Traffic abnormity early warning method based on vehicle-road cloud cooperation

Abstract

The invention discloses a vehicle-road cloud cooperative traffic abnormity early warning method, which comprises the following steps that: a road side unit receives abnormal event data reported by a vehicle side, checks the abnormal event data in combination with road side data, and uploads the abnormal event data to a cloud side; the cloud receives the data reported by the road side unit, performs fusion in combination with the cloud data, and evaluates the risk level after the occurrence of the abnormal event: constructs a space-time diagram neural network, and calculates the vulnerability index of the road network based on the traffic road network diagram structure; constructing a Bayesian network, updating probability distribution of the network by using detected data corresponding to the abnormal event as an observation value, and predicting a conduction path of the abnormal event; carrying out preliminary risk grade assessment according to the vulnerability index and the predicted conduction path; and performing traffic intervention simulation based on the abnormal event, calculating an intervention index improvement rate according to a simulation result, and correcting a risk level. According to the method, deep fusion is carried out on the vehicle and road cloud data, the traffic abnormal event identification capability is improved, and the dynamic risk level evaluation of the whole road network is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic control systems, and in particular to a vehicle-road-cloud collaborative traffic anomaly warning method. Background Art

[0002] Traditional traffic control anomaly warning systems are hampered by the triple constraints of data fragmentation, shallow models, and rigid systems. Traffic anomaly events involve multimodal data, including vehicle-side data (video, GPS, and CAN bus), roadside data (radar and signal phase), and cloud data (meteorological data and historical event libraries). The lack of a unified spatiotemporal benchmark and standardized interfaces makes cross-modal semantic alignment difficult. This is particularly true for long-tail scenarios such as rainstorms and accidents. Traffic anomalies exhibit spatiotemporal propagation (e.g., congestion spreading to upstream intersections). Traditional RNN / LSTM models lag in modeling the temporal correlations of congestion propagation and ignore the spatial cascading effects of road network topology, resulting in a high rate of misjudgment of anomalies. Existing solutions often rely on centralized cloud-based processing, which suffers from extended data transmission times and low utilization of edge computing resources. Furthermore, the coordination mechanisms between the vehicle, road, and cloud layers are loose, lacking dynamic resource scheduling capabilities, and thus failing to meet the requirements for real-time warnings and accurate decision-making. These shortcomings collectively hinder improvements in traffic situational awareness and emergency response efficiency.

[0003] The "Multi-dimensional Traffic Safety Anomaly Analysis and Warning System, Method, and Program Product," disclosed in Chinese patent documents with publication number CN119107808A and publication date December 10, 2024, includes a server and a mobile device. The server comprises a multi-dimensional data acquisition module, a data preprocessing module, a feature extraction module, an anomaly detection module, a behavior analysis module, a risk assessment module, and a risk evaluation module. The risk assessment module is configured to identify the location information corresponding to the anomaly on the current road based on the output road environment anomaly data and vehicle behavior analysis data, determine the risk level corresponding to the current road based on a preset risk assessment strategy, and formulate a corresponding warning strategy based on the corresponding risk level; and the transmission module is configured to transmit a corresponding warning signal to the mobile device within a preset range of the anomaly location based on the formulated warning strategy. Although this technology also uses multi-dimensional data for traffic safety anomaly analysis and warning, its multi-dimensional data is primarily based on historical data and vehicle-side data from different mobile devices. It does not perform collaborative analysis of vehicle-side data, roadside data, and cloud data, lacks dynamic resource scheduling capabilities, and cannot meet the needs of real-time warning and accurate decision-making. Summary of the Invention

[0004] The present invention aims to overcome the problems in the existing technology of loose coordination mechanism among vehicle-road-cloud layers, lack of dynamic resource scheduling capability, inability to meet the needs of real-time warning and accurate decision-making, high misjudgment rate and low timeliness, and provides a traffic anomaly warning method with vehicle-road-cloud coordination.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A vehicle-road-cloud collaborative traffic anomaly warning method, comprising: The roadside unit receives abnormal event data reported by the vehicle, verifies it with roadside data, and uploads it to the cloud; The cloud receives data reported by the roadside unit and integrates it with the cloud data to assess the risk level after an abnormal event occurs: Construct a spatiotemporal graph neural network and calculate the vulnerability index of the road network based on the traffic network graph structure; Construct a Bayesian network and use the data corresponding to the detected abnormal events as observations to update the network's probability distribution and predict the conduction path of the abnormal events; Conduct preliminary risk assessment using vulnerability index and predicted transmission pathways; Traffic intervention simulation is carried out based on abnormal events, the intervention indicator improvement rate is calculated according to the simulation results, and the risk level is revised.

[0006] The present invention improves the ability to identify traffic anomalies by deeply integrating vehicle-road-cloud data. Combined with the reasoning ability of the cloud-based multimodal large model, it realizes dynamic risk assessment of the entire road network and formulates a corresponding hierarchical response mechanism, which can handle abnormal events accurately and in real time. As for the method of traffic risk assessment, a preliminary risk assessment is first performed based on the vulnerability index of the road network and the prediction results of the abnormal event transmission path, and then the intervention index improvement rate is obtained according to the simulation of the intervention effect of the abnormal event. The preliminary risk assessment result is corrected with the intervention index improvement rate to obtain a risk assessment result that is more in line with the actual road conditions, avoiding secondary traffic anomalies caused by the resource mismatch between the assessed risk level and the actual road status.

[0007] Preferably, the verification in combination with roadside data includes: Map vehicle-side data and roadside data into a unified map coordinate system for consistency verification; If the consistency check result does not meet the consistency threshold requirement, recheck; If the consistency check result meets the consistency threshold requirement, data within a certain time period before and after the abnormal event occurs will be intercepted and uploaded to the cloud.

[0008] Preferably, the traffic network graph structure is established by taking intersections as nodes and road sections connecting upstream and downstream intersections as edges; The vulnerability index of the road network is the weighted sum of several vulnerability indicators; The vulnerability indicators include at least the environmental risk exposure that reflects the comprehensive risk level faced by the road network, and the emergency resource coverage that measures the adequacy of emergency resources provided after an emergency.

