【作者】 杨柳；

【导师】 李克平；

【作者基本信息】 北京交通大学 ， 系统科学， 2021， 博士

【摘要】 安全永远是轨道交通运输的重要主题。轨道交通通常是指国家铁路系统、城际轨道交通和城市轨道交通。保证轨道交通运输安全需要工作人员对列车和铁轨做日常监测、定期检测和维护。为此,积累了大量相关的非结构化文本数据。目前,对于非结构化文本数据往往采取省略、替换或删除的策略。由于非结构化文本数据相对结构化数据的信息更加抽象,不能直接被计算机处理,因此开发难度大。随着机器学习方法的进步,非结构化文本数据在自然语言处理领域得到发展,一些将非结构化文本数据转化成结构化数据的方法被提出。同时,轨道交通信息化要求的提出让轨道交通非结构化文本数据开发得到重视,迫切需要有效方法对报告或记录等形式的非结构化文本数据进行挖掘,提取出有效信息,从而进一步构建致因模型或事故模型对轨道交通安全进行风险分析。因此,本文主要针对轨道交通的非结构化文本数据进行研究,提出了一种基于网络的方法对文本进行关键词提取,进而识别出事故致因文本,实现事故的致因分析和风险分析,满足轨道交通信息化和智能化的发展需求。本文主要内容和创新点如下:1.基于复杂网络的关键词提取方法。为了帮助轨道交通相关人员从大量的事故/故障文本中获得关键内容,考虑到现有的基于网络理论的关键字提取方法仅考虑文档中词语之间的联系,而忽略句子影响,本文提出了一种新的网络模型NWS(New Word-Sentence),构建了由词网和句网组成的双层文本网络,该网络考虑了文本中句子对词语的影响。实验结果证明NWS方法提取出的关键词准确率、召回率和F值都优于经典TF-IDF(Term Frequency-Inverse Document Frequency)方法、传统的MF(the Most Frequent)方法和同样基于网络的Word-net方法和Text Rank方法,其中NWS的准确率、召回率和F值分别比Word-net方法高7.95%、8.27%和6.54%,NWS平均准确率也比TF-IDF方法的结果高17.56%。2.利用关键词提取的轨道交通事故致因文本识别方法。为能够准确地从轨道交通文本数据中识别轨道交通事故的致因文本以进行之后的致因分析和风险分析,提出了基于NWS关键词提取的致因文本识别方法。该方法首先使用NWS方法获得原因文本的特征,结合文本预处理过程,生成与致因文本有关的用户词典和停用词表来实现致因文本的识别。为验证该方法的有效性,通过改变方法中不同条件构成实验组与所提出方法进行效果对比。实验结果表明,致因文本识别方法在考虑关键词特征、停用词功能和文本相似度的条件下,识别效果最好,得到的准确率、召回率和F值最高。单就准确率来说,与其他实验差值最高达到8.72%。实验验证,本文通过对北京市地铁事故报告进行致因文本的识别,完成构建bow-tie模型,最后提出相应的预防措施。3.面向文本数据的动态贝叶斯网络事故风险分析。本文提出了一种基于文本数据的贝叶斯网络构造方法,即从事故文本中识别出致因文本,并根据文本特征和词语关联实现了故障树的构建。基于文本数据建立的故障树以及图模型,最后得到的贝叶斯网络对事故进行风险分析。实验数据选自美国联邦铁路局2008年—2018年列车脱轨文本数据集,本文提出的方法识别出事故的致因文本,从致因文本中提取出致因关系构建故障树,从而进一步获得相应贝叶斯网络的先验概率以及得到贝叶斯网络顶层事故发生的概率结果,计算出列车发生脱轨事故的概率为2.24E-04。4.基于地铁文本数据的致因链风险分析方法。本文使用网络爬虫技术获得2006-2020年“北京地铁”发布在微博平台上的地铁运营数据,结合本文已提出的关键词提取方法和致因文本识别方法,考虑地铁线路因素等有效信息对致因网络的影响,构建致因双层网络。该网络由致因网络和文本句网络共同构成,其中致因网络的节点是识别出的致因文本,文本句网络的节点是记录每条事故故障的句子。由于该双层网络具有文本属性,能够计算得到致因节点的综合特征值,将其作为致因节点的危险值。对致因网络采用相继故障方法得到事故最大可能致因链后,将致因节点的危险值代入公式中计算出各个致因因素和致因链的风险值。结果表明,风险最大的致因链为C9→D4,风险值达到27.717,它表示由于拥堵造成地铁设备故障引发事故的风险大,能够演变成较大事故,比如造成电梯损坏事故,因此应当及时地疏导地铁站内人流量。更多还原

