Research On EEG Emotion Recognition Based On Deep Learning

Emotions are complex psychological states that involve human responses to external stimuli and the physiological reactions associated with them.Emotions contain both emotional elements and rational thinking.In various human-machine interaction scenarios,accurate identification of the emotional state of the human can improve the interaction and communication between the human and the machine,making it more friendly and natural,and thereby preventing and stopping certain emergencies.The hotness of cognitive analysis and classification of emotions has been a major focus in many disciplines,such as neuroscience,psychology,cognitive science,computer science,and artificial intelligence,due to the advancement of science and technology.In recent years,deep learning techniques have demonstrated remarkable proficiency in emotion classification.However,current EEG feature extraction methods still rely on manual design,and extracting recognizable emotional features remains a relatively complex process.Most of the existing studies ignore the serious influence of baseline signals and the differences of EEG signals among different volunteers,resulting in low classification accuracy of the models.This paper delves into the pre-processing methods of baseline signals and the use of deep neural networks for emotion classification,with a concentration on three primary areas.(1)Two pre-processing methods for baseline signals based on the dimensionality reduction of baseline signals is proposed.Principal component analysis algorithm is employed by the first method to extract the principal components of the baseline signal as baseline features,thereby eliminating effects of baseline signals and noise on the experimental signal during its acquisition.At the same time,in view of the large amount of baseline signal data collected in some datasets and the high computational complexity when using principal component analysis algorithm,it is proposed to use singular value decomposition algorithm to extract baseline features by directly matrix decomposition of baseline signals and then dimensionality reduction of baseline data,which can effectively improve the computational efficiency.(2)A multidimensional parallel convolutional neural network model is proposed to tackle the issue of extracting feature patterns and single dimensions from existing network models.This model extracts time features and space features of EEG signals using convolutional modules of different dimensions from the perspective of time and space,and splices the extracted features to form spatio-temporal features for emotion classification.In the spatial feature extractor,a convolution method with the convolution size decreasing step by step is designed to accommodate the differences of EEG signals between different volunteers,fully extract spatial feature information,and improve the classification performance of the model.(3)A capsule network model based on attention mechanism guidance is presented,which can effectively address the issue of the traditional convolutional neural network’s pooling layer losing essential features.At the same time,channel attention mechanism is used for adaptive coding,so as to enhance the importance expression of each electrode channel.This model can excavate hidden features in the multi-channel EEG signal and simplify the design of the traditional capsule network model.It can reduce the running time and improve the computational efficiency without affecting the classification performance of the model.