Expressway emergency management and control method and system

Abstract

The invention discloses an expressway emergency management and control method and system, and the method comprises the steps: building a multi-dimensional heterogeneous data fusion model through a time-space diagram neural network according to the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, and carrying out the dynamic prediction of the risk level of a road segment, and obtaining a real-time risk level map of a whole road network; based on the real-time risk level map, outputting a dynamic shunting scheme containing variable lane marks and a speed limiting strategy; according to the dynamic shunting scheme, generating a vehicle trajectory optimization strategy; and outputting an anti-interference optimization control strategy based on a vehicle trajectory optimization strategy in combination with sudden obstacle data polled by the unmanned aerial vehicle in real time. According to the embodiment of the invention, rapid and accurate path adjustment and cooperative control can be carried out, and the emergency management level of the expressway is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic control, and in particular to a method and system for emergency control of a highway. Background Art

[0002] As vital transportation infrastructure, highways play a key role in ensuring regional economic development, navigating the flow of people, and providing emergency rescue services. However, with the continued growth in the total number of vehicles and the increasingly complex traffic environment, highway traffic management faces numerous challenges. These include the frequent occurrence of emergencies such as accidents, severe weather, and sudden natural disasters, which pose a significant threat to driving safety and traffic efficiency. Traditional highway emergency management relies on manual scheduling and static emergency plans, resulting in a slow response and difficulty adapting to real-time traffic and weather conditions. While existing monitoring systems can provide certain accident information and traffic data, they lack comprehensive awareness and prediction of dynamic risks, resulting in delayed emergency responses and the inability to achieve optimal scheduling and traffic flow. Summary of the Invention

[0003] The purpose of the present invention is to provide a highway emergency management and control method and system to address the deficiencies in the existing technology, enable rapid and accurate path adjustment and coordinated control, and improve the level of highway emergency management.

[0004] An embodiment of the present application provides a highway emergency management method, the method comprising: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network. Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output; According to the dynamic diversion scheme, the path adjustment information is broadcast to the vehicle terminals through the communication system, and the vehicle trajectories are collaboratively controlled using a distributed robust optimization algorithm to generate a vehicle trajectory optimization strategy; Based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data of real-time drone inspections, a digital twin model is used to perform dynamic conflict simulation. The conflict path planning is corrected through a game theory decision model, and an anti-interference optimization control strategy is output.

[0005] Optionally, based on the traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of a road section and obtain a real-time risk level map of the entire road network, including: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, the multi-source heterogeneous data is spatiotemporally aligned using chaos theory coding methods to generate spatiotemporally synchronized feature tensors. Based on the spatiotemporal synchronized feature tensor, the spatiotemporal convolutional layer of the spatiotemporal graph neural network is used to extract the spatiotemporal dependencies of local road sections and obtain the spatiotemporal fusion feature matrix. Based on the spatiotemporal fusion feature matrix, the contribution of different meteorological factors to the risk level is dynamically adjusted using the meteorological impact weight allocation algorithm to generate a risk-sensitive feature vector. Based on the risk-sensitive feature vector, the risk index of each road section is calculated through a dynamic evolution prediction model, and a real-time risk level map of the entire road network is generated in combination with the road network topology.

[0006] Optionally, based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output, including: Based on the road section risk index of the real-time risk level map, the state space and action space of multi-agent reinforcement learning are constructed. The state space includes risk distribution and lane occupancy rate, and the action space is a set of variable lane signs and speed limit instructions. Based on the state space and action space, a dual-objective reward function is designed. The first objective is the negative inverse of the emergency response time, and the second objective is the logarithmic function of the road section traffic efficiency. This generates the agent training signal. Based on the agent training signals, the agents are trained using a master-slave collaborative training strategy, where the master agent generates global diversion instructions and the slave agents optimize local lane control strategies to output a preliminary diversion instruction set using the trained master and slave agents. Based on the preliminary diversion instruction set, the multi-agent decision conflicts are eliminated through the game equilibrium algorithm to generate a conflict-free diversion instruction sequence; The conflict-free diversion instruction sequence is encoded into variable lane marking control signals and gradient speed limit strategies to generate a dynamic diversion plan.

[0007] Optionally, according to the dynamic diversion scheme, the path adjustment information is broadcast to the vehicle terminals through the communication system, and the vehicle trajectories are collaboratively controlled using a distributed robust optimization algorithm to generate a vehicle trajectory optimization strategy, including: parsing the variable lane identification information in the dynamic diversion scheme and broadcasting lane change instructions, target speed limits, and expected travel time windows to affected vehicles via the V2X communication system; Receive real-time position, speed, and communication delay data from the vehicle terminal, build a robust optimization objective function for the vehicle kinematic model, and define the trajectory tracking error tolerance threshold; A distributed alternating direction multiplier method is used to solve the robust optimization objective function and generate a vehicle trajectory offset sequence that satisfies spatiotemporal constraints; The vehicle trajectory offset sequence is smoothed and filtered to eliminate the risks of sharp turns and speed mutations, and a vehicle trajectory optimization strategy containing a timestamp-coordinate-speed triple is output.

[0008] Optionally, the vehicle trajectory optimization strategy is combined with the sudden obstacle data from the real-time inspection of the UAV, and a digital twin model is used to perform dynamic conflict simulation. The conflict path planning is corrected through a game theory decision model, and an anti-interference optimization control strategy is output, including: Receive the sudden obstacle point cloud data uploaded by the UAV inspection system, perform spatiotemporal matching with the vehicle trajectory optimization strategy, and identify potential conflict areas between the trajectory and obstacles; Loading road network 3D topology, vehicle trajectory, and obstacle data into the digital twin model, conducting multi-agent collaborative simulation, and detecting vehicle-obstacle collision probabilities and congestion-derived risks. Construct a multi-party game decision-making model, define the utility functions of the traffic management department, social vehicles, and emergency vehicles, and generate the optimal path priority rules through reverse induction; Reconstructing vehicle trajectory constraints based on the optimal path priority rule, replanning the vehicle path in the conflict area using a model predictive control algorithm, and generating a revised trajectory instruction set; The corrected trajectory command is integrated with the original diversion plan to output an anti-interference optimization control strategy that includes emergency lane preemption authority, dynamic road right allocation, and secondary accident warning rules.

[0009] Optionally, the three-dimensional topology of the road network, vehicle trajectory, and obstacle data are loaded into the digital twin model to perform multi-agent collaborative simulation to detect the probability of vehicle-obstacle collisions and congestion-derived risks, including: Based on the 3D topological data of the road network, a dynamic voxel partitioning algorithm is used to discretize the road space into spatiotemporally correlated voxel units. Each voxel contains position coordinates, lane attributes, and capacity parameters, generating a dynamic voxelized road network model. Based on the dynamic voxelized road network model, the vehicle trajectory data is mapped into a moving voxel sequence with a time stamp, and the obstacle point cloud data is converted into a static dangerous voxel set to obtain a spatiotemporally coupled traffic element distribution map; Based on the traffic element distribution map, a multi-agent collaborative simulation engine was constructed. The vehicle agent's motion decision-making rules were defined as a risk-aware obstacle avoidance strategy, and the obstacle agent's attribute was defined as an immovable threat source. This generated an interactive simulation scenario. The Monte Carlo method is used to simulate the motion trajectory of the vehicle agent in the interactive simulation scene. The contact probability between the vehicle and the dangerous factors is calculated through the spatiotemporal propagation model, and the collision risk probability matrix is ​​output. Based on the collision risk probability matrix and combined with the spatiotemporal distribution data of vehicle density, the probability of secondary congestion is evaluated through the congestion ripple effect prediction algorithm to generate a comprehensive risk map that includes collision hotspots and congestion-derived risk levels.

[0010] Optionally, the multi-party game decision model is constructed to define the utility functions of the traffic management department, social vehicles, and emergency vehicles, and to generate the optimal path priority rules through the reverse induction method, including: The utility function of the traffic management department is defined as the weighted sum of the global traffic efficiency of the road network and the priority guarantee of emergency vehicles. The utility function of social vehicles is defined as the product of travel time saving rate and path stability. The utility function of emergency vehicles is defined as the path reliability index. This generates a prototype of the three-party utility function. Based on the three-party utility function prototype, a dynamic game tree structure is constructed, wherein the main node of the game tree represents the decision moment, the branches represent the possible action combinations of each party, and the leaf nodes are associated with the utility function value; The backward induction method is used to backtrack from the game tree leaf nodes to solve the problem, calculate the cumulative utility value of each decision path, and filter the non-inferior solution set through the Pareto optimality criterion to obtain the candidate path priority rule set; Conflict resolution is performed on the candidate rule set. The conflict intensity index is defined as the product of the path overlap rate and the priority difference. The Hungarian algorithm is used to match the optimal rule combination to generate a preliminary draft of the conflict-free path priority. The draft of the conflict-free path priority is input into the reinforcement learning fine-tuner, and the strategy is optimized with the average delay reduction rate of the road network as the reward signal, and the optimal path priority rule including the emergency lane preemption threshold, social vehicle yielding rule and dynamic road right allocation table is output.

[0011] Another embodiment of the present application provides a highway emergency management and control system, the system comprising: The prediction module is used to construct a multi-dimensional heterogeneous data fusion model based on traffic flow, meteorological data and accident reports collected by roadside sensing equipment through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network; An output module is configured to output a dynamic diversion solution including variable lane markings and speed limit strategies based on the real-time risk level map using a multi-agent reinforcement learning algorithm with the optimization goals of minimizing emergency response time and maximizing traffic efficiency; a control module configured to broadcast path adjustment information to vehicle terminals via a communication system according to the dynamic diversion scheme, perform collaborative control of vehicle trajectories using a distributed robust optimization algorithm, and generate a vehicle trajectory optimization strategy; The correction module is used to perform dynamic conflict simulation based on the vehicle trajectory optimization strategy and the sudden obstacle data of the real-time inspection of the drone using a digital twin model, correct the conflict path planning through the game theory decision model, and output the anti-interference optimization control strategy.

