A Study On Wearable Emotional State Monitoring Technology Based On Cross-Mutation Point TinyML

The COVID-19 pandemic has led to increased attention towards physical and psychological health conditions,with anxiety and depression being on the rise.Emotions are important indicators of physical and mental health,as they are psychological and physiological states that accompany consciousness and cognitive processes.However,accurately and objectively understanding emotions and assessing mood changes can be challenging.Wearable devices,when combined with Tiny Machine Learning(TinyML)technology,can continuously and accurately monitor individuals’ emotional and mental health in real-time based on various physiological signals.In this paper,we propose a wearable monitoring technology for sudden changes in emotional states based on these two technologies.To build on the wearable monitoring technology proposed earlier,we further investigate the method for detecting multidimensional psychological mutation points.The proposed method defines emotion as a relationship between an individual and their environment,with emotional states being a transition from one state to another.Accurately capturing this transfer is crucial for determining an individual’s positive or negative emotional state.To achieve this,we construct a vector representing emotional states by extracting high-dimensional features from multimodal physiological signals,mapping those features to a low-dimensional space,and applying the MK mutation test to detect and localize sudden changes in emotional states.The results demonstrate that the proposed method is capable of detecting and locating emotion mutations with greater accuracy,with a minimum deviation from mutation detection of only 1 second and a maximum deviation of 7 seconds.Furthermore,we also investigate the TinyML classification method for crossmutation point emotional states in this thesis.To overcome the limitations of traditional methods in emotion classification tasks,such as high computational overhead and inability to monitor individual emotion states in real-time,an emotion classification decision tree network based on embedded devices is established.Unlike most emotion elicitation experiments,our study does not use audiovisual materials to induce emotional state changes.Instead,we use dating social scenes to induce negative emotions in a positive context,allowing subjects to produce emotional state changes naturally.Experimental results demonstrate that the emotional state mutation detection and classification models established in this thesis achieved 84% detection rate and 90.54%classification accuracy in the negative emotion mutation detection task,respectively.