Please wait a minute...
浙江大学学报(工学版)  2025, Vol. 59 Issue (5): 1083-1091    DOI: 10.3785/j.issn.1008-973X.2025.05.021
土木工程、交通工程     
网络级道路交通运行状态的深度学习识别方法
罗义凯1(),辛苡琳2,徐金华1,陈桂珍1,李岩1,*()
1. 长安大学 运输工程学院,陕西 西安 710064
2. 比亚迪汽车有限公司,陕西 西安 710018
Deep learning method for recognizing network-level road traffic state
Yikai LUO1(),Yilin XIN2,Jinhua XU1,Guizhen CHEN1,Yan LI1,*()
1. College of Transportation Engineering, Chang’an University, Xi’an 710064, China
2. BYD Automobile Limited Company, Xi’an 710018, China
 全文: PDF(1388 KB)   HTML
摘要:

为了精准、实时、高效地掌握道路网各区域交通运行状态,基于网约车轨迹数据提取相关运行参数,对研究区域进行时空单元划分,构建将特征提取与聚类过程融合的深度聚类网络模型,对交通状态进行分类. 对聚类结果量化获取类别标签,结合集成学习、贝叶斯优化和轻量梯度提升机,提出交通状态识别模型. 西安市网约车数据测试的结果表明,道路运行状态可以分为畅通、缓行、轻度拥堵、中度拥堵和严重拥堵5种类型,严重拥堵路段占比在早晚高峰时段明显增加,平峰时段有所减少. 所提聚类模型的效果均优于对比模型,交通状态识别模型计算的精确率、召回率、F1分数和准确率分别为0.982 1、0.984 4、0.983 3、0.983 9.

关键词: 网络级道路交通运行状态深度聚类轨迹数据轻量梯度提升机    
Abstract:

The research area was divided into spatiotemporal units, and a deep clustering network model that integrated feature extraction and clustering process was constructed based on the trajectory data of online car-hailing to extract relevant operation parameters to identify traffic states in order to accurately, real-time and efficiently grasp the traffic operation state of various areas in the road network. The clustering results were quantified to obtain category labels, and a traffic state identification model was proposed combining integrated learning, Bayesian optimization and light gradient boosting machine. The test results of Xi'an online car-hailing data show that road operation states can be divided into 5 types: smooth, slow, mild congestion, moderate congestion and severe congestion. The proportion of severely congested road sections increases significantly during morning and evening peak periods and decreases during off-peak periods. The proposed clustering model performs better than the comparison models, with the precision, recall, F1-score and accuracy of the traffic state identification model being 0.982 1, 0.984 4, 0.983 3 and 0.983 9 respectively.

Key words: network-level road    traffic operation state    deep clustering    trajectory data    light gradient boosting machine
收稿日期: 2024-02-27 出版日期: 2025-04-25
CLC:  U 491  
基金资助: 国家自然科学基金资助项目(51408049);陕西省自然科学基础研究计划资助项目(2020JM-237).
通讯作者: 李岩     E-mail: lyk@chd.edu.cn;lyan@chd.edu.cn
作者简介: 罗义凯(1998—),男,博士生,从事深度学习和智能交通的研究. orcid.org/0009-0007-9024-4656. E-mail:lyk@chd.edu.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章  
罗义凯
辛苡琳
徐金华
陈桂珍
李岩

引用本文:

罗义凯,辛苡琳,徐金华,陈桂珍,李岩. 网络级道路交通运行状态的深度学习识别方法[J]. 浙江大学学报(工学版), 2025, 59(5): 1083-1091.

Yikai LUO,Yilin XIN,Jinhua XU,Guizhen CHEN,Yan LI. Deep learning method for recognizing network-level road traffic state. Journal of ZheJiang University (Engineering Science), 2025, 59(5): 1083-1091.

链接本文:

https://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2025.05.021        https://www.zjujournals.com/eng/CN/Y2025/V59/I5/1083

