Stress Detection Of Pulse Signal Based On Deep Learning

Among many emotions,stress is a common one in daily life.Psychological research shows that in a variety of emotions,long-term stress can affect people's psychological and physical health,and even lead to depression.Therefore,it is necessary to study the real-time and accurate stress detection algorithm.Because of the difficulties in experimental design and the lack of data sets,there are few studies on stress detection based on physiological signals.At the same time,there are several problems in the existing research: the feature engineering is relatively single,and the time and waveform characteristics of physiological signals are not fully utilized;the recognition algorithms are mostly traditional shallow machine learning algorithms,which have simple mathematical form and are difficult to extract signal information effectively;and the model generalization ability is not strong due to the differences between individual physiological signals.For this reason,based on the blood volume pulse(BVP)signal,this paper does some research on the identification of the pressure-free state.1.Stress detection algorithm for BVP signal.In this paper,in addition to extracting the common statistical features,we use the knowledge of nonlinear dynamics and wavelet transform to build a feature engineering including time domain,frequency domain and waveform characteristics.Based on the new feature engineering,a feature map with time-frequency characteristics is established by using the time correlation of physiological signals.In order to make full use of the time-frequency characteristics of the two-dimensional feature map,this paper introduces the 2D convolution neural network,and combines the time-frequency information contained in the signal into the multi-layer features of the neural network,which effectively improves the recognition ability of the model.The experiment results show that the convolutional neural network can make full use of both time and frequency information of signals and improve the recognition accuracy.Among them,experiments based on WESAD achieved 73.92% and 86.72% accuracy in three and two classification respectively.2.Adversarial analysis to eliminate individual differences.Relevant studies have shown that the pulse signal contains information unique to the individual.Therefore,in order to improve the generalization of the model,it is necessary for the network to learn individual-independent features.Referring to the idea of adversarial system,this paper adds an adversarial module to combat individual differences.The results on the WESAD and SCUT datasets show that the antagonistic network effectively confuses identity differences,achieves 80.11% accuracy of three classification,and reduces the standard deviation of the model's recognition accuracy on each subject.The algorithms and experiments presented in this paper provide ideas for the study of realtime pressure identification.In addition,the research and Analysis on eliminating individual physiological signal differences can also provide reference for the design of pressure recognition models with generalization.