Extraction Of Nonlinear Features Based On Physiological Signals In Basic Emotion

Affective detection is one of the basic and core issues in affective computing. Compared with all kinds of explicit indicators such as facial expression and speech, physiological signals can be more authentic and reliable in user’s inner affective experience, so the affective detection based on physiological signals attract much attention. Recently most affective detection based on physiological signals is to extract the statistic feature in time domain and frequency domain; few studies have focused on the nonlinear characteristic of physiological signals. In order to further extend previous researches, this paper mainly examines the affective detection based on the nonlinear features of physiological signals.In this paper the below research work is carried out:(1) For the difficulties about collecting emotional physiological signals, the memory experimental paradigm is adopted to induce the participants’ emotion. At the same time, BIOPAC MP 150 is used to collect the participants’GSR and HR signals. The characteristics of the experiment are by the recalling method we can gather the participants’ subjective evaluation of the effect, obtain their inner experience about whether emotional induced or not and the intensity. In the experiment,200 pjysiological datum are collected, and the final effective datum are:happiness is 96, sadness is 87, anger is 36, and fear is 59.(2) In order to extract the nonlinear features of GSR and HR, we firstly confirm that these two physiological signals do possess nonlinear characteristic. So the surrogate data method is adopted to verify the existence of non-linear characteristics. Firstly we produce the surrogate data of all kinds of affective physiological signals, then select time reversal amount T1 and higher-order autocorrelation function T2 to be as statistical test items. The result shows that the GSR and HR of the four kinds of emotion are significantly different, so we can prove that the GSR and HR contain reliable nonlinear components.(3) For the GSR and HR, which include nonlinear component, several kinds of nonlinear numerical analysis methods are used to calculate the nonlinear features of physiological signals. These methods and features include:phase space reconstruction, C-C methods adopted to compute embedding dimension and delay time, wolf method used to compute largest Lyapunov exponent, GP algorithem adopted to compute correlation dimension, modified lempel-Ziv complexity adopted to compute first-order and two-order lempel-Ziv complexity and approximate entropy, detrended moving average algorithm used to compute multifractal spectrum and relevant spectrum features, recurrence quantification analysis adopted to compute a series of features about recurrent plot and calculate the poincare map of HR. According to the above calculation, we get 45 nonlinear features. Finally the principal component analysis is used to reduce dimension, and only ten main features are used to later classification.(4) For the problem of original data-set imbalanced (happiness have 96, sadness have 87, anger have 36, fear have 59), we propose an improved algorithms named IBFSVM that mainly resolves imbalanced dataset classification,which is based on the classic fuzzy support vector machine,. The IBFSVM algorithm integrates the cost-sensitive ideal to the membership function, which can enhance the recognition performance in minority class. By running three artificial datasets and six UCI datasets, we compare the performance of C-SVM, FSVM and IBFSVMThe result shows that in the case of little difference in positive and negative sample size, the performance of three kinds of classifier has little difference. With the increasing imbalance ration of positive and negative class, the superiority of IBFSVM gradually reflects. In the Abanole dataset, in which the imbalance ratio is 39.55, the recognition rate in minority of IBFSVM is 78.25%, while that of C_SVM and FSVM only are 5.16% and 5.79% respectively. From the above data we can find that when there is large difference in sample number between minority class and majority class, IBFSVM can significantly improve the classification performance in minority class. IBFSVM is adopted to classify four kinds of affective physiological data, the g-mean respectively are: happiness is 97.33%, sadness is 94.18%, fear is 87.9%, anger is 83.51%.(5) Based on IBFSVM, the recognition rate of traditional statistic features and nonlinear features is compared, and the result shows that the nonlinear features get better recognition rate. of happiness the recognition rate of statistic features is 80%, and that of nonlinear features is 97.33%. of sadness the recognition rate of statistic features is 77.41%, and that of nonlinear features is 94.18%. of anger the recognition rate of statistic features is 78.76%, and that of nonlinear features is 83.51%. of fear the recognition rate of statistic features is 80.1%, and that of nonlinear features is 87.9%.