Expressway intelligent decision method and system fusing knowledge base and large model reasoning

Abstract

The application provides an expressway intelligent decision-making method and system fusing a knowledge base and large model reasoning, and relates to the technical field of intelligent transportation. The application solves the problem of long artificial time consumption by fusing multi-source sensing data and extracting a unified feature vector, and realizes rapid and accurate sensing of event information. Through a three-level hierarchical retrieval strategy composed of knowledge graph structure retrieval, vector semantic retrieval and graph reordering, the most relevant decision context information of the current event is quickly and accurately located and screened from massive dispersed knowledge resources, overcoming the low efficiency of manual review and information omission defects. Then, a large language model is used to reason and generate multiple candidate schemes in combination with the feature vector and the context information, shortening the scheme generation time. Finally, through multi-dimensional quantitative evaluation and global conflict resolution, resource and measure conflicts can be automatically coordinated in a multi-event concurrent scenario. The application improves the disposal efficiency of expressway events.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to an intelligent decision-making method and system for highways that integrates knowledge base and large model reasoning. Background Technology

[0002] Highways are characterized by high traffic volume, high speed, and closed operation. If accidents, breakdowns, or debris spills occur, they can easily cause congestion and secondary accidents if not handled quickly.

[0003] Various sensing devices, including video surveillance, radar detection, traffic flow detection, and meteorological monitoring, are deployed along the highway. However, these devices belong to different systems with varying data formats and transmission protocols. After an incident occurs, operations personnel need to simultaneously check multiple terminals and manually compare data from different sources to confirm the authenticity and specific location of the incident, a time-consuming process. Especially at night or in severe weather conditions, the reliability of a single sensing method decreases, and the repeated verification process further prolongs the incident confirmation time. Furthermore, traffic flow and meteorological data remain at a simple statistical level, lacking quantitative analysis methods to assess the impact of the incident and the constraints on response, meaning the sensing results cannot be directly used for subsequent decision-making.

[0004] Furthermore, incident handling involves a large number of emergency plans, historical cases, control protocols, and organizational contact information. However, these knowledge resources are scattered across different documents and systems, lacking a unified organization and retrieval mechanism. Decision-makers need to review relevant documents one by one, or rely on personal memory to find applicable clauses and similar cases, depending on the type of incident and the situation on site. Information retrieval is time-consuming and prone to omissions. The formulation of response plans mainly relies on human experience for deduction, requiring repeated weighing of the combined effects of different response measures and communication between departments to confirm resource allocation and control measures, resulting in a highly sequential process. When multiple incidents occur simultaneously or their impact overlaps, the difficulty of coordinating plans increases exponentially, and the decision-making time is significantly prolonged.

[0005] Therefore, current decision-making processes for handling highway incidents suffer from long response times and low efficiency. Summary of the Invention

[0006] This invention provides a smart decision-making method and system for highways that integrates knowledge base and large model reasoning, which can solve the problems of long response time and low efficiency in the current highway incident handling decision-making, and improve the efficiency of highway incident handling.

[0007] In a first aspect, the present invention provides an intelligent decision-making method for highways that integrates knowledge base and large-scale model reasoning. The method includes: acquiring multi-source sensing data at the time of a highway event; extracting features from the multi-source sensing data to obtain feature vectors for each event; performing a three-level hierarchical retrieval process—knowledge graph structure retrieval, vector semantic retrieval, and graph reordering—based on the feature vectors of each event and a pre-defined knowledge base to obtain decision context information for each event; performing analysis and reasoning using a large-scale language model based on the feature vectors and decision context information of each event to generate multiple candidate handling schemes for each event; and performing multi-dimensional quantitative evaluation and global conflict resolution based on the multiple candidate handling schemes for each event to generate a collaborative handling scheme for the highway event.

