基于施工场景数据的地铁工人行为安全研究

郭圣煜

摘要

中国地铁高速度、高密度和高强度的建设任务，加上地铁施工工人流动性强、受教育程度普遍不高，接受专业技能培训不足的现状，使得因工人的不安全行为而导致的施工事故时有发生，工人的行为安全已经成为地铁建设中重要的研究课题。本文以行为安全（Behavior-Based Safety, BBS）方法为基础，采用大数据时代背景下的图像识别技术，存储技术和个性化推荐技术，结合关联规则、聚类分析等统计分析方法，观察、分析和矫正工人的不安全行为，深入了解工人的行为安全模式，探究工人不安全行为规律，改善工人的不安全行为。具体研究内容如下：

1、地铁施工工人不安全行为场景观察方法研究。从地铁施工安全标准、操作规程、事故案例和专家经验中提炼工人常见的不安全行为，再结合地铁工程划分标准列出的施工工作分解结构（Work Breakdown Structure, WBS），统一分类编码形成行为风险知识库。实时动态采集施工现场监控视频、现场照片等施工现场场景图像数据，一方面，利用摄像头行为分析技术矢量解析视频序列中反映知识库内的部分工人不安全行为场景；另一方面，通过工作危害性分析（Job Hazard Analysis, JHA）找出工人不安全行为的影响因素，分别利用向量空间模型（Vector Space Model, VSM）和多层次融合语句相似度算法矢量解析现场照片中反映的工人不安全行为场景，提取语义信息，匹配知识库内对应条目。从这两方面保证现场工人不安全行为场景全样本数据采集，实现自动观察记录，然后利用分布式文件管理系统（Hadoop Distributed File System, HDFS）在基于Hadoop的大数据云平台中存储数据；

2、地铁施工工人不安全行为场景数据规律研究。以获取的大量工人不安全行为场景数据为基础，基于人类动力学的研究方法，统计地铁施工工人出现不安全行为的时间间隔，分析不安全行为的时间统计特性，包括不安全行为时间间隔分布规律，以及不安全行为的阵发性与记忆性。并进一步构建关联规则挖掘数据库，通过Apriori算法挖掘不安全行为、工种岗位和施工阶段的关联关系，得到各个工种岗位的工人在各个施工阶段所出现不同不安全行为的频率，由此构建地铁施工工人不安全行为知识地图；

3、地铁施工工人个性化行为矫正模型构建研究。收集地铁施工工人特征信息，包括个人基本信息（工种岗位）和工人动态作业信息（施工任务），依据不安全行为场景数据统计规律，通过数据驱动的方式研究工人行为矫正内容个性化混合推荐机制，包括基于内容的个性化推荐和基于MapReduce的协同过滤个性化推荐，建立工人行为矫正内容个性化精准推送模型。然后，开发地铁施工工人个性化行为矫正系统，以工人不安全行为场景图像数据作为推送对象，针对不同工人特点实施个性化行为矫正；

4、地铁施工工人行为矫正效果评价研究。通过问卷调查、统计分析和实地观察等方法获取工人行为矫正的结果数据，从工人参与行为矫正满意度、工人参与行为矫正成绩分析和工人现场行为安全表现三个方面，评价地铁施工工人个性化行为矫正效果。满意度采用调查问卷的方式分析，从矫正系统功能体验、矫正内容设计和矫正效果体验三个方面设计多个问卷问题进行调研；工人参与行为矫正成绩通过系统统计，采用对比和聚类两种方式进行分析；工人现场行为安全通过现场实地观察，采取前后对比和有无对比分析方法，以不安全行为率（S）作为指标，分析地铁施工工人参与行为矫正前后，已参与和未参与行为矫正的现场行为安全表现。

