Research On Local Spatio-Temporal Field And Deep Learning Methods For Eeg-Based Emotion Recognition

Electroencephalogram(EEG)signals contain abundant information about human emotionality.Emotion plays an important role in human cognitive activities and behaviors,and it has a significant influence on our life and work.Emotionality comes from the brain and affects the body through the nerve and endocrine systems.Therefore,it is of great significance in human-computer interaction and mental disease monitoring to recognize emotions by mining effective information from EEG signals.In recent years,with the rapid advancement of brain-computer interfaces,more and more researchers have joined in the affective computing research.However,there are many problems in emotion recognition based on EEG signals,such as drastic signal changes,substantial noise interference,extremely weak emotional information,etc.All of these limit the further development of EEG emotion recognition.This paper carries out related research on critical links such as feature extraction and deep learning classification in EEG emotion recognition.We adopt relevant methods of signal processing to extract features and deep learning methods for classification.Then,on this basis,we propose a new approach,"Spatio-Temporal Field",And explore the effectiveness of different methods for EEG emotion recognition.The main content of this article is as follows:The first part is based on the rich frequency domain information of EEG signals and studies the emotional information in its five frequency bands: γ 、θ、α、β、γ.Based on these five frequency bands,four features are extracted respectively: Power Spectral Density features,Rational Asymmetry of Spectral Power features,Differential Entropy features,and Rational Asymmetry features.The properties of the symmetrical distribution of electrodes in the process of EEG signal acquisition are integrated into the Rational Asymmetry of Spectral Power features and Rational Asymmetry features.After that,based on these features,we propose three deep learning models: One-Dimensional Convolutional Neural Network,Long Short-Term Memory,and Bidirectional Long Short-Term Memory for emotion classification.We carry out experiments on DEAP and DREAMER,classify valence and arousal into two categories,and compare their performance with each other.Based on the experimental results,we can discover more effective features and deep learning models to provide the reference for the model building in the next section.The second part is based on the deep learning method of the first part,which integrates a new local field method with an appropriate feature and classifier model.It proposes a "Spatio-Temporal Field" model,which includes four stages: signal serialization,feature representation,space localization,and emotion classification.The feature representation and emotion classification in this method refer to the combination of feature and classifier in the deep learning method。The local field incorporates the "divide-and-conquer" idea.It can effectively reduce the data complexity in the divided local areas of the feature space,despite of severe noise interferences and serious data variations.Moreover,the dynamic temporal information exploited by the Bi-LSTM is also complementary with the local spatial information explored by the local field.So,it can better tackle the weak signal classification problem in emotion recognition based on EEG.Finally,we conduct experiments on DEAP and DREAMER databases,and compare with other advanced methods,which proves the effectiveness of our method.In addition,the performance of our method under different data division ratios and the loss function curve during the training process are also analyzed.