[0008] In a second aspect, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0010] This invention provides a smart decision-making method and system for highways that integrates knowledge base and large-scale model reasoning. By fusing multi-source sensing data and extracting unified feature vectors, this invention solves the problem of long confirmation times caused by manual comparison of data from multiple terminals, achieving rapid and accurate perception of event information. Through a three-level hierarchical retrieval strategy consisting of knowledge graph structure retrieval, vector semantic retrieval, and graph reordering, it quickly and accurately locates and filters the most relevant decision context information from massive, dispersed knowledge resources, overcoming the low retrieval efficiency and information omissions caused by traditional manual document retrieval. Based on this, a large language model is used to combine feature vectors and contextual information for reasoning and generate multiple candidate solutions, automating and parallelizing the solution formulation process that originally relied on manual experience deduction and serial communication, shortening the solution generation time. Finally, through multi-dimensional quantitative evaluation and global conflict resolution, it can automatically coordinate resource and measure conflicts in multi-event concurrent scenarios, avoiding the problem of exponentially increasing difficulty of manual coordination. This invention solves the problems of long response times and low efficiency in existing highway event handling decisions, improving the efficiency of highway event handling. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a highway intelligent decision-making method that integrates knowledge base and large model reasoning, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] like Figure 1 As shown, the present invention provides a highway intelligent decision-making method that integrates knowledge base and large model reasoning, the method including steps S101-S105.

[0015] S101. Acquire multi-source sensing data when a highway incident occurs.

[0016] In some embodiments, multi-source sensing data includes video surveillance data, radar detection data, traffic flow data, event alarm data, and meteorological data.

[0017] Among them, video surveillance data comes from fixed cameras and PTZ cameras deployed along the route and is transmitted in the form of video streams; radar detection data comes from millimeter-wave radar or lidar and is output in the form of target point clouds or trajectory data; traffic flow data comes from loop detectors or microwave detectors deployed on the road surface and is reported in the form of statistical indicators such as cross-sectional flow, speed, and occupancy; event alarm data comes from roadside emergency telephones, vehicle-mounted automatic alarm systems, or manual reporting channels; and meteorological data comes from meteorological monitoring stations deployed along the route.

[0018] For example, embodiments of the present invention can connect in real time with video surveillance systems, radar detection systems, traffic flow detection systems, event alarm systems and meteorological monitoring systems deployed along highways through a data acquisition interface. The raw data is received in accordance with the transmission protocols specified by each system (such as RTSP video stream protocol, MQTT message queue protocol, HTTP API interface protocol, etc.), and aligned and cached according to a unified timestamp to ensure that the data processed subsequently has a consistent time base.

[0019] S102. Based on multi-source sensing data, perform feature extraction to obtain the feature vector of each event.

[0020] In some embodiments, the feature vector is a unified numerical representation formed by extracting, encoding and fusing multi-source heterogeneous sensing data, used to quantitatively characterize the attribute features of an event.

[0021] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1027.

[0022] S1021. Perform target detection on the video surveillance data, identify vehicles, pedestrians, debris, and maintenance construction areas in the video surveillance data, and generate the first event detection result.

[0023] In some embodiments, the first event detection result includes target type, target location, target quantity, detection time, and first confidence level.

[0024] In some embodiments, video surveillance data refers to real-time video streams captured by cameras along highways. Embodiments of the present invention can employ deep learning-based target detection algorithms (such as YOLO series, Faster R-CNN, etc.) to perform frame-by-frame or frame-by-frame reasoning on the video frames, identifying vehicles, pedestrians, debris (such as scattered goods), and maintenance / construction areas (such as cone-enclosed areas, construction vehicles, etc.) in the image. The detection output is the first event detection result, including the target type (such as "car," "pedestrian," "debris," etc.), the target's position coordinates in the image, the number of targets, the detection timestamp, and a first confidence level (range 0~1) characterizing the detection reliability.

[0025] S1022. Perform target tracking on radar detection data, extract the target vehicle's position, speed, acceleration, and driving trajectory, determine whether there is abnormal driving behavior, and generate a second event detection result.

[0026] In some embodiments, the second event detection result includes an abnormal target vehicle identifier, location, speed, acceleration, abnormal behavior type, and second confidence level.