[0009] Preferably, the environmental risk exposure is a weighted sum of the accident frequency, the real-time environmental factor, and the topological risk factor; The frequency of accidents comes from the accident archive data in the cloud data; The real-time environmental factor is the maximum value of the precipitation factor and the visibility factor, and is derived from the meteorological bureau data in the cloud data; The topological risk factor is the product of betweenness centrality and environmental sensitivity coefficient.

[0010] Preferably, the emergency resource coverage is a weighted sum of time coverage and resource adequacy; Time coverage is the proportion of all emergencies that meet the emergency response requirements; Resource adequacy is the ratio of the number of available resources that meet emergency response conditions to the number of required resources predicted based on the severity of the incident.

[0011] Preferably, the Bayesian network construction includes: taking the external inducement as the root node, the road network state as the intermediate node, the abnormal event as the leaf node, and the edges between the nodes to represent the causal relationship; Each node contains at least a spatiotemporal attribute tag and a severity attribute; at the same time, it is constrained that there are edges between nodes that are adjacent in space or time.

[0012] Preferably, the performing of the preliminary risk level assessment includes: If the vulnerability index is greater than the first index threshold and the transmission path prediction result is greater than or equal to the third transmission level, it is a Level I risk; The vulnerability index is greater than the second index threshold and less than or equal to the first index threshold, and the conduction path prediction result is the second conduction level, which is Level II risk; The vulnerability index is greater than the third index threshold and less than or equal to the second index threshold, and the conduction path prediction result is the first conduction level, which is a Level III risk.

[0013] Preferably, the modified risk level includes: For Level I risk, if the improvement rate of several pre-indicators exceeds the upper limit of Level I risk, it will be downgraded to Level II risk; For Level II risk, if the improvement rate of some pre-indicators exceeds the upper limit of Level II risk, it will be downgraded to Level III risk; if the improvement rate of some pre-indicators is less than the lower limit of Level II risk, it will be upgraded to Level I risk; For Level III risk, if the improvement rate of several pre-indicators is greater than the upper limit of Level III risk, the warning will be lifted; if the improvement rate of several pre-indicators is less than the lower limit of Level III risk, it will be upgraded to Level II risk.

[0014] Preferably, the intervention indicator improvement rate is a weighted sum of the congestion duration reduction rate and the impact range contraction rate, and the corresponding weights are calculated by an entropy weight method; The congestion duration reduction rate is the ratio of the congestion duration reduction after the intervention simulation to the actual duration without intervention; The impact range shrinkage rate is the ratio of the reduction in the impact range after the intervention simulation to the actual impact range without intervention.

[0015] Preferably, the step of taking a road segment connecting an upstream and downstream intersection as an edge includes updating the weight of the edge; The weight of an edge is the weighted sum of several weight indicators of the road segment corresponding to the edge; the weight indicators include: Real-time traffic density distribution, the ratio of the actual number of vehicles to the number of vehicles that can be accommodated; vehicle delay index, the ratio of actual travel time to free flow time; lane availability, the proportion of lanes closed due to abnormal events; weather impact factor, the traffic efficiency attenuation coefficient dynamically adjusted based on meteorological data.

[0016] The present invention has the following beneficial effects: deep integration of vehicle-road-cloud data improves the ability to identify traffic anomalies, combines the reasoning ability of the cloud-based multimodal large model, realizes dynamic risk assessment of the entire road network, and formulates a corresponding hierarchical response mechanism, which can accurately and in real time handle abnormal events; for the method of traffic risk assessment, a preliminary risk assessment is first performed based on the vulnerability index of the road network and the prediction results of the abnormal event transmission path, and then the intervention index improvement rate is obtained according to the simulation of the intervention effect of the abnormal event. The preliminary risk assessment result is corrected with the intervention index improvement rate to obtain a risk assessment result that is more in line with the actual road conditions, avoiding secondary traffic anomalies caused by the resource mismatch between the assessed risk level and the actual road status. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a vehicle-road-cloud collaborative traffic anomaly warning method in the present invention.

[0018] Figure 2 This is a flow chart of the present invention for performing preliminary risk level assessment using vulnerability index and predicted conduction path.

[0019] Figure 3 This is a flow chart of correcting risk levels based on intervention indicator improvement rates in the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, a vehicle-road-cloud collaborative traffic anomaly warning method includes: The roadside unit receives abnormal event data reported by the vehicle, verifies it with roadside data, and uploads it to the cloud; The cloud receives data reported by the roadside unit and integrates it with the cloud data to assess the risk level after an abnormal event occurs: Construct a spatiotemporal graph neural network and calculate the vulnerability index of the road network based on the traffic network graph structure; Construct a Bayesian network and use the data corresponding to the detected abnormal events as observations to update the network's probability distribution and predict the conduction path of the abnormal events; Conduct preliminary risk assessment using vulnerability index and predicted transmission pathways; Traffic intervention simulation is carried out based on abnormal events, the intervention indicator improvement rate is calculated according to the simulation results, and the risk level is revised.

[0022] It should be noted that the present invention improves the ability to identify abnormal traffic events by deeply integrating vehicle-road-cloud data. Combined with the reasoning ability of the cloud-based multimodal large model, it realizes the dynamic risk assessment of the entire road network and formulates a corresponding hierarchical response mechanism, which can handle abnormal events accurately and in real time. For the method of traffic risk assessment, a preliminary risk assessment is first performed based on the vulnerability index of the road network and the prediction results of the abnormal event transmission path, and then the intervention index improvement rate is obtained according to the simulation of the intervention effect of the abnormal event. The preliminary risk assessment result is corrected with the intervention index improvement rate to obtain a risk assessment result that is more in line with the actual road conditions, avoiding secondary traffic anomalies caused by the resource mismatch between the assessed risk level and the actual road status.

[0023] It is worth noting that the specific steps mainly include three parts: multi-data fusion of vehicle-side, roadside and cloud side, preliminary risk level assessment based on vulnerability index and transmission path on the cloud side, and risk level correction based on intervention simulation results.

