Research On EEG Emotion Recognition Based On Domain Adaptive Few-Shot Learning Networks

With the maturity of science and technology,brain science has become an important research field for human beings to explore the mystery of life.In order to analyze the discharge activity of the brain,Electroencephalogram(EEG)is widely used in different applications of brain science.Because EEG signals are not easy to camouflage,emotion recognition using EEG signals has attracted extensive attention of researchers in recent years.However,due to individual age,mentality,environment and other factors,EEG signals have great individual differences,resulting in the poor generalization performance of EEG emotion recognition model among different subjects.Moreover,due to the sample limitation that new individuals can be used for training,EEG emotion recognition technology is difficult to be applied to the actual scene.Domain adaptation(DA)method is the mainstream method to solve the above problems by migrating knowledge between different domains,so that it does not need to meet the requirements of the same distribution of training and test samples.Few-shot learning(FSL)method does not rely on large-scale samples,and can realize low-cost and rapid model deployment for a new task.Based on domain adaptation,this thesis explores the possibility of combining few-shot learning method,and studies EEG emotion recognition models in singlesource and multi-source domain,which provides a new idea for solving the problem of crosssubject EEG emotion recognition.The main research contents of this thesis are as follows:Firstly,from the perspective of reducing the demand for new individual samples,a singlesource domain adaptive few-shot learning(SDA-FSL)method is proposed.This method combines multiple source domain data into a single source domain,uses only a small number of target calibration samples and source domain data to carry out supervised domain adaptive few-shot learning,and extracts the common features of source and target.By improving the sampling strategy of prototypical network and designing an instance-attention mechanism,it can better adapt to cross-domain task.The experimental results show that the method is outperform the mainstream domain adaptation method and few-shot learning method,which proves that it has good practical application value.Then,for the problem that most existing single-source domain methods ignore the diversity of multi sources,this thesis also proposes a multi-source selective domain adaptive few-shot learning(MSDA-FSL)method,which trains feature extractor and classifier for each individual seperately,and selects the source domains closest to the target domain for the following multisource domain adaptive few-shot learning.In addition to extract the common features,the domain-specific features are also extracted,and a domain component attention mechanism is designed to combine the two kinds of features.Finally,the domain weight is calculated based on the similarity ranking between the source and the target to integrate the classification results.The experimental results show that this method can improve the utilization efficiency of source domains,also improve the accuracy of EEG emotion recognition of new subject.