Eeg Emotion Recognition Based On Multi Feature And IGWO-SVM

This paper focuses on the multi-scale fuzzy entropy feature and CNN feature of EEG signals,and the improved grey wolf algorithm to optimize support vector machine classification method to realize EEG emotion recognition.The main work of the thesis is as follows:(1)The study compared several EEG features,feature extraction and classification methods.After analysis,we selected the segmented composite multi-scale fuzzy entropy features for small samples and the CNN features for large samples to obtain highly discriminative features,and we selected the improved grey wolf algorithm to optimize support vector machine classification to improve the EEG emotion recognition rate.(2)An improved multi-scale fuzzy entropy feature(PCMSFE)is proposed for small samples.At the same time,the support vector machine classification optimized by grey wolf algorithm is used.The multi-scale fuzzy entropy feature used segmental coarse granulation and composite multi-scale fuzzy entropy to solve the problem of data loss and inaccurate calculation caused by excessive scale factor.The support vector machine improved by grey wolf algorithm optimized(IGWO-SVM)could adjust the internal parameters,balance the global and local search of the grey wolf algorithm by the cosine convergence factor,and select the combination of the harmonic mean and the static mean to update the grey,so that the wolf could get the best position faster.For the large sample,the convolutional layer,the pooling layer and the fully connected layer of the convolutional neural network are selected to extract the data features,and the improved power spectral entropy is used for preprocessing,it removed the Fourier transform step and determined the probability value to improve calculation efficiency and accuracy.Before the feature data is input to the IGWO-SVM model,the Relief feature selection was performed.The feature selection algorithm is used to calculate the feature weight,which plays a good role in further improving the emotion recognition rate and shortening the running time.(3)The algorithm proposed in this paper is compared with some similar studies in the open DEAP dataset.The emotion recognition rates of approximate entropy,sample entropy and PCMSFE feature are discussed.The emotion recognition rates of the full-band and ?-and ?-band characteristics are compared.The effect of improved grey wolf algorithm and several other algorithms is improved.Experiments show that the recognition rate of PCMSFE and IGWO-SVM is the highest,and the average recognition rate of valence,arousal,dominance and liking can reach more than 87%.And we test the recognition rate of valence and arousal under high/low liking.The experiment shows that the emotion classification rate is high when the liking is low.In addition,whether to preprocess and improvement of power spectral entropy,whether to use feature selection algorithm and the use of different thresholds,CNN classification and IGWO-SVM classification of emotion recognition rate are compared.At the same time,the emotion recognition rate of the above model and KNN,decision tree and Bayesian network are discussed.The experiment shows that the EEG emotion recognition rate is the highest when adopting the CNN feature and the IGWO-SVM model,and the average recognition rate in each dimension can reach more than 94%.In addition,the accuracy and efficiency of PCMSFE features and CNN features are compared.The results show that using PCMSFE features in small samples can improve the emotion recognition rate.And using CNN extraction features in large samples can improve the recognition efficiency.