基于因果推断的图神经网络沥青路面车辙预测方法  

Causal inference-based graph neural network method for predicting asphalt pavement performance

作　　者：陈凯 王小荷 时欣利 曹进德 CHEN Kai;WANG Xiaohe;SHI Xinli;CAO Jinde(School of Cyber Science and Engineering,Southeast University,Nanjing 211189,China;Key Laboratory of Transport Industry of Comprephensive Transportation Theory(Nanjing Modern Multimodal Transportation Laboratory),Nanjing 211135,China;School of Mathematics,Southeast University,Nanjing 211189,China)

机构地区：[1]东南大学网络空间安全学院,江苏南京211189 [2]综合交通运输理论交通运输行业重点实验室(南京现代综合交通实验室),江苏南京211135 [3]东南大学数学学院,江苏南京211189

出　　处：《南通大学学报(自然科学版)》2025年第1期18-27,50,共11页Journal of Nantong University(Natural Science Edition)

基　　金：国家重点研发计划项目(2020YFA0714300);综合交通运输理论交通运输行业重点实验室(南京现代综合交通实验室)开放课题资助项目(MTF2023004)。

摘　　要：为了提升沥青路面车辙预测的精度,通过整合图神经网络与因果推断方法,对多变量时间序列的长短时间模式以及变量之间的相互依赖进行建模,提出了端到端的多元时间序列预测模型。该模型由全局特征提取模块、局部特征提取模块、因果推断模块和双通图卷积模块4个模块组成。在全局特征提取模块,利用注意力机制及门控循环单元(gated recurrent unit,GRU)模型去捕获变量内部长时间模式特征;在局部特征提取模块,使用扩张卷积的方式,用不同大小的卷积神经网络(convolutional neural networks,CNN)卷积核去捕获变量内部不同维度的短时间模式特征;在因果推断模块,采用基于信息理论的因果分析来获取相关性,在使用传递熵来表述变量间因果关系的基础上,整合关系系数矩阵,通过权重配比的方式呈现变量之间的相关性揭示变量间的复杂关系;在双通图卷积模块,将高通滤波器引入传统低通图卷积神经网络,构建双通图卷积预测模块,从低通和高通双通道去同时捕获节点信号或特征的低频分量与高频分量,以提升模型的预测精度。最后,采用交通运输部公路科学研究院轨道数据集RIOHTrack,将提出的模型与经典统计学模型VARIMA、浅层学习模型SVR、深度学习模型GRU、结合注意力机制的GRU和TE-GCN等基准模型进行比较。结果表明,该模型在所有沥青路面结构类别上都取得了最佳的预测性能,且与传统统计学模型相比,基于深度学习的模型更加有效稳定,添加注意力机制的GRU模块可以捕获长期依赖性,从而使模型获得更好的预测性能。该模型为沥青路面的车辙预测提供了一种高效的方法,有望用于未来路面结构和养护方案设计,以提升路面使用寿命。To enhance the prediction accuracy of asphalt pavement rutting,this study introduces an end-to-end multivariate time series prediction model that integrates graph neural networks(GNN)with causal inference methodologies.The proposed model aims to effectively capture long-term and short-term temporal patterns as well as interdependencies among multiple variables.The model comprises four modules:global feature extraction,local feature extraction,causal inference,and dual-channel graph convolution.The global feature extraction module employs attention mechanisms and gated recurrent units(GRU)to capture long-term temporal dependencies within variables.The local feature extraction module utilizes dilated convolutional neural networks(CNN)with various kernel sizes to extract short-term temporal patterns at different scales.In the causal inference module,relationships among variables are identified using transfer entropy based on information theory,resulting in a relationship coefficient matrix that quantifies complex dependencies among variables.The dual-channel graph convolution module extends traditional low-pass graph convolutional neural networks by integrating a high-pass filter,simultaneously capturing low-frequency and high-frequency components of node signals or features to potentially improve prediction accuracy.The proposed approach was evaluated using the RIOHTrack dataset from the Research Institute of Highway Track,with comparisons conducted against several benchmark models,including the classical statistical model VARIMA,shallow learning model SVR,deep learning model GRU,attention mechanism-enhanced GRU,and TE-GCN.Experimental results indicate that the proposed model achieves superior predictive performance across various categories of asphalt pavement structures.Compared to traditional statistical models,deep learning-based models are more effective and stable,and the GRU module enhanced with attention mechanisms can capture long-term dependencies,further enhancing predictive performance.Overall,the propo

关 键 词：因果推断 多元时间序列预测 双通图卷积神经网络 沥青路面 车辙 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]