Research On Brain Signal Emotion Recognition Based On Transfer Learning

Emotion is an integrated state comprised of people’s ideas,cognition,feelings,and behaviors.It plays an essential role in how people interact with each other,make decisions about events and interpret information.In recent years,the accuracy of identifying human emotions using information technology tools has received wide attention from experts in various fields.The powerful techniques involved in analyzing human emotions include computer vision,which is capable of analyzing facial expressions,and natural language processing,which is capable of analyzing emotions expressed in texts.Electroencephalogram(EEG)has received much attention as a method of analyzing human emotions by observing the physiological activity of nerve cells in the cerebral cortex.However,for the emotion recognition based on EEG,there is a large individual difference between the EEG signals of different subjects,that is,each subject’s EEG signals are different from others,resulting in cross-subject emotion identification being a difficult problem.Therefore,how to improve the ability of emotional prediction in crosssubject experiments is the focus of studying work in this article.In this thesis,deep learning and transfer learning methods are used to identify and classify the emotions of EEG signals.Firstly,to solve the problem of the insufficient generalization ability of fixed prior knowledge and artificially pre-defined brain network structure in cross-subject experiments,we proposed an emotion recognition model(DOGNN),which can dynamically construct different graph structures based on input features.Based on the data of different subjects,the model constructs the adjacency matrix,and adaptively generates different graph structures.Moreover,the introduction of the domain adversarial method further reduces the difference in EEG signals across subjects.Secondly,for the existing study,when dealing with different subjects’ EEG signals,previous studies just simply matched the feature space of source domain and target domain by the transfer learning and ignore the in-domain category mismatch of the two aligned domains which is caused by the instability of the EEG signals.This thesis,we further present a decision boundary-based emotion recognition model(SMCD),which can generate soft labels based on a self-training method.Additionally,the SMCD model can spatially align source and target domains under considering a task-specific category so that the features from the target domain can be included in the source domain feature distribution of its corresponding category.Finally,the detailed experiments of the two models that have been conducted in two public datasets: SEED and SEED-IV,and the results show that these methods have a significant effect on emotion recognition.