Study On Emotion Recognition Of Pulse Single Based On Feature Evaluation And Selection

Emotion is the attitude experience and behavioral responses to external objective things occurring in the process of exploration and plays a big import role in the individual exploring,logic reasoning,analysis,plan and decision.Emotion recognition is the process that collecting human physiological signals or behavioral characteristics which contain the emotional information and then mining the features expressing the difference of emotional status from signals,in order to building emotional model with highly separable and general ability.Accurately,rapidly and predictively recognizing,expressing and having emotion,and moreover controlling emotions for constructive purpose have much theory and practice significance in clinical medicine,social science and engineering application.Pulse signal have mature analysis methods and havecollecting approach in much ways and have the advantages of portability,non-invasion,continuity,objectivity and reliability.In this thesis,six kinds of emotional pulse signals are collected,and then statistical analysis,information theory and machine learning techniques were used to analyze,quantify,evaluate and screen the emotional information components of that to build pulse signal emotion matrix,in the aim of mapping relationship between the matrix and the affective state and establishing emotion classification models with high performance based on the optimal learning algorithm.The main research contents and results are as follows:(1)Through the design of the appropriate emotional pulse signal acquisition program,a reliable environment for the implementation of the experiment is established to collect six kinds of emotional state of the pulse data of 60 college students.After noise filtering,the pulse signal,through the feature point,time sequence,feature layer mapping,252 features with a certain physical meaning are extracted in linear and nonlinear space to construct characteristic matrix to describe the change rule of the pulse signal in order to realize the quantitative morphological characteristics of emotional state of the pulse.(2)Statistical test technology is used to analysis of the distribution of characteristic sequences.The nonparametric methods are used to explore the different and correlation between the characteristics of sequence and emotional status in aims of concerning the emotional characteristics in pulse signals.(3)The normalized mutual information method is used to quantify the correlation degree between the feature sequences and the relationship between the feature and the emotional state and the redundant information of the feature matrix.Based on that,using the minimum redundancy and maximum correlation algorithm,random forest algorithm and statistical test methods to search and construct a feature matrix of maximum resolution of emotional information and minimum redundancy in the feature space,and then quantify the contribution characteristics of emotion recognition value.(4)Combining statistical analysis technology,information theory,machine learning algorithm and statistical methods to establish feature selection system,in order to analysis,mining,evaluate and weight the features to construct a feature subset with smaller dimensions and greater ability to distinguish the emotion status.(5)Using random sampling technique,10-folds cross-validation and grid search algorithm to build the emotion recognition models with the optimal subset and the optimum parameters and subset,with aim of achieving high classification performance in one-one and one-much emotional states recognition.The performance of the model is evaluated by the combination of random sampling and cross examination.(6)Based on the model for identifying the state of human emotion based on the minimal subset and the optimal parameters of the random forest algorithm,a research method on the relationship between the performance of the model and the data partition is designed.And then the framework of research on mine construction is constructed to quantitate the relationship between the performance of model recognition and the size of training set in order to find a training set with minimal size.Through the above research,we find that the emotion information contained in the pulse signal and change of the emotional state will cause the change of pulse signal,and the features extracted in this thesis can be used to quantitatively characterize these changes.It is also found that the resolution ability of linear feature is significantly higher than that of nonlinear feature,which means the changes of pulse shape caused by affective factors are mainly reflected in the online space,especially in the frequency domain.In this thesis,the three level feature selection systems can effectively reduce the dimensionality of the emotional features matrix and the complexity of the model,and improve operation efficiency and generalization ability of models.Emotion classification models based on optimal feature subset and tune parameters have high accuracy and strong robustness and generalization ability.Moreover,under the premise of ensuring the accuracy of classification,the training set is minimized in this study.