基于机器学习的公交驾驶员事故风险识别及影响因素研究

doi:10.16265/j.cnki.issn1003-3033.2023.02.0034

摘要/Abstract

摘要：

为从公交驾驶员群体中识别出易发生事故的风险公交驾驶员,结合某市公交公司营运安全管理系统数据库、百度应用程序接口(API)及网络爬取技术,并应用K近邻算法补充缺失值,获取42条线路及1 893名驾驶员的数据;基于驾驶员、车辆、线路特征、违规行为、事故、管理等基本特征变量构造派生变量;采用包括递归特征消除、有惩罚项的逻辑回归、随机森林的集成方法选择特征;采用极致梯度提升(XGBoost)等6种机器方法分别建立分类模型,并采用贝叶斯方法优化超参数。结果表明:在构建的6个分类模型中,XGBoost方法构建的模型其受试者工作特征(ROC)曲线下的面积(AUC)评估结果最佳;运用贝叶斯方法优化模型,可以在一定程度上提升ROC的AUC指标;对于风险公交驾驶员预测准确率达到98.66%,运营单位还可以根据自身情况权衡虚报率与命中率代价。此外,车辆服役时间、违规次数等特征对于事故风险具有明显的非线性影响。

关键词: 风险公交驾驶员, 机器学习, 事故风险, 极致梯度提升(XGBoost), SHapley加性解释(SHAP)值

Abstract:

In order to identify the bus drivers who are about to cause accidents, the data set was obtained by combining the Bus Safety Management System database, the Baidu application programming interface(API) and web crawling technology. K-Nearest Neighbor algorithm was used to supplement the missing values and data from 1893 drivers in 42 lines was obtained. The basic characteristic variables included driver, vehicle, route characteristics, violations, accidents, management, and further construct derived characteristics on this basis. An integrated method, including recursive feature elimination, logistic regression with penalty terms, random forest and others, was designed and used for feature selection. The model was built using 6 machine methods like XGBoost and optimized for the hyper-parameters using Bayesian methods. The results indicate that among the six classification models constructed, the model constructed by XGBoost method has the best area under receiver operating characteristic(ROC) area under curve(AUC) evaluation results. Bayesian optimization can improve the AUC of ROC to a certain extent. For the accident driver's prediction accuracy rate reaches 98.66%, the operating unit can also weigh the false positive rate and true positive rate according to its own situation. Moreover, the nonlinear influence effect of features is found in the model results. The characteristics of vehicle service time, driving age, violations, punishment and other characteristics have a very obvious role in the accident risk.

Key words: risky bus drivers, machine learning, accident risk, extreme gradient boosting(XGBoost), SHapley additive explanation (SHAP) value

朱彤, 秦丹, 魏雯, 任杰, 冯移冬. 基于机器学习的公交驾驶员事故风险识别及影响因素研究[J]. 中国安全科学学报, 2023, 33(2): 23-30.

ZHU Tong, QIN Dan, WEI Wen, REN Jie, FENG Yidong. Research on accident risk identification and influencing factors of bus drivers based on machine learning[J]. China Safety Science Journal, 2023, 33(2): 23-30.

图/表 8

表1

特征变量及编码

特征类型	特征名称	类型	编码
因变量	$T i + 1$ 时段是否发生责任事故	定序	0=未发生;1=发生
驾驶员特征	性别	定类	[1,0]=男;[0,1]=女
	年龄	定序	1=≤25;2=26-35;3=36-45;4=46-55;5=≥56
	驾龄	定序	1=≤2;2=3-5;3=6-10;4=11-15;5=16-20;6=≥21
	文化程度	定序	1=小学;2=初中;3=职高中专及技校;4=高中;5=大专;6=大学及以上
	用工形式	定类	[1,0,0]=合同工;[0,1,0]=计划外聘用;[0,0,1]=计划外社会
车辆特征	车层	定序	1=单层;2=双层
	车辆动力类型	定类	[1,0,0]=纯电动;[0,1,0]=混合动力;[0,0,1]=天然气动力
	车辆服役时间	连续	=年
线路特征	线路站点数	连续	=站
	线路类型	定序	[1,0,0]=建成区外;[0,1,0]=跨建成区内外;[0,0,1]=建成区内
	线路长度	连续	=公里
违规行为特征	违规次数	连续	=次
	累计违规等级	连续	=级
	平均违规等级	连续	=级
	距上次违规时长	定序	1=未发生过违规;2=1个月以内;3=3个月以内;4=半年以内; 5=1年以内;6=2年以内;7=3年以内;8=超过3年
	违规晴天占比	定序	1=未发生过;2=T_i时段内晴天低于非晴天;3=T_i时段内晴天等于非晴天;4=T_i时段内晴天高于非晴天
	违规早晚高峰占比	定序	1=未发生过违规;2=早晚高峰小于平峰;3=早晚高峰等于平峰;4=早晚高峰大于平峰
	违规去程占比	定序	1=未发生过违规;2=去程小于回程;3=去程等于回程; 4=去程大于回程
	历史同月是否发生违规	定类	0=未发生;1=发生
	历史同季是否发生违规	定类	0=未发生;1=发生
事故特征	事故次数	连续	=次
	距上次事故时长	定序	1=之前未发生过事故;2=1个月以内;3=3个月以内;4=半年以内; 5=1年以内;6=2年以内;7=3年以内;8=超过3年
	事故晴天占比	定序	1=未发生过事故;2=T_i时段内晴天低于非晴天;3=T_i时段内晴天等于非晴天;4=T_i时段内晴天高于非晴天
	事故早晚高峰占比	定序	1=未发生过事故;2=早晚高峰小于平峰;3=早晚高峰等于平峰;4=早晚高峰大于平峰
	事故去程占比	定序	1=未发生过违规;2=去程小于回程;3=去程等于回程; 4=去程大于回程
	历史同月是否发生事故	定类	0=未发生;1=发生
	历史同季是否发生事故	定类	0=未发生;1=发生
	T_i时段是否发生事故	定序	0=未发生;1=发生
	$T i - 1$ 时段是否发生事故	定序	0=未发生;1=发生
	$T i - 2$ 时段是否发生事故	定序	0=未发生;1=发生
	$T i - 3$ 时段是否发生事故	定序	0=未发生;1=发生
	$T i - 4$ 时段是否发生事故	定序	0=未发生;1=发生
管理特征	培训总次数	连续	=次
	处罚总次数	连续	=次
	平均处罚等级	连续	=级
	批评教育总次数	连续	=次