[0012] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0013] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0014] Compared with the existing technology, the present invention provides a highway emergency management and control method. Based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map of the entire road network; based on the real-time risk level map, a dynamic diversion plan including variable lane markings and speed limit strategies is output; based on the dynamic diversion plan, a vehicle trajectory optimization strategy is generated; based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of real-time drone inspections, an anti-interference optimization management and control strategy is output, thereby enabling rapid and accurate path adjustment and coordinated control, thereby improving the level of highway emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A hardware structure block diagram of a computer terminal for a highway emergency control method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of a highway emergency control method provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of a highway emergency management and control system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0017] The embodiment of the present invention first provides a highway emergency management and control method, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers, etc.

[0018] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal of a highway emergency control method provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause the processor to execute any one of the highway emergency management and control methods.

[0020] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0021] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any highway emergency control method.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0023] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0024] See also Figure 2 , an embodiment of the present invention provides a highway emergency management method, which may include the following steps: S201: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed using a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network. Roadside sensing equipment collects multidimensional traffic data (volume, weather, and accidents) in real time, and a fusion model is constructed using a spatiotemporal graph neural network. This model captures the spatiotemporal evolution of traffic flow and the coupled influence of meteorological factors, integrating discrete local risk predictions into a dynamic risk assessment for the entire road network. Specifically, the model uses a graph structure to represent road network topology and quantifies the risk transmission effect between road sections through node embedding and edge weight assignment. This enables a leap from fragmented data to systematic risk assessment, providing a scientific basis for emergency decision-making. The spatiotemporal graph neural network overcomes the shortcomings of traditional methods in modeling spatiotemporal correlations. The risk level map can intuitively display the distribution and evolution trends of high-risk areas, supporting preventive management and control.

[0025] S202: Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output; Based on the risk map, a multi-agent reinforcement learning framework is employed to coordinate global diversion and local control. The master agent coordinates network-level resource allocation, while the slave agent focuses on lane-level command generation. A dual-objective reward function balances emergency response efficiency with guaranteed traffic capacity. A game equilibrium algorithm ensures consistency in the decisions of each agent, ultimately outputting dynamically adjustable lane control and speed limit strategies. This overcomes the limitations of static diversion solutions and achieves optimal spatiotemporal allocation of emergency resources. This multi-agent architecture ensures global optimization while retaining local response flexibility. The combination of variable lanes and gradient speed limits effectively alleviates congestion on bottleneck roads.

[0026] S203, broadcasting path adjustment information to vehicle terminals via a communication system according to the dynamic diversion solution, using a distributed robust optimization algorithm to collaboratively control vehicle trajectories and generate a vehicle trajectory optimization strategy; Through V2X communication, the traffic diversion solution is broken down into vehicle-level control commands, and distributed robust optimization is employed to address vehicle trajectory planning. The algorithm considers real-world constraints such as communication latency, employing an alternating direction multiplier method for coordinated collision avoidance between vehicles. Smoothing filtering eliminates sudden motion fluctuations, generating a safe and feasible trajectory sequence. This precisely translates macro-control strategies into micro-vehicle behaviors, improving solution execution reliability. Robust optimization enhances adaptability to uncertainty, and trajectory smoothing prevents secondary accidents caused by abrupt lane changes.

[0027] S204, based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data of the real-time inspection by the UAV, a digital twin model is used to perform dynamic conflict simulation, and the conflict path planning is corrected through the game theory decision model to output an anti-interference optimization control strategy.

[0028] Trajectory plans are simulated and verified in a digital twin environment, using drone inspection data. Potential conflicts are identified through collision probability calculation and congestion derivative analysis. Game theory is used to model the interests of multiple parties, ultimately outputting an anti-interference strategy that includes priority rules and right-of-way allocation. This achieves a closed-loop control system of "prediction-simulation-optimization," significantly improving system resilience. The game-theoretic decision-making model balances emergency response efficiency and social equity, while dynamic priority rules ensure the rapid passage of special vehicles while minimizing disruption to public traffic.

[0029] Specifically, based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network, including: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, the multi-source heterogeneous data is spatiotemporally aligned using chaos theory coding methods to generate spatiotemporally synchronized feature tensors. Traffic flow data (vehicles per minute), weather data (wind speed, visibility, rainfall), and accident reports (location, type, severity) collected by roadside sensing devices (such as cameras, millimeter-wave radar, and weather sensors) can experience time stamp discrepancies and spatial resolution differences. For example, if a camera captures 30 frames per second, while a weather sensor updates data every 5 seconds, the upload of an accident report may be delayed by 10 seconds.

[0030] The chaos theory coding method uses Logistic mapping to generate chaotic sequences, which are used to unify the time base of multi-source data: Time alignment: Using the accident report as the reference time (e.g., t = 1000ms), the timestamps of other data are aligned using chaotic sequence interpolation. For example, meteorological data sampled at t = 995ms and t = 1005ms are weighted using chaotic sequence weights (e.g., 0.4 and 0.6) to generate an equivalent value at t = 1000ms.

[0031] The initial value x0 of the chaotic sequence is determined by the device ID hash value, ensuring that the time interpolation weights of different devices are unique.

[0032] Spatial alignment: Discrete sensor data (such as lane-level traffic flow) is mapped to a unified road grid grid (10m×10m grid), and missing grids are filled using bilinear interpolation. For example, if a camera covers three grids and traffic flow data is [50, 60, 55] vehicles per minute, the adjacent uncovered grids are calculated using distance weights (for example, the flow rate of a grid with a distance of 0.5 times the weight of the center grid is 57.5).

[0033] Data fusion: Encode the time-space aligned data into a four-dimensional feature tensor (time × space grid × data category × sensor ID). For example, at a certain moment, the tensor dimensions are 1 × 200 × 3 × 5 (200 grids, 3 types of data, 5 sensors).

[0034] Output example: In the spatiotemporal synchronized feature tensor, the traffic flow at grid (15, 20) at t = 1000ms is 65 vehicles / minute, the visibility is 800 meters, and there are no accident marks.

[0035] Based on the spatiotemporal synchronized feature tensor, the spatiotemporal convolutional layer of the spatiotemporal graph neural network is used to extract the spatiotemporal dependencies of local road sections and obtain the spatiotemporal fusion feature matrix. The spatiotemporal graph neural network (STGNN) consists of a spatiotemporal convolutional layer (ST-Conv Block), the core of which is the cascade of the graph convolutional network (GCN) and the temporal convolutional network (TCN): Graph structure definition: Node: Each road grid grid (10m×10m) is a node, with a total of N=200 nodes; Edge: Adjacent grids (up, down, left, right, and diagonal) are connected, and the edge weight is determined by the road connectivity (for example, the weight of the main road is 0.9, and the ramp is 0.6).

[0036] Spatiotemporal convolution operation: Spatial convolution: GCN aggregates neighboring node features. For example, the traffic feature of node A is updated to the weighted sum of its own value (65) and neighboring nodes B (60) and C (70): 65×0.5 + 60×0.3 + 70×0.2 = 64.5.

[0037] Temporal Convolution: TCN captures temporal dependencies through dilated causal convolution (dilation = 2). For example, the current time feature is convolved with the features at time t-2 and t-4 to extract periodic congestion patterns.

[0038] Feature fusion: The spatial and temporal convolution outputs are fused through a gating mechanism (GLU), and the output feature matrix dimension is N×D (N=200 nodes, D=128-dimensional features).

[0039] Output example: The spatiotemporal fusion features of the grid (15,20) include: traffic trend (+0.3%), visibility decay rate (-5% / minute), and accident risk probability (0.15).

[0040] Based on the spatiotemporal fusion feature matrix, the contribution of different meteorological factors to the risk level is dynamically adjusted using the meteorological impact weight allocation algorithm to generate a risk-sensitive feature vector. The weather impact weight distribution algorithm is based on the combination of entropy weight method and attention mechanism: Entropy weight method calculates static weight: The correlation between various meteorological factors (wind speed, visibility, rainfall) and accident rates was statistically analyzed based on historical data: when visibility is <500 meters, the accident rate increases by 300%; when rainfall is >20mm / h, the accident rate increases by 200%; wind speeds >10m / s have little impact, and the accident rate increases by 50%.

[0041] The weight of each factor is calculated using the entropy value (visibility 0.5, rainfall 0.3, wind speed 0.2).

[0042] Dynamic adjustment of attention mechanism: Real-time weather data is fed into a bidirectional LSTM, which outputs a dynamic attention score: When the current visibility is 800 meters, the rainfall is 5 mm / h, and the wind speed is 8 m / s, the attention scores are [0.6, 0.3, 0.1]; if the rainfall suddenly increases to 30 mm / h, the scores become [0.4, 0.5, 0.1].

[0043] Weight fusion and feature enhancement: The final weight = entropy weight method weight × attention score, which is normalized to obtain [visibility 0.4, rainfall 0.45, wind speed 0.15]. The spatiotemporal fusion feature matrix is ​​multiplied element by element by the meteorological weight matrix to generate a risk-sensitive feature vector.

[0044] Output example: The risk-sensitive eigenvector of the grid (15, 20) is: [traffic trend 0.3%×0.4=0.12%, visibility attenuation -5%×0.45=-2.25%, accident risk 0.15×0.4=0.06].

[0045] Based on the risk-sensitive feature vector, the risk index of each road section is calculated through a dynamic evolution prediction model, and a real-time risk level map of the entire road network is generated in combination with the road network topology.

