Method and apparatus for handling implementation of multi-point coupling events for a highway

Abstract

The application provides a method and device for handling highway multi-point coupling events, coupling relationship modeling based on traffic events on the highway, and then constructing a directed weighted graph. According to the directed weighted graph, the resource state vector and the environment parameter vector, a structured prompt text is generated, and an initial decision scheme is generated through a large model. The current scene is matched with the historical cases in the case library to determine the matching historical cases. Based on the initial decision scheme and the matching historical cases, and through the multi-objective optimization model, the optimization solution is obtained under the set target loss function and constraint condition to obtain the refined handling scheme. In the scheme, the coupling relationship modeling of the traffic event breaks through the traditional task independence assumption, provides a theoretical basis for global collaborative optimization, and optimizes the matching historical cases as the initial solution, solves the defects that the traditional multi-agent reinforcement learning training cycle is long and cannot meet the real-time requirements of online scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems, and more specifically, to a method and apparatus for handling multi-point coupled events on highways. Background Technology

[0002] With the integration of Large Language Models (LLM) into traffic control, transportation infrastructure is being endowed with "cognitive" capabilities. Future highways will not merely be physical roads, but rather social networks composed of countless "intelligent agents" with communication, memory, and reflection abilities. Multi-agent systems, through the collaborative work of multiple heterogeneous agents, can handle complex distributed decision-making problems. However, existing multi-agent methods often assume that tasks are independent, neglecting the complex coupling relationships between tasks in multi-event scenarios on highways. These coupling relationships manifest as spatial propagation (upstream events affect downstream traffic flow), temporal dependence (event handling is subject to temporal constraints), resource competition (limited resources are allocated among multiple tasks), and causal relationships (one event may trigger other events). Furthermore, traditional multi-agent reinforcement learning methods require long-term offline training (typically several days to weeks), making it difficult to meet the real-time requirements of online emergency scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for handling multi-point coupled events on highways, so as to achieve global collaborative optimization and meet the high real-time requirements of online scenarios.

[0004] In a first aspect, the present invention provides a method for handling multi-point coupled events on highways, the method comprising:

[0005] Based on the current traffic events on the highway, a coupling relationship model is performed. Based on the coupling relationship modeling results, a directed weighted graph is constructed. The directed weighted graph includes a set of nodes, a set of edges, and an edge weight function. Each node in the set of nodes represents a traffic event. Each edge in the set of edges represents a coupling relationship between two traffic events. The edge weight function characterizes the coupling strength between the two traffic events.

[0006] Structured prompt text is generated based on the directed weighted graph, resource state vector, and environmental parameter vector. An initial decision scheme is then generated based on the structured prompt text and a large model.

[0007] The current scenario is matched with historical cases in the case library to determine the matching historical cases;

[0008] Based on the initial decision scheme and matching historical cases, a refined solution is obtained by optimizing the solution using a multi-objective optimization model under the set objective loss function and constraints.

[0009] Secondly, the present invention provides a device for handling multi-point coupled events on highways, the device comprising:

[0010] The module is used to model the coupling relationship of traffic events in the current scenario on the highway, and to construct a directed weighted graph based on the coupling relationship modeling results. The directed weighted graph includes a set of nodes, a set of edges, and an edge weight function. Each node in the set of nodes represents a traffic event, each edge in the set of edges represents a coupling relationship between two traffic events, and the edge weight function characterizes the coupling strength between the two traffic events.

[0011] The generation module is used to generate structured prompt text based on the directed weighted graph, resource state vector, and environmental parameter vector, and to generate an initial decision scheme based on the structured prompt text and a large model.

[0012] The matching module is used to perform similarity matching between the current scene and historical cases in the case library to determine the matching historical cases;

[0013] The disposal implementation module is used to obtain a refined disposal plan based on the initial decision scheme and matched historical cases, and through a multi-objective optimization model under the set objective loss function and constraints.

[0014] This invention provides a method and apparatus for handling multi-point coupled events on highways. It models the coupling relationships of traffic events in the current highway scenario and constructs a directed weighted graph based on the modeling results. Structured prompt text is generated based on the directed weighted graph, resource state vectors, and environmental parameter vectors. An initial decision plan is then generated based on the structured prompt text and a large model. The current scenario is matched with historical cases in a case library to identify matching historical cases. Based on the initial decision plan and the matching historical cases, a multi-objective optimization model is used to optimize the solution under set objective loss functions and constraints to obtain a refined handling plan. This solution overcomes the traditional task independence assumption by modeling the coupling relationships of traffic events, providing a theoretical basis for global collaborative optimization. Furthermore, by using matching historical cases as initial solutions for optimization, it addresses the shortcomings of traditional multi-agent reinforcement learning, such as long training cycles and inability to meet the real-time requirements of online scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the implementation method for handling multi-point coupled events on highways provided in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of the system architecture of the central intelligent agent and the edge intelligent agent provided in the embodiments of the present invention;

[0018] Figure 3 A schematic diagram of the hierarchical structure of the three-layer rapid decision-making architecture module provided in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram of the processing logic of the multi-agent collaborative learning mechanism module provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0021] Existing technologies for handling multiple events on highways are mainly divided into two categories: manual scheduling schemes and rule-based automated schemes.

[0022] In the manual dispatching scheme, human decision-making at the monitoring center is central. Monitors observe the road network status in real time through the video surveillance system. Upon discovering a traffic incident, they contact relevant departments via telephone or walkie-talkie. Incident priority assessment relies on the monitor's personal experience, and resource allocation decisions are made manually by the monitor based on current available resources. Each emergency response department independently conducts on-site handling operations upon receiving instructions, with inter-departmental coordination communicated verbally via telephone or walkie-talkie.

[0023] The typical process of this plan includes the following steps: incident occurrence, monitoring personnel discovery, manual priority assessment, telephone notification of relevant departments, emergency resource deployment, on-site handling, and verbal coordination. During the handling process, each emergency unit operates largely independently, lacking a system-level global collaborative optimization mechanism.

[0024] In rule-based automation solutions, an automated rule engine is used for resource scheduling. The system automatically detects traffic events through roadside sensor networks or video intelligent analysis algorithms and calculates a priority score for each event based on preset rules. Priority scores are typically based on a single-dimensional indicator, such as the severity level of the event or the number of lanes affected. The system sorts all detected events from highest to lowest priority score, forming a processing queue. Resource allocation adopts a first-come, first-served strategy; the system allocates available resources to events sequentially according to the queue order, with higher-priority events receiving resources first, and lower-priority events waiting in the queue.

[0025] The typical process of this scheme includes automatic event detection, priority calculation, queue sorting, sequential resource allocation, and independent event handling. Although the scheme achieves a certain degree of automation, it still treats each event as an independent object to be handled, without considering the coupling effects between events through mechanisms such as traffic flow propagation and resource competition.

[0026] Existing technologies have many shortcomings in dealing with concurrent multi-event scenarios on highways. These scenarios mainly have the following characteristics:

[0027] Concurrency of events: During peak traffic hours or under adverse weather conditions, multiple traffic incidents may occur simultaneously on the same section of highway. These incidents may coexist or partially overlap in time, be spatially distributed at different station locations, and be triggered independently but mutually influential.

[0028] Limited resources: The quantity of emergency response resources is subject to rigid constraints. In scenarios with multiple concurrent events, the contradiction between resource supply and demand is prominent, making it impossible to allocate dedicated resources to each event. Therefore, it is necessary to optimize scheduling to achieve efficient resource utilization.

[0029] Task coupling: Multiple concurrent events exert coupled influence through mechanisms such as traffic flow propagation, shared rescue channels, and temporal dependencies. The handling decisions of a single event can affect the handling of other events, and local decision-making errors may trigger systemic chain reactions.

[0030] Environmental dynamism: New events may occur at any time during the handling process, resource status may change in real time, traffic flow may fluctuate dynamically, and weather conditions may deteriorate, requiring the system to have online replanning capabilities.

[0031] Existing technologies have the following main drawbacks when dealing with concurrent multi-event scenarios on highways with the aforementioned characteristics:

[0032] The lack of modeling of coupling relationships between tasks leads to suboptimal global performance: Existing technologies assume that each event is independent, which is generally valid in single-event scenarios, but deviates significantly from reality when multiple events occur concurrently. In real-world scenarios, multiple events generate complex coupling effects through various mechanisms.

[0033] Existing technologies suffer from a significant decline in decision-making quality in concurrent multi-event scenarios due to a lack of modeling of the coupling relationships between multiple events. In manual scheduling schemes, monitors, limited by cognitive abilities, struggle to fully assess complex coupling relationships and are prone to employing local optimization strategies. Rule-based schemes calculate priorities for each event independently and process them by score, neglecting the impact of coupling. While high-priority events are handled first, they may occupy critical resources (such as the only large clearing vehicle) for up to 60 minutes, causing other events to queue and increasing overall system congestion losses. Experimental data shows that in the independent handling mode, the average system congestion loss in a 4-event concurrent scenario is 1350 vehicle-hours, significantly higher than the collaborative optimization scheme that considers coupling relationships (approximately 990 vehicle-hours).