表1

表2

评价指标及含义

判断指标	指标含义	评估依据	评估标准
准确率	预测正确的样本占总样本的比例	(TP+TN)/(TP+TN+FP+FN)	值越高越好
精准率	预测为正例的样本中真实正例的比例	TP/(TP + FP)	值越高越好
命中率	真正例被预测为正例的比例	TP/(TP + FN)	值越高越好
F₁值	调和平均的查准率和查全率	$2 × 精确率 × 命中率 (精确率 + 命中率)$	越接近1越好
AUC	ROC曲线下的面积	$12 ∑ i = 1 m - 1 (x i + 1 - x i) · (y i + y i + 1)$	值越高越好
虚报率	假正例被预测为反例的比例	FP/(TN + FP)	值越低越好
未命中率	假反例被预测为正例的比例	FN/(TP + FN)	值越低越好

表2

表3

表4

表5

图1

图2

图3

参考文献 20

[1]	BENTAMA A, KHATORY A, MILLOT M. Spatial analysis of bus accidents in France[C]. Proceedings of 2017 International Colloquium on Logistics and Supply Chain Management, 2017: 124-129.
[2]	WU Xiaolin, ZHANG Huimin, XIAO Wangxin, et al. Are bus company regulations associated with crash risk? findings from a retrospective survey in four Chinese cities[J]. International Journal of Environmental Research and Public Health, 2019, 16(8): DOI:10.3390/ijerph16081342. doi: 10.3390/ijerph16081342
[3]	GELAU C, GASSER T M, SEECK A. Fahrerassistenz und verkehrssicherheit[M]. Wiesbaden: Vieweg+Teubner Verlag, 2009: 24-32.
[4]	KAPLAN S, PRATO C G. Cyclist-motorist crash patterns in denmark: a latent class clustering approach[J]. Traffic injury prevention, 2013, 14(7): 725-733. doi: 10.1080/15389588.2012.759654 pmid: 23944832
[5]	TSENG C M. Social-demographics, driving experience and yearly driving distance in relation to a tour bus driver's at-fault accident risk[J]. Tourism Management, 2012, 33(4):910-915. doi: 10.1016/j.tourman.2011.09.011
[6]	JAVADREZA V, AFSHIN S M, ZAHRA T, et al. Aberrant driving behaviour, risk involvement, and their related factors among taxi drivers[J]. International Journal of Environmental Research and Public Health, 2018, 15 (8): DOI: 10.3390/ijerph15081626. doi: 10.3390/ijerph15081626
[7]	GOH K, CURRIE G, SARVI M, et al. Factors affecting the probability of bus drivers being at-fault in bus-involved accidents[J]. Accident Analysis and Prevention, 2014, 66:20-26. doi: 10.1016/j.aap.2013.12.022
[8]	USECHE S A, CENDALES B, ALONSO F, et al. A matter of style? Testing the moderating effect of driving styles on the relationship between job strain and work-related crashes of professional drivers[J]. Transportation Research Part F: Traffic Psychology and Behaviour, 2020, 72: 307-317. doi: 10.1016/j.trf.2020.05.015
[9]	AREZES P M, SÉRGIO MIGUEL A. The role of safety culture in safety performance measurement[J]. Measuring Business Excellence, 2003, 7(4):20-28. doi: 10.1108/13683040310509287
[10]	王晓勇, 罗珅, 任杰, 等. 公交驾驶员违规间隔时间及影响因素研究[J]. 中国安全科学学报, 2019, 29(6):128-133. doi: 10.16265/j.cnki.issn1003-3033.2019.06.021
	WANG Xiaoyong, LUO Shen, REN Jie, et al. Research on interval time of traffic violations for bus drivers and its influencing factors[J]. China Safety Science Journal, 2019, 29(6):128-133. doi: 10.16265/j.cnki.issn1003-3033.2019.06.021