[0046] The dynamic evolution prediction model uses the Graph Attention Recurrent Network (GA-RNN): Model structure: Graph Attention Layer: Calculates the risk propagation weight between nodes. For example, the influence weight of the upstream grid (15, 20) on the downstream grid (16, 20) is 0.8. GRU Layer: Models the temporal evolution of the network, with a hidden layer dimension of 64 and a time step of 5 minutes. Output Layer: A fully connected layer maps the network to a risk index (0-1).

[0047] Risk Communication Rules: Congestion ripple effect: If the upstream risk index is >0.7, the downstream risk will decay according to distance (10% decay every 10 meters); accident chain reaction: the risk index within 100 meters around the accident point will increase by 0.2.

[0048] Road network topology integration: The road connection relationships (such as intersections and ramps) are encoded into an adjacency matrix to limit the risk propagation direction (only allowing diffusion along the vehicle's travel direction).

[0049] Risk level map generation: Grading standards: Low risk (0~0.3): Green, no control required; Medium risk (0.3~0.6): Yellow, speed limit recommended; High risk (0.6~1): Red, diversion required.

[0050] Visual output: Displayed on the map in the form of a heat map, with high-risk areas marked as red patches and real-time traffic flow arrows superimposed.

[0051] Output example: At t=10:00, visibility on section A (grids 15-25) is reduced due to rainfall, with a risk index of 0.72 (red); section B (grids 30-40) is clear, with a risk index of 0.25 (green).

[0052] Specifically, based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency. The algorithm outputs a dynamic diversion solution that includes variable lane markings and speed limit strategies, including: Based on the road section risk index of the real-time risk level map, the state space and action space of multi-agent reinforcement learning are constructed. The state space includes risk distribution and lane occupancy rate, and the action space is a set of variable lane signs and speed limit instructions. The construction of state space and action space is the core foundation of multi-agent reinforcement learning. First, the state space needs to accurately represent the real-time risk distribution and traffic dynamics of the road network. The dimensions of the state space are designed as follows: Risk distribution features: The risk level of each road section is quantified as an index from 0 to 100 (0 indicates safe, 100 indicates extremely high risk), extracted based on a real-time risk level map. For example, if a road section has a risk index of 85 due to an accident, the corresponding risk feature value in the state vector is 0.85.

[0053] Lane occupancy rate: The vehicle density in each lane is measured in real time using roadside millimeter-wave radar and calculated as the ratio of the number of occupied vehicles to the lane's maximum capacity (0-1). For example, on a three-lane highway, if the occupancy rates for the left, center, and right lanes are 0.6, 0.8, and 0.3, respectively, the state vector corresponding to the three dimensions is [0.6, 0.8, 0.3].

[0054] Composite state encoding: The entire road network is divided into 500-meter segments. The state vector for each segment consists of a risk index (1 dimension) and a lane occupancy rate (N dimensions, where N is the number of lanes). Normalization is performed to generate a uniformly dimensional input. For example, the state vector for a six-lane, two-way road segment has a dimension of 1 (risk) + 6 (lanes) = 7.

[0055] The action space is defined as the set of control instructions that the agent can execute: Variable lane marking control: including lane opening / closing, direction switching (such as tidal lanes), emergency lane activation, etc. For example, the action code is: 0: Maintain the current lane configuration; 1: Open the left lane for emergency passage; 2: Close the middle lane and switch to the wrong direction.

[0056] Speed ​​limit policy instruction: Set dynamic speed limits (such as 60km / h and 70km / h) in steps of 10km / h. For example, the action code [60, 70, 80] corresponds to three optional speed limit gears.

[0057] Composite action space: The action space of each agent (corresponding to a road section unit) is the Cartesian product of the above two types of instructions. The total number of actions is 3 (lane instructions) × 3 (speed limit instructions) = 9 combinations.

[0058] State-action mapping example: When the risk index of a road section is 75 and the lane occupancy rate is 0.9, the agent may choose an action (close the middle lane and limit the speed to 60 km / h) to reduce the risk of increased congestion.

[0059] Based on the state space and action space, a dual-objective reward function is designed. The first objective is the negative inverse of the emergency response time, and the second objective is the logarithmic function of the road section traffic efficiency. This generates the agent training signal. The dual-objective reward function needs to balance emergency response speed and traffic efficiency. The specific design is as follows: Negative countdown of emergency response time: Definition: Emergency response time Tresponse refers to the average time (in minutes) from the occurrence of an incident to the arrival of rescue vehicles. Reward calculation is , where the smaller the Tresponse, the higher the reward.

[0060] Data acquisition: Arrival times are calculated in real time using emergency vehicle GPS trajectories. For example, if the diversion strategy reduces the Tresponse time from 15 minutes to 10 minutes, R1 improves from 0.0625 to 0.0909.

[0061] Logarithmic function of road section traffic efficiency: Definition: Efficiency = Actual traffic volume / Theoretical maximum traffic volume (unit: vehicles / hour). Rewards are calculated as R² = log(1 + Efficiency) to prevent drastic fluctuations in rewards when efficiency approaches zero.

[0062] Calculation example: The theoretical capacity of a road section is 2,000 vehicles per hour, and the actual number of vehicles passing through is 1,800. Then the efficiency = 0.9, and R2 = log(1.9)≈0.278.

[0063] Comprehensive reward function: Rtotal = w1*R1 + w2* R2, weights w1 = 0.7, w2 = 0.3, focusing on emergency response optimization.

[0064] Dynamic weight adjustment: When the average risk index of the road network exceeds the threshold (such as 70), it is temporarily adjusted to w1=0.9 to give priority to emergency passage.

[0065] Training Signal Generation: Temporal Difference Error (TD Error): Using the Q-learning algorithm, the difference between the predicted Q value and the actual Q value is calculated to guide the agent's strategy update. For example, if the expected reward for an action is 5 and the actual reward is 6, the TD Error = 1, driving the Q value to be revised upward.

[0066] Prioritized experience replay: Experience samples with high TD Error (such as emergency vehicle obstruction events) are prioritized for training to accelerate learning of key scenarios.

[0067] Based on the agent training signals, the agents are trained using a master-slave collaborative training strategy, where the master agent generates global diversion instructions and the slave agents optimize local lane control strategies to output a preliminary diversion instruction set using the trained master and slave agents. The master-slave agent architecture achieves global-local collaboration through hierarchical decision-making: Main agent (global decision-making layer): Input: Risk heat map of the entire road network (compressed into a 128-dimensional feature vector); Output: Macro traffic diversion instructions, such as "reduce north-to-south traffic by 20%"; Training algorithm: Proximal Policy Optimization (PPO) is used, with a learning rate of 0.00025 and a discount factor of γ = 0.99.

[0068] From the agent (local execution layer): Input: State vector of the road section (7 dimensions); Output: Lane control and speed limit instructions (9 actions); Training algorithm: DQN (Deep Q-Network), experience replay buffer capacity of 10,000, batch size of 64.

[0069] Collaborative training process: Global command issuance: The main agent generates diversion commands every 5 minutes, such as "reduce traffic volume on road section AC by 15%"; Local strategy optimization: The slave agent adjusts its action space based on the global command. For example, if the master command requires a 15% flow reduction, the slave agent will increase the weight of the speed limit action and prioritize the lower speed limit value. Joint training: The master agent receives the local reward mean of the slave agents (e.g., average Rtotal = 0.65) and adjusts the global strategy. If the mean falls below a threshold (e.g., 0.5), the master agent increases the diversion intensity.

[0070] For example: During a training session, the master agent instructs "Reduce traffic flow on road section XY by 10%." The slave agent achieves a 12% traffic reduction by closing one lane and limiting the speed to 60 km / h, receiving a reward R_total = 0.72, and the strategy is retained.

[0071] Based on the preliminary diversion instruction set, the multi-agent decision conflicts are eliminated through the game equilibrium algorithm to generate a conflict-free diversion instruction sequence; Multi-agent decision conflicts are primarily manifested as conflicting instructions between adjacent road sections (e.g., opening an emergency lane on section A leads to increased congestion on section B). The game equilibrium algorithm solves this problem in the following steps: Conflict detection: Type 1: Lane configuration conflict (e.g., lane closure on road section A causes congestion at the entrance to road section B); Type 2: Speed ​​limit change conflict (e.g., road section A has a speed limit of 60 km / h, while the adjacent road section B has a speed limit of 80 km / h, resulting in a rear-end collision risk).

[0072] Game model construction: Players: Agents involved in the conflict (e.g., A and B); Strategy Set: Set of optional actions for each agent (e.g., A's 9 actions); Payoff Matrix: Calculates the joint reward for each action combination , conflict penalty coefficient λ=0.5, R_A: the independent benefit value obtained by the road section agent A under the current strategy combination, R_B: the independent benefit value obtained by the road section agent B under the current strategy combination.

[0073] Nash equilibrium solution: A fictitious play algorithm is used to iteratively search for an equilibrium point. For example, after 10 rounds of iteration, agents A and B converge on a strategy (A speeds 60 km / h, B stays in the lane). At this point, neither agent has the incentive to unilaterally change its strategy.

[0074] Conflict resolution rules: Priority coverage: instructions for high-risk sections are executed first (for example, instructions for accident section A override conflicting instructions for adjacent section B); spatiotemporal coordination: scheduled scheduling of cross-section traffic (for example, notifying section B 1 minute in advance to prepare to receive traffic from section A).

[0075] For example, the instruction "Close the middle lane" on road section A (risk index 80) conflicts with the instruction "Open all lanes" on road section B (risk index 50). The game equilibrium algorithm determines that instruction A takes priority, and instruction B is adjusted to "Open the two left lanes."

[0076] The conflict-free diversion instruction sequence is encoded into variable lane marking control signals and gradient speed limit strategies to generate a dynamic diversion plan.

