| With the progress of society and economic development,more and more people are concerned about the influence of emotions in life and work.In the fast-paced life of modern people,they are often faced with heavy mental stress.Long-term bad emotions can easily lead to insomnia and increase the incidence of mental illnesses such as anxiety and depression,threatening people's health and even life.On the other hand,with the development of artificial intelligence,emotional intelligence is becoming an indispensable part of the field of artificial intelligence.Identifying emotions accurately and efficiently has important theoretical and practical value in clinical medicine,social science,and engineering practice.Designing and developing a machine or wearable device with emotion recognition function,real-time monitoring physiological signals of patients with mental diseases,perceiving changes in patient emotions,and giving corresponding feedback in time has become a research hotspot in the treatment of mental diseases.In addition,emotion recognition has wide application prospects in online education,criminal investigation,traffic safety,and leisure and entertainment.In the current research of emotion recognition based on physiological signals,feature extraction is not widely used.This paper proposes a new method for pulse and respiratory feature extraction from the extraction method of predecessors.For the pulse signal,the pulse morphological time series is constructed by identifying the feature points,and the pulse characteristics are extracted from the time domain and the frequency domain by combining the intrinsic modal functions obtained by the integrated empirical mode decomposition.For the respiratory signal,the multi-scale entropy algorithm was used to extract the nonlinear characteristics of the respiratory signal.Combined with time and frequency domain analysis,the respiratory characteristics related to different emotion states were extracted from various aspects.The main research contents of this article are as follows:(1)Design and implement a reasonable emotional evoked experiment by collecting the pulse and respiratory signals of sixty undergraduate students through multiple-guide physiological instruments.After signal re-sampling,wavelet filtering was used to remove the noise of the pulse signal and the second-order IIR peak filter to increase the main frequency of the respiratory signal.(2)The peak point of the pulse signal is detected by the adaptive differential threshold method.The pulse rising branch is enhanced by the SSF function,the pulse starting point is detected by the differential threshold algorithm,and the pulse morphology time series is constructed based on the characteristic points.Using the integrated empirical mode decomposition to obtain the local modal function in different frequency bands,the lower frequency and smaller amplitude components are removed,and the component with higher similarity to the pulse signal is retained.By calculating the time and frequency domain features of the above time series,a total of 84 pulse characteristics depicting different emotional states were obtained.(3)Multi-scale entropy algorithm is used to calculate the complexity of respiratory signals under different emotional states,and the relationship between resampling frequency and scale factors is explored,and different multiscale entropy algorithms using different threshold factors are determined through experimental comparison.Combined with the temporal and frequency domain features of respiratory signals,a total of 57 respiratory characteristics characterizing different emotional states were obtained.(4)The use of ReliefF algorithm to evaluate the pulse and respiration characteristics can be used to positively influence emotion classification.Based on the optimal feature subsets that are positively related to emotions,ten-fold cross validation and grid optimization are used to determine random forest algorithm parameters,and one-to-one and one-to-many emotion recognition models are established.Among them,the highest recognition rate of the one-on-one emotion recognition model can reach 84%,and the recognition rate of the multiple-to-multiple emotion recognition model based on pulse and breath combination features is 72%.This result indicates that the features extracted in this paper can be used for emotion recognition.(5)Based on the established optimal emotion recognition model,the characteristics of each type of model are analyzed,and the five most important features for emotional recognition modeling are selected using the Gini coefficient gain as an indicator.The results show that the pulse waveform factor and frequency domain characteristics The maximum value of breath and the ratio of mean and multi-scale entropy are good for emotion classification.At the same time,when the pulse and respiratory signals are used together for emotion recognition,the respiratory characteristics contribute more than the pulse characteristics. |
