The Affective Recognition Of GSR Based On Nonlinear Prediction

In the research of affective recognition based on physiological signals, feature selection is the key of the research. In order to improve the recognition rate of affective recognition, this paper focuses on the extraction of effective features which can reflect different affective states. In this paper, the reliable galvanic skin response (GSR) had been intercepted after the subject’s four true emotions (anger, fear, happy and grief) were aroused successfully, to creat the affective physiological signal database.According to the nonlinear characteristic of GSR, this paper establishes a nonlinear prediction model based on the method of volterra series nonlinear prediction to predict the nonlinear regularity of the GSR under different affective states. The neural network approach is presented to solve each kernel of the volterra series for the drawbacks which the prediction filter coefficients is difficult to convergence and the predict result exists delay. This method not only can improve the calculation of coefficient vector, but also can speed up the convergence of filter coefficients, even improve the prediction accuracy. Then the effective nonlinear characteristics can be extracted based on the improved prediction model to be used for affective recognition. And the good recognition effect can prove the the proposed method for the affective recognition is effective. In the end, ant colony optimization algorithm and random forest classifier are used to optimize the linear and nonlinear characteristics which were extracted before. This method can reduce the feature dimension at the same time ensure the recognition rate. Finally the correct recognition rates of four kinds of affective states all surpass eighty percent, in which the recognition rate of grief affective reaches95%. This result is an significant improvement compared with previous studies. The implementation process is described as follows:(1) Make the plan of data collection and establish a sample database of affective physiological signal. In order to fully stimulate the subjects’four kinds of affective states, we choose the video clips with higher arouse degrees and the video without any affective for video editing. Then we document the process in detail and analysis the collected data. Finally the affective sample database are established by capturing the signals which be induced well.(2) The nonlinearity and predictability analysis of GSR. The nonlinear characteristics of GSR are proved by the method of surrogate data. By comparing the prediction results of nonlinear signal and random signal, we prove that the method of nonlinear prediction can restored the original rule, and there is no predictability at all in random signal.(3) The design of prediction model based on volterra series. We restore the chaotic attractor in the high dimensional space by space reconstructing of the GSR time series, construct new function to approach the original one by the expansion of the volterra series. At the same time, the evolution hiding in chaotic attractor has been found and the existing data could be described. Then we could describe the next moment state easily by the current state of the system to construct the prediction model of the volterra series.(4) Improve the prediction model, and construct the time series which contains strong affective rules. This paper solves each kernel of the volterra series by the method of neural network, improves the coefficient vector calculation methods so that the convergence of filter coefficients can be speeded up and the prediction accuracy can be improved. The nonlinear time series which contain obvious emotional change rule would be reconstructed for the feature extraction to ensure all the features which are mentioned in this paper are real and effectiveness.(5) Analyze and extract the nonlinear features of GSR with affection. After the pretreatment such as filtering and normalizing for the collected GSR, the time series based on the improved predict model can be calculated. And we adopt the nonlinear method to analyse the nonlinear characteristics which come from the original GSR signal and the reconstruction time series after prediction. Meanwhile we compare the entropy, dimension and other nonlinear parameters, and prove that these nonlinear features can be used for affective recognition.(6) Feature selection based on ant colony optimization algorithm. Some nonlinear characteristics are added in the linear features. And the ant colony algorithm is used to selecte the effective feature in all of the above features. Then Random Forest classifier is used for affective classification.The experimental results show that:(1) GSR signal is chaotic. It has strong nonlinear rule and strong predictability.(2) The method of volterra series prediction can make a better short-term prediction of the trend of GSR with affection, and restore the affective rules in a certain extent, which directly leads to the difference of the nonlinear feature calculation between the original signal and the predicted signal.(3) The GSR time series strengthen its law of the nonlinearity after the procession of volterra series prediction. The application of nonlinear characteristics after prediction can improve the correct recognition rate of the target affect; this indicates that the volterra series prediction has an important research value and significance in the field of affective recognition.(4) Ant colony optimization algorithm is used for the feature selection of linear and nonlinear characteristics in this paper. In the result, the selected probability of the nonlinear feature extracted from the GSR which has improved the prediction method is higher.(5) The rules of linear and nonlinear have been described in detail in this paper through all the characteristics of the GSR signal. And finally we get a better recognition rate of happy (80.44%).angry (81.21%), grief (95.33%) and fear (84.85%)...