[0024] A time synchronization system based on the PTP protocol is established to achieve us-level time synchronization between the vehicle-side OBD module, roadside MEC equipment and cloud server through hardware timestamps.

[0025] A lightweight MobileNet model is deployed on the vehicle side, combined with vehicle-side data to achieve real-time abnormal event detection and upload structured data via V2X communication. Vehicle-side data includes vehicle sensor data, vehicle status data, and V2X communication data. After receiving abnormal event data reported by the vehicle side, the roadside unit (RSU) combines it with roadside equipment data, performs a multi-source data verification process, and uploads it to the cloud after compression. Roadside equipment data includes traffic flow monitoring data, signal control data, and environmental perception data. The cloud combines the abnormal event data reported by the roadside unit with cloud data, enabling a large multimodal model to fuse abnormal event data and achieve spatiotemporal feature inference in complex scenarios. Cloud data includes historical databases, dynamically fused data, and cross-domain collaborative data.

[0026] A spatiotemporal graph neural network model is constructed, using intersections as nodes and road sections connecting upstream and downstream intersections as edges. The vulnerability index of the entire road network is calculated in real time. A Bayesian network structure is constructed, using detected abnormal events and their corresponding attributes as observations. The probability distribution of the entire network is updated, and the transmission paths of abnormal events are predicted. A preliminary risk level assessment is conducted by combining the vulnerability index of the entire road network and the transmission paths of abnormal events.

[0027] Based on abnormal events that have occurred, counterfactual reasoning and intervention simulation effect evaluation are performed to generate the optimal response strategy. The improvement rate of intervention indicators is calculated based on the optimal corresponding strategy, and the risk level and response mechanism are dynamically adjusted.

[0028] As a specific embodiment, multi-data fusion of vehicle-side, roadside-cloud includes four parts: time synchronization, vehicle-side data processing, roadside data processing, and cloud-side data processing.

[0029] For time synchronization, a PTP Grandmaster master clock is deployed in the cloud. This clock obtains nanosecond-level reference time from GPS / Beidou satellite signals or atomic clocks. Roadside units (RSUs) serve as PTP boundary clocks, synchronizing with the cloud master clock via a dedicated fiber-optic network. This is also extended to vehicles using the 802.1AS-Rev onboard PTP protocol. This ensures nanosecond-level time synchronization between vehicle, roadside, and cloud data after multi-source data is integrated into the cloud, facilitating the fusion of diverse data.

[0030] Vehicle-side data processing operations include vehicle sensor data, vehicle status data, and V2X communication data. The acquired vehicle sensor data includes real-time video streams from onboard cameras, millimeter-wave / LiDAR 3D target detection data, and high-precision GPS fusion trajectory data. Vehicle status data includes abnormal status signals such as sudden braking, ABS activation, airbag triggering, sudden steering wheel angle changes, and frequent lane changes. V2X communication data includes vehicle speed, heading angle, brake status broadcast data, eCall emergency call signals, and cooperative collision warning.

[0031] The lightweight MobileNet model uses depthwise separable convolution technology. When it detects events such as traffic accidents (such as vehicle collisions and sudden braking), road obstacles (such as spilled objects), or abnormal traffic lights, it encapsulates structured feature data, including event type, bounding box coordinates, confidence level, and timestamp, into a lightweight JSON format and then uploads it to the roadside unit (RSU) via the V2X communication module. The upload strategy uses a dynamic threshold trigger mechanism: transmission is initiated only when the detection confidence exceeds a preset threshold and the event duration exceeds the threshold. Events are then graded based on their urgency. For example, major events (such as fires) immediately trigger direct transmission over 5G high-speed channels, while more common events are cached locally and compressed in batches for transmission over LTE.

[0032] Roadside data processing operations include traffic flow monitoring data, signal control data, and environmental perception data. Traffic flow monitoring data includes lane-level flow, occupancy, and average speed collected by geomagnetic or radar sensors, as well as queue lengths detected by video recognition or radar. Signal control data includes signal phase status such as the remaining green light duration, phase switching logs, and bus / ambulance priority signal requests. Environmental perception data includes road surface temperature, water depth, visibility, and roadside camera / dome video streams.

[0033] The roadside unit receives abnormal event data reported by the vehicle and verifies it in combination with the roadside data, including: Map vehicle-side data and roadside data into a unified map coordinate system for consistency verification; If the consistency check result does not meet the consistency threshold requirement, recheck; If the consistency check result meets the consistency threshold requirement, data within a certain time period before and after the abnormal event occurs will be intercepted and uploaded to the cloud.

[0034] Specifically, after receiving abnormal event data reported by the vehicle, the roadside unit initiates a data consistency verification process. Using SLAM (Simultaneous Localization and Mapping) technology, the vehicle's GPS coordinates and detection results from roadside cameras and millimeter-wave radar are mapped to a unified high-precision map coordinate system (with an error of less than 0.3 meters). Verification within this same map coordinate system eliminates spatial offsets between devices and establishes a consistent observation environment in both time and space.

[0035] Kalman filtering is used to generate a comprehensive trajectory and speed curve, and the roadside data and vehicle-side data are compared for consistency. The existing methods that can perform consistency verification can achieve the purpose of verification in this implementation, and there are no specific selection requirements. As an example of this embodiment, the trajectory consistency is judged by calculating the point sequence similarity through the dynamic time warping algorithm. Speed ​​verification is performed by comparing the speed difference estimated by radar Doppler speed measurement and video optical flow method (difference threshold <10%). The acceleration matching is analyzed by cross-checking the inertial measurement unit data and the second-order derivative of the trajectory.

[0036] When using the dynamic time warping algorithm to calculate point sequence similarity for trajectory consistency judgment, the sequence points are aligned by bending the time axis to adapt to the speed change scenarios of the traffic section, and consistency is determined by a standardized Euclidean distance threshold. When speed verification is performed by comparing the speed difference estimated by radar Doppler speed measurement and video optical flow method, radar Doppler speed measurement can directly measure speed using microwave phase difference, while video optical flow method can use the LK algorithm to track pixel displacement to measure speed. Speed ​​consistency is verified based on the speed difference between the two measurements. The difference threshold can be appropriately relaxed in high-speed scenarios. When analyzing acceleration matching by cross-checking inertial measurement unit data and the second-order derivative of the trajectory, the inertial measurement unit data can obtain acceleration values ​​through a three-axis accelerometer, and the second-order derivative of the trajectory can be obtained by performing a difference fit on the trajectory and then taking the derivative.

