军事知识服务平台
基于多模因果感知注意力的服务区驶入判别模型
Multimodal causal-aware attention-based discriminative model for vehicle drive in expressway service area
摘要 [目标]面向高速公路服务区动态优化需求,针对驶入行为预测中多源特征融合不足、因果关联缺失和解释性较弱的问题,提出基于多模态因果感知注意力的服务区驶入判别模型,旨在融合车辆、环境、行程和资源信息,提升车辆驶入概率预测精度与决策支持能力.[方法]基于黑龙江省 4 条高速公路 20 个服务区 21 600 辆车调查数据,采用高斯混合模型刻画连续行驶距离多峰分布,并经 AIC/BIC 验证构建五维吸引力指标体系.设计车辆属性、时空环境、行程状态与服务区资源四模态编码器进行特征提取与层归一化融合,引入先验稀疏因果图和分层因果注意力机制,结合自适应特征交叉网络生成80维增强特征并输入 XGBoost 分类器实现驶入概率预测.[结果]所提模型测试集准确率为89.96%,较基准模型提升20.15%,五折交叉验证标准差低于0.004 2,具有较好的预测精度和稳定性.消融试验显示,多模态编码器、因果注意力和特征交叉模块分别贡献19.0%,8.45%和3.48%的准确率增益.[结论]所提模型能够融合多源异构特征,并通过因果约束强化特征权重调节,使判别过程更符合服务区驶入行为机理,弥补了传统模型在特征融合、因果表达和解释性方面的不足,可为服务区规划布局、动态运营优化和资源配置提供支撑.
Abstract:
[Objective]This study proposes a multimodal causal-aware attention-based discriminative model for vehicle drive in service area,meeting the demand for dynamic optimization of expressway service areas.The model aims to address the issues of insufficient multi-source feature fusion,lack of causal correlation,and weak interpretability in vehicle entry behavior prediction;as well as integrate the information of vehicle,environment,trip,and resource,improving the prediction accuracy of vehicle entry probability and decision support capabilities.[Method]Survey data from 21 600 vehicles were collected in 20 service areas along four expressways in Heilongjiang Province.The multi-peak distribution of continuous driving distance was characterized by Gaussian mixture model.A five-dimensional attractiveness indicator system was then constructed and validated using AIC and BIC.Four modality-specific encoders were designed to extract features from vehicle attributes,spatiotemporal environment,travel status,and service-area resources,which were fused through layer normalization.A prior-guided sparse causal graph and a hierarchical causal-aware attention mechanism were further introduced.An adaptive feature-crossing network generated 80-dimensional enhanced features,which were input into XGBoost classifier for entry probability prediction.[Result]The proposed model achieved an accuracy of 89.96%on the test set,which was 20.15%higher than that with baseline model.The standard deviation in five-fold cross-validation was below 0.004 2,indicating good accuracy and stability.The ablation test indicated that the multimodal encoder,causal-aware attention mechanism,and feature-crossing module contributed accuracy gains of 19.0%,8.45%,and 3.48%respectively.[Conclusion]The proposed model can effectively integrate multi-source heterogeneous features and enhance feature-weight adjustment through causal constraints.Its discrimination process is therefore more consistent with the behavioral mechanism of vehicle drive in service area.The model overcomes the limitations of traditional methods in feature fusion,causal representation,and interpretability.The findings provide support for expressway service area planning,dynamic operation optimization,and resource allocation.
刊 名 公路交通科技 ISTICPKU
年,卷(期): 2026, 43(9)
分类号: U491
机标分类号: TP391 F253.9 TP181
在线出版日期: 2026年10月22日
基金项目: 2023ZYZX2009:中央引导地方科技发展资金项目2023AB14001:广西重点研发计划项目,