Emotion Recognition Of EEG Signals Based On Transfer Learning

With the rapid development of artificial intelligence technology,research on affective computing has become a hot spot.It is a very challenging task to enable machines to automatically recognize changes in human emotional characteristics and make human-computer interaction more natural and smooth.Among them,the EEG signals collected by the brain-computer interface are widely used in the research of affective computing because they contain a large amount of physiological information and are undisguisable.The thesis use large open EEG datasets SEED and DEAP to make experiments.My research work on EEG signal sentiment classification includes feature selection and extraction of EEG signals,selection of domain adaptive algorithms,fusion of deep neural networks and transfer learning methods to build classification models,etc.The specific research work includes the following three aspects:(1)In order to reduce the distribution difference between the source domain and the target domain EEG signal,a balanced distribution adaptation(BDA)algorithm is used to classify EEG signals for emotion.The model adds a balance factor to adaptively adjust the weights of the edge distribution and conditional distribution of EEG data in the source domain and target domain.This enhances the interpretability of transfer learning based on a distribution.The proposed method uses differential entropy features on SEED datasets for affective tricategorization experiments.73.14%accuracy was achieved in experiments with different time factors.At the same time,compared experiments with three classical transfer learning methods.It is verified that transfer learning has good performance in processing EEG signal emotion classification tasks.(2)In order to enhance the interconnection of joint distribution between adaptation layers in the network,a new electroencephalogram(EEG)emotion recognition method based on deep convolutional joint adaptation network(CNN-JAN)is presented,which incorporates the idea of joint adaptation in transfer learning into deep convolutional networks.Firstly,the model uses a rectangular convolution kernel to extract the deep emotion-related spatial features between EEG data channels.Then,the extracted spatial features are fed into the adaptation layer with multi-kernel joint maximum mean discrepancy(MK-JMMD)for transfer learning to reduce the distribution differences between the source and target domains.The experiments are carried out on differential entropy features and differential causality features of EEG data from the SEED dataset to verify the effectiveness and advantages of our proposed method.As a result,the within-subject emotion classification accuracy on differential entropy features reaches 84.01%,and the crosssubject emotion classification accuracy is also improved compared with other current popular transfer learning methods.(3)In order to avoid the confusion and loss of data in the process of transfer learning,a transfer learning method combining attention mechanism and subdomain adaptation(ASAN)is proposed.The main idea of this model is to align the features of each sub-domain to complete the global data alignment.First of all,ASAN uses the spatial attention mechanism to enhance the significance of EEG features between channels.Then CNN is used for deep feature extraction.Finally,the extracted depth features are measured using the local maximum mean discrepancy algorithm to measure the subdomain distribution of the same label between different domains.This method minimizes the difference in distribution between each subfield,thereby narrowing the difference between the two global domains and completing transfer learning at a more fine-grained level.The experiments are performed using differential entropy and Pearson correlation coefficient features on SEED and DEAP datasets.The proposed ASAN model achieves 87.56%accuracy in the cross-time cross-validation experiment of the SEED dataset and 80.03%accuracy in the cross-subject experiment on DEAP.Further experiments are carried out on cross-dataset subdomain transfer learning,and compared with related algorithms for verification.The ASAN model achieved 69.22%triclassification accuracy in the transfer learning experiment from DEAP to SEED dataset,and the performance was the best for all comparison experiments.These verify the accuracy and robustness of the model.In this paper,an EEG signal sentiment classification model integrated with deep learning is proposed from the whole domain and subfield aspects of transfer learning.The proposed method significantly enhances the data characteristics in the deep network and improves the performance of model classification.These have laid a good foundation for subsequent research on human-computer interaction related applications based on EEG signals.