[0034] Slow response speed and lack of dynamic adaptability: Existing technologies have inherent limitations in terms of response speed and dynamic replanning.

[0035] The response delay in manual dispatching primarily stems from manual operations. From the moment a monitor detects an event to the mobilization of relevant departments, the entire process takes an average of 8 to 12 minutes. When multiple events occur simultaneously, monitors must process multiple information sources and coordinate resources from various parties, leading to an exponential increase in decision-making complexity. Practice shows that when more than three concurrent events occur, monitors are prone to decision-making confusion, coordination errors, and a significant decline in the quality of handling the situation.

[0036] While rule-based solutions offer a fast response time (1 to 2 minutes), their static optimization strategy cannot handle dynamic changes during the response process. New events may arise during the response, increasing the number of tasks; recovery vehicles may arrive later than expected due to congestion; rescue teams may need to return temporarily to obtain additional equipment; and deteriorating weather conditions (visibility decreasing from 200 meters to 50 meters) may affect operational safety. These changes render the initial plan ineffective, but rule-based solutions lack a replanning mechanism and continue to execute the original plan, resulting in poor response outcomes.

[0037] While traditional multi-agent reinforcement learning (MARL) can theoretically solve multi-task cooperative optimization problems, it requires long-term offline training in a simulation environment. A typical MARL training cycle takes several days to several weeks, during which the agent needs to interact with the environment millions of times to learn the optimal policy. This offline training mode is unsuitable for online emergency scenarios, as these scenarios require the generation of response plans immediately after an event occurs (usually within one minute).

[0038] Based on the above research findings, this invention provides a solution for handling multi-point coupled events on highways. By modeling the coupling relationship of traffic events, it breaks through the traditional assumption of task independence, provides a theoretical basis for global collaborative optimization, and uses matching historical cases as initial solutions for optimization, thereby solving the shortcomings of traditional multi-agent reinforcement learning, such as long training cycles and inability to meet the real-time requirements of online scenarios.

[0039] Please see Figure 1 This is a flowchart illustrating a method for handling multi-point coupled events on highways, provided by an embodiment of the present invention. This method can be executed by a device for handling multi-point coupled events on highways. This device can be implemented in software and / or hardware and can be configured in an electronic device, such as a computer, server, or, for example, a server used for traffic management monitoring. The detailed steps of this method for handling multi-point coupled events on highways are described below.

[0040] S11, perform coupling relationship modeling based on traffic events in the current scenario on the highway, and construct a directed weighted graph based on the coupling relationship modeling results;

[0041] A directed weighted graph consists of a set of nodes, a set of edges, and an edge weight function. Each node in the set of nodes represents a traffic event, each edge in the set of edges represents a coupling relationship between two traffic events, and the edge weight function characterizes the coupling strength between two traffic events.

[0042] S12, generate structured prompt text based on the directed weighted graph, resource state vector and environmental parameter vector, and generate an initial decision scheme based on the structured prompt text and the large model;

[0043] S13, perform similarity matching between the current scene and historical cases in the case library to determine the matching historical cases;

[0044] S14, based on the initial decision scheme and matching historical cases, and through a multi-objective optimization model, optimizes the solution under the set objective loss function and constraints to obtain a refined disposal scheme.

[0045] The method for handling multi-point coupled events on highways provided in this embodiment breaks through the traditional assumption of task independence by modeling the coupling relationship of traffic events, providing a theoretical basis for global collaborative optimization. Furthermore, it uses matching historical cases as initial solutions for optimization, solving the shortcomings of traditional multi-agent reinforcement learning, such as long training cycles and inability to meet the real-time requirements of online scenarios.

[0046] The specific implementation methods of each of the above steps will be explained in detail below.

[0047] like Figure 2As shown in the illustration, the processing implementation method provided in this embodiment is applied to a collaborative processing system comprising a central (scheduling) intelligent agent and multiple edge (dedicated) intelligent agents. The central intelligent agent performs global collaborative processing, while each edge intelligent agent can perform task processing for specific tasks, including event detection, application processing, traffic guidance, congestion prediction, and drone scheduling. Each edge intelligent agent can communicate with its corresponding actuator to achieve data and instruction interaction. The central and edge intelligent agents can be devices with data and image processing capabilities, such as computers and servers.

[0048] Specifically, the central agent is deployed on a cloud server (configuration: 32-core CPU, 128GB RAM, 4 GPUs), responsible for global coordination, three-layer decision-making, experience sharing pool management, and federated aggregation. Each edge agent is deployed on an edge server (configuration: 8-core CPU, 32GB RAM), responsible for local perception, task execution, and model training. All agents communicate via a dedicated 5G network, with cloud-edge communication latency <50ms and bandwidth >100Mbps.

[0049] At the system level, the system architecture used to implement this disposal method mainly includes three core modules: a task spatiotemporal coupling network module, a three-layer rapid collaborative decision-making architecture module, and a multi-agent collaborative learning mechanism module.

[0050] Among them, the task spatiotemporal coupling network module can explicitly characterize the complex coupling relationship between traffic events by constructing a task spatiotemporal coupling network, breaking through the traditional assumption of task independence and providing a theoretical basis for global collaborative optimization.

[0051] Traffic incidents on highways can occur concurrently, and these concurrent incidents can have complex coupling relationships. In this embodiment, coupling relationship modeling is performed based on traffic incidents in the current highway scenario. Specifically, this can be achieved in the following way:

[0052] For each traffic event in the current scenario on the highway, construct the attribute vector of the traffic event; for any two traffic events, perform spatial coupling modeling, temporal coupling modeling, resource coupling modeling and causal coupling modeling based on the attribute vectors of the two traffic events respectively.

[0053] Spatial coupling modeling results are used to determine the impact of the handling of one traffic event on the other traffic event; temporal coupling modeling results are used to determine whether the two traffic events are executed serially or in parallel; resource coupling modeling results are used to determine the resource allocation order and dynamic task switching strategy; and causal coupling modeling results are used to determine the handling order of traffic events and the chain reactions they induce.

[0054] Specifically, spatial coupling manifests as adjacent events propagating through traffic flow and generating mutual geographic influences. When an upstream event causes a complete closure, vehicles cannot pass, resulting in downstream clearing or rescue vehicles being blocked in the congested area. For example, a complete closure of the K10 marker causes a clearing vehicle that would normally reach the K15 marker in 5 minutes to be stuck upstream of K10, extending its actual arrival time to 25 minutes, severely impacting the timeliness of handling the K15 event.

[0055] Temporal coupling manifests as dependencies and constraints on event handling over time. For example, in the event of a truck overturning, the rescue team must first confirm the safety of the scene before large recovery vehicles can enter to work. If the rescue team is assigned to other events, the recovery vehicles must wait even if they arrive, creating a sequential constraint.

[0056] Resource coupling manifests as competition for limited emergency resources among multiple events. For example, a road segment may have two clearing vehicles, but three events all require clearing support, resulting in a resource supply-demand ratio of 1:1.5, indicating a significant shortage. Resource allocation decisions directly impact the timing of event response and the overall system performance.

[0057] Causal coupling manifests as the mishandling of one event potentially triggering or exacerbating other events. For example, if a complete road closure lasts too long, upstream vehicles may queue up, creating long traffic jams. Frequent sudden braking by vehicles dramatically increases the risk of rear-end collisions. Statistics show that the probability of a secondary rear-end collision at the tail of a congested queue is approximately 15 times higher than under normal driving conditions.

[0058] In this embodiment, the directed weighted graph constructed based on the coupling relationship modeling results can be represented as follows: ,in:

[0059] For a set of nodes, each node This represents a traffic incident;

[0060] Let be the set of edges, each edge Indicates a traffic incident Traffic incidents There is a coupling relationship between them;

[0061] Let the edge weight function be... This indicates the coupling strength between two traffic events.

[0062] For any node in the node set Define its attribute vector as a six-tuple:

[0063]

[0064] The meanings of each attribute component in the attribute vector are as follows: Event spatial location (in terms of station number, unit: kilometers); ∈ {rear-end collision, rollover, spillage, malfunction, weather}: Event type; Severity level: 1 indicates mild, 5 indicates severe; : Time of the event (unit: minutes); : Estimated processing time (unit: minutes); : Affected area (number of lanes or length of road segment, unit: lanes or kilometers).

[0065] Based on the attribute vectors of traffic events, spatial coupling modeling, temporal coupling modeling, resource coupling modeling, and causal coupling modeling are performed respectively.

[0066] Spatial coupling modeling based on the attribute vectors of two traffic events can be achieved in the following way:

[0067] The geographical distance between two traffic events is calculated based on their spatial locations; a traffic flow propagation function is constructed based on the spatial locations and handling methods of the two traffic events; and the spatial coupling strength is calculated based on the geographical distance, the traffic flow propagation function, and the set impact attenuation characteristic distance.

