Expressway traffic anomaly event detection model based on dual spatio-temporal masks and bidirectional Mamba state space modeling
Article Text (Baidu Translation)
-
摘要: 为了有效处理交通异常事件检测中的挑战，提出了一种基于双重时空掩码与双向Mamba状态空间建模(DSTBM)的交通异常事件检测模型；引入了基于异常事件时空范围和影响范围的掩码，学习异常事件对交通流在时间和空间上的动态影响，自适应遮掩事件发生时段和区域及其影响范围，帮助模型关注流量异常特征；设计了双向Mamba状态空间模型，结合多尺度特征提取与双向状态空间建模，进一步扩展状态空间建模的方向性，捕捉异常事件的短时动态特性和多尺度时空依赖关系，同时缩短检测延迟；构建了结合孤立森林算法与K-means聚类的异常评分方法，根据流量时空特征计算异常得分，区分一般与异常流量状态；将所提出的模型在3个真实数据集上进行评估。研究结果表明：DSTBM模型在SCJ和OKA数据集上均超越了对比试验中的基线模型，其中评估模型准确性的指标精确率、召回率和F1分数分别至少提高了15.32%、14.98%和15.61%，衡量模型响应速度的指标延迟率降低了19.79%；国内案例研究表明，DSTBM在复杂城市道路环境中依然具备有效的异常检测能力，在K=10%时召回率达0.607、总延迟为12.2 min，展示出其在国内高密度、复杂路网中的应用潜力和改进空间。所提出的DSTBM模型能够有效捕捉复杂时空依赖性，并缩短对交通异常事件的检测延迟。Abstract: To address the challenges in traffic anomaly event detection, a traffic anomaly event detection model was proposed based on dual spatio-temporal masks and bidirectional Mamba state space modeling (DSTBM). Masks on the basis of the temporal and spatial extent of anomaly events and their influence range were introduced. The spatio-temporal dynamic effects on traffic caused by anomalies were learned, and the time periods and spatial regions of mask anomalies and their influence range were adapted automatically. A bidirectional Mamba state space model was designed to integrate multi-scale feature extraction with bidirectional state space modeling. The directionality of state space modeling was extended to capture the short-term dynamic features and multi-scale spatio-temporal dependencies of anomalies, and to reduce detection latency. Additionally, an anomaly scoring method was developed through the combination of the isolation forest algorithm with K-means clustering. Anomaly scores were calculated based on spatio-temporal traffic features to distinguish between normal and abnormal traffic states. The proposed model was evaluated on three real datasets. Experimental results demonstrate that DSTBM outperforms all baseline models in the comparative experiments on the SJC and OKA datasets. The indicator precision rate, recall rate, and F1 score, as the evaluation metric for model accuracy, are enhanced by at least 15.32%, 14.98%, and 15.61%, respectively. The indicator latency rate for measuring the model response speed is reduced by 19.79%. In addition, a domestic case study shows that DSTBM maintains effective anomaly detection capabilities in complex urban road environments. The recall rate is 0.607 and the total latency is 12.2 min when K=10%. This highlights its potential applicability and room for enhancement in high-density and intricate road networks. The proposed DSTBM model can effectively capture complex spatio-temporal dependencies and reduce detection latency for traffic anomalies.