[0027] For example, radar detection data refers to target point cloud or detection list data output by millimeter-wave radar or lidar. Embodiments of this invention can employ multi-target tracking algorithms (such as Kalman filtering, joint probabilistic data association, etc.) to associate and track targets in consecutive frames, extracting the position (latitude and longitude or relative coordinates), speed (magnitude and direction), acceleration, and historical driving trajectory of each target vehicle. Based on this, a preset abnormal behavior discrimination rule (such as whether the vehicle speed is below the minimum speed limit of the road segment, whether sudden deceleration occurs, whether serpentine driving occurs, etc.) is used to determine whether abnormal driving behavior exists. The detection output is a second event detection result, including the abnormal target vehicle identifier, position, speed, acceleration, abnormal behavior type (such as "sudden braking," "lane departure," "low-speed driving," etc.), and a second confidence level.

[0028] S1023. Analyze the event alarm data, extract the alarm event type, alarm location and alarm time, and generate the third event detection result.

[0029] In some embodiments, the third event detection result includes alarm event type, alarm location, alarm time, alarm source, and third confidence level.

[0030] For example, event alarm data refers to structured or semi-structured alarm information from channels such as roadside emergency phones, vehicle-mounted automatic alarm systems, mobile APP reports, or manual telephone alarms. This embodiment of the invention can extract fields according to a predefined alarm data format to obtain the alarm event type (e.g., "traffic accident," "vehicle malfunction," "fire," etc.), alarm location (GPS coordinates or station number), and alarm time. The detection output is a third event detection result, including the alarm event type, alarm location, alarm time, alarm source (e.g., "roadside phone," "automatic alarm," "manual report," etc.), and a third confidence level.

[0031] S1024. Cross-validate the detection results of the first event, the second event, and the third event to generate a complete perception record.

[0032] In some embodiments, a complete perception record includes event type, event location, affected lane, affected area, detection source, and overall confidence level.

[0033] For example, embodiments of the present invention can comprehensively compare and verify the detection results of the first event, the second event, and the third event. Specifically, this includes: location consistency verification (determining whether the locations detected by different sources are within a preset error range), type consistency verification (determining whether the event types detected by different sources match or are compatible), and time consistency verification (determining whether the detection times of different sources are within a reasonable time window). Based on the confidence level and consistency of each source, a comprehensive confidence level is calculated using methods such as weighted voting or Bayesian fusion, and then merged to generate a complete perception record, including event type, event location, affected lane, affected area, list of detection sources, and comprehensive confidence level.

[0034] S1025. Analyze the traffic flow data, extract the lane-level traffic flow, time occupancy, space occupancy and average speed upstream and downstream of the event location, and calculate the traffic flow mutation coefficient.

[0035] In some embodiments, the traffic flow abrupt change coefficient is used to characterize the impact of an event on traffic flow.

[0036] For example, traffic flow data refers to statistical indicators such as traffic flow, time occupancy, space occupancy, and average speed reported by lane-level detectors upstream and downstream of the event location. Embodiments of this invention can extract the values ​​of these indicators for each lane upstream and downstream of the event location, calculate the changes before and after the event, and then calculate the traffic flow abrupt change coefficient. For instance, this invention can compare the average flow / speed over a period after the event with the baseline value for the same period before the event, calculate the relative rate of change, and weight it by considering the changes in occupancy to obtain a dimensionless coefficient value used to characterize the impact of the event on traffic flow.

[0037] S1026. Analyze meteorological data, extract weather phenomena, visibility values, road surface temperature, road surface humidity, and wind force at the location of the event, classify and quantify each meteorological parameter, and generate meteorological impact factors.

[0038] In some embodiments, the meteorological impact factor is used to characterize the degree to which meteorological conditions constrain event handling.

[0039] For example, meteorological data refers to parameters such as weather phenomena, visibility values, road surface temperature, road surface humidity, and wind force reported by meteorological monitoring stations at the location of the event. This embodiment of the invention can extract the numerical values ​​of the above meteorological parameters and then encode each parameter according to a preset graded quantification coding table. For example, visibility is divided into four levels: "Excellent (>200m)," "Good (100-200m)," "Medium (50-100m)," and "Poor (<50m)," and each is assigned a coded value; road surface humidity is coded according to states such as dry, damp, wet, and waterlogged. The coded values ​​of each parameter are weighted and fused according to preset weights to generate a meteorological influence factor, which is used to characterize the degree of constraint of meteorological conditions on event handling.