[0068] Spatial coupling characterizes the geospatial interactions between traffic events. For any two traffic event nodes... and The spatial coupling strength is defined as:

[0069]

[0070] In the formula: Traffic incident Traffic incidents Geographical distance (station difference, unit: kilometers); The distance that affects the attenuation characteristic is usually taken as... kilometer; Spatial coupling strength coefficient, usually taken as ; : Traffic flow propagation function, with a value range of [0, 1].

[0071] Traffic flow propagation function The relative positions of traffic incidents are determined based on their individual spatial locations, and this is combined with the handling methods.

[0072]

[0073] This function characterizes the propagation impact of upstream events on downstream traffic flow: when an upstream event is fully closed, downstream traffic is completely blocked, and the propagation impact is the strongest (φ = 1.0); when a partial lane closure is adopted, downstream traffic is partially restricted, and the propagation impact is moderate (φ = 0.4); downstream events have a very small impact on upstream events (φ = 0.1).

[0074] Spatial coupling modeling enables the system to identify that when the upstream K10 chainage event is handled with full closure, it will block the rescue channel to the downstream K15 chainage. In this case, the handling method of the K10 event should be adjusted (change to occupying the emergency lane and keeping the main lane open) or priority should be given to dispatching rescue resources that are closer to K15.

[0075] Furthermore, time-coupled modeling based on the attribute vectors of two traffic events can be achieved in the following way:

[0076] Temporal coupling describes the dependencies and constraints of event handling in the time dimension. A directed acyclic graph of task temporal dependencies is constructed. ,in:

[0077] (1) Deterministic temporal dependency: If traffic events The handling of traffic incidents must be done Before starting, establish directed edges. The edge weights are set as follows:

[0078]

[0079] Typical deterministic dependencies include: rescue team confirms site safety → clearing vehicle enters work area; drone completes site reconnaissance → emergency response plan is developed.

[0080] (2) Triggering probability modeling: If a traffic event Improper handling may trigger traffic incidents The edge weights are set to the induced probability:

[0081]

[0082] in Indicates a traffic incident Traffic incidents under conditions of failure to handle The probability of occurrence, estimated using historical statistical data. Typical precipitating relationships include:

[0083] Rear-end collision handling time exceeding 30 minutes → traffic congestion behind triggers secondary rear-end collisions. A truck overturned, causing a hazardous materials leak that was not promptly controlled, potentially leading to an environmental pollution incident. .

[0084] Temporal coupling modeling enables the system to identify which tasks must be executed sequentially, which tasks can be processed in parallel, and how to reduce the risk of secondary accidents through rapid response.

[0085] Furthermore, resource coupling modeling based on the attribute vectors of two traffic events can be achieved in the following way:

[0086] Obtain the total number of resource types and the total available amount of each type of resource in the current scenario, and obtain the demand for each type of resource for each of the two traffic events; construct resource allocation constraints based on the demand for each type of resource by each traffic event and the total available amount of each type of resource; define the arrival time function of each type of resource to each traffic event based on the spatial location of each traffic event and the location distribution of each type of resource.

[0087] Resource coupling describes the competition and constraint relationships of limited resources among multiple tasks. A resource demand matrix is ​​defined. With resource availability vector ,in: Total number of traffic incidents; Total number of resource types; Traffic incident For resource types The demand (an integer representing the quantity of resources required); Resource type The total available amount (integer).

[0088] Resource allocation constraints are constructed as follows:

[0089]

[0090] in For binary decision variables, Indicates the resource type Assigned to traffic events .

[0091] Considering the location distribution of resources, define resources Traffic incident Arrival time function:

[0092]

[0093] In the formula: Current time (unit: minutes); :resource Current location (station number, unit: kilometers); Traffic incident Spatial location (station number, unit: kilometers); : The function for the distance traveled between two points; Average driving speed, usually taken as km / h; Delay time caused by traffic congestion (unit: minutes) is calculated in real time by the congestion prediction module.

[0094] Resource coupling modeling enables the system to identify how to optimize resource allocation order and dynamic task switching strategies when multiple traffic events compete for limited resources, so as to avoid critical resources being occupied by a single event for a long time, which would cause delays in the handling of other events.

[0095] Finally, causal coupling modeling based on the attribute vectors of the two traffic events can be achieved in the following way:

[0096] Calculate the spatiotemporal correlation evidence score, semantic correlation evidence score, and domain rule evidence score between two traffic events; sum the spatiotemporal correlation evidence score, semantic correlation evidence score, and domain rule evidence score according to their weights to obtain the causal correlation probability.

[0097] Causal coupling describes the causal chains and chain reactions between events. Constructing causal reasoning networks. It is used to identify root cause events and derivative events.

[0098] For any two traffic events and Calculate the probability of causal association:

[0099]

[0100] In the formula: Spatiotemporal correlation evidence score; Semantic relevance evidence score; Domain rule evidence score; Weighting coefficients, satisfying , usually take , , .

[0101] Among them, spatiotemporal correlation evidence: if a traffic incident Traffic incident Then, and satisfying:

[0102]

[0103] This suggests a possible spatiotemporal correlation. Typically, we take... kilometer, Minutes. The spatiotemporal correlation evidence score is calculated as follows:

[0104]

[0105] Semantic association evidence: Utilizing large-scale language models to analyze event description text and identify semantic causal relationships. For example, in traffic events... and traffic incidents The text description is input into the model, and the model outputs a semantic association evidence score. .

[0106] Domain-Specific Rule Evidence: Establishing a Set of Causal Rules Based on a Traffic Safety Domain Knowledge Base Typical rules include:

[0107] Rule 1: Full closure and handling time > 30 minutes → Upstream congestion → Increased risk of rear-end collision; Rule 2: Rainy or foggy weather and vehicle speed > 80 km / h → Increased risk of rear-end collision; Rule 3: Hazardous materials transport vehicle accident → Secondary environmental pollution accident.

[0108] If a traffic incident and traffic incidents If a causal rule is matched, then ,otherwise .

[0109] when (The threshold is usually set to 0.6) When a directed edge is established in a causal network, The edge weight is .

[0110] Causal coupling modeling enables the system to identify: root cause events should be dealt with first, and the handling of derivative events should take into account the impact of the root cause events; at the same time, it can predict the chain reaction that may be induced by improper handling of a certain event and formulate preventive measures in advance.

[0111] In this embodiment, the spatial coupling modeling, temporal coupling modeling, resource coupling modeling and causal coupling modeling constructed above are comprehensively reflected in the directed weighted graph constructed based on the modeling results of coupling relationships.

[0112] It should be noted that in this embodiment, coupling modeling can be performed based on any one or more of the above-mentioned coupling modeling methods. For example, only spatial coupling modeling and resource coupling modeling can be established, while temporal coupling modeling and causal coupling modeling are ignored to reduce computational complexity. Spatial coupling modeling can be easily quantified through geographical distance (e.g., Resource coupling occurs if the resource requirement is within a certain range (e.g., strong / medium / weak), otherwise it is not, and so on. Resource coupling is expressed through a resource demand matrix. The coupling strength is represented by a discretized value (strong / medium / weak) rather than a continuous value, simplifying the modeling process.

[0113] Thus, it has low computational overhead and is simple to implement, making it suitable for edge computing devices. However, the accuracy of coupled modeling is reduced, the optimization effect is slightly inferior to the complete solution, and the reduction in congestion loss is weakened.

[0114] Building upon the above, a three-layer rapid collaborative decision-making architecture module is utilized to integrate three decision-making paradigms: large-model reasoning, case retrieval, and mathematical optimization. This enables real-time collaborative decision-making within minutes, addressing the issue of long training cycles in traditional multi-agent reinforcement learning. Combined with... Figure 3 As shown, the three-layer rapid collaborative decision-making architecture module mainly includes three layers: large model reasoning layer (Layer 1), case retrieval layer (Layer 2), and mathematical optimization layer (Layer 3).

[0115] The large-scale model inference layer utilizes a pre-trained traffic-specific large-scale model for semantic reasoning and task decomposition, enabling online inference without offline training. Specifically, it generates structured prompt text based on a directed weighted graph, resource state vectors, and environmental parameter vectors. Based on this structured prompt text, an initial decision scheme is generated using the large-scale model. This step can be implemented as follows:

[0116] The design incorporates a toolset that includes spatial analysis tools, temporal reasoning tools, and risk assessment tools. Structured prompt text is generated based on directed weighted graphs, resource state vectors, and environmental parameter vectors. Based on this structured prompt text, the tools within the toolset are processed through a large model to obtain the analysis results of each tool. Semantic reasoning is then performed on the analysis results of each tool to generate an initial decision-making scheme.

[0117] In this embodiment, the required information elements include: a task-coupled network (directed weighted graph). Resource state vector ; Environmental parameter vector .

[0118] In addition, a tool enhancement mechanism was designed to improve the ability of large models to handle task coupling relationships, and a toolset was designed. ,include:

[0119] Spatial analysis tools for calculating the geographical extent of influence between events; A time-series reasoning tool that identifies dependencies and parallel opportunities; Risk assessment tool to predict the probability of secondary accidents.

