Neural Network Model Of EEG Time Series

Emotion is a psychophysiological process triggered by the human body based on the perception of the environment,which plays a vital role in human daily cognition and life.With the continuous progress of artificial intelligence technology in recent years,emotion recognition has been widely used in educational planning,medical insurance,public opinion monitoring and safe driving.Compared with human’s external behavior characteristics,EEG has better spontaneity and objectivity.It is most closely related to the brain that produces human emotions and a good carrier to express emotional information.However,there are still some urgent problems to be solved in the study of emotional EEG signals,such as the difficulty of nonlinear characteristic expression,the low correlation between EEG channels and emotions,the large computation of neural network,and the poor biological interpretability of the models.Therefore,this paper focuses on how to effectively extract the emotional features of EEG signals,reduce the redundancy of EEG channels,and efficiently and accurately identify emotional states.The following studies are carried out from three aspects: spikes encoding algorithm of EEG signals,channel selection method and classification model:1.Traditional signal analysis methods is difficult to reflect the nonlinear characteristics of EEG signals,the pulse neural network cannot be used analog signals as input data,this paper used the BSA encoding algorithm based on the technology of stimulus estimates to encode the original EEG signals of DEAP dataset,and discrete spikes time series with a size of40*32*60*128 was generated by deconvolution calculation of a single subject’s continuous EEG signal and FIR linear filter,then the coding results were evaluated and optimized by combining SNR and RMSE.Meanwhile,grid search was used to improve the hyperparameter optimization function of BSA coding to solve the problem of significant difference of EEG signals between different subjects.The optimal coding parameter groups of different subjects were constructed to improve the information expression efficiency and accuracy of coding pulses.2.Aiming at different EEG channels in emotion recognition task have emotional representation differences and channel redundancy problem,this paper puts forward a channel selection algorithm based on the ReliefF analysis of channel characteristics.Firstly,calculates the average spike intensity of 32 EEG channels as the characteristic value.Secondly,channel feature weights under different neighbor numbers were calculated based on ReliefF features to describe the emotional representation ability of each EEG channel to four emotional dimensions of Valence,Arousal,Dominance and Valence-Arousal in the DEAP dataset.Finally,the EEG data was reduced from the original 32 channels to 8,10,10 and 16 channels respectively based on the result of feature weight evaluation.By dividing time slices,three dichotomy datasets of Valence,Arousal and Dominance dimension and a four categoric datasets of Valence-Arousal dimension were constructed.3.The traditional artificial neural network has many problems,such as poor biological interpretability,large model parameters,complex network structure and huge computation.In this paper,a spiking neural network model based on negative feedback suppression is proposed,whose input layer,excitatory layer and inhibition layer are constructed by LIF neuron model with adaptive membrane potential threshold.At the same time,STDP learning rules based on impulse response gain mechanism were combined to train the emotional state classification of the model.Finally,the validity of the model was verified by the DEAP dataset,the dichotomous accuracy of Valence,Arousal and Dominance dimension was82.54%,78.31% and 85.17%,respectively,and the accuracy of four categories of ValenceArousal dimension was 73.98%.The classification accuracy has been significantly improved compared with the same category models.