Research On Psychological Stress Recognition Algorithm Based On Deep Learning

People live in an environment full of stress and anxiety for a long time,which lead to physical illness,as well as affect cognitive function and mental health.Therefore,it is especially important to identify the stress and decompress it.Traditional stress measurement methods mainly measure the subjective feeling of pressure through the stress perception scale,which ignores the objective physiological indicators.Electroencephalography(EEG),as an objective physiological data,has a strong association with different psychological and physiological conditions.The method combining traditional feature extraction algorithm with machine learning requires a great deal of prior knowledge.Deep learning does not require prior knowledge to dig deep into the deep features of data.Therefore,this thesis combines the physiological data obtained by electroencephalography and the subjective data obtained by the pressure perception scale to identify and analyze stress.The main contents of this thesis include:(1)In order to induce stress,identify and decompress stress,a stress-induced experiment based on a mental arithmetic task is designed.In the experiment,EEG signals are collected from 23 undergraduate and graduate studentsby non-invasive EEG acquisition equipment.The experiment is divided into four parts,namely the resting state task,the mental arithmetic task without music,the mental arithmetic task under the condition of relaxation music,and the mental arithmetic task under the condition of natural rhythm music.After completing each task,a stress perception scale test is performed.The collected EEG signals are preprocessed by Independent Component Analysis(ICA)to remove noise and obtain pure EEG signals.The power spectrum analysis of the EEG signal is performed by the Welch method,and the Frontal Alpha Asymmetry(FAA)index is calculated.(2)In order to identify pressure,a Transformer-based stress EEG signal classification model is proposed.The Transformer model in deep learning is explored,and the encoder module in the Transformer model is applied to the analysis of EEG signals.Adaptive improvements are made and parameters are optimized for EEG signal analysis.Experimental results show that the binary classification of EEG signals with and without stress achieves an accuracy of 93.78%.In order to verify the validity of the model on other data sets,two EEG signals public data sets are tested,and the binary classification accuracy reaches 84.36% and 78.87%,respectively.(3)In order to further improve the accuracy of pressure recognition,a stress EEG signal recognition model based on the self-attention mechanism is proposed.The self-attention mechanism and the EEGNet model in deep learning are explored,and improvements are made to the EEGNet model.The self-attention mechanism is introduced in the EEGNet model,the self-attention module is added,and the parameters are optimized.In the experimental data,the binary classification of EEG signals with and without stress achieves an accuracy rate of 97.69%,which is 1.46% higher compared with the EEGNet model.Finally,mean analysis is carried out on behavioral data such as the average answering time and answering accuracy of the subjects in the mental arithmetic task,and a paired T test is performed on the stress perception scale.