Research On Multi-band Emotion Classification Methods Based On Multiple Physiological Signals

Emotion plays a vital role in human life.With the continuous development of artificial intelligence technology,computers have better technical support in the area of human-computer interaction and personalized services.Accurate identification of people's emotional state through computer is a research hotspot in data mining,brain-computer interface,bioinformatics and other fields.Current research on emotion recognition can be divided into the following aspects:speech-based emotion recognition,facial expression-based emotion recognition,physiological signal-based emotion recognition,and motion-based emotion recognition,etc.Since physiological signals can reflect people's emotional state more objectively and realistically,more and more researchers begin to study emotional recognition based on physiological signals.At present,there are still many problems in this research process,such as emotional labeling of physiological signals requires a lot of manpower and material resources;the non-linear and non-stable characteristics of physiological signals bring difficulties to feature extraction;the standardized emotion classification model based on physiological signals has not been established;there is no emotional classification model for independent users and so on.This paper focuses on the problems of feature extraction and model stability in the process of emotional classification of physiological signals.The main contents are as follows:(1)Aiming at the non-linear and non-stationary characteristics of EEG signals,a multi-scale feature emotion classification method based on EEG signals is proposed.Firstly,a feature extraction method based on empirical mode decomposition(EMD)and autoregressive model(AR)is proposed.The Alpha and Beta bands in the EEG signal are extracted by filtering,and the intrinsic mode function(IMF)is obtained by EMD decomposition.The AR coefficient is calculated to form the feature set,and the support vector machine is used for recognition.In the two-class task of arousal dimension and valence dimension,the classification accuracies of 87.35%and 83.12%are obtained,respectively.Secondly,a feature extraction method based on empirical wavelet transform(EWT)and AR is proposed.The highest classification accuracy is only 72.04%,which is lower than that of EMD and AR method.(2)Aiming at the unstable result of emotion classification of physiological signals,an ensemble emotion classification method based on multi-band of multiple physiological signals is proposed.By extracting Theta,Alpha,Beta,and Gamma bands and calculating Hjorth parameters from EEG signals,EOG signals,and EMG signals,four different feature sets are formed,namely,EEG feature sets;EEG and EOG feature sets;EEG and EMG feature sets;EEG,EOG and EMG feature sets.Then,K-nearest neighbor,random forest,decision tree single classification model and the ensemble classification model are used to perform the emotional two-class task and the emotional four-class task on the arousal dimension and the valence dimension.The results show that the highest classification accuracy can be obtained when the EEG,EOG,EMG combination feature set is adopted and the ensemble classification model is used.For the two-class task,the best results on arousal and valence are 94.42%,the best result on valence level is and 94.02%,respectively.For the four-class task,the highest average classification accuracy is 90.74%,and it shows good stability.