[11]	林庆丰, 邓院昌, 张圆, 史晨军. 公交驾驶员驾驶愤怒量表的编制及初步应用[J]. 中国安全科学学报, 2018, 28(6):49-54. doi: 10.16265/j.cnki.issn1003-3033.2018.06.009
	LIN Qingfeng, DENG Yuanchang, ZHANG Yuan, SHI Chenjun. Development of driving anger scale for bus drivers and its preliminary application[J]. China Safety Science Journal, 2018, 28(6):49-54. doi: 10.16265/j.cnki.issn1003-3033.2018.06.009
[12]	闭小梅, 闭瑞华. KNN算法综述[J]. 科技创新导报, 2009(14): 31.
	BI Xiaomei, BI Ruihua. Review of KNN algorithms[J]. Science and Technology Innovation Herald, 2009(14):31.
[13]	CHANDRASHEKAR G, SAHIN F. A survey on feature selection methods[J]. Computers & Electrical Engineering, 2014, 40(1): 16-28.
[14]	DIETTERICH T G. Ensemble methods in machine learning[C]. International Workshop on Multiple Classifier Systems, 2000: 1-15.
[15]	何清, 李宁, 罗文娟, 等. 大数据下的机器学习算法综述[J]. 模式识别与人工智能, 2014, 27(4): 327-336.
	HE Qing, LI Ning, LUO Wenjuan, et al. A survey of machine learning algorithms for big data[J]. Pattern Recognition and Artificial Intelligence, 2014, 27(4): 327-336.
[16]	CHAWLA N, BOWYER K, HALL L, et al. SMOTE: Synthetic minority over-sampfing technique[J]. Journal of Artificial Intelligence Research, 2002, 16: 321-357. doi: 10.1613/jair.953
[17]	FRIEDMAN J H. Greedy function approximation: a gradient boosting machine[J]. Annals of Statistics, 2001, 29 (5) :1189-1232. doi: 10.1214/aos/1013203450
[18]	DRUMMOND C, HOLTE R C. Explicitly representing expected cost: an alternative to ROC representation[C]. Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2000: 198-207.
[19]	NASROLLAHZADEH A A, SOFI A R, RAVANI B. Identifying factors associated with roadside work zone collisions using machine learning techniques[J]. Accident Analysis & Prevention, 2021, 158(10): DOI: 10.1016/j.aap.2021.106203. doi: 10.1016/j.aap.2021.106203
[20]	FIORENTINI N, LOSA M. Handling imbalanced data in road crash severity prediction by machine learning algorithms[J]. Infrastructures, 2020, 5(7): DOI: 61. 10.3390/infrastructures5070061. doi: 61. 10.3390/infrastructures5070061

特征类型	特征名称	平均值
违规特性	违规次数	0.527
	平均违规等级	0.376
	违规时间间隔离散程度	0.184
	距离上次违规时长	0.133
	违规晴天占比	0.429
	违规早晚高峰占比	0.442
	违规去程占比	0.509
	历史同月是否发生违规	0.163
	历史同季是否发生违规	0.196
事故特性	事故次数	0.211
	Ti是否发生事故	0.202
	Ti-1是否发生事故	0.209
	Ti-4是否发生事故	0.133
培训次数	处罚总次数	0.372
	平均处罚等级	0.564
	批评教育总次数	0.152
	培训总次数	0.144
驾驶员特性	用工形式	0.107
	性别	0.205
	年龄	0.196
	驾龄	0.172
	文化程度	0.159
线路特性	线路长度	0.222
线路特性	线路站点数	0.181
车辆特性	车辆动力类型	0.173
车辆特性	车辆服役时间	0.406

模型	准确率	精准率	命中率	F₁值	AUC
逻辑回归	99.81	99.84	99.98	99.91	90.96
朴素贝叶斯	76.57	99.99	76.54	86.71	89.28
决策树	99.84	100.00	100.00	100.00	77.84
随机森林	99.84	100.00	100.00	100.00	88.04
GBDT	99.77	99.84	99.93	99.89	85.83
XGBoost	98.66	99.91	98.75	99.33	96.31

真实情况	预测结果
真实情况	发生事故	未发生事故
发生事故	9	12
未发生事故	160	12 680