图 1  网络级道路交通运行状态识别方法的流程图
图 2  考虑航向和距离约束的滑动窗口地图匹配算法的流程
图 3  网约车轨迹点缺失和异常的示意图
图 4  基于LSTM的深度自编码神经网络
算法1 TS-DEC深度聚类
输入:经数据处理的高维时序数据
步骤:
//依据DEC算法中KL散度(Kullback-Leibler divergence, KL)的聚类方法
1)通过LSTM深度自编码神经网络,获取特征分布空间$ \boldsymbol{Z} $及特征映射$ {f}_{\theta } $.
2)采用K-means算法,确定$ \boldsymbol{Z} $中的初始聚类中心$ {\boldsymbol{\mu }}_{{j}}\left\{{\boldsymbol{\mu }}_{{j}}\in \boldsymbol{Z},j=1,2, \cdots ,k\right\} $.
3) while 总损失$ L $未趋于稳定 do
4)计算特征空间中提取的特征与聚类中心的软分配$ q $,使用学生t分布来衡量特征向量$ {\boldsymbol{z}}_{{i}} $与聚类中心$ {\boldsymbol{\mu }}_{{j}} $之间的相似性,采用$ {\boldsymbol{z}}_{{i}} $$ {\boldsymbol{\mu }}_{{j}} $之间的归一化相似性进行软分配.
5)计算软分配分布$ q $与辅助目标分布$ p $之间的KL散度,将二者的KL散度作为聚类损失$ {L}_{{\mathrm{C}}} $,以KL散度最小化作为聚类目标.
6)模型的总损失$ L $由特征提取过程产生的重构误差$ {L}_{\mathrm{R}\mathrm{E}} $和聚类过程产生的聚类误差$ {L}_{{\mathrm{C}}} $组成.
7) 迭代更新模型.
8) end while
输出:交通状态特征矩阵、聚类类别
  
算法2 Bagging-Bo-LightGBM识别模型
输入:LightGBM超参数向量$ \boldsymbol{x}=[{\mathrm{num}}\_{\mathrm{leaves,max}}\_ {\mathrm{depth}},\eta ,{\mathrm{n}}\_ {\mathrm{estimators}}] $、训练数据集$ D $(包含提取后的特征和标签向量)
步骤:
//贝叶斯优化lightGBM超参数过程
1)初始化:超参数、贝叶斯网络结构、迭代次数$ T $.
2) 定义目标函数:准确率$ f\left(\boldsymbol{x}\right) $ .
3)定义超参数空间联合先验分布$ P\left(\boldsymbol{x}\right)={\prod }_{i=1}^{d}P\left({x}_{i}\right) $.
4)选择初始参数点$ ({x}_{1},{x}_{2}{,\cdots ,x}_{n}) $, 计算目标函数值.
5)使用高斯过程(Gaussian process, GP)拟合代理模型,$ \hat{f}\left(\boldsymbol{x}\right)\sim{{\mathrm{G}}}{{\mathrm{P}}}\left(m\left(\boldsymbol{x}\right),k\left(\boldsymbol{x},\bar{\boldsymbol{x}}\right)\right) $.
6) for $ t=\mathrm{1,2},\cdots ,T $ do
7) 依据高斯过程后验分布找到下一个参数点 $ {\boldsymbol{x}}_{\boldsymbol{n}+1} $.
8) $ {\boldsymbol{x}}_{\boldsymbol{n}+1}={\mathrm{argmax}}\;\left\{\mu \left(\boldsymbol{x}\right)+k\sigma \left(\boldsymbol{x}\right)\right\} $.
9) 使用$ {\boldsymbol{x}}_{\boldsymbol{n}+1} $计算目标函数$ f\left(\boldsymbol{x}\right) $值.
10) 更新代理模型.
11)end for
//在优化超参数的基础上,使用Bagging优化分类过程
12)初始化:子采样率、子模型数量$ m $.
13)for $ i=\mathrm{1,2}, \cdots ,m $ do
14) 对训练集$ D $中所有样本进行i次有放回的抽样,生成训练子集$ {D}_{i} $.
15) 在子训练集$ {D}_{i} $上训练基本模型$ {M}_{i} $.
16) end for
17)使用所有基本模型$ {M}_{i} $进行预测,并统计预测结果.
18)采用软投票(soft voting, SV),对分类结果加权平均.
19)最终的分类结果$ {C}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}} $可以表示为$ {C}_{\mathrm{fi}\mathrm{n}\mathrm{a}\mathrm{l}}={\mathrm{argmax}}_{{{C}}} \left(\sum _{i=1}^{m}{m}^{-1}{P}_{i}\left({{C}}\right)\right) $,其中$ {P}_{i}\left({{C}}\right) $为预测概率.
输出:Bagging和Bo共同优化后的LightGBM集成模型
  