-
图 3 不同K下6种方法的检测延迟曲线
Figure 3. Detection delay curves of 6 methods with different K values
图 4 空间距离衰减系数对总召回率的影响
Figure 4. Effects of spatial distance decay coefficients on total recall rate
表 1 2个真实世界的PeMS数据集
Table 1. Two real-world PeMS datasets
参数 SJC OKA 道路总长度/km 227.24 75.99 传感器数量 245 146 事故数量 3 313 3 618 事故类型数量 4 4 事故平均持续时间/min 45.25 47.63 时间范围 2018年5月1日到2018年10月30日 2018年1月1日到2018年6月30日 
下载: 导出CSV
表 2 在SJC数据集的精确率和F1分数评估结果
Table 2. Precisions and F1 results on SJC dataset
方法 在K的总精确率 在K的F1分数 0.2% 1% 2% 3% 5% 10% 0.2% 1% 2% 3% 5% 10% K-means 0.029 0.068 0.099 0.134 0.186 0.243 0.036 0.085 0.124 0.168 0.233 0.304 OmniAnomaly 0.035 0.091 0.129 0.178 0.241 0.303 0.044 0.114 0.162 0.221 0.301 0.379 Anomaly Transformer 0.071 0.124 0.174 0.184 0.286 0.351 0.090 0.155 0.218 0.243 0.357 0.439 STGAN 0.065 0.109 0.172 0.235 0.288 0.362 0.081 0.136 0.215 0.294 0.360 0.452 CST-GL 0.073 0.123 0.175 0.229 0.295 0.371 0.090 0.154 0.219 0.291 0.369 0.464 DSTBM 0.090 0.143 0.221 0.285 0.369 0.454 0.113 0.179 0.276 0.357 0.461 0.568 DSTBM模型相对基线模型中最佳指标提升比例/% 23.29 15.32 26.29 21.28 25.08 22.37 24.72 15.61 26.06 21.33 25.05 22.48
下载: 导出CSV
表 3 在OKA数据集的精确率和F1分数评估结果
Table 3. Precisions and F1 results on OKA dataset
方法 在K的总精确率 在K的F1分数 0.2% 1% 2% 3% 5% 10% 0.2% 1% 2% 3% 5% 10% K-means 0.019 0.059 0.084 0.099 0.125 0.140 0.024 0.074 0.105 0.124 0.156 0.175 OmniAnomaly 0.032 0.068 0.136 0.174 0.217 0.261 0.040 0.093 0.170 0.222 0.276 0.329 Anomaly Transformer 0.068 0.127 0.185 0.185 0.232 0.285 0.085 0.159 0.231 0.251 0.313 0.385 STGAN 0.062 0.134 0.195 0.245 0.290 0.318 0.078 0.168 0.241 0.306 0.374 0.423 CST-GL 0.059 0.121 0.178 0.230 0.290 0.291 0.074 0.151 0.223 0.287 0.363 0.395 DSTBM 0.091 0.170 0.249 0.312 0.402 0.462 0.114 0.212 0.311 0.390 0.504 0.578 DSTBM模型相对基线模型中最佳指标提升比例/% 33.82 26.87 27.69 27.35 38.62 45.28 33.64 26.67 29.21 27.39 34.61 36.49
下载: 导出CSV
表 4 在SJC数据集的召回率和检测延迟评估结果
Table 4. Recall and detection delay results on SJC dataset
方法 在K的总召回率 在K的总延迟/min 0.2% 1% 2% 3% 5% 10% 0.2% 1% 2% 3% 5% 10% K-means 0.048 0.113 0.165 0.224 0.310 0.405 72.3 44.2 34.7 31.8 27.6 23.5 OmniAnomaly 0.058 0.152 0.216 0.290 0.402 0.506 22.5 30.1 17.2 21.8 19.8 16.2 Anomaly Transformer 0.122 0.205 0.290 0.358 0.476 0.585 28.8 30.4 25.1 21.3 18.3 14.8 STGAN 0.109 0.182 0.286 0.392 0.481 0.602 35.7 25.2 22.4 20.9 17.5 13.8 CST-GL 0.118 0.207 0.292 0.399 0.492 0.618 25.8 15.9 22.6 19.2 17.2 13.7 DSTBM 0.150 0.238 0.367 0.476 0.615 0.758 28.9 21.6 17.8 15.4 12.8 9.1 DSTBM模型相对基线模型中最佳指标提升比例/% 22.95 14.98 25.68 19.30 25.00 22.65 -28.44 -35.85 -3.49 19.79 25.58 33.58
下载: 导出CSV
表 5 在OKA数据集的召回率和检测延迟评估结果
Table 5. Recall and detection delay results on OKA dataset
方法 在K的总召回率 在K的总延迟/min 0.2% 1% 2% 3% 5% 10% 0.2% 1% 2% 3% 5% 10% K-means 0.032 0.098 0.140 0.165 0.209 0.234 52.1 38.4 32.2 27.9 24.6 22.3 OmniAnomaly 0.053 0.145 0.227 0.306 0.379 0.445 28.8 34.1 30.3 25.6 22.8 18.9 Anomaly Transformer 0.114 0.212 0.308 0.391 0.483 0.595 29.5 25.1 28.3 24.8 20.5 16.8 STGAN 0.104 0.224 0.315 0.408 0.527 0.632 40.1 28.9 24.7 23.3 21.5 15.1 CST-GL 0.098 0.202 0.297 0.382 0.485 0.616 29.4 19.2 16.7 21.2 20.1 15.6 DSTBM 0.152 0.283 0.415 0.520 0.674 0.770 29.3 23.1 19.6 16.8 13.9 10.5 DSTBM模型相对基线模型中最佳指标提升比例/% 33.33 26.34 31.75 27.45 27.89 21.84 -1.74 -20.31 -17.37 20.75 30.85 30.46
下载: 导出CSV
表 6 在SJC数据集的不同事件类型评估结果
Table 6. Results on SJC dataset of different event types