【Abstract】 Safety is always an essential theme of rail transportation.Rail transit usually refers to the national railway system,intercity rail transit and urban rail transit.Ensuring rail transportation safety requires staff to do routine monitoring,regular inspection,and maintenance of trains and rails.For this reason,a large amount of relevant unstructured text data has been accumulated.At present,the strategy of omission,replacement or deletion is often adopted for unstructured text data.Since unstructured text data is more abstract than structured data,it cannot be directly processed by a computer,so it is challenging to develop.With the advancement of machine learning methods,unstructured text data has been developed in natural language processing,and some methods of transforming unstructured text data into structured data have been proposed.At the same time,the requirements of rail transit informatization have given attention to the development of rail transit unstructured text data.There is an urgent need for effective methods to mine unstructured text data in reports or records,extract practical information,and further construct the cause.The model or accident model conducts risk analysis on rail transit safety.Therefore,this dissertation mainly focuses on the research of unstructured text data of rail transit and proposes a network-based method to extract keywords from the text,identify the text of the cause of the accident,and realize the cause analysis and risk analysis accident.The development needs of rail transit informatization and intelligence.The main content and innovations of this dissertation are as follows:1.Keyword extraction method based on complex network.In order to help rail transit personnel to obtain key content from a large number of accident/fault texts,considering that the existing keyword extraction methods based on network theory only consider the connection between words in the document,and ignore the impact of sentences,this dissertation proposes a new network model,NWS(New Word-Sentence),constructs a two-layer text network composed of word nets and sentence nets,which considers the influence of sentences in the text on words.The experimental results prove that the keyword accuracy,recall and F value extracted by the NWS method are better than the classic TF-IDF(Term Frequency-Inverse Document Frequency)method,the traditional MF(the Most Frequent)method and the same network-based Word-net method and Text Rank method,the accuracy,recall and F value of NWS are 7.95%,8.27%,and 6.54% higher than Word-net method respectively.The average accuracy of NWS is also 17.56% higher than the result of the TF-IDF method.2.A text recognition method for rail traffic accidents based on keyword extraction.In order to accurately identify the cause text of a rail transit accident from the rail transit text data for subsequent cause analysis and risk analysis,a cause text recognition method based on NWS keyword extraction is proposed.The method first uses the NWS method to obtain the characteristics of the cause text and combines the text preprocessing process to generate a user dictionary and stop vocabulary related to the cause text from realizing the recognition of the cause text.In order to verify the effectiveness of the method,the experimental group was formed by changing the different conditions in the method and the effect of the proposed method was compared.The experimental results show that the causal text recognition method has the best recognition effect and the highest accuracy rate,recall rate,and F value when considering keyword characteristics,stop word function,and text similarity.In terms of accuracy alone,the difference with other experiments is up to 8.72%.Experimental verification,this ar dissertation ticle,through the identification of the cause text of the Beijing subway accident report,completed the construction of the bow-tie model,and finally proposed corresponding preventive measures.3.Dynamic Bayesian network accident risk analysis oriented to text data.This dissertation proposes a Bayesian network construction method based on text data to identify the cause text from the accident text and realize the construction of the fault tree based on the text characteristics and word association.Based on the fault tree and graph model established by the text data,the Bayesian network finally obtained carries out a risk analysis of the accident.The experimental data is selected from the 2008-2018 train derailment text data set of the US Federal Railroad.The method proposed in this dissertation identifies the cause text of the accident and extracts the causal relationship from the cause text to construct the fault tree so as to obtain the corresponding shell further.The prior probability of the Yeesian network and the probability of the accident at the top of the Bayesian network are obtained,and the probability of a train derailment accident is calculated to be 2.24E-04.4.Risk analysis method of cause chain based on subway text data.This dissertation uses web crawler technology to obtain the subway operating data released by the "Beijing Subway" on the Weibo platform from 2006 to 2020,combined with the keyword extraction methods and cause text recognition methods that have been proposed in this dissertation and considers adequate information such as subway line factors.Due to the influence of the network,a cause-causing double-layer network is constructed.The network is composed of a cause network and a text sentence network.The node of the cause network is the identified cause text,and the node of the text sentence network is the sentence that records each accident.Since the two-layer network has text attributes,the total characteristic value of the causal node can be calculated and used as the dangerous value of the causal node.After the successive failure method is adopted for the cause network to obtain the most likely cause chain of the accident,the risk value of the cause node is substituted into the formula to calculate the risk value of each cause factor and the cause chain.The results show that the riskiest causal chain is C9→D4,and the risk value reaches 27.717.It indicates that the risk of accidents caused by subway equipment failure due to congestion is large and can evolve into more significant accidents,such as elevator damage accidents.Therefore,it should be timely reduced the passage flow in subway stations.更多还原