[0120] The large-scale model's inference process is as follows: First, a structured prompt text is constructed based on a directed weighted graph, resource state vectors, and environmental parameter vectors. This prompt text contains attribute information for all traffic events, a coupling matrix, and the status of available resources. Second, the large-scale model autonomously decides which tools to invoke, when to invoke them, and how to combine their outputs based on the structured prompt text. Finally, the large-scale model performs semantic inference based on the analysis results returned by each invoked tool, generating a coarse-grained initial decision scheme.

[0121] Here is an example of a structured prompt text design:

[0122] Scenario description: There are currently 4 concurrent events on the K10-K30 section of the highway;

[0123] - Event 1: K10+500, rear-end collision, blocking the left lane; - Event 2: K15+200, truck overturned, full road closure; - Event 3: K22+800, road debris; - Event 4: K30-K35, rainy / foggy weather;

[0124] Available resources: 2 tow trucks (K5 medium-sized and K20 large-sized), 3 rescue teams;

[0125] Task: Based on the analysis using the following tools, generate suggested solutions;

[0126] Tool 1: Spatial Analysis → Return: Full closure of Event 2 will block the rescue channel of Event 1; Tool 2: Temporal Reasoning → Return: Rescue confirmation is required for Event 2 before obstacle clearing, while Event 1 and Event 3 can be carried out in parallel; Tool 3: Risk Assessment → Return: Delay in handling Event 1 will induce a secondary rear-end collision, with a probability of 0.18;

[0127] Requirements: 1. Identify which events should be prioritized; 2. Suggest a sequence for resource allocation; 3. Identify potential risks and preventative measures.

[0128] The output includes:

[0129] Task grouping suggestions: Identify tasks that should be processed sequentially and those that can be executed in parallel; Key bottleneck identification: Identify high-priority tasks and resource bottlenecks; Preliminary resource allocation suggestions: Initially match resources for each task.

[0130] The following is an example of the initial decision scheme information output by the large model:

[0131] Priority: Event2 > Event1 > Event3 > Event4;

[0132] Reasons: - Event 2, with its complete closure, has the largest impact area and blocks upstream rescue channels, so it must be dealt with first; - Event 1 carries a high risk of secondary rear-end collisions and should be addressed as soon as possible; - Event 3 has a relatively smaller impact, but it can be handled in parallel with Event 1.

[0133] Resource allocation recommendations: - K20 large tow truck → Event 2 (truck rollover requires large equipment); - Rescue Team B → Event 2 (confirm safety first); - K5 medium-sized tow truck → Event 1 (support Event 3 after handling the situation); - Traffic guidance → immediately issue speed reduction warnings at K12-K15 (to prevent secondary rear-end collisions in Event 1);

[0134] Risk warning: - Event 2 may take more than 60 minutes to handle, requiring advance diversion at K10; - Event 4 requires continuous monitoring in rainy or foggy weather, which may affect rescue operations.

[0135] Building upon this, a case retrieval layer searches for similar scenarios in a historical collaborative handling case library to extract reusable handling strategy patterns. Specifically, the current scenario is matched with historical cases in the case library to identify matching historical cases. This step can be achieved in the following way:

[0136] Calculate the structured feature similarity and semantic embedding similarity between the current scene and each historical case in the case library; sum the structured feature similarity and semantic embedding similarity according to weights to obtain the comprehensive similarity; select the historical cases with the highest comprehensive similarity as the matching historical cases; wherein, the structured feature similarity is calculated based on the similarity of the number of events, the distribution of event types, the geographical distribution, and the resource configuration between the current scene and the historical cases, and the semantic embedding similarity is obtained by converting the scene description text of the current scene and the historical cases into embedding vectors and then calculating the cosine similarity between the embedding vectors.

[0137] In this embodiment, the case database DB includes multiple historical cases, each historical case Includes the following information:

[0138] Scene Description Event quantity, type distribution, location distribution, and resource allocation; handling plans. Resource allocation decisions and timelines; outcomes of responses. Congestion losses, response time, and whether secondary accidents occurred; experience summary. Successful experiences or lessons learned from failures.

[0139] Calculate the current scene using a hybrid strategy Historical Cases The overall similarity is calculated using the following formula:

[0140]

[0141] In the formula: :Structural feature similarity; Semantic embedding similarity; Weighting coefficients, satisfying , usually take , .

[0142] Structured feature similarity Calculated based on the following four dimensions:

[0143]

[0144] in: : Event quantity similarity; Event type distribution similarity (based on cosine distance of type vectors); Geographical distribution similarity (based on spatial clustering analysis); Resource configuration similarity.

[0145] Semantic embedding similarity The calculation is performed using a vector database. The scene description texts of the current scene and historical cases are converted into embedded vectors, and the cosine similarity between the embedded vectors is calculated.

[0146] The search output consists of the K historical cases with the highest overall similarity (usually K = 3) as the matching historical cases, denoted as... The corresponding solution is These matching historical cases provide a warm start initial solution for subsequent mathematical optimization layers.

[0147] The following is an example of a case search:

[0148] Current scene feature vector: [

[0150] n_events=4, types=[rear-end collision, rollover, spillage, weather], loc_span=25km, resources=[wreckers: 2, rescue teams: 3] ]

[0152] Search results:

[0153] Case#1287 (Similarity 0.89): - Scenario: 4 events, similar in type, with identical resource configuration; - Solution: Prioritize handling full-width closure events, with medium-sized clearing vehicles dynamically switching to support multiple events; - Result: 820 vehicle-hours of congestion loss, no secondary accidents;

[0154] Case #2156 (Similarity 0.84): - Scenario: 3 events, including a combination of truck rollover and rear-end collision; - Solution: Large tow trucks were prioritized for the truck rollover, with the rescue team and tow trucks working closely together; - Result: 650 vehicle-hours of traffic congestion loss, with no secondary accidents;

[0155] Case #3421 (Similarity 0.81): - Scenario: 5 events, resources are more strained; - Solution: Use dynamic task switching, the clearing vehicle immediately turns to Event 3 after completing Event 1; - Result: 1120 vehicle-hours of congestion loss, 1 secondary accident occurred; - Lesson learned: The delay in handling Event 3 led to the secondary accident, traffic guidance should be implemented in advance.

[0156] Building upon the above, the mathematical optimization layer establishes a multi-objective optimization model based on the outputs of the first two layers to solve for a refined resource allocation scheme. Specifically, based on the initial decision scheme and matching historical cases, the multi-objective optimization model is used to optimize and solve the scheme under the set objective loss function and constraints to obtain a refined disposal plan. This step can be achieved in the following way:

[0157] The system constructs constraints including resource quantity constraints, task timing constraints, coupling effect constraints, and resource arrival time constraints, and constructs a target loss function including congestion loss and risk loss for inducing secondary accidents. Using the disposal plan matching historical cases as the initial solution, a multi-objective optimization model is used to optimize the solution based on the initial decision plan and the initial solution, under the constraints and guided by the target loss function, to obtain a refined disposal plan. The refined disposal plan includes resource allocation decision, disposal timing arrangement, disposal method selection, and expected effect evaluation.

[0158] In this embodiment, the decision variables are defined as follows: Binary variable Indicates the resource type Assigned to traffic events ; Continuous variable, representing traffic events The start time of the treatment (unit: minutes); Discrete variables representing traffic events The handling method.

[0159] The objective loss function is to minimize the total congestion loss of the system, as shown below:

[0160]

[0161] in: Traffic incident Consumption losses (unit: vehicle-hours); Traffic incident Risk loss due to secondary accidents (unit: vehicle-hour).

[0162] Congestion Losses Calculated using traffic flow models:

[0163]

[0164] In the formula: Traffic incident The moment when the disposal process ends; : Traffic incidents Length of the resulting congestion queue (in kilometers); : Average waiting time for vehicles in the congestion queue at any given time (in hours).

[0165] Risks and losses that could trigger secondary accidents Defined as:

[0166]

[0167] in: Traffic incident The probability of triggering a secondary accident is provided by the causal coupling model; Secondary accident penalty coefficient, usually taken as Car and Time.

[0168] The constraints constructed include the following four categories:

[0169] (1) Resource quantity constraints:

[0170]

[0171] (2) Task timing constraints:

[0172]

[0173] (3) Coupling effect constraints:

[0174]

[0175] in The spatial coupling strength threshold is typically taken as... This constraint prevents adjacent events from simultaneously triggering full closures, which could completely block rescue access.

[0176] (4) Resource arrival time constraints:

[0177]

[0178] This constraint ensures that processing operations can only begin once resources arrive.

[0179] Historical matching cases retrieved using the aforementioned case retrieval layer Disposal plan As the initial solution. Extract Assigning values ​​to decision variables Inputting it into a multi-objective optimization model (such as the Gurobi optimization solver) as a warm start initial solution shortens the time it takes for the model to search for the optimal solution.

[0180] The optimization solution adopts the branch and bound algorithm. The optimization solver performs local search and global optimization based on the initial solution and initial decision scheme in the warm start, and finally outputs a refined solution.

