Research On Stress Recognition Based On Self-Supervised Learning And Unsupervised Domain Adaptation

In the modern living environment where work and life stress are increasing day by day,the problems caused by psychological stress to people’s physical and mental health are becoming more and more obvious.Long-term stress can even cause depression.Therefore,the accurate recognition and evaluation of people’s stress is particularly important.Stress recognition based on physiological data is reliable and effective,and it is not affected by subjective factors.Now it has received extensive attention from researchers at home and abroad.However,the existing research on stress recognition has many problems:(1)The hand-crafted features are limited,and its ability to represent stress is insufficient;(2)The physiological data collected by different stress-induced paradigms are different,and when the stress recognition model established by the data collected under a specific paradigm is used in other paradigms,the generalization ability is not strong,and it is difficult to apply in real life to cope with stress diversity.Based on this,this paper has carried out relevant research work in the field of stress recognition based on photoplethysmography(PPG),and carried out the following work:(1)Research and compare existing stress-inducing experimental paradigms,choosing Stroop test,stress recall experiment and social stress test as psychological stress inducing methods,designing a physiological data collection process,and recruiting 73 subjects to establish a multi-paradigm stress dataset based on PPG signals.(2)Aiming at the insufficiency of traditional manual features in PPG signals for stress representation,this paper introduces self-supervised learning,designs self-supervised pretexttasks according to the structure and characteristics of the data,and performs several signal transformations on the original samples to bring supervision information to enable the convolutional neural network to learn how to extract features through unlabeled data.Extracted self-supervised deep features can be effectively used for stress recognition.Compared with traditional features,it has more advantages and a more robust stress classification ability.These two features can be fused to make up for the shortcomings of traditional features and bring better stress recognition performance.(3)Aiming at the problem of insufficient generalization ability in cross-paradigm stress prediction by the stress recognition model established by the data collected under a specific paradigm,this paper introduces a deep unsupervised domain adaptive method,and proposes a method based on Maximum Mean Discrepancy domain adaptation and improved class-aware domain adaptation model.The stress recognition model improves the generalization ability under the new unlabeled paradigm by reducing the difference between different experimental paradigms.Experimental results show that the stress recognition model based on unsupervised domain adaptation shows superior recognition performance in cross-paradigm tasks,with an average increase of 3.76% in accuracy and an average increase of 9.41% in recall.(4)Finally,this paper builds an online PPG signal stress recognition system with functions such as physiological data collection,signal visualization,stress state recognition,and stress recognition result display,which can provide references for future application research.In summary,this paper establishes a physiological signal stress dataset,and conducts certain explorations in feature extraction.At the same time,it provides new ideas for solving cross-paradigm stress prediction,which is conducive to the application of stress recognition technology in real life.