方法 在10%的总召回率 在10%的总延迟/min 事故 故障 封路 天气 事故 故障 封路 天气 K-means 0.412 0.398 0.379 0.345 23.4 23.8 24.6 25.3 OmniAnomaly 0.478 0.463 0.441 0.423 16.3 16.9 17.8 18.4 Anomaly Transformer 0.603 0.589 0.574 0.561 14.9 15.3 15.7 16.1 STGAN 0.637 0.621 0.613 0.604 13.7 13.9 14.3 14.8 CST-GL 0.649 0.634 0.623 0.619 13.5 13.6 13.9 14.1 DSTBM 0.769 0.762 0.758 0.743 9.3 9.5 9.7 10.1
下载: 导出CSV
表 7 不同变体在SJC数据集的召回率结果
Table 7. Recall results of different variants on SJC dataset
K% 0.2 1 2 3 5 10 w/o MST 0.124 0.197 0.306 0.396 0.514 0.642 w/o MEX 0.138 0.218 0.331 0.426 0.551 0.679 w/o DST 0.092 0.149 0.244 0.326 0.442 0.583 w/o BM 0.135 0.226 0.342 0.454 0.578 0.716 DSTBM 0.150 0.238 0.367 0.476 0.615 0.758
下载: 导出CSV
表 8 有无状态空间建模变体在SJC数据集的效率评估结果
Table 8. Efficiency evaluation results of the variants with or without SSM on SJC dataset
变体 在K的总延迟/min 最大内存使用量/GB 训练周期至收敛轮次 0.2% 1% 2% 3% 5% 10% w/o BM 34.5 27.8 23.1 19.5 16.4 12.7 11.2 30 w/o SSM 41.2 34.6 30.9 27.8 24.5 21.2 17.8 42 DSTBM 28.9 21.6 17.8 15.4 12.8 9.1 14.5 28
下载: 导出CSV
表 9 DSTBM在不同国家数据集的评估结果
Table 9. Evaluation results of DSTBM on different country datasets
数据集 在K的总召回率 在K的总延迟/min 0.2% 1% 2% 3% 5% 10% 0.2% 1% 2% 3% 5% 10% BjTT 0.116 0.192 0.301 0.385 0.491 0.607 36.4 30.8 25.1 20.9 17.3 12.2 SJC 0.150 0.238 0.367 0.476 0.615 0.758 28.9 21.6 17.8 15.4 12.8 9.1 OKA 0.152 0.283 0.415 0.520 0.674 0.770 29.3 23.1 19.6 16.8 13.9 10.5
下载: 导出CSV
-
[1] PAN D, HAMDAR S. From traffic analysis to real-time management: a hazard-based modeling for incident durations extracted through traffic detector data anomaly detection[J]. Transportation Research Record: Journal of the Transportation Research Board, 2024, 2678(2): 389-400. doi: 10.1177/03611981231174445 [2] 陈华, 刘伟. 城市道路交通异常事件检测研究综述[C]//中国公路学会. 2022世界交通运输大会(WTC2022)论文集. 北京: 中国公路学会, 2022: 587-593.CHEN Hua, LIU Wei. A review of research on abnormal traffic incident detection in urban roads[C]//China Highway and Transportation Society. Proceedings of 2022 World Transport Convention (WTC2022). Beijing: China Highway and Transportation Society, 2022: 587-593. [3] 姚磊, 刘渊. 基于ANFIS的交通事件持续时间预测[J]. 计算机工程, 2014, 40(2): 189-192, 198.YAO Lei, LIU Yuan. Traffic incident duration prediction based on adaptive neural-fuzzy inference system[J]. Computer Engineering, 2014, 40(2): 189-192, 198. [4] OUNOUGHI C, YAHIA S B. Data fusion for ITS: a systematic literature review[J]. Information Fusion, 2023, 89: 267-291. doi: 10.1016/j.inffus.2022.08.016 [5] SATTAR S, LI S N, CHAPMAN M. Developing a near real-time road surface anomaly detection approach for road surface monitoring[J]. Measurement, 2021, 185: 109990. doi: 10.1016/j.measurement.2021.109990 [6] KALAIR K, CONNAUGHTON C. Anomaly detection and classification in traffic flow data from fluctuations in the flow-density relationship[J]. Transportation Research Part C: Emerging Technologies, 2021, 127: 103178. doi: 10.1016/j.trc.2021.103178 [7] 王媛媛, 吴光业, 吴克文, 等. 考虑数据异质性和掩蔽问题的高速公路异常交通状况检测[J]. 交通工程, 2024, 24(6): 94-99, 112.WANG Yuan-yuan, WU Guang-ye, WU Ke-wen, et al. Freeway abnormal traffic situation detection considering data heterogeneity and masking[J]. Journal of Transportation Engineering, 2024, 24(6): 94-99, 112. [8] YAN Y, ZHANG S, TANG J J, et al. Understanding characteristics in multivariate traffic flow time series from complex network structure[J]. Physica A: Statistical Mechanics and Its Applications, 2017, 477: 149-160. doi: 10.1016/j.physa.2017.02.040 [9] 姚俊峰, 何瑞, 史童童, 等. 