[0077] The final diversion solution needs to be converted into executable physical control instructions: Variable lane marking code: LED control protocol: Adopts the NTCIP (National Transportation Communications for ITS Protocol) standard. The command format is [location ID, lane status, effective time]. For example, the command [Lane_102, Emergency_Open, T+30s] means that the emergency access function of lane 102 will be enabled in 30 seconds.

[0078] Dynamic switching logic: Dynamically adjusts the switching frequency based on the risk index. For example, if the risk is greater than 70, the lane status will be refreshed every 5 minutes, and if the risk is less than 30, the lane status will be refreshed every 15 minutes.

[0079] Gradient rate limit strategy generation: Basic speed limits are selected based on the action space (e.g., 60 / 70 / 80 km / h). Smooth transitions are implemented by setting speed limit gradients between adjacent road sections (e.g., Section A: 60 km / h → Section B: 70 km / h, with a 500-meter gradient in between, at 2 km / h intervals every 100 meters). Dynamic adjustments are implemented by sending speed limit updates in real time via V2X communication. For example, if a drone detects a new obstacle, it can immediately reduce the speed limit from 70 km / h to 50 km / h.

[0080] Solution packaging and delivery: Command sequence format: JSON structure containing timestamp, road section ID, lane instruction, speed limit value, and validity period; Communication protocol: Using C-V2X (Cellular Vehicle-to-Everything) broadcast, latency <100ms, reliability >99.9%.

[0081] Specifically, according to the dynamic diversion scheme, the path adjustment information is broadcast to the vehicle terminals through the communication system, and the vehicle trajectories are collaboratively controlled using a distributed robust optimization algorithm to generate a vehicle trajectory optimization strategy, including: parsing the variable lane identification information in the dynamic diversion scheme and broadcasting lane change instructions, target speed limits, and expected travel time windows to affected vehicles via the V2X communication system; The variable lane marking information in the dynamic diversion scheme includes lane function switching instructions (such as temporary opening of the emergency lane), speed limit gradient adjustment parameters (such as reducing from 100km / h to 60km / h), and time windows (such as 10:00-10:30). The parsing process uses a semantic segmentation algorithm to identify the coding rules of electronic road signs: Lane marking analysis: For example, the code "L3-EMG-OPEN" means that lane 3 is open to the emergency lane; Speed ​​limit extraction: The code "SPD-GRAD[80,60,30s]" means a linear decrease from 80 km / h to 60 km / h within 30 seconds; Time window mapping: Convert the UTC timestamp to the local time zone (such as UTC+8) and associate it with the weather warning period.

[0082] The V2X communication broadcast mechanism is designed based on the 3GPP standard: Message encapsulation: ASN.1 encoding format is used, and the data packet structure is: Header: message type (0x0A indicates lane change), priority (0-7, 7 is the highest); Payload: Lane ID, target speed limit, and effective time (accuracy 0.1 second); CRC check: 32-bit cyclic redundancy check code.

[0083] Broadcast strategy: Regional targeting: Through the geo-fencing function of the RSU (roadside unit), only the affected road section (such as 2 kilometers upstream and downstream of the accident point) is covered; Redundant transmission: Each command is sent three times in succession (with an interval of 50ms) to combat channel fading; Priority preemption: High-priority instructions (such as fire truck passage) can interrupt low-priority broadcasts.

[0084] For example, when the RSU detects an accident 3 kilometers upstream, it immediately broadcasts the message: "L2-OPEN, SPD 60km / h, 10:00:00-10:15:00", and 200 private vehicles within the coverage area receive and interpret the command within 100ms.

[0085] Receive real-time position, speed, and communication delay data from the vehicle terminal, build a robust optimization objective function for the vehicle kinematic model, and define the trajectory tracking error tolerance threshold; The vehicle kinematic model uses a bicycle model to simplify vehicle dynamics: State vector: x, y, v, θ (position x / y, velocity, heading angle); Control input: front wheel angle δ (-30°~+30°), acceleration a (-3m / s²~2m / s²).

[0086] The robust optimization objective function is designed as a multi-objective weighted sum: Tracking error minimization: the sum of the squares of the lateral deviation (|y-y_ref|) and the heading angle deviation (|θ-θ_ref|), with weights of 0.7 and 0.3 respectively; Control smoothness: Penalty terms for the rate of change of corners (|Δδ / Δt|) and the rate of change of acceleration (|Δa / Δt|), with a weight of 0.2; Communication delay compensation: Define the delay compensation factor λ=1 / (1+τ), where τ is the delay time (in seconds), and apply λ scaling to the control quantity.

[0087] The trajectory tracking error tolerance threshold is dynamically adjusted according to the vehicle type: small car: lateral deviation ≤ 0.3m, heading angle deviation ≤ 5°; truck: lateral deviation ≤ 0.5m, heading angle deviation ≤ 8°; emergency vehicle: lateral deviation ≤ 1.0m (priority is given to ensuring the right of way).

[0088] For example, for an ambulance traveling at 80 km / h (v = 22.22 m / s) with a communication delay of τ = 0.2 seconds, the position prediction error compensation is 22.22 m / s × 0.2s = 4.44 m. Therefore, an additional position compensation term must be added to the optimization function.

[0089] A distributed alternating direction multiplier method is used to solve the robust optimization objective function and generate a vehicle trajectory offset sequence that satisfies spatiotemporal constraints; Decomposition and coordination process of distributed alternating direction multiplier method (ADMM): Problem decomposition: Each vehicle is treated as an independent node, and the local optimization problem is to minimize the objective function. The constraints include: safe distance: maintain a headway of ≥ 2 seconds from the vehicle in front (for example, 44.4 meters at 80 km / h); lane boundary: lateral position deviation ≤ 0.5 times the lane width (for a standard lane of 3.75 meters, the offset is ≤ 1.875 meters).

[0090] Global coordination: The central coordinator (RSU) receives trajectory proposals from all vehicles and calculates the Lagrange multiplier update for the global consistency constraint: Multiplier update formula: μ^(k+1) =μ^k +ρ(Ax - z), where ρ=1.0 is the penalty coefficient, Ax is the local solution, and z is the global average; Iteration termination condition: primal residual ||Ax - z||2<0.1 or dual residual ||μ^(k+1) - μ^k||2<0.01.

[0091] Solver parameter settings: Maximum number of iterations: 50 times (90% of cases converge within 30 times in actual measurements); Step size adaptation: Dynamically adjust ρ (0.5~2.0) according to the residual decrease rate; Asynchronous tolerance: Allow 10% node delay update (to cope with communication packet loss).

[0092] For example, in a certain optimization, 100 vehicles reached a global consensus after 20 iterations, generating a trajectory offset sequence such as: t = 0s: Vehicle A deviates 0.5m to the right and decelerates to 70km / h; t = 2s: Vehicle B deviates 1.2m to the left and accelerates to 65km / h.

[0093] The vehicle trajectory offset sequence is smoothed and filtered to eliminate the risks of sharp turns and speed mutations, and a vehicle trajectory optimization strategy containing a timestamp-coordinate-speed triple is output.

[0094] The smoothing filter algorithm uses piecewise cubic Hermite interpolation (PCHIP) and Kalman filtering fusion: PCHIP treatment: Input: discrete trajectory points output by ADMM (e.g., 1 point per second); Interpolation rules: Ensure the continuity of velocity and acceleration to avoid overshoot (such as maximum curvature ≤ 0.1m -1 ); Output: 10Hz high-density trace (one point every 0.1 seconds).

[0095] Kalman filter: Equation of state: x k = F*x k-1 + B*u k + w k , where F is the state transfer matrix, B is the control input matrix, and wk is the process noise (covariance Q = 0.01I); Observation equation: z k = H*x k + v k , observation noise v k Covariance R=0.1I; Filter gain: Dynamically adjusted to suppress GPS positioning jitter (typical value 1~3Hz).

[0096] Sharp turn detection and suppression: Curvature threshold: When the instantaneous curvature κ = Δθ / Δs > 0.15m⁻¹, it is judged as a sharp turn; Suppression strategy: insert transition trajectory points, limit curvature to ≤ 0.1m⁻¹, and extend turning time by 50%.

[0097] The final trajectory is output as a timestamp-coordinate-velocity triplet in CSV format, for example: Timestamp (ms), X (m), Y (m), Speed ​​(km / h); 1630000001000, 1023.45, 456.78, 72.5; 1630000001100, 1023.88, 457.12, 71.8.

[0098] Specifically, based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data from real-time drone inspections, a digital twin model is used to perform dynamic conflict simulation. The conflict path planning is corrected through a game theory decision model, and an anti-interference optimization control strategy is output, including: Receive the sudden obstacle point cloud data uploaded by the UAV inspection system, perform spatiotemporal matching with the vehicle trajectory optimization strategy, and identify potential conflict areas between the trajectory and obstacles; Obstacle point cloud data collected by the drone using LiDAR and vision sensors is uploaded to the edge computing node at a frequency of 10 frames per second. Each frame of data contains 3D coordinates (x, y, z), obstacle type (such as falling rocks or scattered objects), and a timestamp (accuracy ±10ms). The spatiotemporal matching algorithm uses spatiotemporal hash coding technology to align the timestamp-coordinate-speed triplet used in the vehicle trajectory optimization strategy (for example, a vehicle's position at t = 123456ms is (120.5, 30.2) and its speed is 60km / h) with the time-space window of the obstacle point cloud. The specific process is as follows: Time Window Alignment: The matching window is centered on the timestamp of the vehicle trajectory and extended by 50ms before and after. For example, if the vehicle trajectory time is t=1000ms, the obstacle data collected by the drone between t=950ms and 1050ms will be matched.