[0181] The output includes: resource allocation decisions Resource allocation plans for each traffic incident; Response timeline. The start time and estimated end time of each traffic incident; the choice of handling method. : Handling methods for each traffic incident (full closure / partial closure / emergency lane); Expected effect assessment: estimated total system congestion losses, and secondary accident risks for each incident.

[0182] The following is an example of the output of the mathematical optimization layer:

[0183] Optimal solution:

[0184] Event 2 (Truck rollover, K15+200): - Resources: Large tow truck (K20), Rescue Team B; - Start Time: T+3 minutes (Rescue Team first); - Handling Method: Full closure (mandatory); - Estimated Duration: 60 minutes;

[0185] Event 1 (Rear-end collision, K10+500): - Resources: Medium-sized tow truck (K5), Rescue Team A; - Start Time: T+2 minutes (concurrent with Event 2); - Handling Method: Occupy the emergency lane (keep the main lane open); - Estimated Duration: 20 minutes;

[0186] Event 3 (Spilled Material, K22+800): - Resource: Medium Wrecker (K5, dispatched after Event 1); - Start Time: T+25 minutes (after Event 1); - Handling Method: Block main lane; - Estimated Duration: 15 minutes;

[0187] Coordinated measures: - Traffic guidance: Issue speed reduction guidance at K8, K12, and K14 at T+0 minutes; - Drones: Head to Event2 for reconnaissance at T+1 minutes, and turn to Event3 at T+20 minutes; - Congestion prediction: Continuously monitor congestion at K10-K15, and trigger diversion when necessary;

[0188] Expected results: - Total system congestion loss: approximately 890 vehicle-hours; - Secondary accident risk: 4.2% (low risk); - Total handling time: 65 minutes.

[0189] In this embodiment, the above-mentioned large model reasoning layer (Layer 1), case retrieval layer (Layer 2), and mathematical optimization layer (Layer 3) take approximately 10 seconds, 5 seconds, and 12 seconds respectively. Task decomposition and agent matching take approximately 3 seconds, for a total time of approximately 30 seconds, which meets the 1-minute real-time requirement.

[0190] It should be noted that in situations where historical case data is insufficient or during the cold start phase, a two-layer decision-making architecture can be adopted: A coarse-grained decision scheme is generated through the Layer 1 large model inference layer. Then, the Layer 3 mathematical optimization layer directly optimizes the scheme based on the Layer 1 output (without a warm start).

[0191] The advantages of this alternative are a simplified architecture and less reliance on historical data; the disadvantages are that the Layer 3 optimization layer lacks a warm start initial solution, which significantly increases the solution time from 12 seconds to minutes, extending the total response time and leading to more congestion losses and the risk of secondary accidents.

[0192] Existing manual scheduling schemes and rule-based automated schemes both lack collaborative learning mechanisms among agents, resulting in low learning efficiency. Specifically, in existing technologies, each emergency response unit adopts an independent learning mode, lacking experience sharing and collaborative evolution mechanisms, leading to low learning efficiency.

[0193] The experience barrier is a significant issue. Each response unit maintains its own independent knowledge base and decision-making model. Success stories, lessons learned from failures, and improvement suggestions are accumulated only within each unit and cannot be shared between units. For example, the "three-step method for rapid clearance of rear-end collisions in rainy nights" summarized by the clearing team was not shared with the rescue team, forcing the rescue team to explore the optimal strategy from scratch in similar scenarios. This experience barrier results in a large amount of repetitive learning.

[0194] Furthermore, the learning cycle is long, and deploying new units is difficult. Each processing unit relies solely on local data to train its model, requiring the accumulation of a large number of case samples to achieve stable performance. Practice shows that for a new processing unit to achieve an 85% decision accuracy rate from deployment, it typically needs to accumulate more than 700 processing cases, with a learning cycle of up to 3 months. When deploying a new unit, it cannot utilize the historical experience of existing units and must go through three stages from scratch: an initial exploration period (extensive trial and error), a rapid growth period (performance improvement), and a stabilization period (performance convergence).

[0195] Furthermore, there is a lack of collaborative reflection mechanisms from a collective perspective. Each processing unit makes decisions and summarizes experiences from its own perspective, pursuing optimal local performance. However, in multi-unit collaborative scenarios, the overall system performance depends not only on the individual capabilities of each unit but also on the quality of cooperation between units. Existing technologies lack mechanisms for collective review and collaborative reflection from a system-wide perspective, making it difficult to identify problems at the collaboration level (such as information transmission delays, resource allocation conflicts, and improper timing arrangements), and unable to extract generalizable collaboration patterns, resulting in slow improvement in the quality of collaboration.

[0196] Based on the above research findings, please refer to the following: Figure 4 The multi-agent collaborative learning mechanism module provided in this embodiment realizes the leap from independent learning to collaborative learning of heterogeneous agents by constructing a group learning framework that integrates experience sharing, collaborative reflection and federated training.

[0197] Based on this, the processing method provided in this embodiment may further include the following steps:

[0198] After completing their handling plans, each edge agent uploads relevant case information to a shared pool. The central agent evaluates the value of each case in the shared pool and determines whether to store or filter cases based on the evaluation results. Each edge agent learns from the case information in the shared pool to optimize its own model. The central agent collects complete data on the collaborative handling of multiple traffic accidents, evaluates performance deviations based on the complete data, and extracts collaborative pattern rules and anti-pattern rules based on the performance deviation evaluation results for each edge agent to update its own knowledge base. The central agent collects the updated model parameters uploaded by each edge agent after training its decision model based on local data, aggregates the updated model parameters to obtain a global model, and distributes the global model to each edge agent so that each edge agent can update its local decision model based on the global model.

[0199] In this embodiment, the multi-agent collaborative learning mechanism module mainly includes three mechanisms: experience sharing pool, collaborative reflection, and federated training.

[0200] The experience sharing pool is maintained by the central scheduling agent and serves as a global knowledge base accessible to all edge agents.

[0201] Data structure design, shared pool It includes four categories of knowledge items: A collection of collaborative handling case studies, each case including a scenario description, handling plan, and effect evaluation; Best practice rule sets, proven and efficient collaboration patterns; A collection of lessons learned from failures, and decision-making patterns (anti-patterns) that lead to adverse consequences. : A set of collaborative mode templates, which are reusable templates for task allocation and resource scheduling.

[0202] Case Upload and Evaluation Process: After completing the handling plan, the edge agent uploads the relevant case information to the shared pool. The central agent evaluates the value of each uploaded case; the evaluation function is defined as follows:

[0203]

[0204] In the formula: Effectiveness metrics are quantified based on factors such as congestion losses and response time. Innovation metrics measure whether a case study contains novel strategies; Reusability metrics measure the universality of a case study; Weighting coefficients, satisfying .

[0205] When the case value Below the threshold When the similarity is typically 0.6, cases are filtered out and not stored. For cases that pass the value assessment, the system performs deduplication: it calculates the similarity between the new case and the already stored cases; if the similarity exceeds a certain threshold... If the value is 0.9 (usually taken as 0.9), it is considered a duplicate case. In this case, a new case that does not duplicate the already stored cases is stored.

[0206] Case retrieval and learning: When the edge agent encounters a new task, it retrieves similar cases from the shared pool. The retrieval employs the same hybrid strategy as the case retrieval layer described above, returning Top-K similar cases. The agent learns successful strategies from the retrieved cases, avoiding repeating the failures of others, thereby accelerating the convergence of its own model (local model).

[0207] The collaborative reflection mechanism is initiated by the central agent after each multi-event collaborative handling process, organizing all peripheral agents to conduct collective reflection.

[0208] The reflection process includes five steps:

[0209] Step 1: Data Collection. Collect complete data for this collaborative response, including the initial situation (event description, resource status, environmental parameters), decision-making process (intermediate outputs of the three-layer architecture, decision-making basis of each agent), execution process (resource scheduling records, task completion time, emergencies), and final results (congestion losses, response time, whether secondary accidents occurred, expert scores).

[0210] Step 2: Performance Evaluation. Compare the actual results with the expected results to quantify the performance deviation:

[0211]

[0212]

[0213] in For actual congestion losses, To account for anticipated congestion losses, This refers to the actual processing time. The expected processing time is calculated. The causes of deviations are analyzed, such as resource allocation delays, inaccurate traffic flow forecasts, and unexpected new events.

[0214] Step 3: Problem Identification. Examine the collaboration process from a group perspective and identify problems in the collaboration. Evaluation dimensions include:

[0215] Task allocation rationality: Are there any idle resources, excessive usage, or allocation conflicts?

[0216] Appropriate timing of coordination: Can serial tasks be parallelized? Do parallel tasks compete for resources?

[0217] Timeliness of information transmission: Whether each intelligent agent is informed of changes in the task status in a timely manner;

[0218] Emergency response flexibility: Whether the plan can be quickly adjusted in the event of an emergency.

[0219] Step 4: Experience Extraction. Extract generalizable collaboration pattern rules or anti-pattern rules from the group debriefing. Collaboration pattern rules are described as triples:

[0220]

[0221] The Scenario describes the applicable scenario, the Strategy describes the collaborative strategy, and the Effect describes the expected outcome. Anti-pattern rules are also described as triples, but the Effect represents the negative consequences.