图 5  网约车轨迹数据的采集区域
订单编号车辆编号经度/(°)纬度/(°)时间戳
79b55f7533……e14c06bccc0bcb8012……f81a5827108.9460134.25298153912798
79b55f7533……e14c06bccc0bcb8012……f81a5827108.9460234.25296153912801
79b55f7533……e14c06bccc0bcb8012……f81a5827108.9460734.25293153912804
79b55f7533……e14c06bccc0bcb8012……f81a5827108.9460834.25292153912807
79b55f7533……e14c06bccc0bcb8012……f81a5827108.9460834.25293153912811
表 1  网约车轨迹数据的样例
图 6  深度聚类模型的损失函数
图 7  时空单元的聚类数目
图 8  深度聚类特征的三维可视化
模型$ {S}_{{\mathrm{c}}} $$ {C}_{{\mathrm{h}}} $
K-means0.492 265 713.11
FCM0.469 168 899.26
RF-K-means0.684 592 241.37
AE-K-means0.711 6117 002.29
TS-DEC0.886 3204 168.57
表 2  不同模型对时空单元的聚类效果评估
图 9  对应交通运行状态指标的分布特征
类别等级运行状态
A4中度拥堵
B1畅通
C5严重拥堵
D3轻度拥堵
E2缓行
表 3  交通运行状态的量化结果
图 10  各等级路段时间占比的分布特征
图 11  道路交通运行状态的全局空间分布云图
图 12  各交通运行状态识别模型的混淆矩阵
分类模型PR$ {F}_{1-{\mathrm{score}}} $$ {A}_{{\mathrm{cc}}} $
SVM0.815 40.819 20.817 30.817 7
DT0.868 50.883 20.873 80.874 9
XGBoost0.908 30.919 60.913 60.915 4
所提模型0.982 10.984 40.983 30.983 9
表 4  各交通运行状态识别模型的评价结果
1 张琦, 陈红, 周继彪, 等 道路开口对临近交叉口交通安全的影响[J]. 浙江大学学报: 工学版, 2021, 55 (4): 720- 726
ZHANG Qi, CHEN Hong, ZHOU Jibiao, et al Effect of roadway access on traffic safety at adjacent intersection[J]. Journal of Zhejiang University: Engineering Science, 2021, 55 (4): 720- 726
2 XIE J Y, GIRSHICK R, FARHADI A. Unsupervised deep embedding for clustering analysis [C]// International Conference on Machine Learning . New York: JMLR, 2016: 478-487.
3 黎茂盛, 李杭聪 基于交叉口车牌识别数据的网络交通状态分类方法[J]. 哈尔滨工业大学学报, 2023, 55 (11): 82- 90
LI Maosheng, LI Hangcong Classification method of network traffic state based on electronic police data at intersections[J]. Journal of Harbin Institute of Technology, 2023, 55 (11): 82- 90
4 李岩, 王泰州, 徐金华, 等 面向动态交通分配的交通需求深度学习预测方法[J]. 交通运输系统工程与信息, 2024, 24 (1): 115- 123
LI Yan, WANG Taizhou, XV Jinhua, et al Traffic demand prediction method based on deep learning for dynamic traffic assignment[J]. Journal of Transportation Systems Engineering and Information Technology, 2024, 24 (1): 115- 123
5 JAVANI B, BABAZADEH A, CEDRER A Path-based capacity-restrained dynamic traffic assignment algorithm[J]. Transportmetrica B: Transport Dynamics, 2019, 7 (1): 741- 746
doi: 10.1080/21680566.2018.1496861
6 郭璘, 周继彪, 董升, 等 基于改进K-means算法的城市道路交通事故分析[J]. 中国公路学报, 2018, 31 (4): 270- 279
GUO Lin, ZHOU Jibiao, DONG Sheng, et al Analysis of urban road traffic accidents based on improved k-means algorithm[J]. China Journal of Highway and Transport, 2018, 31 (4): 270- 279
7 XU J H, LUO W B, LI Y, et al A multi-directional recurrent graph convolutional network model for reconstructing traffic spatiotemporal diagram[J]. Transportation Letters, 2024, 16 (5): 405- 415
doi: 10.1080/19427867.2023.2198829
8 苏俊杰, 柴干, 季文韬 基于投影寻踪的快速路交织区交通状态识别方法[J]. 交通信息与安全, 2019, 37 (2): 114- 119
SU Junjie, CHAI Gan, JI Wentao Traffic state discrimination method for interchange areas in expressways based on projection pursuit[J]. Journal of Transport Information and Safety, 2019, 37 (2): 114- 119
9 李晓璐, 于昕明, 杜崇, 等 基于权值优化的 FCM-MSVM 算法及其在高速公路状态识别中的应用[J]. 北京交通大学学报, 2018, 42 (4): 72- 78
LI Xiaolu, YU Xinming, DU Chong, et al FCM-MSVM algorithm based on weight optimization and its application in highway state discrimination[J]. Journal of Beijing Jiaotong University, 2018, 42 (4): 72- 78