[0037] After separately assessing the consistency of trajectory, velocity, and acceleration, the system outputs a consistency coefficient ranging from 0 to 1 based on a confidence scoring model (such as DS evidence theory). If the coefficient falls below the consistency threshold of 0.7, revalidation or flagging of the data as low-confidence is triggered. If the coefficient exceeds the consistency threshold of 0.7, the system automatically triggers an intelligent data filtering mechanism, capturing only 10 seconds of video footage from the core period of the abnormal event (5 seconds before and after the abnormal event, or 2 seconds before and 8 seconds after the abnormal event, or other preset time slices) and associated metadata, including the accident location, vehicle ID, and environmental parameters. Using feature summary generation technology, the original data is compressed to 15% before uploading to the cloud. This strategy effectively ensures the integrity and real-time availability of abnormal event data.

[0038] For cloud data processing operations, cloud data includes historical databases, dynamic fusion data, and cross-domain collaborative data. The acquired historical database includes an accident archive database and a traffic pattern database. The accident archive database contains information such as the time, location, type, and casualty data of historical accidents, and the traffic pattern database contains information such as traffic patterns on weekdays / holidays and characteristics of morning and evening peaks. Dynamic fusion data includes meteorological data such as rainfall intensity, wind speed, and road icing warning levels accessed through the Meteorological Bureau API, high-precision map data such as lane-level topology, slope curvature, and construction area electronic fence coordinates, as well as urban event data such as large-scale events and road construction closures. Cross-domain collaborative data includes bus dispatch data such as real-time bus arrival times and priority route requirements, as well as emergency rescue data such as 120 and 119 dispatch route planning and rescue channels that require priority protection.

[0039] After the cloud data is fused with the abnormal event data uploaded by the roadside unit, it is used as data support for subsequent traffic risk assessment and traffic intervention simulation, realizing the correlation and interaction of 2D visual images, 3D radar point clouds and text report data in the vehicle-side, roadside and cloud data.

[0040] As an optional embodiment, the method for fusing cloud data with abnormal event data uploaded by roadside units includes the following: after the cloud accesses multi-source data, it first uses the PTP protocol to ensure nanosecond time synchronization of vehicle-side, roadside, and cloud data, and uses SLAM technology to map radar point clouds, camera video streams, meteorological sensor data, etc. to a unified high-precision map space. Then, feature fusion is initiated: in the early fusion stage, the 2D visual detection box and the 3D radar point cloud are fused through spatial projection and target association algorithms to construct cross-modal target features; in the semantic fusion stage, the CLIP model is used to align text reports with visual content to verify logical consistency. Mid-term fusion adopts a multimodal Transformer architecture and uses a cross-modal attention mechanism to interact visual features with radar features to enhance the robustness of anomaly detection in complex scenarios such as heavy rain and at night.

[0041] Specifically, the Precision Time Protocol (PTP) enables nanosecond-level time synchronization between the vehicle (onboard sensors), roadside units (RSUs), and the cloud, eliminating data misalignment caused by sensor clock drift (for example, time deviation between radar point clouds and video frames), laying the foundation for multi-source data fusion. SLAM (Simultaneous Localization and Mapping) technology takes input data such as radar point clouds, camera video streams, and roadside weather data. Through point cloud registration (using the Iterative Closest Point (ICP) algorithm to align the radar point cloud with the map), and visual-to-map matching (using feature points such as SIFT and ORB to associate the image with the HD map semantic layer), all sensor data is mapped to a unified HD map coordinate system.

[0042] After completing spatiotemporal synchronization and spatial mapping, early fusion and semantic fusion are performed. Early fusion enables cross-modal object-level association, linking 2D visual objects (such as vehicles and pedestrians) with 3D radar point cloud objects. This process includes spatial projection, which projects the visual bounding box (2D Bounding Box) into 3D radar space using camera extrinsics; object association, which uses the Hungarian algorithm or Intersection over Union (IoU) to match the projected bounding box with the radar bounding box; and feature construction, which fuses visual texture / color features (RGB) with radar spatial / velocity features (point cloud reflectivity + Doppler velocity) to generate a cross-modal object vector.

[0043] Semantic fusion enables text-visual consistency verification. The CLIP model inputs text and video image data, such as roadside weather reports (e.g., "Heavy rain reduces visibility") and camera video frames. The text encoder uses the text data to generate corresponding semantic vectors. The image encoder extracts visual feature vectors from the video image data. After feature extraction, a cosine similarity comparison verifies the logical consistency between the text description and the image content (for example, determining whether heavy rain has actually affected visibility). The model also outputs a consistency score, which serves as a confidence weight for subsequent decision-making.

[0044] The cross-modal attention mechanism of the multimodal Transformer architecture can input the results of previous early fusion and semantic fusion, and interact with visual and radar features to improve detection robustness in complex scenarios. During this cross-modal interaction, visual features can be used to help the radar focus on key areas (such as occluded targets or key areas). Radar detection data can also be used to supplement scenarios where visual visibility fails (such as at night or in foggy conditions). Dynamic learning of the contribution weights of each modality in complex scenarios (for example, increasing the weight of radar features and decreasing the weight of visual features in heavy rain or fog; or reducing the weights of low-confidence data features based on the confidence weights after semantic fusion; and other weight adjustment methods) can be used to obtain more accurate results and complete target information from multiple angles.

[0045] As a specific embodiment, the preliminary risk level assessment based on the vulnerability index and the transmission path in the cloud includes the calculation of the vulnerability index, the prediction of the transmission path, and a method for combining the two to perform preliminary risk level assessment.

[0046] First, we need to define the structure of the traffic network graph. The traffic network graph consists of nodes and edges, where each intersection is defined as a node. The attributes of the node include real-time traffic flow, average vehicle speed, signal phase status, and historical accident frequency. The road sections connecting upstream and downstream intersections are defined as edges.