[0222] Step 5: Knowledge Consolidation. The extracted experience is consolidated into executable rules and updated to the knowledge base of each edge agent.

[0223] Successful experiences → Best practice rules Prioritize; lessons learned from failure → anti-pattern warning Triggering an alarm mechanism; New collaboration mode → Template library Expand for future reuse.

[0224] The federated training framework allows each edge agent to independently train a decision model on local data, and only uploads the updated model parameters to the central agent for federated aggregation, thus achieving group knowledge sharing while protecting data privacy.

[0225] Local training phase: Each edge agent In local dataset Training decision model The training objective is to minimize the local loss function:

[0226]

[0227] in: Status (including event description, resource status, etc.); Actions (decision-making options); : Rewards (evaluation of the effectiveness of the measures); : Parameters are Decision-making models; Loss function, such as mean squared error or cross-entropy.

[0228] Local training ensures that each edge agent can learn specific capabilities from its own unique experience, and the raw data does not leave the local device, thus protecting data privacy.

[0229] Federated aggregation phase: After each edge agent has completed training, the model parameters are updated. Uploaded to the central agent. The central agent uses a weighted average strategy to aggregate parameters:

[0230]

[0231] In the formula: Total number of edge agents; Intelligent agent The number of local training samples; Total sample size; : No. Global model parameters of the wheel.

[0232] In addition to parameter aggregation, the central agent also extracts global knowledge rules from the experience of each peripheral agent:

[0233] Statistically analyze the high-frequency decision-making patterns of each agent to identify common strategies; identify common patterns across agents (e.g., "the risk of rear-end collision increases by 40% in rainy weather"); construct a global causal graph and integrate the local causal knowledge of each agent.

[0234] Global model distribution phase: The central agent will distribute the aggregated global model Distribute the model to each edge agent, and each agent downloads and updates its local decision model:

[0235]

[0236] The global model incorporates the collective wisdom of all agents, enabling each agent to learn from the experiences of others and avoid redundant exploration.

[0237] Iterative optimization: Repeat the "local training → federated aggregation → global distribution" loop until the global model converges or reaches a preset performance threshold. The convergence criterion is:

[0238]

[0239] in The convergence threshold is usually set to [value missing]. .

[0240] The combined effects of the collaborative learning mechanism are as follows: Experience sharing pools enable knowledge flow, allowing agents to learn from others' experiences; collaborative reflection extracts group patterns, enabling the system to discover areas for improvement at the collaborative level; and federated training drives capability evolution, allowing agents to benefit from collective wisdom. These three mechanisms work synergistically to shift heterogeneous agents from independent evolution to collaborative evolution, improving learning efficiency by 150% and reducing the number of local cases required to achieve stable performance from 700 to 280.

[0241] It should be noted that in scenarios where data privacy requirements are not high or network bandwidth is sufficient, centralized learning can be used instead of federated training: each edge agent uploads its original training data to the central server; the central server performs centralized training to generate a unified model; and each edge agent downloads the unified model.

[0242] The advantages of this alternative are higher training efficiency (centralized training converges faster than federated learning); the disadvantages are weaker data privacy protection, higher network transmission overhead, and the potential weakening of the personalized capabilities of edge agents due to the averaging effect of the global model.

[0243] The above-mentioned alternatives may be considered as suboptimal options under certain constraints, but they are inferior to the main solution in terms of protection scope and technical effectiveness.

[0244] To verify the superiority of the processing implementation method provided by this invention, experimental verification was conducted using the processing implementation method provided by this invention, as follows:

[0245] A comparative test was conducted in a simulated 110-kilometer highway scenario. The experimental group used the method of this invention, while the control group used the traditional independent treatment mode.

[0246] Experimental scenarios covered: concurrent multi-event scenarios during evening rush hour (18:00-20:00); multi-event scenarios under severe weather conditions (rain and fog); and high-traffic scenarios during holidays (traffic flow exceeding 5000 vehicles / hour).

[0247] Experimental data: 18 concurrent scenarios were processed, with a total of 72 concurrent events; the average congestion loss in the experimental group was 990 vehicle-hours, while that in the control group was 1350 vehicle-hours, a reduction of 26.7%; the secondary accident rate in the experimental group was 4.4%, while that in the control group was 13.5%, a reduction of 67.4%; the average response time in the experimental group was 3.2 minutes, while that in the control group was 8.5 minutes, a reduction of 62%.

[0248] In summary, the method for handling multi-point coupled events on highways provided by this invention constructs a task spatiotemporal coupled network to explicitly characterize the complex relationships between events, designs a three-layer rapid decision-making architecture integrating large-model reasoning, case retrieval, and mathematical optimization to achieve minute-level collaborative decision-making, and establishes an experience-sharing and collaborative learning mechanism to achieve collective intelligent evolution. Specifically, it mainly includes several improvements:

[0249] 1. Task-Spatiotemporal Coupled Network Modeling Method:

[0250] The core innovation of this invention lies in the first systematic modeling of task coupling relationships in traffic incident handling, constructing a task network model that includes four types of coupling: spatial coupling, temporal coupling, resource coupling, and causal coupling.

[0251] Spatial coupling: Quantify the geospatial interactions between tasks through geographical influence factors and traffic flow propagation coefficients; Temporal coupling: Distinguish between deterministic temporal dependencies and induced probabilistic relationships, and construct a task temporal dependency graph; Resource coupling: Consider resource location and arrival time constraints through resource demand matrices and availability vectors; Causal coupling: Integrate three types of evidence—spatial-temporal correlation, semantic correlation, and domain rules—to identify root cause events and derivative events.

[0252] This technology breaks through the traditional assumption of task independence, provides a theoretical basis for global collaborative optimization, and is the core mechanism for achieving a 27% reduction in system congestion losses.

[0253] 2. A three-tiered rapid decision-making architecture integrating large-scale model reasoning, case retrieval, and mathematical optimization:

[0254] This invention designs a three-layer fast decision-making architecture, which solves the problems of long training cycles and inability to meet the real-time requirements of online scenarios in traditional multi-agent reinforcement learning.

[0255] Layer 1 Large Model Inference Layer: Leveraging the semantic understanding and reasoning capabilities of the large traffic vertical model, an innovative tool enhancement mechanism is designed to enable the large model to autonomously invoke coupling analysis tools to assist in reasoning and generate coarse-grained decision schemes.

[0256] Layer 2 Case Retrieval Layer: Employs a hybrid retrieval strategy (structured features + semantic embedding) to retrieve similar scenarios from the historical case library, extract reusable task allocation patterns and resource scheduling strategies, and provide an initial warm start solution for Layer 3.

[0257] Layer 3 Mathematical Optimization Layer: Based on the output of the first two layers, a multi-objective optimization model is established. A warm start mechanism is used to shorten the solution time and solve a refined resource allocation scheme under constraints.

[0258] This technology achieves a balance between decision quality and real-time performance, with a total response time of approximately 30 seconds, which is more than 10,000 times faster than the traditional MARL response speed.

[0259] 3. Large model tool enhancement mechanism:

[0260] In the Layer 1 large model inference layer, this invention designs a tool enhancement mechanism that enables the large model to autonomously call external tools to assist in task analysis and decision generation.

[0261] Toolset design: including spatial analysis tools, temporal reasoning tools, risk assessment tools, etc., each tool encapsulates specific domain knowledge and algorithms.

[0262] Tool invocation strategy: The large model autonomously decides which tools to invoke, when to invoke them, and how to combine tool outputs based on task requirements.

[0263] Prompt design: Transform the task-coupled network into a structured Prompt, guide the large model to focus on key information, and provide tool interface descriptions.

[0264] This technology enhances the reasoning ability of large models in traffic incident handling scenarios, enabling them to handle complex task coupling relationships and generate high-quality coarse-grained decision solutions.

[0265] 4. Experience sharing pool and collaborative reflection mechanism:

[0266] This invention constructs an experience-sharing pool and a collaborative reflection mechanism to achieve collaborative learning among heterogeneous intelligent agents.

[0267] Experience sharing pool: Maintained by a centrally scheduled intelligent agent, it stores collaborative handling cases, best practice rules, lessons learned from failures, and collaborative mode templates to enable knowledge flow across intelligent agents.

[0268] Collaborative reflection mechanism: a five-step process (data collection → effect evaluation → problem identification → experience extraction → knowledge consolidation) that examines the collaborative process from a group perspective and extracts scalable collaborative patterns or anti-patterns.

[0269] This technology improves the learning efficiency of intelligent agents by 150%, reduces the number of local cases required to achieve stable performance from 700 to 280, and shortens the deployment cycle of new intelligent agents from 3 months to 1 month.

[0270] 5. Federated Training Framework:

[0271] This invention designs a federated training framework that allows each edge agent to train independently on local data, and only uploads model updates to the central agent for federated aggregation, thus achieving a balance between privacy protection and knowledge sharing.

