The Study And Application Of Affective Computing Based On Bio-Signals

Emotion is a complex psychological and physiological phenomenon of human beings.Emotion exists in all aspects of human existence and life.The purpose of the study of affective computing through bio-signals is to measure the emotion state of people objectively.The main innovations in this article can be summarized as follows:(1)Combining Hilbert Huang Transform and time window,we extract power spectrum as feature from EEG signal for emotion recognition analysis.After that,we use Gauss kernel function Support Vector Machine(SVM)to classify different emotion states.The validation analysis on DEAP dataset shows a better classification accuracy than previous studies and other methods.According to the analysis of classification results,it is found that different subjects have different emotions for the same stimulus.In the awake state of the subjects,the higher frequency wavelet corresponds to the higher accuracy of emotion classification.(2)We propose a method to fuse the multidimensional features of EEG signals through pictures.At the same time,we proposed a feature extraction method for these pictures and an affective recognition method based on this.The multidimensional features of EEG signals include frequency domain dimensions,spatial dimensions,and time dimensions.The fusion method includes three steps.In the first step,the frequency domain characteristics of the EEG signals are extracted within different time windows.In the second step,the frequency domain features are mapped to formed a twodimensional picture according to the spatial distribution of EEG electrodes.In the third step,the generated two-dimensional pictures are arranged according to the order of the time window,thereby forming a picture sequence to express the characteristics of EEG signal under different stimuli.The multi-dimensional EEG feature extraction method uses a deep convolutional neural network for automatic feature extraction,and the extracted features are analized by LSTM recurrent neural network to recogonize the different emotion states.This study fuses deep convolutional neural network and LSTM recurrent neural network to form a hybrid deep neural network.The model validation phase uses the DEAP data set as a training set,test set,and validation set.In order to increase the training sample,we adopt the method of adding noise to the images to expand the original training set.By comparing the accuracy of emotion recognition,the results show that our method is effective.A further analysis of the experimental results revealed that the EEG feature sequence generated with a time window of 3 seconds corresponds to a higher accuracy of sentiment classification;secondly,in the EEG feature map,the details of changes in the FP1 and FP2 points have a greater impact on the entire affective recognition accuracy.(3)The fourth chapter discusses the problem of emotion recognition through the fusion of multimodal physiological signals.In the first stage of feature fusion,the features are extracted from different physiological signals respectively.Then the stacked autoencoder neural network fuses the different features.After that,the fusion features are classified with LSTM recurrent neural network for emotion recognition.After comparing the accuracy of emotion recognition with multi-modal fusion feature and the single modal feature,it is found that the multi-modal fusion feature corresponds to the higher accuracy than singal-modal.It proves that the multimodal fusion classification method is effective.(4)We designed an EEG signal acquisition system based on wearable EEG collection equipment and smart mobile devices.The main function of the system is to record the EEG signal and the POMS scale of the subjects for a long time.Then,we analyzed the correlation of EEG features and the POMS scales,in order to find out the emotion changes reflected by EEG signals in people's dayly life without stimula.Eight Subjects participate in the experiment,and the resting state EEG signal for 7 days were collected.After that,the correlation coefficient between the EEG features and the POMS scale is obtained by correlation analysis.The experimental results show that there is a significant correlation between the EEG features of the FP1 position and the POMS scale.For example,there is a significant positive correlation between POMS anger component and EEG features.There are four significant negative correlations between POMS self-components and EEG features,while no significant correlation is found between other POMS components and EEG features.