Emotion Recognition Of Electromyography Signal Based On Random Forest Algorithm

In this paper, EMG is used to research emotion recognition about sad, happy, disgust, surprise, fear angry. Firstly experiment need to design a perfect emotion induced plan and establish a database of emotional physiological signals. According to the characteristics of the electromyography signal, we choose the wavelet transform method to eliminate noise of the original signal and do the original feature extraction. Original feature contribution is sorted by random forest algorithm which accord to the classification and recognition result, to choose the important features. The improved discrete binary quantum particle swarm optimization (PSO) algorithm and Fisher classifier which constructed based on kernels is used to do feature selection. And then, parameter selection of kernel function is discussed, to determine the optimal kernel parameters. At last the emotion model is set up.This article main research content:1The acquisition scheme is fixed and the database is established. Participant will watch rich emotional color film, and fully inspire his joy, disgust, surprise, grief, angry, fear six kinds of feeling, the physiological signals is recorded at the same time. Then, according the questionnaires that filled out by participants to analyze the reliability of data, the abnormal data will be removed and the original database will be established.2The electromyography signal feature set access method. Then wavelet transform method was used for de-noised effective EMG signals, after noise reduction, statistical features in time-domain were obtained, Daubechies5wavelet with orthogonality and compact support was adopted as basic function to do5-layer decomposition of EMG signal after noise, then21statistical features of each layer’s detail coefficients were extracted. So we can obtain126original features in total.3The feature selection methods. However, not all features make contributes to emotion recognition, so it is necessary to find affective features from them. Feature selection in emotion recognition is a combinatorial optimization problem thus a NP problem. If using exhaustive method with the existing conditions obviously it can’t achieve, so to find a fast effective feature selection method is an important problem. This paper is divided into two steps to achieve this goal. First step random forest algorithm is used, according to the characteristics of the evaluation to the original features of contribution, to choose the larger contribution to the classification. The second step it is used that improved discrete binary quantum particle swarm optimization (PSO) combined with Fisher classifier which constructed base on kernels to choose the most representative electromyography signal feature subset. Then kernel function parameter value is discussed in order to determine the optimal kernel parameters. The difference will be know if compared with similar emotional recognition model. The optimal emotion recognition model is established.4The emotional model is constructed based on the optimal kernel parameters, and testing it.The result:Using random forest algorithm for dimension reduction is feasible. Those big contribution features will be choose by using random forest algorithm which accord to his characteristic. By using the improved discrete binary quantum particle swarm optimization (PSO) and a Fisher which based on Gaussian radial basis kernel, the optimal combination of feature will be choose. Through discussion, the optimal kernel parameters will be selected.This system is superior than a system to use normal Fisher classifier.