Emotion Recognition Based On EEG Signals

Rapid and accurate recognition of multiple emotions is becoming a research hot spot in the field of Brain-Computer Interface and Human-computer interaction currently.Emotion is an internal and external manifestation of physical and psychological fluctuations when a person is stimulated by the outside world and it affects the communication distance between people in daily life.With the development of human-computer interaction technology,a kind of algorithm that exerts influence on emotions is produced.The purpose of this algorithm is to allow machines to perceive,recognize,analyze,and infer human emotions automatically,so that machines have better intelligent interactions.Due to the many shortcomings of traditional non-physiological signal emotion recognition,and emotions are closely related to the brain,EEG signals can be used as a bridge between the brain and emotions to reflect the internal changes of the brain in different emotional states,and can be an assist treatment of the emotional state of patients with mental illness.In this paper,we focuse on a variety of emotional classification problems and the factors that cause individual differences,design a kind of film segment inducing experiment,aiming to study the whole process of emotion recognition based on EEG signals.Through the comparison of the identification methods and the selection of the optimal channel set and identify the emotional state of different participants quickly and accurately.The specific content of this article is summarized as follows:(1)A non-intrusive BP EEG device was used to construct a data set of EEG signals of 9 participants under four emotions: fear,anger,sadness and happiness.The movie fragments of the complete plot were used to induce different emotions of the participants,and the collected emotional EEG was processed for noise reduction.(2)We used time domain,frequency domain and nonlinear dynamics feature extraction methods to extract and normalize a total of 15 different effective features,according to the root mean square feature and the scatter plot of the three-dimensional time domain feature,it can verify the distinction between the four emotions.Meanwhile,we tooks the average accuracy rate as the evaluation index of classification and recognition,the K-nearest neighbor algorithm and support vector machine were applied to calculate the average classification accuracy of the overall EEG characteristics of 9 participants.The comparison shows that the classification speed of the K nearest neighbor algorithm is faster and the classification accuracy of the support vector machine algorithm is higher;the two classification accuracy rates obtained by using the Gamma band feature of the Welch method in the frequency domain feature are the best among the 15 features;The two classification accuracy differences obtained by the mean value features of absolute value of first order difference are the smallest,and the classification stability is the best.(3)According to the characteristics of Wel?5,we draw the average power spectrum topographic maps of different genders and four emotions,afterwards,analyzed the brain power differences in the EEG Theta,Alpha,Beta and Gamma band signals.The result shows that the Gamma frequency band has the highest power value in the whole brain area of the male and female genders and the four emotions;the male Beta frequency band power in the happy and fearful emotions is significantly higher than that of females in all brain areas except for the bilateral temporal lobes;Gender Alpha band and Theta band have lower power in most brain areas,males Theta band under sad and angry emotions have the highest power in the parieto-occipital lobe,and the power in the frontal lobe of the male Alpha band and Theta band in fear is the highest.Moreover,females under happy and fear emotions have the highest power in the right temporal lobe.(4)Based on the particle swarm optimization algorithm and ion motion algorithm,the channel optimization selection was performed for each participant,we found the optimal channel set and common channel under different selection results.The result shows that the average classification accuracy of the optimal channel set SVM in IMO channel optimization is higher than 95%,and fewer channels are selected.The weight ranking method can find the optimal channel combination more accurately;From the power spectrum topographic map of Wel?5 characteristic Gamma band,we found that the common channels selected by the two methods are mainly in the frontal and temporal lobes.