Research On EEG Emotion Recognition Based On Deep Learning

In recent years,emotion recognition has become a research hotspot in the field of artificial intelligence.As a new technology,EEG emotion recognition has attracted more and more attention.The traditional EEG emotion recognition method has some limitations.Therefore,researchers began to explore the use of new technologies such as deep learning for the research of EEG emotion recognition.EEG data,which is different from text and image data,exists widely and needs special acquisition equipment to obtain.The shortage of EEG data is an urgent problem to be solved.As a one-dimensional time series,EEG signals need to choose appropriate representation methods,and appropriate data representation can reduce complexity and training costs.In addition,different EEG data sets may have different sampling rates,channel numbers and signal lengths,which are difficult to be uniformly trained.In view of these problems,this thesis studies from three aspects:1.The EEG data set is constructed from the EEG collection experiment,and the EEG emotion data of multiple subjects is created with audio and video stimulation as the task goal.The EEG data collection experiment is designed and carried out.A complete set of EEG collection experiment specifications is developed from the preparation work in the early stage of EEG collection,the experimental process in the middle stage,and the data induction and sorting content in the later stage.The data collection helps to supplement the problem of insufficient training data.2.The characteristic content of one-dimensional time series data representation is difficult to be significantly reflected in emotion recognition,and the potential information of simple EEG signals is difficult to be discovered.Therefore,the one-dimensional time series is expressed in the form of multi-dimensional graph,converted into MDWVG(Multiple Directed Weighted Visible Graph),the weighted clustering coefficient is used to characterize the MDWVG structure,and the multi-dimensional characteristic information is strengthened by constructing the EEG link matrix,Emotional recognition of EEG signals is carried out through convolution neural network.3.In the face of different data sets,different models need to be built for training.The application cost is high,and there is no pre-training model to adapt to cross-data set recognition.Therefore,a cross-data set pre-training model is built,and Transformer is built as a framework by mask prediction,and a self-monitoring pre-training method is designed to enhance data characteristics.Self-supervised training false tags are generated by clustering the density of EEG signals to reduce manual workload.The pre-training model can adapt to different EEG data sets.The research carried out experiments on SEED and DEAP data sets,and the results showed that the accuracy of the self-supervised pre-training model on cross-data sets was86.5%,which exceeded the benchmark score by 9%(F1 score).The accuracy rate of the complex network method in the SEED verification of the open data set reached 93.85%,and the accuracy rate of emotion recognition was 9.4% higher than that of the single-variable visible image.The method is also applicable to cross-subject data,and has wide application in practice.