[0099] Spatial hash mapping: Divide the road space into 1m×1m grid cells, and map the vehicle trajectory points and obstacle point cloud coordinates to the same grid. For example, the vehicle position (120.5, 30.2) is mapped to the grid (120, 30). If there is an obstacle point within this grid, it is marked as a potential conflict.

[0100] Dynamic expansion processing: Dynamically expands the conflict zone based on vehicle speed and safety distance (e.g., a 2-second following distance). For example, when a vehicle is traveling at 60 km / h (16.67 m / s), the safety distance is 33.34 m. The algorithm marks all obstacle cells within 33.34 m in front of the vehicle as high-risk areas.

[0101] In the final conflict area map, each grid is labeled with a conflict probability (0-1). For example, a grid with a conflict probability of 0.8 indicates that there is an 80% probability of a vehicle-obstacle collision in the area within the next 2 seconds.

[0102] Loading road network 3D topology, vehicle trajectory, and obstacle data into the digital twin model, conducting multi-agent collaborative simulation, and detecting vehicle-obstacle collision probabilities and congestion-derived risks. The digital twin model is built based on BIM (Building Information Modeling) and GIS (Geographic Information System) data, and includes details such as lane curvature, slope, and traffic sign locations. A dynamic voxel partitioning algorithm discretizes the road space into space-time voxels of 0.5m × 0.5m × 0.1s (time step), with each voxel storing the following parameters: Capacity: Based on lane type (e.g. emergency lane capacity is 1, social lane capacity is 0.8); Threat level: The threat level of obstacle pixels is 1, and the threat level of no obstacle is 0; Vehicle density: Real-time statistics of the number of vehicles in each voxel. When the density is greater than 0.3 vehicles / m², the voxel is marked as a congestion risk voxel.

[0103] Two types of agents are defined in the multi-agent collaborative simulation engine: Vehicle agent: The motion strategy is based on the IDM (Intelligent Driving Model), with parameters including a maximum acceleration of 1.5 m / s² and a safe distance of 2 seconds. Obstacle Agent: A static threat source that triggers the vehicle's obstacle avoidance logic (such as changing lanes or slowing down).

[0104] The Monte Carlo simulation was run 1000 times, each time injecting random perturbations (e.g., ±20ms communication delay, ±0.5m sensor error). The collision probability was calculated as the percentage of overlaps between the vehicle's trajectory and obstacle pixels. For example, if a collision occurred 150 times out of 1000 simulations on a given road section, the collision probability would be 15%.

[0105] Congestion ripple effect prediction uses a spatiotemporal propagation model: Input: initial congestion voxel position, vehicle density gradient, average speed; Output: Congestion spread range within the next 5 minutes. For example, if an accident causes the density within 500 meters upstream to exceed the threshold, the congestion is predicted to spread to 1 km in 10 minutes.

[0106] The simulation results generate a comprehensive risk map, including areas of high collision probability (red) and congestion diffusion paths (yellow arrows).

[0107] Construct a multi-party game decision-making model, define the utility functions of the traffic management department, social vehicles, and emergency vehicles, and generate the optimal path priority rules through reverse induction; The tripartite utility function is defined as follows: Traffic management department: Utility = 0.6 × global traffic efficiency (average vehicle speed / speed limit) + 0.4 × emergency vehicle arrival rate (the proportion of emergency vehicles arriving at the accident site on time); Social vehicle: Utility = 0.7 × travel time saving rate (1 - actual time / shortest time) + 0.3 × route stability (number of route changes ≤ 1); Emergency vehicles: Utility = route reliability (1 - proportion of congested road sections).

[0108] Dynamic game tree construction: Decision node: Every 30 seconds is a decision moment; Action branch: Traffic management department: open / close emergency lane, adjust speed limit (±10km / h); private vehicles: maintain lane / change lanes, accelerate / decelerate; emergency vehicles: choose the shortest path or detour.

[0109] The reverse induction method backtracks from the leaf node (such as t=300 seconds) to solve: Leaf node utility calculation: For example, at t = 300 seconds, if the emergency vehicle chooses a detour route (reliability 0.9), the average time saved by social vehicles is 15%, and the traffic efficiency of the traffic management department is 0.85, then the three-party utilities are 0.85, 0.7 × 0.15 + 0.3 × 1 = 0.405, and 0.9 respectively; Parent node strategy selection: compare the Pareto optimal solutions of all child nodes (such as maximizing the utility of three parties at the same time) and eliminate dominated strategies; Conflict resolution: When the path overlap rate is greater than 30% (for example, private vehicles and emergency vehicles compete for the same lane), the Hungarian algorithm is used to match the optimal action combination. For example, the emergency lane is prioritized for the emergency vehicle, and private vehicles switch to the adjacent lane.

[0110] The optimal path priority rules finally generated include: Emergency lane preemption threshold: When the emergency vehicle is less than 2km away from the accident site, it automatically obtains lane priority; Rules for private vehicles to give way: Private vehicles within 200m behind the emergency vehicle must change lanes within 10 seconds; Dynamic road right allocation table: During peak hours (7:00-9:00), the priority of the emergency lane is increased by 20%.

[0111] Reconstructing vehicle trajectory constraints based on the optimal path priority rule, replanning the vehicle path in the conflict area using a model predictive control algorithm, and generating a revised trajectory instruction set; Model Predictive Control (MPC) uses the next 15 seconds as the prediction horizon, with a control cycle of 0.5 seconds. The optimization goal is to minimize trajectory deviation and collision risk: State variables: vehicle position (x, y), velocity v, heading angle θ; Controlled variables: acceleration a, steering wheel angle δ; Constraints: Speed ​​limit: 0≤v≤speed limit+10% (elastic buffer); Acceleration range: -3m / s²≤a≤2m / s²; Safe distance: Maintain a distance of ≥ 2 seconds from the vehicle in front.

[0112] Solver configuration: Use IPOPT (Interior Point Optimizer) to handle nonlinear constraints; Rolling optimization window: solve the trajectory of the next 5 control cycles (2.5 seconds) each time.

[0113] For example, if a social vehicle's original trajectory requires crossing a conflict grid (with a probability of 0.8), MPC replans the path so that it changes lanes to the right, reducing its speed from 60 km / h to 50 km / h, and generates a corrected trajectory instruction: Timestamp (ms) Coordinates (x,y) Speed ​​(km / h) 123500 (120.7,30.5) 55 124000 (120.9,30.8) 50 The corrected trajectory command is integrated with the original diversion plan to output an anti-interference optimization control strategy that includes emergency lane preemption authority, dynamic road right allocation, and secondary accident warning rules.

[0114] Strategy fusion is divided into three layers: Emergency lane control: Dynamically adjusts road rights based on preemption thresholds. For example, when an emergency vehicle is 1.5 km from an accident site, the emergency lane is forcibly cleared via V2X broadcast. Upon receiving the command, private vehicles must complete the lane change within 15 seconds. Dynamic road right allocation: Calculate lane capacity based on real-time traffic flow (e.g., updated every 5 minutes) and allocate weights: emergency lane: weight 0.6; left social lane: 0.3; right social lane: 0.1.

[0115] Secondary accident warning: The LSTM model is trained using historical data to predict the probability of a secondary accident within 500 meters of the accident site. When the threshold exceeds 30%, a warning is triggered (such as reducing the speed limit by 20% or increasing drone patrols).

[0116] Examples of anti-interference strategies: Scenario: A rockfall occurred at K12+300 on the highway, and the drone detected a 5m³ obstacle; Response: Emergency vehicle A obtains exclusive use of the emergency lane from K12+200 to K12+500; the speed limit for private vehicles in the section from K12+100 to K12+600 is reduced to 60 km / h; LSTM predicts that the probability of a secondary accident in the next 10 minutes is 35%, triggering an orange warning, and the roadside screen displays "Drive with caution."

[0117] The final strategy is broadcast to vehicle terminals via 5G-V2X, with an end-to-end latency of <20ms, enabling coordinated control of the entire road network.

[0118] Specifically, the digital twin model is loaded with 3D road network topology, vehicle trajectory, and obstacle data to conduct multi-agent collaborative simulation and detect vehicle-obstacle collision probabilities and congestion-derived risks, including: Based on the 3D topological data of the road network, a dynamic voxel partitioning algorithm is used to discretize the road space into spatiotemporally correlated voxel units. Each voxel contains position coordinates, lane attributes, and capacity parameters, generating a dynamic voxelized road network model. The core of the dynamic voxel partitioning algorithm is to divide the three-dimensional space (length, width, and time) of the highway into regular grid cells (voxels), each voxel representing a space-time unit. The voxel size is dynamically adjusted according to the characteristics of the road section: Spatial Dimensions: Length: The voxel length of the main road section is 5 meters (covering the average braking distance of a vehicle), and the ramp is shortened to 3 meters to cope with frequent lane changes; Width: Adjusted according to the number of lanes, the width of a single lane voxel is 3.75 meters, and the emergency lane is widened to 4.5 meters; Time dimension: The acceleration time slice is 0.1 second, matching the vehicle terminal reporting frequency (10Hz).

[0119] Each voxel stores the following parameters: Position coordinates: using WGS84 coordinate system, accuracy 0.1 meter; Lane attributes: coded as 4-bit binary (e.g. 0011 means "normal lane + construction status"); Traffic capacity: Dynamically calculated based on lane type (emergency / normal), slope (0%~10%), and curvature radius (>500 meters for straight roads). For example, the traffic capacity of a normal lane on a straight road is 2,000 vehicles / hour, while that on a curved road is reduced to 1,800 vehicles / hour.

[0120] Dynamic adjustment mechanism: Traffic flow trigger: When the traffic volume on a certain road section exceeds 90% of the capacity within 5 minutes (e.g. 1,800 vehicles on a straight road), the voxel length is automatically compressed to 3 meters to improve the resolution; Accident impact diffusion: If an accident point is detected, the voxel attributes are expanded according to the ripple model with the accident voxel as the center (for example, the voxels within 100 meters around the accident point are marked as "high risk").

