【作者】 李娜；

【导师】 侯义斌；

【作者基本信息】 北京工业大学 ， 计算机应用技术， 2013， 博士

【摘要】 可穿戴健康监测系统是可穿戴计算在医疗领域的典型应用，它将改变我国远程医疗和家庭保健医疗中终端用户传统的“被动”监测模式，实现低生理和心理负荷下人体生理信号自动、连续、动态地获取。国内外学者已在该研究领域做了大量的工作，然而现有的研究往往没有考虑实际应用中人体生理特征和运动状态相关联的特点，仅仅从生理数据就对用户的健康情况作出判断，缺乏当时的运动状态信息，造成一定程度的误判。因此将两者有效结合，研究基于人体运动状态识别的可穿戴健康监测系统具有重要的现实意义。本文依据用户活动自由的需要，设计可穿戴健康监测马甲以获取人体生理特征值和运动参数，并在运动状态实时识别的基础上对生理状态进行诊断，以提高日常运动环境下个人健康监测的准确性。论文主要从系统架构、人体运动状态识别、跌倒动作识别、系统能量管理策略这四个方面进行深入的研究。主要创新性研究成果如下：（1）提出基于人体运动状态识别的可穿戴健康监测系统架构。针对多种类型设备和多种传输方式共存的情况，建立基于Agent的系统架构模型，并对其通信协议、交互方式进行描述和定义。该架构独立于具体的硬件单元，便于系统的扩展和相关软件的配置和部署。（2）提出基于单个三轴加速度传感器的人体运动状态识别算法。根据人体日常活动具有短时持续性的特征，将其运动状态划分为稳定状态和非稳定状态。将三轴加速度矢量值转换为加速度幅值变化量，消除了传感器坐标系的佩戴相关性，使用卡尔曼滤波实时识别出稳定或非稳定状态；同时采用自适应阈值法对稳定状态时的跑步、走路动作进行识别。实验结果表明：该算法在状态识别方面达到较高的准确率；对于跑步、走路动作的识别准确率优于决策树识别算法。（3）提出一种基于单个三轴加速度传感器的危险性跌倒动作（相对一段时间不能自我恢复活动）识别算法。通过提取超重强度、持续失重时间、倾斜角度、持续静止时间为特征值，消除传感器坐标系的佩戴相关性并减少计算复杂性；根据运动状态动态调整传感器的重力参照值，提高测量准确度。实验结果表明：该算法的识别准确率优于一阶支持向量机算法和基于多轴向阈值的识别算法。（4）提出一种基于事件驱动的系统能量管理策略。为了解决移动终端持续感知情况下系统的能耗问题，建立事件模型：将人体正常情况下持续静止状态作为系统的休眠触发事件；将人体运动状态转换、生理特征异常作为系统的唤醒事件，采用事件触发机制，根据人体自身状态来自适应地调整系统工作周期。实验结果表明，该方法在保持系统实时性和准确性前提下，与不采用能量管理策略相比，节省约25%的电量。本文以人体运动状态识别为基础，结合可穿戴技术、信号处理技术、无线通信技术，使用加速度传感器、生物传感器、蓝牙模块、智能手机和后台服务器搭建“基于人体运动状态识别的可穿戴健康监测系统”。通过对志愿者的日常穿戴测试，系统在实现人体运动状态实时识别的基础上，对危险性跌倒动作和不同运动状态下的生理信号异常发出报警，并能根据用户状态采用能量管理策略节约系统能量，证明了基于人体运动状态识别的可穿戴健康监测架构的可靠性。更多还原

【Abstract】 Wearable health-monitoring system is a typical application of wearablecomputing in healthcare field. It will change traditional "passive" usage mode oftelemedicine system and family healthcare in our nation, by providing continualmonitoring of physiological signals of end user with slight mental and physicalburden. International and local researchers have done much work in this field, butcurrent researches usually determine health status by only physiological signals,which ingore the relation between human physiological characteristics and activities.Lack of activity information at that time may causes misdiagnosis. Therefore, witheffective combination, it will show more practical significance to study wearablehealth-monitoring system based on human activity recognition.Meeting the requirement of user to move freely, this dissertation designed awaistcoat for wearable health-monitoring to get physiological and movement signals,and then determined health status based on recognition of activities in real time. Thismethod improves precise of human health monitoring in daily life. In thisdissertation, four aspects will be researched: system architecture, human activityrecognition, falling detection, and power management strategy for system. The majorcontributions of this dissertation are stated as follows:(1) Provide a general architecture of wearable health-monitoring system basedon human activity recognition. To address the issues with different type of devicesand communication methods, we construct system architecture based on agent model,and define its communication protocol, interaction processes. This architecture isindependent of hardware units, which makes the system more scalable and therelated softwares easily be deployed.(2) Propose an algorithm to classify activity states of human with singleaccelerometer. According to the constancy feature of daily movements in short time,activity states are divided into steady and unsteady ones. We transform raw datameasured from the three axis into changes of signal vector magnitude to avoiddependence on wearing coordinate, and apply Kalman filter to classify the above twostates in real time.Meanwhile, we use thresholds which are automatically adapted fordifferent users to recognize activities of running and walking when they are onsteady state. Experiment results showed that the algorithm got better performance inaccuracy of activity recognition. It performed higher accuracy for running and walking activities than the decision tree algorithm.(3) Propose an algorithm to recognize dangerous fall of human body with singleaccelerometer, where “dangerous fall” means subject could not return to his/hernormal behaviors after impacting on the ground. Features of overweight, continuousweightless time, tilting angle and continuous still time are abstracted, which are allindependent of the sensor orientation with respect to the body, and simplifycomplexity of computing. To improve accuracy of measurement, the referencedgravity output value will be adapted with activity states. Experimental resultsshowed higher accuracy than one-class SVM algorithm and the algorithms based onmulti-axial directions.(4) State the event-driven strategy for power management of system. In order toreduce energy consumption on mobile device when continuous sensing, an eventmodel is built which regards continually being still on healthy situation as sleepevent, activity state transitions and abnormal physical signals as waken event.Duration of the system working cycle can be adapted automatically according to thestate of subjects. Experimental results demonstrated that, keeping performance inreal time and accuracy, the system could save25%energy than that without thispower management strategy.This dissertation combines physiological monitoring with human activityrecognition to construct "wearable health-monitoring system based on recognition ofhuman activity state (WHMSHAR)", which applies technology of wearablecomputing, signal processing, wireless communication with accelerometer, physicalsensors, Bluetooth, and runs on smart phone and server. Worn by volunteers in dailylife, the tested system can successfully send out alert in case of dangerous fall eventsand abnormal physiological signals in different activity states, which is evidence ofthe reliability of wearable health-monitoring architecture based on human activityrecognition.更多还原