基于ANN识别施工人员跌落险兆事故的研究  

Approach to detecting and identifying the near-miss falls of the construction workers via the ANN system

作　　者：张明媛[1] 曹天卓 赵雪峰[2] ZHANG Ming-yuan;CAO Tian-zhuo;ZHAO Xue-feng(Department of Construction Management,Dalian University of Technology,Dalian 116024,Liaoning,China;School of Civil Engineering,Dalian University of Technology,Dalian 116024,Liaoning,China)

机构地区：[1]大连理工大学建设管理系,辽宁大连116024 [2]大连理工大学土木工程学院,辽宁大连116024

出　　处：《安全与环境学报》2018年第5期1703-1710,共8页Journal of Safety and Environment

基　　金：十三五重点研究计划项目(2016YFC0802002-03)

摘　　要：建筑施工现场的跌落事故可视为跌落险兆事故的进一步演化,为深入认知跌落事故发生机理和主动预防跌落事故发生,需对各种环境条件下发生的跌落险兆事故进行探究。以三轴加速度和三轴角度信息描述施工人员不同运动状态的可能性,探究基于智能手机和ANN（人工神经网络）识别施工人员跌落险兆事故的应用潜力。鉴于跌落险兆样本的传统获取方式往往低效且易受人为主观因素的干扰,以平衡板模拟跌落险兆发生时的失、复稳状态,构建静、动态跌落险兆转换模型,为ANN（人工神经网络）的训练提供充足、可信的样本。结果表明,ANN识别跌落险兆事故的准确率高达90. 50%。同时,获取阈值H20%和H10%,以定量的方式描述跌落险兆事故与跌落事故间的相关性,并从概率角度解释其应用意义,可为提升建筑施工安全管理水平提供新的视角。This paper is aimed at making an exploration of the ways to detect and identify the workers＇ near-miss fail risks by utilizing the smartphone as a data acquisition tool based on ANN （ Artificial Neural Network） system. To achieve the purpose, it is necessary to describe the workers＇ different motion status-in-situ in business with the help of the accelerometers and gyroscopes embedded in the smartphones for demonstration. The feasibility of adopting three-axis acceleration and angle data. What is more, since the traditionally sampling acquisition method of near-miss falls proves to be often inefficient and easily affected by the subjective factors and lead to the environmental loss-of-balance （LOB） , we have managed to create and cultivate evaluation experiments and simulate the static near-miss fall scenarios by establishing artificially a balance board in our simulation practice. Thus, based on the large-scale authentic sampling analysis, it has made us possible to identify the. universal difference in the accel- eration between the static and dynamic near-miss falls. And, then, we have managed to calculate and work out the average data of the three-axis acceleration produced by taking the walking situation of the various subjects as the compensation data so as to realize the transition between the static and dynamic near-miss falls. Thus, the mean absolute difference in the acceleration features between the compensated static and dynamic near-miss falls of all the subjects should be kept within 13% , which is helpful to improve the similarity between the simulated and the actual scenes of near-miss falls. At the same time, the above introduced method can provide abundant training samples in an efficient and low-cost way without the interference of the subjective factors. The results of our evaluation demonstrate that the recognition ac- curacy can be made as high as up to 90. 50%. And, in case of checking the testing samples, it would be possible to build up the number of near-miss falls versus the

关 键 词：安全工程 跌落险兆 动作识别 智能手机 机器学习 人工神经网络 

分 类 号：X947[环境科学与工程—安全科学]