[0040] S1027. The complete perception record, traffic flow mutation coefficient, meteorological influence factors and GPS station location information are fused to generate a feature vector for each event.

[0041] For example, GPS station location information refers to the latitude and longitude coordinates of the location where the event occurred and its corresponding highway mileage station. In this embodiment of the invention, each field in the complete perception record (event type, location, affected lane, affected area, comprehensive confidence level, etc.) can be converted into numerical codes, and then concatenated with the calculated traffic flow mutation coefficient, meteorological influence factor, GPS latitude and longitude coordinates, and station values ​​to form a fixed-dimensional numerical vector.

[0042] S103. Based on the feature vector of each event and the preset knowledge base, perform a three-level hierarchical retrieval of knowledge graph structure, vector semantic retrieval, and graph reordering to obtain the decision context information of each event.

[0043] In some embodiments, the preset knowledge base is a pre-constructed knowledge set used to support highway event decision-making. It consists of two parts: a unified vector knowledge base and a transportation domain knowledge graph, which are linked through a mapping relationship. The unified vector knowledge base is obtained by vectorizing multi-source knowledge resources such as emergency plan databases, traffic control standard databases, organizational and contact information databases, historical event handling case databases, policy and regulatory documents, and technical standards and specifications. The transportation domain knowledge graph is a graph-structured knowledge network constructed by extracting entities such as event types, handling measures, equipment resources, organizational structures, and personnel roles from the above knowledge resources, as well as the relationships between entities such as causal relationships, temporal relationships, and hierarchical relationships.

[0044] In some embodiments, the three-level hierarchical retrieval is a progressive retrieval strategy consisting of three levels: knowledge graph structure retrieval, vector semantic retrieval, and graph reordering. The first level is knowledge graph structure retrieval: starting with the event type and location in the event's feature vector, a breadth-first traversal is performed in the knowledge graph to obtain knowledge nodes and reasoning paths that have a direct causal, temporal, or hierarchical relationship with the current event. The second level is vector semantic retrieval: using the event's feature vector as the query vector, semantic similarity is calculated in a unified vector knowledge base, and the data is sorted from high to low according to similarity scores to obtain semantically relevant knowledge fragments. The third level is graph reordering: utilizing the structured information of the knowledge graph, the retrieved knowledge fragments are reordered, moving forward knowledge fragments that are structurally more relevant and have more logical paths.

[0045] In some embodiments, decision context information refers to the set of knowledge most relevant to the current event, selected from a preset knowledge base after a three-level hierarchical retrieval process, and constructed by weighting and fusing knowledge according to knowledge type, similarity score, and source reliability. This information includes five categories of knowledge: emergency plans, historical cases, regulatory clauses, resource information, and contact information.

[0046] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1034.

[0047] S1031. Based on the event type and event location in the feature vector of each event, perform a breadth-first traversal in the preset knowledge base, starting from each event node, to obtain knowledge nodes and reasoning paths that have a direct causal relationship, temporal relationship, or hierarchical relationship with each event.

[0048] For example, embodiments of the present invention can first parse the event type code (e.g., code 001 corresponding to "traffic accident") and event location information (GPS coordinates or mileage markers) from the feature vector of each event, and convert this information into a query statement for a graph database. Using the event node as the starting node, a breadth-first traversal is performed along preset "causal relationship" edges (e.g., "accident" leads to "congestion"), "temporal relationship" edges (e.g., "occurred first," "handled later"), and "hierarchical relationship" edges (e.g., "event type" belongs to a "major category" or "specific subcategory"). A reasonable upper limit for the number of layers is set during the traversal to prevent excessive expansion of the search scope. The search output consists of knowledge nodes connected to the starting event node (e.g., related contingency plan nodes, historical case nodes, equipment resource nodes) and a sequence of relationships along the connecting paths. These reasoning paths provide a structured logical chain for subsequent solution generation.

