Study Of Emotion Classification Method Based On Multiple Physiological Signals

Physiological signals,as the most easily acquired signals by human body tlirough sensors,contain a large amount of important physiological and psychological information of human body.Acquisition and recognition of a variety of physiological signals are of great significance to computer recognition of human emotions.Limited by the complexity of physiological signals such as EEG,GSR,RSP and PPG,it has been difficult to extract useful features from these physiological signals and make accurate classification and prediction of emotions.With the concept of "affective computing" proposed,scholars are committed to the concept of emotion mathematical,so that the computer can recognize and process,and make the recognition and classification of emotional states.The traditional emotion classification method is usually to manually extract relevant features and use machine learning model for classification.With the development of deep learning and other models that have the ability of representation learning,some methods of automatic learning and classification based on deep learning model have emerged.This paper makes some relevant explorations on these research directions,among which the main research contents and contributions include:The methods of physiological signal analysis modeling and feature extraction are reviewed in detail.Secondly,based on the DEAP dataset,physiological signals such as EEG,GSR,RSP and PPG were used to extract wavelet entropy,sample entropy,statistical features and other features,and KNN,SVM and Xgboost were used to conduct binary classification experiments on emotions.The results showed that the classification accuracy of valence was 9.7% higher than that of DEAP paper,and the accuracy of arousal was 1.4% higher than that of DEAP paper.In order to improve traditional emotion recognition methods,physiological signals need to be deeply understood and relevant features need to be manually extracted.A multi-grain scanning learning method based on deep forest was proposed,and the time-domain data of EEG,GSR,RSP and PPG in DEAP dataset were extracted automatically.At the same time,a method to fuse the feature vectors of 32 channels EEG signals is proposed,and the feature is further studied through cascade forest.The results showed that the classification accuracy of valence was 14.9% higher than that of DEAP paper,and the accuracy of arousal was 6.1% higher than that of DEAP paper.The experiment shows that deep forest can extract the feature of the signal automatically and predict the signal classification to some extent.