[0272] Local training: Each agent trains its own capabilities on local data, protecting data privacy; Federated aggregation: A strategy combining gradient aggregation and global knowledge extraction is used to generate a global model; Global model distribution: Each agent downloads the global model to update its local model, learning from collective wisdom.

[0273] This technology achieves a balance between data privacy protection and collective intelligence evolution, and is suitable for scenarios involving multi-party collaboration where data sharing is inconvenient.

[0274] 6. Multi-objective optimization model based on task coupling constraints:

[0275] In the Layer 3 mathematical optimization layer, this invention establishes a multi-objective optimization model that considers task coupling constraints.

[0276] Objective function: Minimize the total congestion loss of the system (including congestion loss and the risk of secondary accidents); Constraints: Resource quantity constraints, task timing constraints, coupling effect constraints (such as avoiding simultaneous full closure of adjacent events), and resource arrival time constraints; Warm start mechanism: Use decision variables from similar cases as initial solutions to shorten the solver search time.

[0277] This technique explicitly incorporates task coupling relationships into optimization constraints, ensuring that the generated solution is physically feasible and has optimal system performance.

[0278] Based on the above improvements, the processing method provided by the present invention has at least the following beneficial effects:

[0279] 1. System congestion losses reduced by 27%:

[0280] This effect is a result of the combined effect of task spatiotemporal coupled network modeling and global collaborative optimization mechanism.

[0281] First, spatial coupling modeling enables the system to identify situations where a fully enclosed response to an upstream event would block downstream rescue channels. Based on spatial coupling strength calculations, when... At that time, the system determines a traffic incident. Traffic incidents There is a strong spatial influence, which in turn adjusts traffic events. The approach to handling upstream incidents was changed from full closure to using the emergency lane while keeping the main lanes open, thus avoiding complete blockage of rescue routes. Experimental data showed that after replacing full closure with using the emergency lane, the arrival time of downstream recovery vehicles was reduced from 25 minutes to 7 minutes, the handling delay was reduced by 18 minutes, and the congestion loss was reduced by approximately 15%.

[0282] Second, resource coupling modeling enables the system to optimize resource allocation order and dynamic task switching strategies. By establishing a resource demand matrix R and a resource availability vector C, the system identifies resource supply and demand imbalances. When the resource supply-demand ratio is less than 1, the system adopts a dynamic task switching strategy: after completing a high-priority event, the clearing vehicle immediately switches to a lower-priority event instead of returning to base to wait. This strategy reduces resource idle time, increases resource utilization by 25%, and further reduces overall system congestion losses by approximately 8%.

[0283] Third, the Layer 3 mathematical optimization layer uses minimizing the total system congestion loss as the objective function to solve for the globally optimal solution under resource constraints and task coupling constraints. Compared to the greedy strategy (processing events sequentially according to priority), global optimization considers the coupling effects between events, avoiding the global performance degradation caused by local optimal decisions. Experimental data shows that global optimization further reduces congestion loss by approximately 4% compared to the greedy strategy.

[0284] Based on the above three improvements, the system congestion loss decreased from 1350 vehicle-hours to 990 vehicle-hours, a reduction of 26.7%. Calculated at an economic loss of 30 yuan per vehicle-hour, this translates to a saving of approximately 10,800 yuan in congestion costs per concurrent scenario. During the three-month experiment, the system handled 18 concurrent scenarios, accumulating savings of approximately 194,000 yuan in congestion costs, demonstrating significant economic benefits.

[0285] 2. Secondary accident rate reduced by 67%:

[0286] This effect arises from the combined effect of causal coupling modeling and multi-agent collaborative prevention mechanisms.

[0287] First, causal coupling modeling identifies potential causal chains by integrating three types of evidence: spatiotemporal correlation, semantic correlation, and domain rules. At that time, the system determines a traffic incident. May trigger traffic incidents This will lead to traffic incidents Marking an event as high-risk and prioritizing its handling. For example, the system identifies the causal chain of "excessive handling time for a rear-end collision → traffic congestion behind → sudden braking of vehicles → increased risk of a rear-end collision" and calculates the probability of a secondary rear-end collision. This was determined to be a high-risk scenario.

[0288] Second, multi-agent collaboration enables rapid execution of preventative measures. The traffic guidance agent immediately publishes speed reduction guidance information (from 80 km / h to 50 km / h) on multiple upstream information boards; the congestion prediction agent monitors traffic flow evolution in real time; the emergency response agent quickly clears obstacles, shortening accident exposure time; and the drone agent conducts advance reconnaissance, saving information acquisition time. This collaborative prevention and control network allows following vehicles to slow down in advance, shortens response time, and reduces the risk of secondary accidents. Down to .

[0289] Third, the collaborative learning mechanism extracts lessons from historical secondary accident cases and solidifies them into anti-pattern warnings. The system summarizes the anti-pattern of "rainy night & rear-end collision & failure to provide timely guidance → secondary accident," automatically triggering alarms in similar scenarios and mandating early intervention by traffic guidance systems to avoid repeating mistakes.

[0290] By combining the above three mechanisms, the secondary accident rate decreased from 13.5% to 4.4%, a reduction of 67.4%. Specifically, the secondary accident rate for single-incident handling decreased from 10.0% to 4.0%, and the secondary accident rate for multiple concurrent incidents decreased from 23.1% to 5.2%. Secondary accidents cause casualties and property damage, and reducing the secondary accident rate has significant safety and social value.

[0291] 3. Collaborative decision-making response time is reduced to less than 1 minute:

[0292] This effect stems from the parallel processing and optimization acceleration mechanism of the three-layer fast decision-making architecture.

[0293] First, the Layer 1 large-scale model inference layer utilizes a pre-trained traffic vertical domain large-scale model, enabling online inference without offline training. Large-scale model loading takes approximately 2 seconds, and inference computation takes approximately 8 seconds, totaling approximately 10 seconds. Compared to traditional multi-agent reinforcement learning, which requires several days to weeks of offline training, the large-scale model inference layer represents a paradigm shift from offline training to online inference.

[0294] Second, the Layer 2 case retrieval layer employs a hybrid retrieval strategy, fully leveraging the efficient retrieval capabilities of the vector database. Structured feature calculation takes approximately 2 seconds, and semantic embedding retrieval takes approximately 3 seconds, totaling approximately 5 seconds. The retrieval results provide a warm start initial solution for Layer 3, significantly shortening the optimization solution time.

[0295] Third, the Layer 3 mathematical optimization layer employs a warm start mechanism, using decision variables from similar cases as initial solutions input into the Gurobi solver. Experimental data shows that after adopting warm start, the solver's search time for the optimal solution is reduced from 30 seconds to 12 seconds, an acceleration of approximately 60%. This is because the initial solution in a warm start is usually close to the optimal solution, and the solver does not need to perform a global search from scratch, but only needs to perform local optimization near the initial solution.

[0296] The total time for the three-layer architecture is approximately 10 seconds for Layer 1, 5 seconds for Layer 2, 12 seconds for Layer 3, and 3 seconds for task decomposition and agent matching, totaling approximately 30 seconds, meeting the 1-minute real-time requirement. Compared to the traditional 8 to 12 minutes of manual scheduling, the response speed is improved by approximately 82%; compared to the traditional multi-agent reinforcement learning which requires several days to weeks of offline training, the response speed is improved by more than 10,000 times.

[0297] Rapid response enables the system to initiate coordinated handling immediately after an incident occurs, shortening the exposure time of the accident, reducing the risk of congestion spreading, and significantly improving the effectiveness of emergency management.

[0298] 4. The learning efficiency of the intelligent agent is improved by 150%:

[0299] This effect stems from a collaborative learning mechanism that integrates experience sharing pools, collaborative reflection, and federated training.

[0300] First, the experience-sharing pool stores globally collaborative handling cases, best practice rules, and lessons learned from failures, enabling agents to access the experiences of others. When a new agent is deployed, it does not need to learn from scratch; it can directly retrieve similar cases from the shared pool for learning. Experimental data shows that a new agent can achieve a decision accuracy of 85% after 4 weeks of deployment, while it would take 12 weeks to reach the same level using an independent learning model, representing a 3-fold increase in learning speed.

[0301] Second, the collaborative reflection mechanism conducts a collective debriefing after each collaborative action, identifying areas for optimization in collaborative cooperation from a group perspective, and extracting generalizable collaborative patterns or anti-patterns. Once the collaborative patterns are solidified into best practice rules, each agent learns and applies these rules, avoiding redundant exploration. For example, after the "dynamic task switching" pattern is extracted, each agent does not need to explore this strategy independently; simply applying the pattern can improve resource utilization by 25%.

[0302] Third, the federated training framework enables each agent to learn collective wisdom from the global model. The global model aggregates the model parameters and knowledge rules of all agents, containing the experience accumulated by different agents in different scenarios. Each agent downloads the global model to update its local model, which is equivalent to indirectly learning from the experience of others, reducing the number of samples required for local training. According to federated learning theory, with N agents participating in federated training, the number of local samples required to achieve the same performance is approximately [a smaller number] compared to independent learning. When N=5, the sample requirement drops to approximately 45%, which is basically consistent with the experimental observation of 40%.

