Research On EEG Emotion Recognition Combined With Deep Learning And Compressed Sensing

Emotion recognition technology plays a vital role in intelligent human-computer interaction,enabling machines to have the same emotional perception capabilities as humans,and can better serve people’s daily lives.Among the many signals that can be used to study emotion recognition,the EEG signal is a physiological electrical signal generated by the discharge of neurons.It has the advantages of simple measurement,mobility,and not easy to disguise,and can objectively and truly reflect people’s emotional state,so it is widely used in the research of emotion recognition.However,multi-channel electrode collection caps are used for the acquisition of EEG signals,resulting in high-dimensional characteristics of EEG signals.When performing emotion recognition,certain channel redundancy and dimensional disaster will occur.Therefore,this paper uses Compressive Sensing(CS)technology to obtain discrete samples of EEG signals through random sampling under the condition of lower than Nyquist sampling rate,so as to remove redundant information in EEG signals.At the same time,a CS-based emotion recognition model is proposed based on the characteristic that the discriminativeness of the signal in the sparse representation has the function of classification.In addition,in view of the characteristics that the observation matrix in the CS emotion recognition model cannot adapt to the signal,and the problem that the reconstruction algorithm relies on the sparse dictionary,it is proposed to use deep learning to improve the CS emotion recognition model.The main research work of this paper is as follows:(1)The DEAP emotional EEG database was selected as the material for emotion recognition in this paper,and subjects in the database were screened at the same time.Thereby,the EEG signals of subjects that meet the requirements of the experimental balance data set are obtained,and then the selected EEG signals are windowed to obtain EEG emotion samples.Then,the emotional features of the EEG samples are extracted by frequency bands,and the EEG emotional samples are divided into four frequency bands θ,α,β and γ,to extract the power spectral density(PSD)and differential entropy(DE),Differential asymmetry(DASM),rational asymmetry(RASM),asymmetry(ASM),and differential causality(DCAU).Finally,the two frequency bands β and γ that can best reflect the emotional characteristics are selected as the emotional characteristics of the follow-up research through experiments.(2)Due to the high-dimensional characteristics of EEG signals,there are some problems such as channel redundancy and dimensionality disaster in the extracted features.Aiming at this problem,this paper uses compressive sensing technology to observe and sample EEG emotional features,remove redundant information in the features,and improve the efficiency and accuracy of EEG emotion recognition.And based on the discriminativeness of the signal in the sparse representation,a CS-based emotion recognition model is proposed.In this identification model,Gaussian random matrix,Bernoulli matrix,Fourier matrix,sparse random matrix,Toeplitz matrix and Circular matrix are used as its observation matrix,and Orthogonal Matching Pursuit(OMP),Sparsity Adaptive Matching Pursuit(SAMP)and Variable Step Size Sparsity Adaptive Matching Pursuit(VssAMP)are used as reconstruction algorithms.Finally,the recognition performance of the model is verified by comparing with the SVM recognition model.(3)Aiming at the problems existing in CS emotion recognition model: the observation matrix can not adapt to the signal characteristics and the reconstruction algorithm depends on sparse dictionary,a CS emotion recognition model combined with deep learning(CS-DBN-SAE)is proposed.This model utilizes the learning ability of Deep Belief Network(DBN)to transform the traditional random observation matrix into an observation network that can adapt to signal characteristics,by training the Stacked Auto Encoder(SAE)to get rid of the dependence of the OMP,SAMP and VssAMP reconstruction algorithms on sparse dictionaries,thereby improving the recognition rate of EEG emotion.Finally,experiments show that CS-DBN-SAE recognition model can effectively improve the accuracy of emotion recognition compared with CS emotion recognition model.