[0121] Example: A three-lane highway section (1 km long) is divided into 200 spatial voxels (5 meters / voxel) and 1000 temporal voxels (0.1 seconds / voxel), with a total size of 200 × 3 × 1000 = 600,000 voxels. The capacity parameters are updated in real time.

[0122] Based on the dynamic voxelized road network model, the vehicle trajectory data is mapped into a moving voxel sequence with a time stamp, and the obstacle point cloud data is converted into a static dangerous voxel set to obtain a spatiotemporally coupled traffic element distribution map; Vehicle trajectory mapping is achieved through a spatiotemporal encoder: Trajectory preprocessing: Receive GPS data reported by the vehicle terminal (frequency 10Hz, accuracy ±0.5m), eliminate noise through Kalman filtering, and interpolate to fill in missing points; Voxel matching: Map the trajectory point coordinates (e.g., longitude 118.5°, latitude 32.1°) to the nearest voxel center and record the timestamp (e.g., t = 1630000000.1s); Sequence generation: Concatenate voxel IDs in chronological order to form a moving voxel sequence. For example, the trajectory of vehicle A within 1 second is mapped to the element ID sequence [V001, V005, V012,...].

[0123] Obstacle point cloud processing: Point cloud clustering: The obstacle point cloud (density 1000 points / m2) obtained by the drone inspection is clustered into obstacle areas using the DBSCAN algorithm (neighborhood radius 0.5 meters, minimum number of points 10); Voxel conversion: Map the clustered point cloud bounding box (e.g., 2 meters long × 1 meter wide) to a voxel grid, marking all covered voxels as "static hazards"; Threat level labeling: Set the threat level (1 to 5, with 5 being the highest) based on the obstacle type. For example, scattered tire fragments are marked as level 3, and an overturned truck is marked as level 5.

[0124] Output of spatiotemporal traffic element distribution map: Moving voxel layer: stores vehicle trajectory sequences, with each voxel recording vehicle ID, speed (e.g., 60 km / h), and acceleration (±0.3 g). Static hazard layer: stores a set of obstacle pixels, including threat level and estimated duration of existence (e.g., tire debris exists for 2 hours); Metadata index: Establish an R-tree spatial index to support fast query of conflict risks within a certain time and space range.

[0125] For example, near an accident site, the moving voxel sequence of vehicle B overlaps with the static dangerous voxel of the overturned truck at the timestamp 1630000000.5s, triggering a collision warning.

[0126] Based on the traffic element distribution map, a multi-agent collaborative simulation engine was constructed. The vehicle agent's motion decision-making rules were defined as a risk-aware obstacle avoidance strategy, and the obstacle agent's attribute was defined as an immovable threat source. This generated an interactive simulation scenario. The multi-agent simulation engine is extended with the MASON framework and includes two types of agents: Vehicle Agent: State space: position (voxel ID), speed (m / s), heading angle (0-360°), risk perception radius (50 meters); Action space: acceleration / deceleration (±2m / s²), lane change (left / right / hold), path replanning (switching navigation routes); Decision rule: Obstacle avoidance strategy based on risk field, calculation formula:

[0127] When the risk value exceeds a threshold (such as 10), lane change or deceleration is triggered.

[0128] Obstacle Agent: Static properties: position (fixed set of voxels), threat level, geometry (cube / cylinder); Interaction rules: When a vehicle enters a threat system, its utility value is deducted according to the threat level (e.g., a level 5 threat deducts 50% of the traffic efficiency).

[0129] Simulation scene construction process: Road network loading: Import the dynamic voxelized road network model and set the simulation time step to 0.1 seconds; Agent deployment: Instantiate vehicles and obstacles according to the traffic element distribution map; Interaction rule binding: When a vehicle approaches an obstacle, risk perception calculation is triggered, and the obstacle imposes speed restrictions on vehicles within the coverage area (for example, forcing the speed to 30 km / h at threat level 3).

[0130] Example: In the simulation, vehicle C approaches a threat level 4 obstacle at 80 km / h. The risk value rises to 15, and the system forces it to change lanes to the left and reduce its speed to 50 km / h.

[0131] The Monte Carlo method is used to simulate the motion trajectory of the vehicle agent in the interactive simulation scene. The contact probability between the vehicle and the dangerous factors is calculated through the spatiotemporal propagation model, and the collision risk probability matrix is ​​output. Monte Carlo simulation simulates multiple possible trajectories by randomly sampling vehicle behavior parameters: Parameter perturbation settings: Speed ​​fluctuation: normal distribution (mean current speed, standard deviation 5%); Response delay: evenly distributed (0.2~0.5 seconds); Path selection: 30% probability of following navigation suggestions, 70% probability of random lane change.

[0132] Trajectory Generation: 1000 independent simulations were performed for each vehicle, each simulation lasting 10 seconds (100 time steps), and the voxel path of each simulation was recorded.

[0133] Space-time propagation model: Contact detection: Count the number of overlaps between the voxels covered by the vehicle trajectory and the dangerous voxels; Probability calculation: Collision probability = number of overlaps / total number of simulations. For example, vehicle D enters the dangerous system 120 times out of 1000 simulations, with a collision probability of 12%. Matrix construction: The voxels of the entire road network are classified according to collision probability (0%-5% low risk, 5%-20% medium risk, >20% high risk) to generate a three-dimensional (spatial + temporal) risk matrix.

[0134] Visualization output: Heat map mode: Displays risk levels in a color gradient in the digital twin platform, with red representing high-risk factors; Timeline sliding: supports viewing the risk evolution process in 0.1 second steps.

[0135] For example, during heavy rain, visibility on a curved road section is low, and the dispersion of vehicle trajectories increases, causing the collision probability to increase from 5% to 18%. The system marks it as a yellow medium-risk area.

[0136] Based on the collision risk probability matrix and combined with the spatiotemporal distribution data of vehicle density, the probability of secondary congestion is evaluated through the congestion ripple effect prediction algorithm to generate a comprehensive risk map that includes collision hotspots and congestion-derived risk levels.

[0137] Congestion ripple effect prediction uses cellular automaton model: Vehicle density calculation: Count the number of vehicles by volume, density = number of vehicles / voxel volume (e.g. 5m × 3.75m × 0.1s); Propagation rule definition: Congestion trigger: When the density of a certain voxel exceeds a critical value (e.g., 25 vehicles / km) and lasts for 5 seconds, it is marked as a congestion source; Ripple diffusion: Congested voxels spread in the upstream and downstream directions, and the propagation speed = number of lanes × 10% (for example, the propagation speed for three lanes is 0.3 voxels / second); Dissipation Conditions: Density drops below critical value for 10 seconds.

[0138] Calculation of secondary congestion probability: Direct association: High collision risk factors (probability > 20%) directly trigger congestion probability +30%; Indirect association: A logistic regression model is trained based on historical data. Input parameters include collision probability, number of lanes, and weather conditions (such as rainfall intensity), and the output is the secondary congestion probability.

[0139] Comprehensive risk map generation: Collision hotspot layer: marks the location and threat level of high risk factors (probability > 20%); Congestion derivative layer: uses different transparency levels to indicate the probability of secondary congestion (0% to 100%); Overlay display: The red highlighted area represents the dual risks of "collision + congestion" and requires priority handling.

[0140] Example: Due to a 15% collision probability and a reduction in the number of lanes on a construction section, the system predicts a 65% probability of secondary congestion, which is displayed in dark red on the map and triggers an early opening command for the emergency lane.

[0141] Specifically, a multi-party game decision-making model is constructed to define the utility functions of the traffic management department, social vehicles, and emergency vehicles. The optimal path priority rules are generated through the reverse induction method, including: The utility function of the traffic management department is defined as the weighted sum of the global traffic efficiency of the road network and the priority guarantee of emergency vehicles. The utility function of social vehicles is defined as the product of travel time saving rate and path stability. The utility function of emergency vehicles is defined as the path reliability index. This generates a prototype of the three-party utility function. The definition of the three-party utility function needs to quantify the core demands of different parties and achieve mathematical modeling through parametric design.

[0142] Traffic management department utility function: Global traffic efficiency (weight 0.7): Calculate the ratio of the average speed (km / h) of the entire road network to the theoretical maximum speed. For example, if the average speed of the entire road network during a certain period is 80 km / h and the theoretical maximum speed is 120 km / h, then the traffic efficiency is 80 / 120 = 0.67.

[0143] Priority guarantee degree for emergency vehicles (weight 0.3): Statistically calculate the average delay time (seconds) of emergency vehicles (such as ambulances, fire trucks) and compare it with the historical benchmark value (such as 30 seconds). For example, if the actual delay is 20 seconds, then the guarantee degree is 1 - (20 / 30) = 0.33.

[0144] Function form: U_traffic = 0.7×traffic efficiency + 0.3×priority guarantee degree.

[0145] Social vehicle utility function: Travel time savings rate: The proportion of the difference between the actual travel time and the expected time in the unregulated state. For example, if a vehicle actually takes 15 minutes and is expected to take 20 minutes without control, then the savings rate is (20 - 15) / 20 = 25%.

[0146] Path stability: The proportion of the number of path changes (e.g., if the original plan was to change lanes 3 times and actually changed 1 time, the stability is 1 / 3 ≈ 0.33).

[0147] Function form: U_social = savings rate × stability.

[0148] Emergency vehicle utility function: Path reliability index: The probability of the path being unobstructed statistically based on historical data (e.g., if a path is unblocked 90% of the time, the reliability is 0.9).

[0149] Function form: U_emergency = reliability index.