关键词：行为安全；场景数据；行为观察；行为规律；行为矫正；大数据

Abstract

With the high speed, high density and highstrength of Chinese metro construction, also the situation that the metroconstruction construction worker’s strong mobility, low education level and thelack of professional skill training. So accidents frequently happen because ofworker’s unsafe behavior. Worker’s behavioral safety has been an importantresearch subject on metro construction. Based on Behavior-Based Safety (BBS)theory, image recognition, storage and personalized recommendation technologiesare used under the background of Big-Data era, combined with statisticalanalysis methods such as association rule and cluster analysis. So worker’sunsafe behavior can be observed, analysed and modified, worker’s behavioralpattern can be deeply understood, worker’s unsafe behavioral law can beexplored, and finally worker’s unsafe behavior can be changed. The specificresearch contents as follow:

(1) Research on metro construction worker’sunsafe behavioral scenario observation method. Firstly, worker’s criticalunsafe acts are extracted from safety standards, operating instructions,accident cases and expert experience. Combined with a Work Breakdown Structure(WBS) classification standards, worker’s unsafe acts are classified and encodedto form the behavioral risk knowledge base. Secondly, on-site constructionscenario data as monitoring videos and photographs is timely and dynamicallycollected. On the one hand, scenario from monitoring videos reflecting worker’sunsafe acts is vectorially analyzed by camera-based behavior analysistechnology. On the other hand, influence factors of worker’s unsafe behavior isfound by Job Hazard Analysis (JHA) method, the scenario from photographsreflecting worker’s unsafe acts is vectorially analyzed respectively by VectorSpace Model (VSM) and multi-level fusion sentence similarity method. Semanticinformation is then extracted from the scenario to match corresponding unsafebehavior in knowledge base. Finally, data is stored in Big-Data cloud platformbased on Hadoop by Hadoop Distributed File System (HDFS).

(2) Research on metro construction worker’sunsafe behavior scenario data statistical law. Based on obtained much worker’sunsafe behavior scenario data, interevent time statistics of worker’s unsafebehavior are counted on Human Dynamic method. To analyze tatistical propertiesof time, including the interevent time distribution law, also the aroxysmal andmemory of worker’s unsafe behavior. Then, a mining database of association Ruleis built, from which relationship between unsafe behavior, worker type andconstruction phase is found by Apriority algorithm, to build the knowledge mapof worker’s unsafe behavior.

(3) Research on metro construction worker’s personalizedbehavior modification model. Firstly, metro construction worker’scharacteristic information is collected, including personal information (workertype) and dynamic work information (construction ask). According to unsafebehavior scenario data statistical law, personalized ix-recommendationmechanism of worker’s behavior modification content is explored by data-drivenmethod, including content-based recommendation and Map Reduce-based collaborativefiltering recommendation, so as to build personalized precision pushing modelfor worker’s behavior modification content. Then, metro construction worker’s personalizedbehavior modification system is designed, worker’s unsafe behavior scenario datais pushed according to different worker’s characteristic for personalizedbehavior modification.

(4) Research on metro construction worker’sbehavior modification effect evaluation. The result of worker’s behaviormodification is obtained from questionnaire survey, statistical analysis andon-site observation. Based on these data, the effect of metro constructionworker’s personalized behavior modification is evaluated from three aspects: 1)satisfaction degree; 2) score analysis; 3) on-site behavioral safetyperformance. Satisfaction degree is analyzed by questionnaire survey. Thequestionnaire is designedfrom system function experience, modificationcontent design and modification effect experience. The scores are counted fromsystem after worker’s participant in behavior modification, then they areanalyzed by contrastive and cluster analysis. Worker’s behavioral safetyperformance is measured through on-site observation. By setting unsafe behaviorrate (S) as an indicator, the Before and After comparison is used to compareworker’s behavioral safety performance before and after behavior modification,the With and Without comparison is used to compare worker’s behavioral safetyperformance with and without behavior modification.

Key words：Behavioral Safety; Scenarion Data; BehaviorObservation; Behavior law; Behavior modification; Big Data