情绪识别是指采用无生命的传感器和计算机感知测量识别人类情绪状态,其主要环节包括情绪相关信号获取、特征提取以及分类识别.情绪识别可为人类情绪健康监测乃至情绪相关心理精神疾病的初筛提供科学依据.该文构建了多模可穿戴生物信息传感网络测量被测个体的多模情绪相关信号(脑电、脉搏以及血压),经由身体主站将信号传输至远程网络数据中心,并将情绪识别的结果进行网络发布,简化了测量结构,使得被测个体日常情绪监测和远程监控成为可能.由于信号测量和特征提取过程中存在不确定性,该文提出了栈式自编码器(基于深度学习理论)优化的情绪识别算法.71天时间跨度的实验结果表明,栈式自编码器预学习后的特征向量具有更高的一致性与可分性,情绪识别率较相关研究提高了约5％.%Emotional health draws great concern with the enhancement of public health consciousness.Emotional health is closely related to the quality of personal life.Even for some special groups of people,like pilots,soldiers,etc.,their emotional states will have impacts on the stability of communities.Traditionally,to evaluate emotional states of human beings relies on the doctors or psychologists to communicate with subjects and give scores based on various questionnaires.This approach is not scientific enough and leads to the difficulties in the emotional health monitoring in daily-life.Emotion recognition enables lifeless sensors and computers to measure and interpret human emotions.It is a procedure consisting of emotion-related bio signals recording,features extraction and emotional states classification,providing scientific evidence for emotional health monitoring and primary diagnosis of potential mental diseases.In the related works concerning emotion recognition,the application scenarios are usually restricted in the hospitals or labs and the common-used classifiers are not suitable for the daily emotion recognition data set.This paper develops a multimodal biosensor network to simplify the sensing framework so that it can finish emotion recognition tasks when subjects are participating in daily tasks without so many disturbances.Several wearable biosensor nodes record multimodal bio signals (electroencephalography,pulse and blood pressure) and transmit them to a body station employing Arduino UNO3 and its expansion boards.The body station works as a web client connecting to a web data center on the Internet by wireless routers or personal hotspots.The web data center is established on NI-PXI 1065 with a static public IP address.The recognition algorithm is carried out in the data center and the results are displayed for authorized web terminals with the assistance of web publishing service supported by LabVIEW.The multimodal wearable biosensor network can provide emotion-related bio signals from which typical features are extracted based on the existing theories.In particular,due to the uncertainties in signal acquisition and feature extraction,a stacked auto-encoder (based on the deep learning theory) optimized emotion recognition method is proposed to improve the recognition process.The stacked auto-encoder helps to pre-learn the feature vectors and with the fine tuning it generates a better scheme for emotion classification phase.There are 9 emotional states for classification according to the Valence-Arousal dimensional model.A two-layer stacked neural network with a softmax classifier is designed to finish the final classification tasks.The experiment convinces that the feature vectors pre-learned by stacked auto-encoder are of higher quality both in centrality and distinguishability based on the similarity evaluation theory.The final recognition rate is also improved approximately 5％ compared to related works.The main contributions of this paper are the wearable network-based sensing structure,the stacked auto-encoder optimized multimodal emotion recognition algorithm and the quantitative analysis on 71-day experimental data.This is a novel system for daily emotional health monitoring and can provide scientific suggestions for doctors or guardians.However,in the future,large scale of data should be accumulated.Moreover,the dynamic performance and energy efficiency also need improving.
展开▼
