluation criteria of accuracy,precision and F1 score,respectively.And the overall performance of the four sets of fusion features is better than the single control feature.This verifies that the fusion feature approach is more adaptable to the emotion recognition task than a single feature has better classification accuracy than a single feature.On the other hand,the contribution of generalization features to the combined feature accuracy is verified.Experiments show that the average accuracy of the four sets of features is increased by 3.4% after adding generalization features,indicating that the generalization features can effectively fit multi-domain feature information and can effectively improve the accuracy of emotion recognition.However,there is a problem that the recognition rate of generalized features is lower than that of some single features.(2)For the problem that the accuracy of generalized features is lower than that of partial single features,it is proposed that this may be the assumption that shallow classifiers cannot fully utilize the complete feature information of generalized features.At the same time,in order to further verify the recognition effect of the four combination features on the deep learning model,CNN and GRU models were constructed and experiments were carried out respectively.The experimental results show that the combined features of Hjorth parameter,difference asymmetry,wavelet feature and generalization feature once again achieved the best accuracy of 86.14% and84.47%.The generalization feature has a lower accuracy than the Hjorth time-domain feature on the GRU model,Other experimental results are better than other characteristics These results further verify that the generalized features successfully fit multi-domain feature information.This verifies that the deep learning model can better utilize generalization features.But in a single time,series feature,the GRU model exhibits strong performance,with an accuracy rate over 19.1% higher than the lowest feature Wen.(3)Based on a single model,the performance-oriented problem is shown in processing fusion features.At the same time,the purpose is to improve the accuracy of emotion recognition,Therefore,the integrated learning concept is used to construct a network model that integrates CNN and GRU structure.The feature input of the submodel is adjusted for the performance-focused problem.The experimental results show that the fusion model achieves the highest accuracy rate of 92.26% among all comparative literatures,which is higher than the lowest literature.It is verified that the CNN-GRU fusion model has excellent classification performance in emotion recognition,and it also shows that the feature fusion idea combined with the generalized feature LF is accurate and effective.