[0303] By combining the above three mechanisms, the learning efficiency of the intelligent agent is improved by 150%, the number of local cases required to reach stable performance is reduced from 700 to 280, and the learning cycle is shortened from 3 months to 1 month. The improved learning efficiency accelerates the evolution of the overall intelligence level of the system, enabling the system to adapt to new scenarios and new requirements more quickly.

[0304] Based on the same inventive concept, this embodiment of the invention also provides a functional module diagram of a device for handling multi-point coupled events on highways. This embodiment can divide the device into functional modules according to the above method embodiment. For example, each function can be divided into its own functional modules, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0305] The device for handling multi-point coupled events on highways can include a construction module, a generation module, a matching module, and a handling implementation module. The functions of each module of the device for handling multi-point coupled events on highways will be described in detail below.

[0306] The system comprises four modules: a construction module and a generation module. The construction module models the coupling relationships of traffic events in the current highway scenario. Based on this modeling, a directed weighted graph is constructed. This graph includes a set of nodes, a set of edges, and an edge weight function. Each node represents a traffic event, each edge represents a coupling relationship between two traffic events, and the edge weight function characterizes the coupling strength between the two events. The generation module generates structured prompt text based on the directed weighted graph, resource state vectors, and environmental parameter vectors. Based on this structured prompt text, an initial decision plan is generated using a large model. The matching module performs similarity matching between the current scenario and historical cases in the case library to identify matching historical cases. The implementation module, based on the initial decision plan and matching historical cases, optimizes the solution using a multi-objective optimization model under set objective loss functions and constraints to obtain a refined solution.

[0307] The device for handling multi-point coupled events on highways provided in this embodiment can be used to execute the method for handling multi-point coupled events on highways under any of the above embodiments. For details not covered in this embodiment, please refer to the corresponding descriptions in the above embodiments. This embodiment will not elaborate further here.

[0308] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0309] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0310] Furthermore, in the various embodiments of the present invention, the functional modules can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0311] It should be noted that if the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0312] The above are merely embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for handling multi-point coupled events on highways, characterized in that, The method includes: Based on the current traffic events on the highway, a coupling relationship model is performed. Based on the coupling relationship modeling results, a directed weighted graph is constructed. The directed weighted graph includes a set of nodes, a set of edges, and an edge weight function. Each node in the set of nodes represents a traffic event. Each edge in the set of edges represents a coupling relationship between two traffic events. The edge weight function characterizes the coupling strength between the two traffic events. Structured prompt text is generated based on the directed weighted graph, resource state vector, and environmental parameter vector. An initial decision scheme is then generated based on the structured prompt text and a large model. The current scenario is matched with historical cases in the case library to determine the matching historical cases; Based on the initial decision scheme and matching historical cases, a refined disposal scheme is obtained by optimizing the solution through a multi-objective optimization model under the set objective loss function and constraints. The steps for modeling coupling relationships based on traffic events in the current scenario on highways include: For each traffic event in the current scenario on the highway, an attribute vector for each traffic event is constructed. For any two traffic events, spatial coupling modeling, temporal coupling modeling, resource coupling modeling, and causal coupling modeling are performed based on the attribute vectors of the two traffic events, respectively. The spatial coupling modeling result is used to determine the impact of the handling method of one traffic event on the other traffic event. The temporal coupling modeling result is used to determine whether the two traffic events are executed serially or in parallel. The resource coupling modeling result is used to determine the resource allocation order and dynamic task switching strategy. The causal coupling modeling result is used to determine the processing order of traffic events and the induced chain reactions.

2. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The method is applied to a collaborative processing system comprising a central agent and multiple edge agents, and the method further includes: After completing the handling plan, each of the aforementioned edge agents uploads the relevant case information to the shared pool; The central agent evaluates the value of each case in the shared pool and determines whether to store or filter the cases based on the evaluation results. Each edge agent learns the case information in the shared pool to optimize its own model. The central intelligent agent collects complete data on the collaborative handling of multiple traffic accidents, performs performance deviation evaluation based on the complete data, and extracts collaborative pattern rules and anti-pattern rules based on the performance deviation evaluation results, so that each edge intelligent agent can update its own knowledge base. The central agent collects the updated model parameters uploaded by each edge agent after training the decision model based on local data, aggregates the updated model parameters to obtain a global model, and distributes the global model to each edge agent so that each edge agent can update its local decision model based on the global model.

3. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The steps of generating structured prompt text based on the directed weighted graph, resource state vector, and environmental parameter vector, and generating an initial decision scheme based on the structured prompt text and a large model, include: The design includes a toolset comprising multiple tools, such as spatial analysis tools, temporal reasoning tools, and risk assessment tools. Structured prompt text is generated based on the directed weighted graph, resource state vector, and environmental parameter vector; Based on the structured prompt text, and by calling the tools in the toolset through a large model for processing, the analysis results of each called tool are obtained; Semantic reasoning is performed on the analysis results of each tool to generate an initial decision scheme.

4. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The step of performing similarity matching between the current scene and historical cases in the case library to determine matching historical cases includes: Calculate the structured feature similarity and semantic embedding similarity between the current scene and each historical case in the case library; The structured feature similarity and semantic embedding similarity are weighted and summed to obtain the comprehensive similarity. Select the historical cases with the highest overall similarity as the matching historical cases; The structured feature similarity is calculated based on the similarity of the number of events, the distribution of event types, the geographical distribution, and the resource configuration between the current scene and historical cases. The semantic embedding similarity is obtained by converting the scene description text of the current scene and historical cases into embedding vectors and then calculating the cosine similarity between the embedding vectors.

5. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The step of obtaining a refined treatment plan based on the initial decision scheme and matching historical cases, and by optimizing the solution using a multi-objective optimization model under the set objective loss function and constraints, includes: Construct constraints including resource quantity constraints, task timing constraints, coupling effect constraints, and resource arrival time constraints, and construct an objective loss function that includes congestion loss terms and risk loss terms that induce secondary accidents; Using the matching historical case handling plan as the initial solution, a multi-objective optimization model is used to optimize the solution based on the initial decision plan and the initial solution under the constraints and guided by the objective loss function to obtain a refined handling plan. The refined handling plan includes resource allocation decision, handling sequence arrangement, handling method selection and expected effect evaluation.

6. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The steps for spatial coupling modeling based on the attribute vectors of the two traffic events include: The geographical distance between the two traffic events was calculated based on their spatial locations. Construct a traffic flow propagation function based on the spatial location and handling method of the two traffic events; The spatial coupling strength is calculated based on the geographical distance, traffic flow propagation function, and the set influence attenuation characteristic distance. Among them, spatial coupling describes the geographic spatial interactions between traffic events, for any two traffic event nodes. and The spatial coupling strength is defined as: In the formula, Indicates a traffic incident Traffic incidents Geographical distance; Indicates the distance affecting the attenuation characteristic; Indicates the spatial coupling strength coefficient; This represents the traffic flow propagation function, with values ​​ranging from [0, 1]. Traffic flow propagation function Based on the spatial location of each traffic incident, its relative position is determined, and combined with the handling method, the following is determined: The traffic flow propagation function characterizes the propagation effect of upstream events on downstream traffic flow.

7. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The steps for resource coupling modeling based on the attribute vectors of the two traffic events include: Obtain the total number of resource types in the current scenario, the total available amount of each type of resource, and the demand for each type of resource for the two traffic events respectively; Resource allocation constraints are constructed based on the demand for each type of resource in each traffic event and the total available resources of each type. Based on the spatial location of each traffic event and the location distribution of each type of resource, an arrival time function from each type of resource to each traffic event is defined.

8. The method for handling multi-point coupled events on highways according to claim 1, characterized in that, The steps for modeling causal coupling based on the attribute vectors of the two traffic events include: Calculate the spatiotemporal correlation evidence score, semantic correlation evidence score, and domain rule evidence score between the two traffic events; The causal association probability is obtained by weighting and summing the spatiotemporal correlation evidence score, semantic correlation evidence score, and domain rule evidence score.

9. A device for handling multi-point coupled events on highways, characterized in that, The apparatus for implementing the method for handling multi-point coupled events on highways as described in any one of claims 1-8 includes: The module is used to model the coupling relationship of traffic events in the current scenario on the highway, and to construct a directed weighted graph based on the coupling relationship modeling results. The directed weighted graph includes a set of nodes, a set of edges, and an edge weight function. Each node in the set of nodes represents a traffic event, each edge in the set of edges represents a coupling relationship between two traffic events, and the edge weight function characterizes the coupling strength between the two traffic events. The generation module is used to generate structured prompt text based on the directed weighted graph, resource state vector, and environmental parameter vector, and to generate an initial decision scheme based on the structured prompt text and a large model. The matching module is used to perform similarity matching between the current scene and historical cases in the case library to determine the matching historical cases; The disposal implementation module is used to obtain a refined disposal plan based on the initial decision scheme and matched historical cases, and through a multi-objective optimization model under the set objective loss function and constraints.

Images