Research On EEG-based Emotion Recognition Algorithm With Deep Learning

Emotion is a complex physiological and psychological phenomenon produced by the interaction between the internal biochemical system of the biological individual and the external environment.It can reflect the mental state of the biological individual.Emotions have penetrated into all aspects of human life,and related research and applications have also spanned many fields.With the development of machine learning,deep learning and emotion research,emotion recognition based on EEG signals has gradually become a hot topic in various fields.How to extract effective EEG features and how to build more effective models are the key to improving accuracy of emotion recognition of EEG signals,and are also the main problems of EEG emotion recognition.In this thesis,the EEG signals in the DEAP dataset are mainly used as the research data set.Constructed traditional machine learning models,deep neural network models and hybrid deep neural network models to study the emotion recognition of EEG signals in two dimensions of valence and arousal degree.(1)Study the classification performance of traditional machine learning models on EEG signal motion recognition.First,extract the time domain and frequency domain features of EEG signals in DEAP data,and combine the time and frequency domain features to obtain five basic features of EEG.Then,five features are used as input of support vector machine classifier,integrated decision tree classifier,linear classifier and Bayesian linear classifier.Finally,two types of emotion recognition are performed on EEG signals in terms of valence and arousal.The experimental results show that,in terms of features,each classifier has the best recognition performance on the combination features.And in terms of the models,the integrated decision tree is the best classifier,and the average accuracy in terms of valence and arousal reaches 79.71%and 78.28%,respectively.(2)Study the classification performance of EEG signal emotion recognition based on deep neural network models.In the construction of the deep neural network models,three types of CNN models and one type of double-layer LSTM model were selected through continuous optimization of the model parameters to affect the five features of EEG signals in the DEAP data set in terms of valence and arousal classification experiment.The three CNN models use convolution kernels of different scales to learn the local correlation between the time and space of EEG signals.The LSTM model captures the correlation of EEG timing.Compared with the experimental results of traditional shallow classifiers,the deep neural network models can automatically learn and extract deeper temporal and spatial correlation features in EEG signals,thereby improving the accuracy of EEG signal emotion recognition.The CNN model with a 5×5 convolution kernel has the best recognition performance,and the average recognition accuracy of five features in valence and arousal degree is improved by 4.31%and 3.77%respectively compared with the integrated decision tree classifier with the best performance in traditional classifiers.The average recognition accuracy of the five features of the double-layer LSTM in valence and arousal is 5.6%and 3.37%higher than the integrated decision tree classifier respectively.(3)Study the classification performance of EEG emotion recognition based on hybrid deep neural network models.The cascade and parallel models of CNN model and LSTM model are constructed.In the cascade model,first input different features of EEG signals to CNN model to extract the spatial correlation features,and then use the output of the fully connected layer of the CNN model as the input of the LSTM model,and continue to extract the timing correlation features of the EEG signals,and finally through the classification layer,the prediction performances of the valence and the arousal degree are output respectively.The experimental results show that the parallel model has the best recognition performance in the combination features,and the average classification accuracy of valence reaches 99.64%,and the average classification accuracy on arousal reaches 85.88%.In the parallel model,CNN and LSTM simultaneously extract the spatial and temporal features of EEG signals.Then connect the outputs of the fully connected layers of the CNN and LSTM models to form a high-level space-time combination feature.Finally input them to the classification layer for emotion classification.The results show that the CNN and LSTM parallel combination model has the best recognition performance in the combination features,the average recognition accuracy in the valence and arousal reaches 88.76%and 87.14%,which is higher than that of the CNN and LSTM models.Combining the experimental results of the two hybrid models,the hybrid models can effectively improve the accuracy by jointly learning deeper temporal and spatial correlation features of EEG signals.In summary,the deep neural network models and hybrid deep neural network models proposed in this paper are effective,and can learn the deeper EEG features,thereby improving the accuracy of EEG signal emotion recognition.Therefore,the research method in this paper has very good theoretical and practical significance.