[0049] S1032. Using the feature vector of each event as the query vector, perform semantic similarity retrieval in the preset knowledge base, calculate the similarity score between the feature vector and each knowledge vector in the preset knowledge base, sort them from high to low according to the similarity score, and obtain the knowledge fragments that are semantically related to each event.

[0050] For example, in this embodiment of the invention, the feature vector of each event can be directly used as a query vector and input into the vector retrieval engine. This engine pre-stores vector representations corresponding to all knowledge fragments in a unified vector knowledge base. The engine calculates the similarity score between the query vector and each knowledge vector in the database, specifically using cosine similarity or inner product distance methods. After calculating all scores, they are globally sorted in descending order of score, and several top-ranked knowledge fragments are selected as candidate knowledge fragments semantically related to the current event. This step aims to recall knowledge content that is highly relevant to the current event in textual meaning but may not be structurally directly adjacent through semantic-level matching.

[0051] S1033. Using the structured information of the knowledge graph in the preset knowledge base, reorder the semantically related knowledge fragments of each event to generate a reordered knowledge set.

[0052] For example, embodiments of the present invention can utilize the rich structured information contained in the knowledge graph to reorder the knowledge fragments recalled in the previous vector retrieval step. For instance, the present invention can extract the graph entity ID corresponding to each recalled knowledge fragment and reverse-retrieve the shortest path distance between that entity and the initial event node in the knowledge graph, the types of relationships traversed along the path, and the relationship weights. Based on this, a reordering weighted algorithm is designed: entities with a direct relationship to the event node or a short path distance are given a positive gain, increasing their ranking; entities that exist in isolation or have a long path distance are given a negative penalty, decreasing their ranking. After this reordering operation, knowledge fragments that are not only semantically similar but also structurally closely related are preferentially selected, generating a reordered knowledge set.

[0053] S1034. For the reordered knowledge set, perform weighted fusion according to knowledge type, similarity score and source reliability to construct the decision context information for each event.

[0054] In some embodiments, knowledge types include emergency plans, historical cases, regulatory clauses, resource information, and contact information.

[0055] For example, for each piece of knowledge in the reordered knowledge set, a weighted calculation is performed from three dimensions. The first dimension is the knowledge type, with different basic weight coefficients preset according to the category to which the knowledge belongs (such as emergency plans, historical cases, regulatory clauses, resource information, and contact information). For example, regulations and norms have the highest weight, followed by cases. The second dimension is the similarity score, which is the original semantic similarity value calculated in the vector retrieval stage. The third dimension is the source reliability, with a reliability coefficient assigned according to the authority of the knowledge source (such as national-level regulations being higher than local-level regulations, and official historical reports being higher than publicly available online information). Finally, a comprehensive value score is calculated for each piece of knowledge through linear weighted summation or multiplicative fusion. Based on this score, the top-ranked knowledge subset is selected again, and the content of these knowledge fragments is structured, spliced, and formatted to construct decision-making context information.

[0056] S104. Based on the feature vector and decision context information of each event, analyze and reason through a large language model to generate multiple candidate disposal solutions for each event.

[0057] In some embodiments, a candidate disposal scheme refers to a set of multiple differentiated disposal schemes generated in parallel by a large language model for the same event under different optimization orientations.

[0058] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1045.

[0059] S1041. Based on the feature vector of each event, the event type, number of affected lanes, length of affected area, traffic flow mutation coefficient and meteorological influence factor are obtained by analysis.

[0060] For example, embodiments of the present invention can first perform structured parsing on the feature vector of each event. Since the feature vector is formed by concatenating values ​​from multiple dimensions in a fixed order, it is decomposed according to a predefined dimension mapping table. For example, dimensions 1 to 10 are set as one-hot encodings of the event type, dimensions 11 to 12 as normalized values ​​of GPS latitude and longitude coordinates, dimension 13 as the number of affected lanes, dimension 14 as the length of the affected area (in meters), dimension 15 as the traffic flow mutation coefficient, and dimension 16 as the meteorological influence factor. The parsing program extracts the corresponding values ​​from the feature vector one by one according to the mapping table, and then restores them to the actual semantic values ​​through inverse normalization or inverse encoding operations, finally obtaining a structured set of event attributes. For missing or outlier values ​​that may exist in the feature vector, preset default values ​​(such as the number of affected lanes being 1 by default, the meteorological influence factor being 0.5 by default) are used for filling or smoothing to ensure the robustness of the parsing process.

