Research On Emotion Recognition For Brain Signals Based On Riemannian Geometry

Emotion is closely related to human behavior,family and society,so how to identify one’s emotional changes has become an important topic.Brain signals can reflect the basic activities of brain nerves and are not easy to disguise,so emotion recognition based on brain signals can be more authentic and reliable.With the development of brain computer interface technology,emotion recognition based on brain signals has attracted the attention of a large number of researchers.However,limited by brain signals of non-stationarity,individual differences,and higher acquisition costs and inevitable in the process of collecting artifact introduced,problems that how to improve the recognition accuracy of different emotions is still a problem worthy of studying.Based on this,the thesis based on the researches of the emotion recognition and its transference,the main contents are as follows:Firstly,in order to solve the problem of ignoring spatial information based on traditional feature extraction methods of emotion recognition,a new feature extraction method is applied to EEG and the corresponding deep neural network structure is designed.In order to learn the spatial information,Riemannian mean and distance from spatial covariance matrix are obtained on Riemannian manifold.Then tangent space learning is realized by Riemannian kernel method and spatial information is projected from Riemannian geometry to Euclidean space.Then,two fully connected layers are used to learn spatial information embeddings.In addition,the longshort term memory network with attention mechanism is used to learn the temporal information of brain signals from Euclidean space via differential entropy.In order to combine spatial and temporal information,an effective feature fusion strategy was selected to realize the selection decision of specific embedded features through weight learning,and the validity was verified on SEED datasets.Secondly,aim at solving the distribution difference of emotional brain signals,this thesis introduces transfer learning to study the migration of emotional brain signals from two different perspectives.First,based on the long-short term memory neural network designed above,add the improved correlation difference loss based on Riemann measure,reduce the difference between source domain and target domain by minimizing the loss function in back propagation process.Secondly,based on the feature structure,the domain adaptive method based on unsupervised learning is used to transform and reconstruct the source domain features and target domain features by matching the differences between the statistical distribution,and then the previously designed network is used to realize feature fusion and classification.Cross-session and cross-subject experiments were also designed on SEED datasets to verify the effectiveness of the proposed method.