Research On Mental Stress Detection Method Based On Physiological Signals

Physiological signal based mental stress detection can achieve objective,real-time and quantitative recognition of the mental stress condition of human body.Specifically,for medics that are executing the emergency medic rescue missions,it is essential to monitor their mental stress condition and conduct early warning and intervention of high stress,since this can avoid operational errors induced by the decline of judgment caused by excessive mental stress,which is of great significance and value to ensure the smooth progression of the rescue mission.However,existing mental stress recognition algorithm still has the following three problems:(1)The boundary of mental stress is blurred,which leads to the inability of accurate recognition;(2)The individual differences in physiological signals leads to the decrease in the accuracy of cross-individual recognition;(3)The inaccessibility of existing user data makes cross-individual identification difficult.To tackle the aforementioned problems,based on the mental stress detection of medics under emergency medical rescue scenario,the main work of the thesis is listed as follows:1.To tackle the high resemblances among the signals of different stress levels which results in the difficulty in recognition,the thesis collects the physiological data from medics who participate in a simulated emergency medical rescue training program,and build a dataset named the Mental Stress in Emergency Medical Rescue(MSEMR).To effectively distinguish between singals of different stress levels,the thesis proposes a deep learning model named Medics’ ECG-based Mental Stress Detection(MEMSD).The model introduces SENet to learn the feature representations of signals as well as integrates a center loss to optimize the decision boundaries of different stress levels.Experiments on MSEMR dataset demonstrates a recognition accuracy of 82.67% on different mental stress levels.2.To address the subject variability,existing research can not detect the ambiguous target samples located near the decision boundaries of different stress levels,which leads to the decline in recognition accuracy.We therefore propose a Discriminative Clustering Enhanced Adversarial Domain Adaptation(ADA-DC)algorithm which introduces pseudo labels as the additional supervision through weighted-clustering on the extracted features.The pseudo labels are refined iteratively,forcing the ambiguous samples to be correctly classified.ADA-DC achieves an increasing in the recognition accuracies of3.38%,1.79%,1.77% and 1.89% on three public datasets and the MSEMR dataset respectively.3.Aiming at the problem that existing domain adaptation based cross-subject methods are unable to achieve knowledge transfer when they do not have access to the data of source subjects due to the need of privacy protection,we introduce the source-free domain adaptation technique and propose an algorithm named Alleviating Label Shift for Sourcefree Domain Adaptation(ALS-SFDA).By masking and rebuilding the input signals,ALSSFDA can learn feature representations that are insensitive to the label information,thus increasing the cross-subject generalization ability.At the same time,the pseudo labels are refined according to the output confidence of the classifier to yield the final prediction.Experimental results on three public datasets and the MSEMR dataset demonstrate a performance improvement of 0.89%,2.19%,2.51% and 1.96% respectively.