老人跌倒预测系统的研究

发布时间：2018-03-05 10:09

本文选题：跌倒预测　切入点：支持向量机　出处：《苏州大学》2016年硕士论文　论文类型：学位论文

【摘要】：当今世界面临着严重的人口老龄化问题,老人由于身体机能下降、平衡协调能力减弱和视力变差等生理原因,容易发生跌倒。跌倒预测系统可以及时检测跌倒并进行报警,可以减少老人跌倒后等待救助的时间,降低跌倒对人体造成的伤害,减少因跌倒产生的医疗开支,增强老年人独立生活的信心。本文在对当前跌倒检测方法进行总结分析的基础上,采用基于穿戴式设备的方法进行跌倒预测研究。课题的创新点在于结合加速度特征和姿态角特征,同时采用机器学习算法中的支持向量机算法作为分类算法研究跌倒预测,即在人体跌倒碰撞地面前进行跌倒的预判。本文将跌倒预测问题作为一个二分类问题来处理,建立跌倒预测分类模型,实现了跌倒预测和跌倒远程报警,并对跌倒保护装置进行了探讨。本文首先综述了不同的跌倒预测算法,确定支持向量机算法作为分类算法,阐述了支持向量机的理论原理,说明了实现支持向量机算法的过程及参数寻优方法。然后将传感器设备穿戴在人体腰部,建立人体三维坐标系,采集3轴加速度、3轴角速度和3轴磁场数据。定义了4种跌倒行为和10种日常活动行为以及加速度和角速度统计量,分析不同行为过程的加速度特征、角速度特征和姿态角特征,发现采用这些特征进行跌倒预测具有很强的可行性。在特征分析的基础上,进行特征提取,根据加速度和角速度统计量的变化规律,在时域中提取4个加速度特征和4个角速度特征,在捷联惯导系统中解算姿态角,提取姿态角特征,将提取的9个特征组成特征向量。接着对提取的9个特征进行特征选择,采集跌倒行为样本和日常活动行为样本组成训练样本集和测试样本集,对样本采集过程中的变量进行讨论,分别采用序列前向选择方法和序列后向选择方法进行特征选择,综合两种方法得到最优特征组合,最优特征组合包含3个特征,最终得到最优跌倒预测分类模型,最优跌倒预测分类模型对应的跌倒行为检测率为100%,日常活动行为检测率为100%,平均前置时间为291ms,证明了本文提出的跌倒预测算法的有效性和可行性。最后本文对跌倒远程报警及跌倒保护装置进行了研究,采用SIM900A模块实现跌倒远程短信报警,将一款手动充气气囊装置改造成自动充气气囊装置,设计了2种不同的连杆机构连接舵机和气囊装置,并通过实验选择实际耗时较少的连杆机构,发现从检测到跌倒到打开压缩气瓶最少耗时约为133ms,在目前的最优跌倒预测分类模型下,留给气囊的充气时间只有约158ms,而在现有的压缩气瓶条件下,气体完全释放所需的时间约为452ms,所以计算了在满足158ms充气时间条件下的气瓶规格,然后在现有气瓶条件下,通过实验对跌倒保护装置进行了整体测试,实验结果表明跌倒保护装置可以在跌倒碰撞地面前启动,具有一定的可行性。
[Abstract]:The world is facing the serious problem of population aging, the elderly due to the decline in physical function, balance ability weakened and poor eyesight physiological reasons, prone to fall. Fall prediction system can detect falls and alarm, can reduce falls among the elderly wait for the rescue time, falls the harm to human body, reduce the by the fall of the medical expenses, enhance the elderly independent living confidence. Based on the analysis on the current fall detection methods, the prediction method based on the fall of wearable devices. The innovation point is to combine the characteristics of acceleration and attitude angle characteristics, at the same time using machine learning algorithm of support vector machine algorithm the prediction falls as the classification algorithm research, which falls in the body hit the ground before the fall of the pre judgment. This paper will fall as the forecasting problem A two classification problem and establish classification model can fall, fall fall prediction and remote alarm, and the fall protection device are discussed. This paper reviews the different prediction algorithms fall, determine the support vector machine algorithm as the classification algorithm, and expounds the principle of support vector machine, description of the process and the parameters of support vector machine algorithm optimization method. Then the sensor device worn on the waist, the establishment of human 3D coordinates, collecting 3 axis acceleration, 3 axis and 3 axis angular velocity field data. It defines 4 kinds of fall behavior and 10 kinds of daily activities and the acceleration and angular velocity acceleration statistics analysis the characteristics of different behavior process, angular velocity and attitude angle characteristic features, these features are found by the fall prediction is feasible. Based on the analysis, the feature extraction Take, according to the variation of acceleration and angular velocity statistics, 4 acceleration characteristics and 4 angular velocity feature extraction in time domain, in the strapdown inertial navigation system for the calculation of attitude angle, attitude angle feature extraction, 9 feature extraction feature vector. Then the extracted 9 features for feature selection fall, acquisition behavior sample and daily activities behavior of samples consisting of training samples and the test samples, the sample collection process variables are discussed, using the sequential forward selection method and the sequential backward selection method for feature selection respectively, two methods to obtain the optimal combination of features, the best combination of features contains 3 characteristics, finally the optimal fall prediction classification model, optimal prediction fall fall behavior classification models corresponding to the detection rate was 100%, the daily activity detection rate was 100%, the average lead time is 291ms, that the The effectiveness of the proposed fall prediction algorithm and feasibility. Finally we study the fall and fall remote alarm protection device, using SIM900A module to realize the remote alarm message will fall, a manual air bag device into automatic air bag device, designed 2 different linkage mechanism connected with the steering engine and the airbag device, and the actual linkage is selected through experiments with less time-consuming, found from detection to fall to open the compressed gas cylinders least time is about 133ms, at present the optimal classification model to predict falls, balloon inflation time is only about 158ms, while in the condition of compressed gas cylinders, gas release time required is about 452ms. So the calculation to meet the specifications of the cylinders 158ms inflation time conditions, and then in the existing cylinder condition, through the whole test of the fall protection device The experimental results show that the fall protection device can start before the fall and collide on the ground, and it is feasible.

【学位授予单位】：苏州大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TH789

【相似文献】