[0061] S1042. Based on the event type, number of affected lanes, length of affected area, traffic flow mutation coefficient and meteorological influence factor, match the preset highway event classification standard to determine the severity level, response level and resource allocation constraints of each event.

[0062] In some embodiments, the present invention pre-stores a set of highway event classification standards. This table defines the severity level, response level, and resource allocation constraints for different event types using a multi-dimensional combination of conditions. For example, the judgment logic of the classification standards is as follows: if the event type is "traffic accident" and the number of affected lanes is ≥2, the length of the affected area is >200 meters, and the traffic flow mutation coefficient is >0.7, it is judged as a "major event," corresponding to a Level 1 response, requiring at least one clearing vehicle, one patrol car, and one ambulance; if the number of affected lanes is 1, the length of the affected area is ≤100 meters, and the mutation coefficient is ≤0.4, it is judged as a "general event," corresponding to a Level 3 response, requiring only one patrol car. Substituting each attribute value into the classification standards table, the judgment conditions are matched one by one. When all the threshold conditions for a certain level are met, the corresponding event severity level, response level, and the required resource type and quantity minimum are output.

[0063] S1043. Based on the complete perception record of each event, the traffic flow mutation coefficient and meteorological influence factors, as well as the emergency plan, historical cases, control specifications, available resources and contact information in the decision context information, generate a reasoning prompt sequence.

[0064] For example, embodiments of the present invention can combine the complete perception record of the current event, traffic flow mutation coefficient, meteorological influencing factors, and emergency plan content, historical case descriptions, control regulations, available resource lists, and contact information from the decision context information into a structured natural language prompt sequence according to a preset template structure. The construction of the prompt sequence follows a four-part structure of "scenario description - constraints - knowledge reference - task instructions": the first part describes the basic situation of the event (event type, location, scope of impact, traffic impact, meteorological constraints); the second part lists the response levels and resource allocation constraints matched by the grading criteria; the third part lists the relevant plan points, similar historical cases, applicable regulations, available resource details, and contact information in a hierarchical format; the fourth part gives a clear reasoning task instruction, requiring the model to generate a specific handling plan.

[0065] S1044. Based on the inference prompt sequence and the large language model, by setting multiple sets of differentiated target weight parameters, the large model is guided to generate multiple initial handling schemes in parallel along different optimization directions.

[0066] In some embodiments, optimization directions include prioritizing timeliness of disposal, minimizing impact on road traffic, and minimizing disposal resource consumption.

[0067] For example, in embodiments of the present invention, after constructing a basic inference prompt sequence, multiple sets of differentiated target weight parameters can be set to guide the large language model to generate initial disposal schemes along different optimization directions. For instance, explicit target preference descriptions can be added to the input prompt sequence. For example, for the priority of disposal time, an instruction can be added stating that the shortest disposal time should be the primary objective, and resource input can be appropriately increased; for the priority of minimizing road traffic impact, an instruction can be added stating that interference with normal traffic should be minimized, and the scheme with the fewest lane closures should be prioritized; for the priority of minimizing disposal resource consumption, an instruction can be added stating that the principle of minimizing manpower and material consumption should be followed, and lightweight disposal measures should be prioritized. These three sets of instructions independently call the large language model for inference. Upon receiving different target weight parameters, the same model will exhibit differentiated selection tendencies at each step of the autoregressive decoding process in scheme generation, thereby outputting three initial disposal schemes emphasizing different optimization objectives.

[0068] S1045. Based on the normative clauses and emergency plans in the decision-making context information, perform compliance verification and content completion on each initial handling plan to obtain multiple candidate handling plans for each event.