基于机器学习的交通流预测方法综述[J]. 交通运输工程学报, 2023, 23(3): 44-67.YAO Jun-feng, HE Rui, SHI Tong-tong, et al. Review on machine learning-based traffic flow prediction methods[J]. Journal of Traffic and Transportation Engineering, 2023, 23(3): 44-67. [10] WU J, CUI Z M, SHI Y J, et al. Traffic flow anomaly detection based on wavelet denoising and support vector regression[J]. Journal of Algorithms and Computational Technology, 2013, 7(2): 209-225. doi: 10.1260/1748-3018.7.2.209 [11] DJENOURI Y, BELHADI A, LIN J C, et al. Adapted K-nearest neighbors for detecting anomalies on spatio-temporal traffic flow[J]. IEEE Access, 2019, 7: 10015-10027. doi: 10.1109/ACCESS.2019.2891933 [12] KONG X J, SONG X M, XIA F, et al. LoTAD: long-term traffic anomaly detection based on crowdsourced bus trajectory data[J]. World Wide Web, 2018, 21(3): 825-847. doi: 10.1007/s11280-017-0487-4 [13] FITTERS W, CUZZOCREA A, HASSANI M. Enhancing LSTM prediction of vehicle traffic flow data via outlier correlations[C]//IEEE. 2021 IEEE 45th Annual Computers, Software, and Applications Conference (COMPSAC). New York: IEEE, 2021: 210-217. [14] 郑大庆, 林陈威, 王昺杰. 基于注意力-长短期记忆模型的偶发性交通流异常检测[J]. 同济大学学报(自然科学版), 2023, 51(6): 923-931.ZHENG Da-qing, LIN Chen-wei, WANG Bing-jie. Traffic flow occasional anomaly detection based on attention-LSTM model[J]. Journal of Tongji University (Natural Science), 2023, 51(6): 923-931. [15] ZHU G X, ZHAO H B, LIU H Q, et al. A novel LSTM-GAN algorithm for time series anomaly detection[C]//IEEE. 2019 Prognostics and System Health Management Conference (PHM-Qingdao). Qingdao: IEEE, 2019: 1-6. [16] WANG S, ZHANG Y, HU Y L, et al. Knowledge fusion enhanced graph neural network for traffic flow prediction[J]. Physica A: Statistical Mechanics and Its Applications, 2023, 623: 128842. doi: 10.1016/j.physa.2023.128842 [17] LIU T B, ZHANG J D. An adaptive traffic flow prediction model based on spatiotemporal graph neural network[J]. The Journal of Supercomputing, 2023, 79(14): 15245-15269. doi: 10.1007/s11227-023-05261-9 [18] 崔建勋, 要甲, 赵泊媛. 基于深度学习的短期交通流预测方法综述[J]. 交通运输工程学报, 2024, 24(2): 50-64.CUI Jian-xun, YAO Jia, ZHAO Bo-yuan. Review on short- term traffic flow prediction methods based on deep learning[J]. Journal of Traffic and Transportation Engineering, 2024, 24(2): 50-64. [19] JIN M, KOH H Y, WEN Q S, et al. A survey on graph neural networks for time series: forecasting, classification, imputation, and anomaly detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 10466-10485. doi: 10.1109/TPAMI.2024.3443141 [20] ZHENG Y, KOH H Y, JIN M, et al. Correlation-aware spatial-temporal graph learning for multivariate time-series anomaly detection[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 35(9): 11802-11816. [21] HU Y, QU A, WORK D. Detecting extreme traffic events via a context augmented graph autoencoder[J]. ACM Transactions on Intelligent Systems and Technology, 2022, 13(6): 1-23. [22] ZHANG H Y, ZHAO S Y, LIU R H, et al. Automatic traffic anomaly detection on the road network with spatial-temporal graph neural network representation learning[J]. Wireless Communications and Mobile Computing, 2022, 2022(1): 4222827. doi: 10.1155/2022/4222827 [23] DENG L Y, LIAN D F, HUANG Z Y, et al. Graph convolutional adversarial networks for spatiotemporal anomaly detection[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(6): 2416-2428. doi: 10.1109/TNNLS.2021.3136171 [24] WANG C Z, XING S W, GAO R, et al. Disentangled dynamic deviation transformer networks for multivariate time series anomaly detection[J]. Sensors, 2023, 23(3): 1104. [25] QU H H, NING L B, AN R, et al. A survey of Mamba[J]. arXiv, (2025-06-17). https://doi.org/10.48550/arXiv.2408.01129 .