[0047] The vulnerability index of the entire road network is calculated based on the traffic network graph structure. This index is a key indicator that measures the network's resilience and recovery capabilities when faced with emergencies such as traffic accidents, natural disasters, or peak traffic congestion. The vulnerability index is a weighted sum of several vulnerability indicators; the weights corresponding to each vulnerability indicator are adaptive and can be learned and trained using a constructed spatiotemporal graph neural network.

[0048] The process of automatic learning weights of the STGNN model is achieved through spatiotemporal feature fusion + back propagation optimization. Its technical process includes: spatial dependency modeling (graph convolution layer); temporal dynamic modeling (gated timing layer); weight generation (fully connected layer); weight optimization (calculation loss, back propagation).

[0049] Spatial dependency modeling (graph convolutional layer) processes spatial relationships in graph-structured data. Through graph convolution operations (GCN), the model aggregates feature information from adjacent nodes and learns dependency patterns in the spatial dimension. For example, in traffic networks, it can capture the mutual influence of traffic conditions at adjacent intersections. The adjacency matrix defines the spatial relationships between nodes, and the convolution kernel parameters are automatically optimized during training to capture effective spatial patterns.

[0050] Temporal dynamics modeling (gated temporal layers) processes the temporal dynamics of sequence data. Gated recurrent units (GRUs) or temporal convolutional networks (TCNs) are typically used to capture temporal trends, periodicity, and short-term dependencies. GRUs use gating mechanisms (update and reset gates) to control the forgetting and updating of historical information, effectively modeling the evolution of successive time steps. Compared to traditional RNNs, GRUs alleviate the vanishing gradient problem and can learn longer temporal dependencies.

[0051] Weight generation (fully connected layer): After the spatiotemporal features are fused, the fully connected layer integrates and maps them to the target output space (such as predicted values ​​or classification probabilities). This layer contains a set of learnable weight matrices and bias vectors, responsible for transforming high-dimensional spatiotemporal features into final predictions. These weights are randomly set during model initialization and continuously adjusted during training.

[0052] Weight optimization (loss calculation, backpropagation): The model calculates the loss (e.g., mean squared error (MSE)) based on the predicted results and the true labels. Backpropagation propagates gradients from the output layer back to the input layer layer by layer. An optimizer (e.g., Adam) updates the weights of all layers (graph convolutional layers, gated temporal layers, and fully connected layers) based on the gradients. Iterative training minimizes the loss, enabling automatic learning of weights and optimizing model performance.

[0053] It should be noted that vulnerability indicators include at least environmental risk exposure, which reflects the overall risk level faced by the road network, and emergency resource coverage, which measures the adequacy of emergency resources after an emergency. The final vulnerability index is calculated by multiplying the environmental risk exposure by its corresponding weight, adding the product of the emergency resource coverage by its corresponding weight (due to the negative correlation between emergency resource coverage and the vulnerability index, the corresponding weight of the emergency resource coverage can be limited to a negative number during calculation; alternatively, the corresponding weight can be multiplied by taking one minus the emergency resource coverage), and then adding the products of several other commonly used vulnerability indicators and their corresponding weights.

[0054] Specifically, environmental risk exposure reflects the comprehensive risk level faced by a node or edge. It is a weighted sum of accident frequency, real-time environmental factors, and topological risk factors. The corresponding weights are empirical weight coefficients that can be pre-set and dynamically optimized through machine learning. The specific calculation data source is data fused with cloud data.

[0055] Accident frequency is derived from accident archive data in the cloud, calculated as the time-decay-weighted annual number of accidents divided by a standardized benchmark (default is 2). The real-time environmental factor is the maximum of the precipitation factor and the visibility factor, derived from the Meteorological Bureau API data in the cloud. The precipitation factor is the product of the precipitation coefficient (which can be 0.3) and the precipitation intensity (in millimeters per hour), and the visibility factor is one minus the visibility (in kilometers). The topological risk factor is the product of betweenness centrality and the environmental sensitivity coefficient. Betweenness centrality is a conventional vulnerability indicator, while the environmental sensitivity coefficient reflects the degree of impact of the actual environment. Its base value is 1, and the coefficient is adjusted upward or downward based on the actual environment (for example, if the road has low drainage capacity, 0.3 is added to the base value of 1; if the sight distance on a curve is short, 0.4 is added to the base value of 1; if the road has no protective facilities, 0.2 is added to the base value of 1, etc.).

[0056] Specifically, emergency resource coverage measures the ability of emergency resources (such as traffic police, tow trucks, and medical teams) to quickly arrive at the scene and provide adequate services after an emergency occurs. It is a weighted sum of time coverage and resource adequacy. The corresponding weight is an empirical weight coefficient, which can be pre-set and dynamically optimized through machine learning. The specific calculation data source is data fused with cloud data.

[0057] Time coverage is the percentage of all emergencies that meet emergency response requirements (for example, resource arrival time does not exceed 8 minutes). Resource adequacy is the ratio of the number of available resources in the surrounding area that meet emergency response requirements to the number of required resources predicted based on the severity of the incident. The specific definition of the number of available resources is: through the police dispatch system API, emergency center database, and road administration asset management system called by the cloud, the resources in the surrounding area that can arrive at the scene within a certain time are obtained. The specific definition of the number of required resources is: based on historical events and handling situations, a demand resource prediction model is constructed, and the demand resources of the event are predicted in combination with the severity of the accident.

[0058] As an optional embodiment, the vulnerability index of the road network may also include one or more of betweenness centrality, topological connectivity loss, network efficiency loss, and capacity decay rate.

[0059] Specifically, betweenness centrality is a conventional metric used to measure the importance of a node as a shortest path hub. It is an indicator that characterizes the importance of a node by the number of shortest paths passing through it. It refers to the number of times a node acts as an intermediary in the shortest path between two nodes (ODs). The more times a node acts as an intermediary, the greater its betweenness centrality. The higher the betweenness centrality of a node, the more it means that many, or even all, of the shortest paths between other OD pairs must pass through it. If this point disappears, communication between other points will become difficult or even disconnected. In large-scale road networks, the number of shortest paths between node pairs is large, and it is almost impossible for nodes to have completely consistent betweenness centrality, so it has a high degree of discrimination.

