Emotion Recognition Of ECG Signals Based On Machine Learning

Emotion recognition and emotion computing are the key aspects of human-computer emotional interaction,and are also hot research topics.The current emotion recognition is mainly based on speech intonation,facial expression,body posture,text information,physiological signals and so on.Among them,the research on emotion recognition based on physiological signals is particularly prominent,which mainly contains three aspects of research on physiological signal acquisition,feature extraction and feature recognition.The paper mainly focuses on the experimental design of emotion evoking paradigm,Electrocardiogram(ECG)signal acquisition and pre-processing,ECG waveform detection and Heart Rate Variability(HRV)signal acquisition,ECG and HRV feature extraction and emotion recognition of ECG and HRV signals under different emotional states,etc.The paper aims to achieve high accuracy of machine emotion recognition using ECG physiological signals.The main tasks and studies completed were as follows.1)Experimental paradigm design for emotion elicitation.For the four emotions of happiness,pride,fear and anger,the video clips were carefully selected and Scored and evaluated as emotion evoking materials,and the emotion evoking experimental paradigm was designed using Superlab software to induce the subjects to produce the corresponding four positive and negative emotions.2)ECG signal acquisition and pre-processing.Firstly,abnormal ECG analysis was performed using the MIT-BIH Arrhythmia dataset to eliminate pathological ECG;then pure emotional ECG data were obtained after preprocessing the ECG signals in the Augsburg emotional physiology dataset;in addition,the ECG signals of four emotional states were self-acquired by the MP150 physiological signal measuring instrument,and the Acknowledge4.4 software was used to remove the industrial frequency interference,baseline drift and motion artifacts from the collected ECG signals to establish the emotional ECG data set.3)ECG waveform detection and HRV signal acquisition.The wavelet transform is used to extract the modal extrema and trans-zero points of the wavelet coefficients of the ECG signal,set a suitable threshold,detect the ECG R-wave,calculate its first-order difference,and then obtain the heart rate variability signals under different emotional states.4)Feature extraction of emotional physiological signals.Peak detection analysis is used to extract the time-domain statistical features of ECG signals under different emotional states and obtain 7-dimensional ECG feature vectors;then,the time-domain,frequency-domain,time-frequency-domain and non-linear-domain features of HRV signals are extracted to obtain14-dimensional HRV feature vectors.5)Machine learning emotional feature classification.In order to compare the ECG and HRV signals,the ECG and HRV signal features are selected and different classifiers are used for emotion recognition.For the7-dimensional ECG and HRV signal features,a Support Vector Machine classifier modified by the Firefly Algorithm was used to design a machine learning emotion classification model based on ECG features as well as HRV features,and achieved an average correct classification rate of 91.5% and93.5%,respectively;a Subspace K-Nearest Neighbor classifier was designed to achieve emotion recognition of ECG and HRV signals,and achieved The average correct classification rates were 87.16% and 88.78%,respectively;the classification accuracy of 79.19% and 92.97% were achieved by using the tree model classifier algorithm of Random Forest to classify and identify the emotion of ECG and HRV signal features,respectively.In addition,fusing the extracted 14-dimensional Heart Rate Variability signal features,three machine learning classification methods,Support Vector Machine improved by Firefly Algorithm,Subspace K-Nearest Neighbor,and Random Forest,were used for emotion recognition,and the average classification accuracies of 95%,91.84%,and 94.22% were achieved,respectively.(6)Performance evaluation of emotion recognition classifiers.The performance of three classifiers for physiological signal emotion recognition was compared and analyzed,including the effect of improved Support Vector Machine,Subspace K-Nearest Neighbor,and Random Forest methods for emotion recognition.The results show that the correct rate of emotion recognition of HRV signals is higher than that of ECG signals using three different classifier algorithms;the accuracy of the improved Support Vector Machine with Firefly Algorithm is higher than that of Random Forest and Subspace K-Nearest Neighbor algorithms for emotion recognition.In addition,the F1-Score of the improved Support Vector Machine classifier based on the Firefly Algorithm is higher,with an average of 0.94,indicating that the classifier has a stronger generalization ability for emotion recognition.Therefore,extracting the multidimensional features of HRV signals and using the improved Support Vector Machine classification method with the Firefly Algorithm can obtain better classification performance and better achieve physiological signal emotion recognition.The research results of the paper have important Scientific and application values in affective robotics,medical health,psychology,contextual learning,multimedia game development,and business fields.