EEG-based Emotion Recognition Using Support Vector Machine With Kernel Fusion

Along with the development of artificial intelligence,affective computing as a new popular research field has been proposed.On one hand,affective computing can make it possible for computer to recognize human emotion and create a better human-computer environment.On the other hand,by using data mining method researchers from neural science or recognition science can understand the essence of emotion and discover some significant factor which is relevant to emotion from data analysis directly.In past two decades,emotion recognition,the most basic important component in affective computing,has been paid attention and developed quickly.Along with the development of science and sensor technology,electroencephalogram(EEG)has been proposed and can be used to record physiological signal created by ongoing brain activity.Using EEG,researchers can study the relationship between brain activity and emotion directly.Many studies demonstrated that the signals we get by EEG can reflect subject's emotion state well,and some significant regions on brain relevant to emotion have been discovered.Therefore,based on existing researches this paper summarizes lots of paper and construct an EEG-based emotion recognition framework.This framework mainly includes four components: EEG feature extraction,feature selection,classification model.Experiment on real emotion data set based on EEG-based emotion recognition model has been performed to validate our model and find significant features relevant to emotion recognition.In this thesis,the research includes mainly the following three aspects:(1)Data preprocessing and feature extraction on EEG-based signal.Before emotion state classification,we have done signal denoising,downsampling,electrooculogram(EOG)elimination and feature extraction.In feature extraction,we divide whole EEG signal into segments with same 4-second length where there is a 2 seconds overlap between two neighbor segments.According to whether the feature is extracted from whole EEG segment,the features are divided into two groups: local feature and global feature.The local feature is extracted from the signal in the given frequency range,including the power spectrum and the relevant to it.While the global feature is extracted from the signal in the whole frequency range,including the numerical feature and Hjorth parameters.Both two types of feature are applied to data simulation experiment to compare their significance to emotion recognition.(2)Design new classification algorithm and decision function.Based on the performance of support vector machine(SVM)in the application of emotion recognition,a kernel fusion method has been proposed by combining an infinite kernel series.The new kernel created by kernel fusion method increases the input information for SVM as compared with old kernel,which can improve the performance of SVM on small sample.Moreover,the segments from same 1-minute EEG signal should have same labels,but classifier predicts each segment respectively.So we do a ballot decision and select the major label as final label,which makes all labels from a same EEG signal consistent.Based on new classification algorithm and decision function,a model framework is proposed and composed of EEG feature extraction,emotion classification,the significance evaluation of feature.(3)Feature analysis and electrodes reduction.Based on the proposed model framework,a cross validation with different kernels has been done to get the accuracy corresponding to each feature.We compared the results of different kernels and select the best result for feature analysis and electrode reduction.The results show that new kernel created by kernel fusion method has a remarkable improvement as compared with old kernels and reduces the model parameter.The accuracy of emotion state classification reaches 90%,which is higher than 70% of polynomial kernel and radial basis function kernel.According to the feature analysis based on cross validation,the significant feature relevant to emotion recognition is consistent with existing research.All results demonstrate the correctness of our model and the performance improvement of our kernel fusion method.Moreover,statistic analysis on the result of experiment shows that the performance of global features is generally better than the local ones.Finally,sequential floating forward selection is employed to find a minimum electrode set in which we can always find a feature for a given subject's data and the accuracy of new kernel on this feature can reach 70%.The selected minimum electrode set can be used to simplify the experiment process and feature selection in emotion recognition.