突扩巷道模型风流态的LDA与PIV联合测试

针对传统的接触式瞬时速度测量方法的局限性,采用非接触测量技术激光多普勒测速仪(LDA)及粒子成像测速仪(PIV)对平直巷道及断面突扩后风流状态进行试验测试。在巷道试验模型条件下,PIV技术可以瞬时获得巷道突扩流场信息,平直巷道速度流线基本呈平滑直线,突扩隅角有大涡存在,并且涡流区测风方向极不稳定,而且风速很低,风速平均值在0.1~0.2 m/s波动,表明在煤矿井下测风时可以有条件地忽略涡流区。LDA技术测试得到巷道断面各点统计平均流速,由于受突扩涡流及二次流的影响,平直巷道断面风速从壁面以跃迁方式"突变"达到均值,断面风速分布呈近似均等的动态波浪线分布而非准抛物线型分布。结果表明,LDA与PIV测量技术联合应用可以测试以湍流为特征的巷道流场风流变化情况。     

基于文本挖掘的城市景区密集人群风险感知*

为充分挖掘城市景区密集人群风险信息和隐藏规律,以携程、美团、大众点评3个旅游网站上发表的关于城市景区密集人群的游记和评论作为数据源,运用潜在狄利克雷分配(LDA)主题模型及Word2Vec词向量模型挖掘密集人群风险感知主题与关键词及二者之间的关联关系;基于计划行为理论构建城市景区密集人群风险感知模型,结合共现规律对风险感知主题的关联关系分析,明确景区密集人群对个体风险感知的作用规律。研究结果表明:降低个体在密集人群情景下的风险感知水平,将避免公众产生消极行为,降低踩踏事故发生的可能性;同时通过主题共现分析得出影响个体风险感知的核心主题为“网红表演”主题,可通过提高热门表演管制力度来提高个体主观规范对实际行为的正向影响,使得个体风险感知强度处于稳定水平。研究结果可以完善城市景区密集人群风险管理和风险沟通机制,从而有效管理风险。     

民航不安全事件报告危险源识别模型

为了识别民航不安全事件的发生原因,从民航不安全事件报告出发,提出了一种基于LDA主题模型和Word2Vec模型的民航不安全事件报告危险源识别模型。首先,构建危险源标签分类表,选取民航不安全事件报告与Wiki中文数据库作为数据源,然后利用LDA主题模型得出民航不安全事件报告的主题概率分布与词概率分布,最后利用Wiki中文数据库训练Word2Vec模型得到词向量,进而计算文档主题与不同危险源标签之间的相似度,得出民航不安全事件报告的危险源分类。结果表明,该模型的精准率为77.7%、召回率为86.8%、F为82.0%,能够有效识别各民航不安全事件报告的危险源。     

基于LDA的均直巷道断面风速分布规律实验研究

针对传统井下测风方法以及传统接触式流速测量方法的局限性,以矩形均直巷道断面风速实验测试为例,利用非接触式测试技术激光多普勒测速仪（LDA）对断面风速进行实验测试。实验表明:在平均风速为3m/s左右的均直巷道内的稳定流动下,断面风速分布近似呈矩形环状封闭波动曲线,测点风速具有极度的脉动现象,但测点速度大小服从正态分布。在不同断面平均风速下,断面轴线上风速分布均近似服从指数函数形式。研究表明,LDA测试技术能够较好的反映实验巷道流场风流湍流脉动特性,可以作为研究矿井通风一系列复杂流动的实验手段。     

Concentration field and turbulent fluxes during the mixing of two buoyant plumes

In this work an experimental study of mixing of two identical plumes, carried out in a turbulent neutral boundary layer generated in a wind tunnel, is presented. Measurements have been performed with fast flame ionisation detectors (FFIDs) and a two-component Laser-Doppler Anemometer system. Results allow the study of both the average and the fluctuating concentration field, including the turbulent vertical and longitudinal mass fluxes, in single plumes and during the interaction of two identical plumes. This information gives insight into the details of the mixing phase of the two plumes. Results of trajectories and additional rise (due to plume interactions) have been compared with previous measurements carried out in laminar cross-flows, showing similar behaviour. Concentration distributions in plume cross-sections in turbulent cross-flows differ from those measured in laminar cross-flows. Average vertical and longitudinal velocity measurements into the plume core show the strength of the shielding effect of the upwind plume and some details of interaction between the counter-rotating vortex pairs (CVPs). For large values of the alignment angle φ, between the line joining the stacks and the cross-flow, an average negative vertical velocity is measured in the middle of the plume even if its centre of mass is rising. This downward velocity is induced by the slow interaction of the CVPs and generates a vertical stretching of the plume and negative rise enhancement. Vertical turbulent fluxes change sign on the plume centreline and are of opposite sign with respect to the longitudinal turbulent fluxes. Results indicate a good linearity between vertical turbulent fluxes and concentration gradients, with different proportionality for the top and bottom parts of the plume (especially in the near field) indicating that dispersion could be described by a gradient-transfer model.     