[26] YUAN D, XUE J Z, SU J S, et al. ST-mamba: spatial-temporal mamba for traffic flow estimation recovery using limited data[C]//IEEE. 2024 IEEE/CIC International Conference on Communications in China (ICCC). New York: IEEE, 2024: 1928-1933. [27] TAN X, ZHAO M. Self-supervised state space model for real-time traffic accident prediction using eKAN networks[J]. arXiv, (2024-09-09) https://doi.org/10.48550/arXiv.2409.05933 .[28] ZHANG Y Z, TIWARI P, ZHENG Q, et al. A multimodal coupled graph attention network for joint traffic event detection and sentiment classification[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(8): 8542-8554. [29] 苏倩华. 基于多源感知融合数据的交通异常事件检测算法[D]. 大连: 东北财经大学, 2023.SU Qian-hua. Traffic abnormal event detection algorithm based on multi-source perception fusion data[D]. Dalian: Dongbei University of Finance and Economics, 2023. [30] ZHOU W, YU Y H, ZHAN Y F, et al. A vision-based abnormal trajectory detection framework for online traffic incident alert on freeways[J]. Neural Computing and Applications, 2022, 34(17): 14945-14958. [31] SHAO Z Q, BELL M G H, WANG Z, et al. Spatial-temporal selective state space (St-Mamba) model for traffic flow prediction[J]. arXiv, (2024-05-18). https://doi.org/10.48550/arXiv.2404.13257 .[32] CHOI J, KIM H, AN M, et al. SpoT-Mamba: learning long-range dependency on spatio-temporal graphs with selective state spaces[J]. arXiv, (2024-06-17). https://doi.org/10.48550/arXiv.2406.11244 .[33] ZHANG H W, ZHU Y, WANG D, et al. A survey on Visual Mamba[J]. arXiv, (2024-04-26). https://doi.org/10.48550/arXiv.2404.15956 .[34] SUO Y F, DING Z N, ZHANG T. The mamba model: a novel approach for predicting ship trajectories[J]. Journal of Marine Science and Engineering, 2024, 12(8): 1321. [35] HE K, CHEN X, XIE S, et al. Masked autoencoders are scalable vision learners[C]//IEEE. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. New York: IEEE, 2022: 16000-16009. [36] 宫晓婞, 董培信. 基于改进孤立森林算法的交通流异常数据检测模型[J]. 重庆交通大学学报(自然科学版), 2024, 43(5): 61-69, 90.GONG Xiao-xing, DONG Pei-xin. Traffic flow anomaly data detection model based on improved isolation forest algorithm[J]. Journal of Chongqing Jiaotong University (Natural Science), 2024, 43(5): 61-69, 90. [37] LIU C, CHEN J M, LIU H Y, et al. Toward efficient traffic incident detection via explicit edge-level incident modeling[J]. IEEE Internet of Things Journal, 2024, 11(11): 20015-20029. [38] SU Y, ZHAO Y J, NIU C H, et al. Robust anomaly detection for multivariate time series through stochastic recurrent neural network[C]//ACM. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2019: 2828-2837. [39] XU J, WU H, WANG J, et al. Anomaly transformer: time series anomaly detection with association discrepancy[C]//ICLR. Proceedings of the 10th International Conference on Learning Representations (ICLR). Appleton: ICLR, 2022: 1-20. [40] ZHANG C Y, ZHANG Y, SHAO Q T, et al. BjTT: a large- scale multimodal dataset for traffic prediction[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(11): 18992-19003. -
图(5) / 表(9)
计量
- 文章访问数: 212
- HTML全文浏览量: 68
- PDF下载量: 14
- 被引次数: 0