10 LIU J, KHATTAK A J Delivering improved alerts, warnings, and control assistance using basic safety messages transmitted between connected vehicles[J]. Transportation Research Part C: Emerging Technologies, 2016, 68 (3): 83- 100
11 XU J H, LI Y R, LU W, et al A heterogeneous traffic spatio-temporal graph convolution model for traffic prediction[J]. Physica A: Statistical Mechanics and its Applications, 2024, 641: 129746
doi: 10.1016/j.physa.2024.129746
12 丁瑞, 刘俊, 蒋艳, 等 基于车辆加速度数据的互通立交匝道驾驶风险分析[J]. 交通信息与安全, 2021, 39 (1): 17- 25
DING Rui, LIU Jun, JIANG Yan, et al Driving risks of interchange ramps based on vehicle acceleration data[J]. Journal of Transport Information and Safety, 2021, 39 (1): 17- 25
13 张鑫, 张卫华 快速路合流区主线不同交通状态下的安全性分析[J]. 吉林大学学报: 工学版, 2022, 52 (6): 1308- 1314
ZHANG Xin, ZHANG Weihua Safety analysis of main line under different traffic conditions in expressway confluence area[J]. Journal of Jilin University: Engineering and Technology Edition, 2022, 52 (6): 1308- 1314
14 ZENG Q, GUO Q, WONG S C, et al Jointly modeling area-level crash rates by severity: a Bayesian multivariate random-parameters spatio-temporal Tobit regression[J]. Transportmetrica A: Transport Science, 2019, 15 (2): 1867- 1884
doi: 10.1080/23249935.2019.1652867
15 李岩, 曾明哲, 朱才华, 等 基于多点线圈联合数据的高速公路匝道影响范围识别[J]. 交通运输系统工程与信息, 2022, 22 (3): 53- 62
LI Yan, ZENG Mingzhe, ZHU Caihua, et al Identification of freeway ramp influence areas based on multi-point loop data[J]. Journal of Transportation Systems Engineering and Information Technology, 2022, 22 (3): 53- 62
16 张存保, 彭汉辉, 张珊, 等 雾天高速公路实时交通安全状态评价方法[J]. 中国安全科学学报, 2017, 27 (4): 110- 115
ZHANG Cunbao, PENG Hanhui, ZHANG Shan, et al Real-time traffic safety evaluation method for freeway in fog[J]. China Safety Science Journal, 2017, 27 (4): 110- 115
17 MALIN F, NORROS I, INNAMAA S Accident risk of road and weather conditions on different road types[J]. Accident Analysis and Prevention, 2019, 122 (2): 181- 188
18 吴兵, 翟犇, 卢建涛, 等 基于安全风险的恶劣天气下高速公路建议车速确定方法[J]. 同济大学学报: 自然科学版, 2020, 48 (11): 1570- 1578
WU Bing, ZHAI Ben, LU Jiantao, et al Determination of freeway recommended speed based on safety risk under adverse weather conditions[J]. Journal of Tongji University: Natural Science, 2020, 48 (11): 1570- 1578
19 吕能超, 彭凌枫, 吴超仲, 等 基于视频轨迹参数的边缘率减速标线驾驶行为效果评价方法[J]. 安全与环境学报, 2021, 21 (2): 461- 469
LV Nengchao, PENG Lingfeng, WU Chaozhong, et al Valuative judgment of the driving behaviors of the edge rate deceleration bars based on the video trajectory parameters[J]. Journal of Safety and Environment, 2021, 21 (2): 461- 469
20 YUAN J, ABDELATY M, WANG L, et al Utilizing bluetooth and adaptive signal control data for real-time safety analysis on urban arterials[J]. Transportation Research Part C: Emerging Technologies, 2018, 97 (5): 114- 127
21 杨飞, 姜海航, 刘好德, 等 基于GPS轨迹数据的不同交通状态下交通方式识别流程优化方法[J]. 交通运输系统工程与信息, 2020, 20 (4): 83- 89
YANG Fei, JIANG Haihang, LIU Haode, et al Procedure optimization method based on GPS trajectory data for transportation mode recognition under different traffic conditions[J]. Journal of Transportation Systems Engineering and Information Technology, 2020, 20 (4): 83- 89
[1] 曹宁博,陈家辉,赵利英. 智能网联车和人驾车辆混合交通流排队长度估计模型[J]. 浙江大学学报(工学版), 2024, 58(9): 1935-1944.
[2] 王琴,路庆昌,李建宇,王张鑫,张图. 考虑需求的出租车收益路线推荐[J]. 浙江大学学报(工学版), 2024, 58(6): 1233-1242.
[3] 刘超,孔兵,杜国王,周丽华,陈红梅,包崇明. 高阶互信息最大化与伪标签指导的深度聚类[J]. 浙江大学学报(工学版), 2023, 57(2): 299-309.