【作者】 刘畅；

【导师】 陈俊敏；

【作者基本信息】 西南交通大学 ， 安全工程（专业学位）， 2019， 硕士

【摘要】 随着我国“智慧消防”工作的不断推进,要实现“智慧消防”中关于智慧感知、智慧防控和智慧管理的需求,对于火灾风险的精准研判与实时监测至关重要。但目前我国火灾风险评估面临的火灾基础信息累积不足、火灾数据分析不彻底、时效性不强等问题,难以满足当前“智慧消防”工作需求。2008年,Google经研究发现,流感相关关键词搜索量与实际患有流感症状的人数之间存在着密切关系,并根据此相关关系构建了流感趋势在线预警系统,预测效果显著。网络搜索数据应用研究的出现,为提高火灾风险评估的精度与时效性,提供了新的思路。本文旨在研究将网络搜索数据有效应用于城市区域火灾风险评估的思路与方法,拟构建基于网络搜索数据的火灾风险评估模型,以增强火灾风险评估的精度与时效性,为未来基于深度学习的火灾风险评估奠定研究基础。本文主要研究内容与成果如下:(1)基于海因里希的事故因果连锁论与马斯洛需求层次理论,结合火灾风险评估相关概念、发展、作用及特点,对网络搜索数据与火灾风险之间的逻辑关系进行探讨,构建理论框架,认为公众消防安全感会激发消防安全信息的需求,进而驱使消防安全信息搜寻行为。(2)根据已经构建的逻辑框架,从“消防安全常识”、“火灾事故损失”、“消防救援”、“消防安全意识”几个方面,进行搜索关键词的初选及拓展,得到了包含30个关键词的关键词备选库。随后,利用皮尔逊相关系数法和协整检验方法,筛选出含有8个关键词的火灾风险评估的关键词指标体系。(3)基于关键词指标体系,分别运用线性回归、主成分回归、向量自回归以及神经网络四种算法,建立四种火灾风险评估模型。其中,线性回归模型的拟合度为0.744,主成分回归模型的拟合度为0.659,向量自回归模型的拟合度为0.683,神经网络模型的拟合度为0.712;但在模型的选择上不仅要考虑拟合度,更需考虑模型实证检验精度。(4)对已建立的模型进行验证分析后发现,主成分回归模型的平均相对误差为47.1%,优于线性回归模型的95.9%;通过引入误差修正项,优化过后的向量自回归模型平均误差为40.9%,相较于上述两种模型,对于样本中出现的异常点问题得到了改善;通过优化网络参数后的神经网络模型,模型拟合度为0.95,远高于其他三种模型,神经网络模型的平均相对误差为25.3%,说明神经网络所具有的鲁棒性及容错性能,较好的解决了指标间的多重共线性,对于指标间的重复信息有着较高的包容性。考虑到神经网络具有的学习迭代能力,神经网络适合用来构建基于网络搜索数据的火灾风险评估模型。更多还原

【Abstract】 With the promoting smart fire protection campaign,the accurate judgment and real-time monitoring of fire risk is the most important part to achieve fire prevention and control which is comprised of smart FireGuard and smart fire management.But we are facing a host of problems like the insufficient fire basic information,analysing the fire data by halves and the weak on time-effectiveness.It’s hardly to meet the demands of smart fire protection campaign.In 2008,Google found that some search queries have a high correlation with influenza patients.They developed Google Flu Trends(GFT)to indicate influenza activity.The GFT gained some notable achievements.The application of search query data provides a new way to improve timeliness of fire risk assessment.This thesis is about to apply search query data to fire risk assessment effectively and establish a fire risk assessment model based on search query data.It will improve timeliness of fire risk assessment and lay the foundation of deep learning application in fire risk assessment.The contents and results of this thesis are mainly as follows:(1)To begin with,this paper constructs a theoretical framework based on Heinrich’s theory of causality in accidents,Maslow’s hierarchy of needs and the theory of fire risk assessment.The theoretical framework reveals that public fire safety may motivate their demands for fire safety information,and further driving their fire safety information seeking behaviors.(2)From the aspects of fire safety knowledge,fire accident loss,fire rescue and fire safety awareness conduct primary selection and expansion of search keywords.30 keywords are selected in Indicator library,and 8 keywords are used to establish fire risk assessment index system after Pearson correlation coefficient analysis and co-integration test.(3)Four types of fire risk assessment model is established by four methods which are linear regression,principal component regression,vector auto-regression,and artificial neural network.The fitting degrees of the above four models are 0.744,0.659,0.683 and 0.712,respectively.However,in the selection of the model,not only the fitting degree but also the accuracy of the empirical test of the model should be considered.(4)After verification and analysis of the established models,it is found that the average relative error of the principal component regression model is 47.1%,which is better than that of the linear regression model(95.9%).By adding error correction formula,the average relative error of the vector auto-regression model was 40.9%.Compared with the above two models,the outliers in the samples were improved.By optimizing the parameters of the artificial neural network model,the fitting degree of the model is 0.95,much higher than the other three models,artificial neural network model of the average relative error is 25.3%,shows that it has the robustness and fault tolerance performance,better solve the multicollinearity between the indexes,for repetition of information between indexes have higher inclusive.Considering the learning and iteration ability of artificial neural network,it is suitable for building fire risk assessment model based on search query data.更多还原