[0060] Topological connectivity loss reflects the impact of an edge on network connectivity and measures network integrity based on the overall connectivity of the network. This loss represents the ratio of the number of nodes in the largest connected subgraph of the network after an attack, when network units are damaged and canceled, and an edge fails, to the total number of nodes in the original network.

[0061] Network efficiency loss reflects the overall deterioration in the shortest path travel time of a road network. It is the ratio of the difference between the network efficiency in normal conditions and the network efficiency in damaged conditions, divided by the network efficiency in normal conditions. Network efficiency refers to the ease with which nodes in the network can connect with each other. The more convenient the connections between nodes, the higher the network efficiency; the less convenient the connections between nodes, the lower the network efficiency. The network efficiency change rate refers to the change in the overall communication transmission speed of the network caused by changes in the number of network units when the network is disturbed. The core of network efficiency calculation is the shortest distance between nodes. The average shortest distance of all nodes in a network is the network efficiency of that network.

[0062] The capacity decay rate reflects the degree of decay of the road network's capacity. It is calculated by subtracting the current dynamic traffic state from the ideal traffic state of each road section from the ratio of the ideal traffic state to the ideal traffic state, and averaging the ratios of all road sections.

[0063] As an optional embodiment, after defining the traffic road network graph structure, it also includes updating the weight of the edge with the road section connecting the upstream and downstream intersections; the weight of the edge is the weighted sum of several weight indicators of the road section corresponding to the edge; the corresponding weight is an empirical weight coefficient, which is obtained through dynamic optimization of machine learning.

[0064] Weight indicators include: Real-time traffic density distribution, the ratio of the actual number of vehicles to the number of vehicles that can be accommodated; Vehicle delay index, the ratio of actual travel time to free flow time; these two indicators can be calculated using vehicle GPS data or radar trajectory data; Lane availability, the percentage of lanes closed due to abnormal events (due to construction / accidents); this can be obtained through geo-fence data or road construction data; Weather impact factor, a traffic efficiency attenuation coefficient dynamically adjusted based on meteorological data, can be obtained through the Meteorological Bureau API.

[0065] It's important to note that edge weights are the core input of the spatiotemporal graph neural network, used to dynamically quantify the travel risk of each road segment in the road network topology. Their calculated values ​​directly impact: shortest path planning; the calculation of shortest path-related metrics in the vulnerability index; and the accuracy of Bayesian network predictions for transmission paths.

[0066] As a specific embodiment, a Bayesian network is constructed, and the data corresponding to the detected abnormal event is used as the observation value to update the probability distribution of the network and predict the conduction path of the abnormal event.

[0067] Constructing a Bayesian network includes: using external triggers as root nodes, road network status as intermediate nodes, abnormal events as leaf nodes, and edges between nodes to represent causal relationships; each node contains at least spatiotemporal attribute labels and severity attributes; and at the same time, it is constrained that there are edges between nodes that are adjacent in space or time.

[0068] Specifically, weather, construction, and special events are considered root nodes for external factors; road speeds, traffic volume, and queue lengths are considered intermediate nodes reflecting the state of the road network; and accidents, equipment failures, and sudden congestion are considered leaf nodes representing abnormal events. Each node includes attributes such as timestamp, spatial location, and severity. The PC algorithm and the K2 scoring algorithm are combined to automatically learn the dependencies between nodes from historical data, while incremental learning is used to update the network structure in real time.

[0069] It should be noted that after determining the network structure, conditional probability parameters are configured for each node. Continuous variables are first discretized, and then the probability distribution is learned using maximum likelihood estimation or the EM algorithm. To reflect the spatiotemporal characteristics of the transportation system, the model incorporates spatial adjacency constraints and a temporal dynamic model (using an existing dynamic Bayesian network). When an abnormal event is detected, the system converts these observations into evidence and inputs them into the network. A probability propagation algorithm then updates the probability distribution across the entire network. Based on this updated probability distribution, the maximum a posteriori probability path is identified, and the credibility of each path is quantified. The prediction results not only include the spatial path topology but also incorporate temporal dynamic models to provide transmission delay estimates, such as "the impact will spread 2 kilometers upstream within 10 minutes." To ensure the model's timeliness, an online learning mechanism is implemented that dynamically adjusts network parameters by continuously integrating real-time observation data. The deviation between the predicted results and the actual impact is then used for feedback optimization.

[0070] As a specific embodiment, the preliminary risk level assessment based on the vulnerability index and the predicted transmission path includes: If the vulnerability index is greater than the first index threshold and the transmission path prediction result is greater than or equal to the third transmission level, it is a Level I risk; The vulnerability index is greater than the second index threshold and less than or equal to the first index threshold, and the conduction path prediction result is the second conduction level, which is Level II risk; The vulnerability index is greater than the third index threshold and less than or equal to the second index threshold, and the conduction path prediction result is the first conduction level, which is a Level III risk.

[0071] It should be noted that the first transmission level indicates that the abnormal event directly affects adjacent road sections, the second transmission level indicates that the abnormal event's impact spreads to highly correlated parallel / alternative paths, and the third transmission level indicates that the abnormal event triggers regional road network-level paralysis. The preliminary risk level assessment also applies an automatic escalation mechanism for high-transmission scenarios. This means that if the transmission exceeds the transmission limit of the risk level corresponding to the vulnerability index, the escalation mechanism is directly triggered. If the transmission is lower than the transmission level of the risk level corresponding to the vulnerability index, the risk level remains unchanged. For example, if the vulnerability index is between the third index threshold and the second index threshold, but the transmission path is greater than the third transmission level, the risk level is automatically upgraded to Level II and the corresponding response measures are implemented. If the vulnerability index is between the second index threshold and the first index threshold, but the transmission path is at the first transmission level, the risk level remains unchanged at Level II. Table 1 shows an example of each risk level assessment and the corresponding corresponding measures.