[0069] For example, for the three initial response plans generated by the large language model, compliance verification and content completion are performed one by one. For instance, this invention extracts all applicable regulatory clauses and emergency plan content from the decision context information to construct a rule checklist; then, each step in the generated plan is compared item by item with this checklist to check for any violations of regulations (such as recommending the closure of all lanes when visibility is less than 50 meters but failing to set up advance warning signs, or failing to mention ventilation and lighting control in tunnel incidents). If non-compliance is found, the corresponding measures are automatically corrected according to the relevant regulatory clauses. The specific operation for content completion is as follows: checking whether the generated plan contains key necessary elements, including the event confirmation stage, response initiation stage, traffic control measures, obstacle removal and rescue measures, information dissemination measures, on-site cleanup measures, and response termination conditions; for missing elements, corresponding standard action templates are extracted from the emergency plan and inserted to ensure that each candidate plan is structurally complete and has all necessary elements, ultimately outputting multiple candidate response plans.

[0070] S105. Based on multiple candidate handling schemes for each event, conduct multi-dimensional quantitative evaluation and global conflict resolution to generate a collaborative handling scheme for highway events.

[0071] In some embodiments, multi-dimensional quantitative assessment refers to the process of quantitatively scoring each candidate disposal plan from six assessment dimensions: timeliness, safety, resource consumption, traffic impact, feasibility, and historical effects. Each dimension has clear quantitative indicators and scoring standards, and the assessment results are obtained for each plan in each dimension.

[0072] In some embodiments, global conflict resolution refers to the process of detecting and coordinating the resolution of conflicts among the preferred solutions for each event in a multi-event concurrent scenario. Specifically, it includes the detection and resolution of three types of conflicts: resource conflicts (the same resource is requested by multiple events simultaneously), information conflicts (different information is requested to be published on the same information board simultaneously), and spatiotemporal conflicts (overlapping control areas or overlapping control times for different events). Conflict resolution is coordinated based on the priority of each event, prioritizing the resource and control needs of high-priority events.

[0073] In some embodiments, the collaborative handling scheme is a unified coordinated handling plan for multi-event concurrent scenarios, formed after multi-dimensional quantitative evaluation to select the optimal solution for each event and then global conflict resolution. This scheme takes into account the priority, resource availability, and spatiotemporal constraints of each event.

[0074] As one possible implementation, step S105 can be specifically implemented as steps S1051-S1055.

[0075] S1051. For each event, multiple candidate disposal plans are quantitatively evaluated from six evaluation dimensions: timeliness, safety, resource consumption, traffic impact, feasibility, and historical effect, to obtain the evaluation score of each candidate disposal plan in each evaluation dimension.

[0076] In some embodiments, for each candidate response plan for each event, a quantitative score is calculated based on six evaluation dimensions: timeliness, safety, resource consumption, traffic impact, feasibility, and historical effectiveness. The specific quantitative methods for each dimension are as follows: Timeliness is measured by the estimated total time required from the initiation of the plan to the completion of all response measures, obtained by summing the durations of each sub-measure in the plan; the shorter the duration, the higher the score. Safety is measured by the risk level to on-site personnel, passing vehicles, and surrounding facilities during the implementation of the plan, categorized by risk level coding based on factors such as whether lane closures are involved, whether close-range operations are required, and whether hazardous chemicals are present; the lower the risk, the higher the score. Resource consumption is measured by converting the required manpower (person-hours), equipment shifts, and material consumption into a unified cost unit (e.g., yuan); the lower the cost, the higher the score. Traffic impact... The impact dimension uses the degree of reduction in traffic capacity and the duration of the impact during the implementation of the plan as a comprehensive indicator, quantified by simulating the total vehicle-hours of expected vehicle delays; the less the delay, the higher the score. The feasibility dimension uses the degree of matching of each measure in the plan with current resource reserves, personnel allocation, and external conditions (such as weather) as an indicator, checking for any resource gaps or unmet conditions; the fewer the gaps, the higher the score. The historical effect dimension uses the success rate and average handling effect of the plan in similar historical events as a reference; by searching the historical execution records of similar plans in the knowledge base, the success rate and average effect score are calculated; the better the historical effect, the higher the score.