线性判别分析对舟山海洋站降水中常量阳离子含量的年度、季节特征分析

通过对舟山站点2002/2003年度降水中的6种常量阳离子H⁺、NH₄⁺、K⁺、Na⁺、Ca²⁺、Mg²⁺近2a的连续观测,采用稳健统计分析对浓度范围及变化趋势进行了描述.利用线性判别分析对实验数据进行降维处理,得到线性判别曲面,对样本点在新正规变量空间的时间性分布特征进行了描述,依据正规变量的线性组合,解释了决定该样本点时间性特征分布的主要阳离子影响因素,并对造成异常样品点的影响因素进行了分析.     

Invasive species change detection using artificial neural networks and CASI hyperspectral imagery

For monitoring and controlling the extent and intensity of an invasive species, a direct multi-date image classification method was applied in invasive species (salt cedar) change detection in the study area of Lovelock, Nevada. With multidate Compact Airborne Spectrographic Imager (CASI) hyperspectral data sets, two types of hyperspectral CASI input data and two classifiers have been examined and compared for mapping and monitoring the salt cedar change. The two types of input data are all two-date original CASI bands and 12 principal component images (PCs) derived from the two-date CASI images. The two classifiers are an artificial neural network (ANN) and linear discriminant analysis (LDA). The experimental results indicate that (1) the direct multitemporal image classification method applied in land cover change detection is feasible either with original CASI bands or PCs, but a better accuracy was obtained from the CASI PCA transformed data; (2) with the same inputs of 12 PCs, the ANN outperforms the LDA due to the ANN’s non-linear property and ability of handling data without a prerequisite of a certain distribution of the analysis data.     

A method for assisting the accident consequence prediction and cause investigation in petrochemical industries based on natural language processing technology

Risk analysis for production processes in the petrochemical industry is an important procedure for consequence prediction and investigation of accidents. The analyzer must grasp the correlations between the possible causes and consequences. From the potential cause and effect found in risk analysis reports, complete clarification should be obtained. Therefore, this study presents a method for assisting accident consequence prediction and investigation in the petrochemical industry based on risk analysis reports using natural language processing technology. First, a hazard and operability (HAZOP) historical data table is established by filling over 7200 HAZOP analysis data points. Both the causes and consequences in the table are classified into 20 categories each using the Latent Dirichlet Allocation (LDA) models. The LDA clustering results are assigned classification for the cause and consequence topics to the cause and consequences of the HAZOP analysis data. Based on part-of-speech (POS) tagging, all the words in each cause and consequence record are divided into subject and action words. Next, the word combinations of subject and action words with a higher occurrence are considered the key phrases for describing and representing the corresponding cause and consequence topic classifications. The Apriori algorithm is used to determine the frequent item sets, acquire the association rules, and calculate the association degree to obtain the sort order; it can highlight general trends in relational cause and consequence topics. According to the results, the most likely cause of the consequence and the most likely consequence that the cause may lead to are identified. Finally, a visual interface is developed to present the data for the consequence prediction and cause investigation of accidents. The results reveal that the quantity and quality of historic data are important factors that may influence the results. This method can contribute to predicting the accident evolution trend of an abnormal situation, taking preventive measures in advance, improving the accuracy of early warning, and supporting emergency response measures.     

基于数据挖掘的化工生产事故致因主题抽取

为充分挖掘化工生产事故数据中的有效信息和潜在规律,提高对化工事故认知水平,针对某化工集团2010—2016年共1 578起事故数据,利用社会网络分析等方法揭示事故要素间的关联关系;运用潜在狄利克雷分配（LDA）模型进行事故聚类,并抽取到5个事故致因主题。研究结果表明:LDA主题模型等数据挖掘技术能有效挖掘大量事故数据中的潜在信息;5个事故致因主题中,4个涉及到人因或组织层面的缺陷;员工注意力不集中和现场风险管理不足这2个致因主题间具有较强相关性;员工注意力不集中、现场风险管理不足以及设备问题是导致事故发生的主要原因。     

基于文本挖掘的管制运行风险主题分析*

为充分挖掘管制运行风险信息和隐藏规律,实现数据驱动的风险管理。以某管制单位2004—2019年共269条管制原因不安全事件数据为挖掘语料,在考虑上下文语义的基础上,运用潜在狄利克雷分配（LDA）主题模型挖掘管制运行风险主题及关键词,使用Word2Vec挖掘主题之间、关键词之间的关联关系,运用社会网络分析软件UCINET、可视化工具NETDRAW构建语义网络将关联关系进行可视化并进行网络分析。结果表明:LDA主题模型可以通过运行数据实现对管制运行风险的高效提取和深层挖掘,挖掘到管制人为因素、特情处置、地空配合、班组资源管理、组织管理、运行环境、管制指挥共7个主题,其中管制人为因素主题是核心主题,与其他主题都具有较强的相互关联;Word2Vec和语义网络相结合能够更准确地挖掘风险之间的关系,确定主题的重要度排序,识别关键风险。