[0072] Table 1 Risk level assessment results

[0073] The response measures corresponding to each risk level are the baseline framework, and the actual specific corresponding measures are dynamically optimized based on the baseline framework through real-time data. For example, closing ramps or main lines based on real-time traffic, selecting effective strategies based on historical diversion effects, and dynamically narrowing the control range based on camera data.

[0074] Specifically, such as Figure 2 As shown, for the calculated vulnerability index and the predicted transmission path, the vulnerability index is first compared to see if it exceeds the first index threshold. If so, the risk is determined to be Level I. Otherwise, the vulnerability index is determined to be above the second index threshold. If the vulnerability index is above the second index threshold, the risk is first determined to be Level II. Then, a determination is made as to whether the transmission path is above the second transmission level. If so, the risk is upgraded to Level I; otherwise, the risk remains at Level II. If the vulnerability index is below the second index threshold, a determination is made as to whether the vulnerability index is above the third index threshold. If so, the risk is first determined to be Level III. Then, a determination is made as to whether the transmission path is above the first transmission level. If so, the risk is upgraded to Level II; otherwise, the risk remains at Level III. If the vulnerability index is less than or equal to the third index threshold, the risk is upgraded to Level III if the transmission path is above or equal to the first transmission level; otherwise, no risk warning is issued.

[0075] As a specific embodiment, correcting the risk level based on the intervention simulation results includes performing intervention simulation on abnormal events and calculating the intervention indicator improvement rate after the intervention simulation, and correcting the risk level of the preliminary risk assessment results based on the intervention indicator improvement rate.

[0076] It should be noted that, based on the abnormal events that have occurred, a corresponding multi-dimensional set of intervention scenarios is first established, such as 10-minute obstacle clearance and opening of emergency lanes. Then, intervention simulations are conducted on the road network to simulate the scenarios after the abnormal event intervention. The effect of each intervention scenario is quantified, including the actual results, the predicted results after the intervention, and the improvement rate of the intervention indicator. The intervention scenario with the best effect after the multiple intervention simulation results and the corresponding intervention indicator improvement rate are selected as the basis for subsequent risk level revisions.

[0077] It's worth noting that intervention simulation is a common technique used in existing traffic micro-simulation. This application utilizes this existing technology to generate simulation results, which are then combined with vulnerability indices and Bayesian network-based conduction path predictions for a multi-dimensional effectiveness evaluation. Risk levels are then adjusted based on the intervention indicator improvement rate, thereby constructing a three-dimensional vulnerability-conduction-intervention assessment model to dynamically adjust risk levels. Directly using existing intervention simulation methods will not affect the actual effectiveness of this application, so the specific intervention simulation process is not described in detail.

[0078] Specifically, the intervention indicator improvement rate is the weighted sum of the congestion duration reduction rate and the impact range contraction rate, and the corresponding weights are calculated using the entropy weight method.

[0079] The congestion duration reduction rate is the ratio of the shortened congestion duration after the intervention simulation to the actual duration without intervention; the shortened congestion duration after the intervention simulation is the difference between the actual duration without intervention and the congestion duration after the intervention simulation; the actual duration without intervention can be obtained through video recognition or analysis based on the changing characteristics of real-time monitoring data.

[0080] The shrinkage rate of the impact range is the ratio of the reduction in the impact range after the intervention simulation to the actual impact range without intervention; the reduction in the impact range after the intervention simulation is the difference between the actual impact range without intervention and the impact range after the intervention simulation; the actual impact range without intervention can be obtained through analysis of cloud map traffic conditions or real-time monitoring data change characteristics.

[0081] As an optional embodiment, Figure 3 The risk levels shown are modified based on the intervention indicator improvement rate IE, including: For Level I risk, if the improvement rate of several pre-indicators exceeds the upper limit of Level I risk, it will be downgraded to Level II risk; For Level II risk, if the improvement rate of some pre-indicators exceeds the upper limit of Level II risk, it will be downgraded to Level III risk; if the improvement rate of some pre-indicators is less than the lower limit of Level II risk, it will be upgraded to Level I risk; For Level III risk, if the improvement rate of several pre-indicators is greater than the upper limit of Level III risk, the warning will be lifted; if the improvement rate of several pre-indicators is less than the lower limit of Level III risk, it will be upgraded to Level II risk.

[0082] Table 2 shows each risk level and a case study of risk level modification based on the intervention indicator improvement rate (IE).

[0083] Table 2 Risk level correction results

[0084] It is worth noting that the preliminary risk assessment in the present invention is a risk assessment based on the static road network state after the abnormal event occurs and the theoretical conduction path prediction results, and the intervention indicator improvement rate after the intervention simulation reveals the actual controllability of the road network. Based on the actual controllability of the road network, the results of the preliminary risk assessment are corrected to obtain the final risk level, which can make the risk warning results more in line with the actual situation, avoid resource mismatch caused by the difference between the assessed risk level and the actual road conditions, and thus cause traffic anomalies in this life.

[0085] For example, if an intervention simulation shows an improvement rate (IE) significantly higher than expected (i.e., the upper limit for each risk level), this suggests that system resilience has been underestimated: the actual road network capacity is stronger than the theoretical value of the vulnerability index; or that the transmission path has been blocked: the intervention has prematurely suppressed the cascading effect. Maintaining the original risk level in this situation can lead to resource misallocation, such as excessive road closures, which can trigger secondary congestion and traffic anomalies.

[0086] The traffic anomaly warning method of the present invention is described below using a specific case.

[0087] S1. Real-time monitoring of traffic in a city's central area. One day, an abnormal event is detected where two vehicles collide on a road section. The two vehicles initiate an ABS activation signal and upload the accident data to the same roadside unit in the surrounding area through V2X communication.

[0088] S2. After receiving the accident data uploaded by the vehicle, the roadside unit performs a data verification process based on key indicators such as road speed, lane density, and queue length near the accident site to determine the authenticity of the accident, and then uploads the accident data to the cloud.

[0089] S3: The cloud integrates the abnormal event data uploaded by the roadside unit with the cloud data. Using a spatiotemporal graph neural network, the cloud calculates the vulnerability index of the road network at that moment to be 0.75. A Bayesian network also predicts the transmission path: the impact will spread to two intersections upstream within 15 minutes. This corresponds to a preliminary risk assessment of Level II.

