Abstract: Research on the security incident identification method for highway reconstruction and expansion zones is relatively weak, lack of complete and efficient detection methods. A safety incident identification model based on an unsupervised double-layer stacking framework was proposed for the unique requirements of the guided improvement points in highway reconstruction and expansion zone. The proposed model aimed to enhance the accuracy rate of automatic identification of traffic safety incidents. By integrating Bootstrap resampling technology, multiple weak learners, and the XGBoost meta-learner, the challenges posed by changes in road alignment and traffic flow status in reconstruction and expansion zones were effectively addressed. The proposed model demonstrated high precision and accuracy in identifying abnormal states during security incidents. The significant contribution of each component of the double-layer stacking framework and the advantages of the overall model were verified through ablation experiments. Compared with traditional models (SVM, RF, KNN), it is indicated that the proposed model has improved identification accuracy by over 30% and precision by over 7%. This study not only provides an efficient technical solution for traffic safety management in highway reconstruction and expansion areas but also paves new paths for future research and application of safety incident identification in intelligent transportation systems.

Key words: traffic and transportation engineering; incident detection; highway; traffic accidents; ensemble learning; XGboost

摘要： 面向高速公路改扩建区域场景的安全事件识别方法研究相对薄弱，缺少完整且高效的检测方法。针对高速公路改扩建区域导改点场景的特殊需求，提出一种基于无监督双层stacking框架的安全事件识别模型。该模型旨在提高交通安全事件的自动识别准确率，通过融合Bootstrap重采样技术、多种弱学习器以及XGBoost元学习器，成功地应对改扩建区域内道路线型与交通流状态的变化带来的挑战，在识别发生安全事件时的异常状态方面展现出高精确度和准确率。通过消融试验验证了双层stacking框架各组件的显著作用和整体模型的优势。通过与传统模型（SVM、RF、KNN）进行对比，结果表明，该模型在识别准确率上提升了30%以上，在精确率上提高了7%以上。这一研究不仅为高速公路改扩建区域的交通安全管理提供了一种高效的技术解决方案，也为未来智能交通系统中安全事件识别的研究和应用开辟了新路径。

关键词: 交通运输工程；事件检测；高速公路；交通事故；集成学习；XGboost