【作者】 张莹；

【导师】 周备；

【作者基本信息】 长安大学 ， 交通运输工程， 2023， 硕士

【摘要】 随着机动化水平的逐渐升高,交通事故在世界各地频繁发生,为有效对交通事故进行安全防范、降低事故严重程度,围绕交通事故历史数据展开的安全研究也愈发增多。目前交通事故信息采集仍存在困难,多国已经建立交通事故数据库但数据质量参差不齐,常有数据缺失情况。追尾事故作为交通事故的频发类型,其中数据缺失值的存在会导致统计分析结果的准确性下降、偏差增加,并增加模型误判风险、降低模型准确性。生成对抗网络(Generative adversarial network,GAN)是近年来复杂数据分布上无监督学习最具前景的方法之一,其应用范围已逐渐拓展到数据缺失填补领域。本文立足于交通追尾事故中的数据缺失情况,提出利用生成对抗式插补网络(Generative adversarial imputation network,GAIN)进行缺失值填补,形成完整的追尾事故数据集,并在此基础上构建三分类事故严重程度影响因素分析模型,对各特征变量的影响机理进行深入分析。考虑到国内事故数据难以获取,本文选取芝加哥2016-2021年间的101452条追尾事故数据为研究对象,共包含20个自变量和1个因变量,其中自变量涉及驾驶员、车辆、道路及环境信息,因变量为事故严重程度等级,共分为无受伤事故、轻伤事故、重伤或死亡事故三类。在数据缺失情况分析的基础上,利用GAIN实现对缺失数据的填补,并与统计学填补方法中的多重插补(Multiple Imputation by Chained Equations,MICE)和期望最大化(Expectation Maximization,EM)填充算法,以及机器学习方法中的缺失森林(Miss Forest)算法和K最近邻(K-Nearest Neighbor,KNN)算法进行对比,从填补速度、数据方差变化、拟合效果角度验证GAIN的填补效果。结果表明,GAIN算法能够更好地模拟原始数据分布情况,生成结果与原始数据最接近。在数据填补的基础上,分别构建XGBoost和LightGBM的三分类事故严重程度影响因素分析模型,对原始数据集及5种不同算法填补后的数据集进行建模分析。考虑到存在数据不均衡的情况,本文选取随机过采样法对数据进行不均衡处理,并选择准确率(Accuracy)、F1和AUC(Area Under Curve)作为模型的评价指标,结果表明,LightGBM模型的评价指标优于XGBoost模型,且GAIN数据集的评价指标优于其他数据集。在LightGBM模型中,模型评价指标经缺失值填补后得到了一定提高,GAIN数据集相较于原始数据集,其模型准确率提高了0.0456,F1提高了0.0322,AUC提高了0.0543;模型评价指标经数据不均衡处理后得到了小幅改善,GAIN数据集经过数据不均衡处理后,其模型准确率进一步提高了0.0161,F1提高了0.0160,AUC提高了0.0131。最后,利用SHAP分别对全局及事故个例进行模型结果的可视化分析,并以此为依据,从驾驶员、车辆、道路、环境角度提出追尾事故的风险防控策略,以期降低追尾事故的严重程度。本文依托于国家自然科学基金青年基金项目“考虑数据不平衡及模型可解释性的交通事故严重程度致因机理研究”(批准号:52102404)开展研究。更多还原

【Abstract】 As the level of motorization gradually increases,traffic accidents occur frequently around the world,and safety research around traffic accident history data is increasing in order to effectively prevent and reduce the severity of traffic accidents.At present,there are still difficulties in collecting traffic accident information,and many countries have established traffic accident databases,but the quality of data varies,and there are often missing data.As a frequent type of traffic accidents,the presence of missing values in rear-end accidents can lead to a decrease in the accuracy and bias of statistical analysis results,and increase the risk of model misclassification and decrease model accuracy.Generative adversarial network(GAN)is one of the most promising methods for unsupervised learning on complex data distribution in recent years,and its application has been gradually extended to the field of missing data filling.In this paper,based on the missing data in traffic rear-end accidents,we propose to use Generative adversarial imputation network(GAIN)to fill the missing values and form a complete data set of rear-end accidents,and on this basis,we construct a three-classification accident severity influence factor analysis model to analyze the influence mechanism of each feature The analysis of the influence mechanism of each characteristic variable is analyzed in depth.Considering the problem that domestic accident data are difficult to obtain,this paper selects 101,452 rear-end accident data from Chicago during 2016-2021 as the research object,which contains 20 independent variables and 1 dependent variable,among which the independent variables involve driver,vehicle,road and environment information,and the dependent variable is the accident severity level,with three categories: no injury accident,minor injury accident,and serious injury or fatal accident.Based on the analysis of the missing data situation,GAIN is used to achieve the filling of the missing data,and it is compared with the Multiple Imputation by Chained Equations(MICE)and Expectation Maximization(EM)filling methods in statistical filling methods,as well as the machine learning(EM)filling algorithm,and Miss Forest and K-Nearest Neighbor(KNN)algorithms in machine learning methods to verify the filling effect of GAIN in terms of filling speed,variance variation of data,and fitting effect.The results show that the GAIN algorithm can better simulate the original data distribution and generate results closest to the original data.Based on the data filling,the triple classification accident severity influence factor analysis models of XGBoost and LightGBM were constructed to model and analyze the original dataset and the dataset after filling by five different algorithms.Considering the existence of data imbalance,this paper selects the random oversampling method for data imbalance processing,and chooses Accuracy(Accuracy),F1 and AUC(Area Under Curve)as model evaluation indexes,and the results show that the evaluation indexes of LightGBM model are better than XGBoost model,and the evaluation indexes of GAIN dataset are better than other datasets.In the LightGBM model,the model evaluation indexes are improved after filling in the missing data values,and the model accuracy of GAIN dataset is improved by 0.0456,F1 by 0.0322,and AUC by 0.0543 compared with the original dataset;the model evaluation indexes are improved slightly after the data imbalance treatment,and the model accuracy of GAIN dataset is further improved after the data imbalance treatment.Finally,SHAP was used to visualize and analyze the model results for global and accident cases respectively,and based on this,risk prevention and control strategies for rear-end accidents were proposed from the perspectives of drivers,vehicles,roads,and environment,in order to reduce the severity of rear-end accidents.This paper is supported by the National Natural Science Foundation of China(NSFC)under the project "Research on the Causal Mechanism of Traffic Accident Severity Considering Data Imbalance and Model Interpretability"(Grant No.52102404).更多还原