[0150] Parameter example: The weight distribution of the traffic management department can be dynamically adjusted. For example, during peak hours, the weight of the priority guarantee degree is increased to 0.5; the stability threshold of social vehicles is set to 0.5, and if it is lower than this value, path replanning is triggered.

[0151] Based on the above three-party utility function prototype, a dynamic game tree structure is constructed. Among them, the main node of the game tree represents the decision-making moment, the branches represent the combinations of actions available to each party, and the leaf nodes are associated with the utility function values; The construction of the dynamic game tree needs to map the temporal relationship and action space of multi-party decisions.

[0152] Decision-making moment division: The time granularity is set to 5 seconds, and each master node represents a decision window; For example, t=0s, 5s, 10s, etc., covering the entire cycle of emergency control (such as 300 seconds).

[0153] Action space definition: Traffic management department actions: open / close the emergency lane (action A1); adjust the traffic light cycle (A2, such as changing from 60 seconds to 45 seconds); issue a dynamic speed limit instruction (A3, such as reducing from 100 km / h to 80 km / h).

[0154] Actions for other vehicles: Maintain the current lane (B1); change lanes to the left (B2) or right (B3); adjust the following distance (B4, such as increasing the distance from 2 seconds to 3 seconds).

[0155] Emergency vehicle actions: apply for right-of-way priority (C1); switch to an alternative route (C2); request traffic control (C3).

[0156] Game tree construction: Each main node extends to all possible action combinations. For example, at t = 0s, the traffic control department has 3 actions, private vehicles have 4 actions, and emergency vehicles have 3 actions, with a total of 3 × 4 × 3 = 36 branches. Leaf nodes store tripartite utility values. For example, if a branch action is (A1, B2, C1), the corresponding U_Traffic, U_Society, and U_Emergency are calculated.

[0157] Example: At t=0s, if the traffic control department chooses A1 (opening the emergency lane), the private vehicle chooses B2 (changing lanes left), and the emergency vehicle chooses C1 (applying for right of way), then: Traffic efficiency is increased to 0.75, and priority protection is increased to 0.5 → U_Traffic Control = 0.7 × 0.75 + 0.3 × 0.5 = 0.675; The social vehicle saving rate is 20%, and the stability is 0.8 → U_social = 0.2 × 0.8 = 0.16; Emergency vehicle reliability is 0.95 → U_Emergency = 0.95.

[0158] The backward induction method is used to backtrack from the game tree leaf nodes to solve the problem, calculate the cumulative utility value of each decision path, and filter the non-inferior solution set through the Pareto optimality criterion to obtain the candidate path priority rule set; The backward induction method selects the optimal path through reverse deduction and combines Pareto optimality to ensure the validity of the solution set.

[0159] Cumulative utility calculation: Tracing back from the leaf nodes to the root node, the utility values ​​of each party are accumulated layer by layer. For example, if the leaf node utility of a path at t = 300s is (0.6, 0.2, 0.9), and the parent node utility is (0.5, 0.1, 0.8), then the cumulative utility is (0.6 + 0.5, 0.2 + 0.1, 0.9 + 0.8) = (1.1, 0.3, 1.7). The three-party utilities are normalized to avoid differences in numerical dimensions.

[0160] Pareto optimal screening: Define the conditions for solution A to dominate solution B: All utility values ​​of A must be at least equal to or greater than those of B, and at least one must be strictly superior. Select non-dominated solutions from all paths. For example, if the solution set includes (1.1, 0.3, 1.7) and (1.0, 0.4, 1.5), retain the solution if it is superior in terms of traffic control and emergency utility. The final candidate set should be limited to 50-100 solutions.

[0161] Example: After a certain reverse induction, the candidate rule set includes: Rule 1: Mandatory opening of the emergency lane during peak hours (traffic control utility 1.2, social utility 0.3, emergency utility 1.8); Rule 2: Dynamic speed limit + social vehicle cooperative lane change (traffic control utility 1.0, social utility 0.5, emergency utility 1.6).

[0162] Conflict resolution is performed on the candidate rule set. The conflict intensity index is defined as the product of the path overlap rate and the priority difference. The Hungarian algorithm is used to match the optimal rule combination to generate a preliminary draft of the conflict-free path priority. Conflict resolution requires quantifying the mutual exclusivity between rules and achieving matching through optimization algorithms.

[0163] Conflict intensity calculation: Path overlap rate: The ratio of two rules affecting the same road segment in the same time period. For example, if rule 1 affects road segment AB and rule 2 affects road segment BC, the overlap rate is 0%; if both rules affect road segment BC, the overlap rate is 100%; Priority Difference: The absolute value of the difference in priority between rules for the same resource. For example, if rule 1 assigns a priority of 0.9 to the emergency vehicle and rule 2 assigns a priority of 0.7, the difference is |0.9-0.7|=0.2; Conflict intensity: overlap ratio × difference. For example, an overlap ratio of 80% and a difference of 0.2 results in a conflict intensity of 0.16.

[0164] Hungarian algorithm matching: Construct a conflict matrix: a 50×50 matrix (number of candidate rules: 50), where the elements are conflict intensities; The goal is to minimize the total conflict intensity and find the optimal rule combination. For example, choosing rules 1, 3, and 5 to form a set has the lowest total conflict intensity. The Hungarian algorithm solves the problem through steps such as row reduction, column reduction, and covering zero elements, with a time complexity of O(n³).

[0165] Example: After matching, the optimal combination is rules 1, 4, and 7. The total conflict intensity is reduced from the initial 1.2 to 0.3, eliminating 80% of the conflicts.

[0166] The draft of the conflict-free path priority is input into the reinforcement learning fine-tuner, and the strategy is optimized with the average delay reduction rate of the road network as the reward signal, and the optimal path priority rule including the emergency lane preemption threshold, social vehicle yielding rule and dynamic road right allocation table is output.

[0167] Reinforcement learning fine-tuning optimizes rule parameters through trial and error to improve adaptability to actual scenarios.

[0168] State space design: average road network delay (seconds); emergency vehicle arrival rate (vehicles / minute); social vehicle density (vehicles / km).

[0169] Action space design: adjust the emergency lane preemption threshold (e.g., reduce the minimum delay allowed for preemption from 30 seconds to 20 seconds); modify the rules for giving way to non-vehicle vehicles (e.g., increase the giving way distance from 50 meters to 70 meters); and update the dynamic road right allocation table (e.g., increase the road right weight of emergency vehicles during peak hours from 0.7 to 0.8).

[0170] Reward function: For example, if the original delay is 60 seconds and the new delay is 45 seconds, then R = (60-45) / 60 = 0.25.

[0171] Algorithm selection: Deep Q-network (DQN) is used, with a 3-layer fully connected network structure (input 3D state, 64 / 32 hidden layers, output 4D action); the experience replay buffer size is 10,000, the number of batch training samples is 128, and the learning rate is 0.001.

[0172] Training process: Initial 100,000-step exploration phase (ε=0.9), gradually reducing to ε=0.1; evaluate policy performance every 1,000 steps, and retain the top 10% of rules.

[0173] Output results: Emergency lane preemption threshold: Preemption is allowed when the delay is ≥ 25 seconds; Social vehicle yielding rule: Emergency vehicles must yield the left lane if detected within 200 meters; The dynamic road right allocation table is shown in Table 1: Table 1

[0174] It can be seen that based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map of the entire road network; based on the real-time risk level map, a dynamic diversion plan including variable lane markings and speed limit strategies is output; based on the dynamic diversion plan, a vehicle trajectory optimization strategy is generated; based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of real-time drone inspections, an anti-interference optimization control strategy is output, which enables fast and accurate path adjustment and coordinated control, thereby improving the level of emergency management on highways.

[0175] Another embodiment of the present invention provides a highway emergency management system, see Figure 3 , the system may include: Prediction module 301 is used to construct a multi-dimensional heterogeneous data fusion model through a spatiotemporal graph neural network based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, dynamically predict the risk level of road sections, and obtain a real-time risk level map for the entire road network; Output module 302 is configured to output a dynamic diversion solution including variable lane markings and speed limit strategies based on the real-time risk level map using a multi-agent reinforcement learning algorithm with the optimization goals of minimizing emergency response time and maximizing traffic efficiency; A control module 303 is configured to broadcast path adjustment information to vehicle terminals via a communication system according to the dynamic diversion scheme, perform coordinated control of vehicle trajectories using a distributed robust optimization algorithm, and generate a vehicle trajectory optimization strategy; The correction module 304 is used to perform dynamic conflict simulation based on the vehicle trajectory optimization strategy and the sudden obstacle data of the real-time inspection of the drone using a digital twin model, correct the conflict path planning through the game theory decision model, and output the anti-interference optimization control strategy.

[0176] It can be seen that based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map of the entire road network; based on the real-time risk level map, a dynamic diversion plan including variable lane markings and speed limit strategies is output; based on the dynamic diversion plan, a vehicle trajectory optimization strategy is generated; based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of real-time drone inspections, an anti-interference optimization control strategy is output, which enables fast and accurate path adjustment and coordinated control, thereby improving the level of emergency management on highways.

[0177] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0178] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for executing the following steps: S201: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed using a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network. S202: Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output; S203, broadcasting path adjustment information to vehicle terminals via a communication system according to the dynamic diversion solution, using a distributed robust optimization algorithm to collaboratively control vehicle trajectories and generate a vehicle trajectory optimization strategy; S204, based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data of the real-time inspection by the UAV, a digital twin model is used to perform dynamic conflict simulation, and the conflict path planning is corrected through the game theory decision model to output an anti-interference optimization control strategy.

[0179] It can be seen that based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map of the entire road network; based on the real-time risk level map, a dynamic diversion plan including variable lane markings and speed limit strategies is output; based on the dynamic diversion plan, a vehicle trajectory optimization strategy is generated; based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of real-time drone inspections, an anti-interference optimization control strategy is output, which enables fast and accurate path adjustment and coordinated control, thereby improving the level of emergency management on highways.