[0090] S4. Establish a corresponding multidimensional intervention scenario set and simulate the prediction results under different intervention scenarios. The intervention scenarios and prediction results are shown in Table 3.

[0091] Table 3 Intervention scenarios and prediction results

[0092] Based on the above analysis, the optimal solution is for the roadside unit to push avoidance instructions to vehicles within 1 km. When the weights of the congestion duration reduction rate and the impact range reduction rate are both 0.5, the improvement rate of the intervention index exceeds the level II risk upper limit of 0.7.

[0093] S5. Based on the vulnerability index of the entire road network and the transmission path of abnormal events, the preliminary risk assessment result is Level II risk. Based on the condition that the improvement rate of the intervention indicator exceeds the upper limit of Level II risk, the risk level is revised to Level III risk and a Level III risk response is initiated.

[0094] The above embodiments are further elaborations and illustrations of the present invention for ease of understanding, and are not intended to limit the present invention in any way. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A vehicle-road-cloud collaborative traffic anomaly warning method, characterized by: include: The roadside unit receives abnormal event data reported by the vehicle, verifies it with roadside data, and uploads it to the cloud; The cloud receives data reported by the roadside unit and integrates it with the cloud data to assess the risk level after an abnormal event occurs: Construct a spatiotemporal graph neural network and calculate the vulnerability index of the road network based on the traffic network graph structure; Construct a Bayesian network and use the data corresponding to the detected abnormal events as observations to update the network's probability distribution and predict the conduction path of the abnormal events; Conduct preliminary risk assessment using vulnerability index and predicted transmission pathways; Traffic intervention simulation is carried out based on abnormal events, the intervention indicator improvement rate is calculated according to the simulation results, and the risk level is revised.

2. The vehicle-road-cloud collaborative traffic anomaly warning method according to claim 1 is characterized in that: The verification in combination with the roadside data includes: Map vehicle-side data and roadside data into a unified map coordinate system for consistency verification; If the consistency check result does not meet the consistency threshold requirement, recheck; If the consistency check result meets the consistency threshold requirement, data within a certain time period before and after the abnormal event occurs will be intercepted and uploaded to the cloud.

3. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 1 or 2, characterized in that: The traffic road network graph structure is established by taking intersections as nodes and road sections connecting upstream and downstream intersections as edges; The vulnerability index of the road network is the weighted sum of several vulnerability indicators; The vulnerability indicators include at least the environmental risk exposure that reflects the comprehensive risk level faced by the road network, and the emergency resource coverage that measures the adequacy of emergency resources provided after an emergency.

4. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 3 is characterized in that: The environmental risk exposure is the weighted sum of the accident frequency, real-time environmental factors, and topological risk factors; The frequency of accidents comes from the accident archive data in the cloud data; The real-time environmental factor is the maximum value of the precipitation factor and the visibility factor, and is derived from the meteorological bureau data in the cloud data; The topological risk factor is the product of betweenness centrality and environmental sensitivity coefficient.

5. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 3 is characterized in that: The emergency resource coverage is the weighted sum of time coverage and resource adequacy; Time coverage is the proportion of all emergencies that meet the emergency response requirements; Resource adequacy is the ratio of the number of available resources that meet emergency response conditions to the number of required resources predicted based on the severity of the incident.

6. The vehicle-road-cloud collaborative traffic anomaly warning method according to claim 1 is characterized in that: The Bayesian network construction includes: taking the external inducement as the root node, the road network status as the intermediate node, the abnormal event as the leaf node, and the edges between the nodes to represent the causal relationship; Each node contains at least a spatiotemporal attribute tag and a severity attribute; at the same time, it is constrained that there are edges between nodes that are adjacent in space or time.

7. A vehicle-road-cloud coordinated traffic anomaly warning method according to claim 1, 2, 4, 5, or 6, characterized in that: The preliminary risk level assessment includes: If the vulnerability index is greater than the first index threshold and the transmission path prediction result is greater than or equal to the third transmission level, it is a Level I risk; The vulnerability index is greater than the second index threshold and less than or equal to the first index threshold, and the conduction path prediction result is the second conduction level, which is Level II risk; The vulnerability index is greater than the third index threshold and less than or equal to the second index threshold, and the conduction path prediction result is the first conduction level, which is a Level III risk.

8. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 7 is characterized in that: The modified risk levels include: For Level I risk, if the improvement rate of several pre-indicators exceeds the upper limit of Level I risk, it will be downgraded to Level II risk; For Level II risk, if the improvement rate of some pre-indicators exceeds the upper limit of Level II risk, it will be downgraded to Level III risk; if the improvement rate of some pre-indicators is less than the lower limit of Level II risk, it will be upgraded to Level I risk; For Level III risk, if the improvement rate of several pre-indicators is greater than the upper limit of Level III risk, the warning will be lifted; if the improvement rate of several pre-indicators is less than the lower limit of Level III risk, it will be upgraded to Level II risk.

9. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 1 or 8, characterized in that: The improvement rate of the intervention indicator is the weighted sum of the congestion duration reduction rate and the impact range contraction rate, and the corresponding weights are calculated using the entropy weight method; The congestion duration reduction rate is the ratio of the congestion duration reduction after the intervention simulation to the actual duration without intervention; The impact range shrinkage rate is the ratio of the reduction in the impact range after the intervention simulation to the actual impact range without intervention.

10. The vehicle-road-cloud coordinated traffic anomaly warning method according to claim 3, characterized in that: The method of taking the road section connecting the upstream and downstream intersections as the edge includes updating the weight of the edge; The weight of an edge is the weighted sum of several weight indicators of the road segment corresponding to the edge; the weight indicators include: Real-time traffic density distribution, the ratio of the actual number of vehicles to the number of vehicles that can be accommodated; vehicle delay index, the ratio of actual travel time to free flow time; lane availability, the proportion of lanes closed due to abnormal events; weather impact factor, the traffic efficiency attenuation coefficient dynamically adjusted based on meteorological data.

Images