基于Hybrid有限混合模型的交通事故严重程度分析

发布时间：2018-07-01 09:11

本文选题：事故严重程度 + 事故影响因素　；参考：《北京交通大学》2017年硕士论文

【摘要】：随着机动化进程的不断加快,各国交通安全方面的压力也不断增大。在学术界,利用统计学模型对交通事故数据进行分析,探索事故因素与事故严重程度之间的复杂关系已经成为交通领域的一个重要分支。在这个领域中,模型的特性以及拟合的效果直接影响到交通事故数据挖掘的精度与深度,若采用不合理的统计学模型对这种复杂关系进行建模分析,结果可能导致对事故数据的有偏估计和事故因素的错误解释。本文以事故严重程度分类模式为出发点,提出了一种Hybrid有限混合模型(Hybrid Finite Mixture Model,HFM)来对事故因素与事故严重程度的复杂关系进行建模。首先,通过分析导致交通事故发生的影响因素,确定了以人、车、路、环境方面的17个重要的事故因素及其特点,将事故严重程度按照KACBO分类法分为五类,选择美国事故总评系统(General Estimates System,GES)的部分机动车事故数据作为样本。其次,基于对事故数据以及事故严重程度分类模式复杂性的考虑,本文提出了 HFM模型,它由两个不同分类的多元回归模型混合组成,其中一组为多项Logit模型(Multinomial Logit Model,MNL),代表无序化的数据生成过程,另一组为有序Logit模型(Ordered Logit Model,OL),代表有序化的数据生成过程,通过EM算法采用概率选择的方式来决定两个成分的权重。虽然已有学者运用同分类的传统有限混合模型对事故严重程度进行了研究,这些混合模型中的成分均属于同一类模型,在事故严重程度分类模式上还存在着局限性。HFM模型试图规避传统有限混合模型采取单一分类(无序或有序)而导致分类模式失真的问题,利用不同成分有限混合模型对事故严重程度进行分析成为一种新思维。最后,本文将HFM模型与其他四种模型进行了模型评价和弹性分析,包括通过分析模型参数来对事故因素与事故严重程度关系进行解释,以及通过三种不同的评价准则对模型进行综合评价。研究结果表明,HFM模型和其他四种模型均认为30岁以上年龄段、女性驾驶员、车辆追尾会不同程度地加重事故严重程度,而车辆拖尾则会减轻事故严重程度。此外,HFM模型具有无序化特点的事故严重程度分类模式占据着主导地位。相比多元回归模型和传统有限混合模型,HFM模型综合评价结果最佳,可以有效捕捉潜在因素的异质性,能挖掘事故数据中更多的信息。
[Abstract]:With the acceleration of motorization process, the pressure of traffic safety in various countries is also increasing. In academic circles, it has become an important branch of traffic field to analyze traffic accident data by statistical model and explore the complex relationship between accident factors and accident severity. In this field, the characteristics of the model and the effect of fitting directly affect the precision and depth of traffic accident data mining. The results may lead to biased estimation of accident data and misinterpretation of accident factors. Based on the classification model of accident severity, a hybrid finite mixture model (HFM) is proposed to model the complex relationship between accident factors and accident severity. First of all, by analyzing the influencing factors of traffic accidents, 17 important accident factors and their characteristics in people, cars, roads and environment are determined, and the severity of accidents is classified into five categories according to KACBO classification. Select some of the motor vehicle accident data from the General estimates system as a sample. Secondly, based on the complexity of accident data and accident severity classification model, this paper proposes an HFM model, which is composed of two multivariate regression models with different classification. One group is multi-logit model (MNL), which represents the process of unordered data generation, and the other group is ordered logit model (OL), which represents ordered logit model. The weight of two components is determined by means of probability selection through EM algorithm. Although some scholars have studied the severity of accidents using traditional finite mixing models of the same classification, the components of these models all belong to the same class of models. In the classification model of accident severity, there is also the limitation. HFM model tries to avoid the traditional limited mixed model to adopt a single classification (disorder or order), which leads to the distortion of classification mode. It is a new way to analyze the severity of accidents by using finite mixing models with different components. Finally, the HFM model and the other four models are evaluated and analyzed, including the analysis of the model parameters to explain the relationship between the accident factors and the severity of the accident. And through three different evaluation criteria for comprehensive evaluation of the model. The results show that the HFM model and the other four models all think that female drivers and rear-end vehicles will aggravate the severity of the accident in varying degrees, while the trailing of the vehicle will reduce the severity of the accident. In addition, HFM model with disordered characteristics of accident severity classification model occupies a dominant position. Compared with the multiple regression model and the traditional HFM model, the HFM model can effectively capture the heterogeneity of potential factors and mine more information from the accident data.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.31

【参考文献】