[0180] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0181] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0182] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed using a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network. S202: Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output; S203, broadcasting path adjustment information to vehicle terminals via a communication system according to the dynamic diversion solution, using a distributed robust optimization algorithm to collaboratively control vehicle trajectories and generate a vehicle trajectory optimization strategy; S204, based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data of the real-time inspection by the UAV, a digital twin model is used to perform dynamic conflict simulation, and the conflict path planning is corrected through the game theory decision model to output an anti-interference optimization control strategy.

[0183] It can be seen that based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map of the entire road network; based on the real-time risk level map, a dynamic diversion plan including variable lane markings and speed limit strategies is output; based on the dynamic diversion plan, a vehicle trajectory optimization strategy is generated; based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of real-time drone inspections, an anti-interference optimization control strategy is output, which enables fast and accurate path adjustment and coordinated control, thereby improving the level of emergency management on highways.

[0184] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A highway emergency management method, characterized in that: The method comprises: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network. Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output; According to the dynamic diversion scheme, the path adjustment information is broadcast to the vehicle terminals through the communication system, and the vehicle trajectories are collaboratively controlled using a distributed robust optimization algorithm to generate a vehicle trajectory optimization strategy; Based on the vehicle trajectory optimization strategy and combined with the sudden obstacle data of real-time drone inspections, a digital twin model is used to perform dynamic conflict simulation. The conflict path planning is corrected through a game theory decision model, and an anti-interference optimization control strategy is output.

2. The method according to claim 1, characterized in that Based on the traffic flow, meteorological data and accident reports collected by roadside sensing equipment, a multi-dimensional heterogeneous data fusion model is constructed through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network, including: Based on traffic flow, meteorological data, and accident reports collected by roadside sensing equipment, the multi-source heterogeneous data is spatiotemporally aligned using chaos theory coding methods to generate spatiotemporally synchronized feature tensors. Based on the spatiotemporal synchronized feature tensor, the spatiotemporal convolutional layer of the spatiotemporal graph neural network is used to extract the spatiotemporal dependencies of local road sections and obtain the spatiotemporal fusion feature matrix. Based on the spatiotemporal fusion feature matrix, the contribution of different meteorological factors to the risk level is dynamically adjusted using the meteorological impact weight allocation algorithm to generate a risk-sensitive feature vector. Based on the risk-sensitive feature vector, the risk index of each road section is calculated through a dynamic evolution prediction model, and a real-time risk level map of the entire road network is generated in combination with the road network topology.

3. The method according to claim 2, characterized in that Based on the real-time risk level map, a multi-agent reinforcement learning algorithm is used to minimize emergency response time and maximize traffic efficiency, and a dynamic diversion solution including variable lane markings and speed limit strategies is output, including: Based on the road section risk index of the real-time risk level map, the state space and action space of multi-agent reinforcement learning are constructed. The state space includes risk distribution and lane occupancy rate, and the action space is a set of variable lane signs and speed limit instructions. Based on the state space and action space, a dual-objective reward function is designed. The first objective is the negative inverse of the emergency response time, and the second objective is the logarithmic function of the road section traffic efficiency. This generates the agent training signal. Based on the agent training signals, the agents are trained using a master-slave collaborative training strategy, where the master agent generates global diversion instructions and the slave agents optimize local lane control strategies to output a preliminary diversion instruction set using the trained master and slave agents. Based on the preliminary diversion instruction set, the multi-agent decision conflicts are eliminated through the game equilibrium algorithm to generate a conflict-free diversion instruction sequence; The conflict-free diversion instruction sequence is encoded into variable lane marking control signals and gradient speed limit strategies to generate a dynamic diversion plan.

4. The method according to claim 3, characterized in that According to the dynamic diversion scheme, the path adjustment information is broadcast to the vehicle terminal through the communication system, and the vehicle trajectory is collaboratively controlled using a distributed robust optimization algorithm to generate a vehicle trajectory optimization strategy, including: parsing the variable lane identification information in the dynamic diversion scheme and broadcasting lane change instructions, target speed limits, and expected travel time windows to affected vehicles via the V2X communication system; Receive real-time position, speed, and communication delay data from the vehicle terminal, build a robust optimization objective function for the vehicle kinematic model, and define the trajectory tracking error tolerance threshold; A distributed alternating direction multiplier method is used to solve the robust optimization objective function and generate a vehicle trajectory offset sequence that satisfies spatiotemporal constraints; The vehicle trajectory offset sequence is smoothed and filtered to eliminate the risks of sharp turns and speed mutations, and a vehicle trajectory optimization strategy containing a timestamp-coordinate-speed triple is output.

5. The method according to claim 4, characterized in that Based on the vehicle trajectory optimization strategy, combined with the sudden obstacle data of the real-time inspection by the UAV, the digital twin model is used to perform dynamic conflict simulation, the conflict path planning is corrected through the game theory decision model, and the anti-interference optimization control strategy is output, including: Receive the sudden obstacle point cloud data uploaded by the UAV inspection system, perform spatiotemporal matching with the vehicle trajectory optimization strategy, and identify potential conflict areas between the trajectory and obstacles; Loading road network 3D topology, vehicle trajectory, and obstacle data into the digital twin model, conducting multi-agent collaborative simulation, and detecting vehicle-obstacle collision probabilities and congestion-derived risks. Construct a multi-party game decision-making model, define the utility functions of the traffic management department, social vehicles, and emergency vehicles, and generate the optimal path priority rules through reverse induction; Reconstructing vehicle trajectory constraints based on the optimal path priority rule, replanning the vehicle path in the conflict area using a model predictive control algorithm, and generating a revised trajectory instruction set; The corrected trajectory command is integrated with the original diversion plan to output an anti-interference optimization control strategy that includes emergency lane preemption authority, dynamic road right allocation, and secondary accident warning rules.

6. The method according to claim 5, characterized in that The digital twin model is loaded with road network 3D topology, vehicle trajectory, and obstacle data to conduct multi-agent collaborative simulation to detect vehicle-obstacle collision probability and congestion-derived risks, including: Based on the 3D topological data of the road network, a dynamic voxel partitioning algorithm is used to discretize the road space into spatiotemporally correlated voxel units. Each voxel contains position coordinates, lane attributes, and capacity parameters, generating a dynamic voxelized road network model. Based on the dynamic voxelized road network model, the vehicle trajectory data is mapped into a moving voxel sequence with a time stamp, and the obstacle point cloud data is converted into a static dangerous voxel set to obtain a spatiotemporally coupled traffic element distribution map; Based on the traffic element distribution map, a multi-agent collaborative simulation engine was constructed. The vehicle agent's motion decision-making rules were defined as a risk-aware obstacle avoidance strategy, and the obstacle agent's attribute was defined as an immovable threat source. This generated an interactive simulation scenario. The Monte Carlo method is used to simulate the motion trajectory of the vehicle agent in the interactive simulation scene. The contact probability between the vehicle and the dangerous factors is calculated through the spatiotemporal propagation model, and the collision risk probability matrix is ​​output. Based on the collision risk probability matrix and combined with the spatiotemporal distribution data of vehicle density, the probability of secondary congestion is evaluated through the congestion ripple effect prediction algorithm to generate a comprehensive risk map that includes collision hotspots and congestion-derived risk levels.

7. The method according to claim 5, characterized in that The multi-party game decision model is constructed to define the utility functions of the traffic management department, social vehicles, and emergency vehicles, and to generate the optimal path priority rules through the reverse induction method, including: The utility function of the traffic management department is defined as the weighted sum of the global traffic efficiency of the road network and the priority guarantee of emergency vehicles. The utility function of social vehicles is defined as the product of travel time saving rate and path stability. The utility function of emergency vehicles is defined as the path reliability index. This generates a prototype of the three-party utility function. Based on the three-party utility function prototype, a dynamic game tree structure is constructed, wherein the main node of the game tree represents the decision moment, the branches represent the possible action combinations of each party, and the leaf nodes are associated with the utility function value; The backward induction method is used to backtrack from the game tree leaf nodes to solve the problem, calculate the cumulative utility value of each decision path, and filter the non-inferior solution set through the Pareto optimality criterion to obtain the candidate path priority rule set; Conflict resolution is performed on the candidate rule set. The conflict intensity index is defined as the product of the path overlap rate and the priority difference. The Hungarian algorithm is used to match the optimal rule combination to generate a preliminary draft of the conflict-free path priority. The draft of the conflict-free path priority is input into the reinforcement learning fine-tuner, and the strategy is optimized with the average delay reduction rate of the road network as the reward signal, and the optimal path priority rule including the emergency lane preemption threshold, social vehicle yielding rule and dynamic road right allocation table is output.

8. A highway emergency control system, characterized in that: The system comprises: The prediction module is used to construct a multi-dimensional heterogeneous data fusion model based on traffic flow, meteorological data and accident reports collected by roadside sensing equipment through a spatiotemporal graph neural network to dynamically predict the risk level of road sections and obtain a real-time risk level map for the entire road network; An output module is configured to output a dynamic diversion solution including variable lane markings and speed limit strategies based on the real-time risk level map using a multi-agent reinforcement learning algorithm with the optimization goals of minimizing emergency response time and maximizing traffic efficiency; a control module configured to broadcast path adjustment information to vehicle terminals via a communication system according to the dynamic diversion scheme, perform collaborative control of vehicle trajectories using a distributed robust optimization algorithm, and generate a vehicle trajectory optimization strategy; The correction module is used to perform dynamic conflict simulation based on the vehicle trajectory optimization strategy and the sudden obstacle data of the real-time inspection of the drone using a digital twin model, correct the conflict path planning through the game theory decision model, and output the anti-interference optimization control strategy.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 7 when run.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

Images