Abstract: In order to deduce the characteristics of severe traffic accidents on highways from a global perspective, the accidents of different severity levels were incorporated into a unified research system by use of linear programming methods, which was based on the historical accident data of a highway in a city in Southwest China since its opening 10 years ago. The conditional dependency between accident factors and accident severity was simulated from a global perspective by the use of the machine learning-based factor partial dependence analysis method, and the factorial characteristics of severe traffic accidents were reasoned. The research results show that accident time, vehicle type, accident configuration, accident location and weather conditions are pivotal features that precipitate severe accidents. There is a significant nonlinear relationship between accident time, accident location, drivers driving experience and accident severity. The time characteristics of serious accidents are 05:00 and 08:00, with the involved vehicle type predominantly being trucks, and the accident configuration is characterized by rear-end collisions, while the accident locations are typified by the 60-80 km section of the highway. The cloudy weather, drivers with less than one year of driving experience and tunnel sections are identified as distinctive features associated with severe accidents.

Key words: traffic and transportation engineering; safety engineering; accident causation reasoning; accident severity prediction; highway; partial dependence

摘要： 为从全局视角推理高速公路严交通重事故特征，基于西南地区某城市高速公路开通10年以来的历史事故数据，运用线性规划方法将不同严重等级事故纳入统一研究体系，利用机器学习因子偏依赖量分析方法从全局视角模拟事故因素与事故严重度之间的条件依赖关系，推理严重交通事故的因子特征。研究结果表明：事故时间、车辆类型、事故形态、事故位置与天气情况是诱发严重事故的重要特征；事故时间、事故位置、驾驶人驾龄与事故严重度间存在明显的非线性关系，严重事故的时间特征是05:00时和08:00时，涉事车辆类型特征是货车，事故形态特征是追尾，事故位置特征是高速公路里程的60~80 km区段，天气特征是阴天，驾驶人驾龄特征是不足1年，路段类型特征是隧道。

关键词: 交通运输工程；安全工程；事故致因推理；事故严重度预测；高速公路；偏依赖量