[0077] S1052. Based on the evaluation scores of each candidate scheme in each evaluation dimension and the preset weight coefficients, calculate the comprehensive score of each candidate scheme.

[0078] For example, for each candidate solution, its evaluation scores in each of the six dimensions are multiplied by their respective preset weighting coefficients, and then summed to obtain the solution's overall score. The weighting coefficients can be dynamically adjusted according to current management objectives. For instance, in daily operations, traffic impact can be given the highest weight, followed by resource consumption; during major events, safety weight can be increased; and during emergency response drills, timeliness weight can be increased. The sum of the weighting coefficients is always 1. The calculation formula is: Overall Score = Σ (Dimension Score × Dimension Weight).

[0079] S1053. Sort the candidate solutions from high to low according to the comprehensive scores, and select the candidate solution with the highest score as the preferred solution for each event.

[0080] For example, multiple candidate solutions for the same event are sorted from highest to lowest based on their comprehensive scores. The solution with the highest score is the optimal solution for that event under the current target weight configuration. The solution with the highest score is selected as the preferred solution for that event, and its scores in each dimension and its comprehensive score are recorded as the basis for subsequent conflict resolution. If there is a tie for the highest score, the final decision is made according to a preset priority rule (such as the default of the solution with the higher score in the security dimension).

[0081] S1054. Conflict detection is performed on the optimal solutions for each event to obtain the identification results.

[0082] In some embodiments, the identification results include whether there is a resource conflict where the same resource is requested by multiple events simultaneously, an information conflict where the same information board is required to publish different information at the same time, and a spatiotemporal conflict where different event control areas overlap or control times intersect.

[0083] For example, embodiments of the present invention can perform parallel analysis on the preferred solutions for all events to identify three types of potential conflicts. The first type is resource conflict. This involves traversing the equipment resources (such as tow trucks, patrol cars, cranes, ambulances, etc.), personnel teams, and materials involved in each solution, and statistically analyzing the demand distribution of each resource on the time axis. If the total demand for the same resource exceeds the available quantity due to two or more events within the same time period, a resource conflict is determined. The second type is information conflict. This involves checking the information dissemination content of variable message signs or traffic broadcasts in each solution. If the same sign is required to publish different guidance information within the same time period (e.g., one event requires an alert about an accident ahead, please detour; another event requires an alert about construction ahead, please slow down), an information conflict is determined. The third type is spatiotemporal conflict. This involves comparing the traffic control areas (represented by station intervals) and control time windows of each solution. If the impact ranges of two events overlap and their overlapping times intersect, or if the diversion point of one event is located within the control area of ​​another event, a spatiotemporal conflict is determined.

[0084] S1055. Based on the identification results and the priority of each event, conflict resolution is carried out to obtain a collaborative handling plan for highway events.

[0085] For example, embodiments of the present invention can coordinate and resolve conflicts based on a conflict list and the priority of each event. The event priority is determined by the severity level (e.g., major events are higher than general events), urgency level (e.g., personnel casualties are higher than simple vehicle malfunctions), and response level. For instance, for resource conflicts, resources are allocated sequentially from highest to lowest priority, with high-priority events occupying their required resources first, while low-priority events have their conflicting resource items removed from their plans and replaced with alternative resources (e.g., changing a clearing vehicle to a tow truck) or have their activation time delayed until resources become available. For information conflicts, the content published on information boards is determined by priority, with information from high-priority events displayed first, and information from low-priority events adjusted to other information boards or delayed. For spatiotemporal conflicts, low-priority events adjust their controlled area (e.g., reducing the number of closed lanes or shortening the control length) or postpone the start time of control to avoid spatiotemporal overlap with high-priority events. After resolving these conflicts categorized by type, all event plans are adjusted to be executable in parallel and without conflict, ultimately forming a globally oriented collaborative handling plan.

[0086] This invention provides an intelligent decision-making method for highways that integrates knowledge base and large-scale model reasoning. By fusing multi-source sensing data and extracting unified feature vectors, it solves the problem of long confirmation times caused by manual comparison of data from multiple terminals, achieving rapid and accurate perception of event information. Through a three-level hierarchical retrieval strategy consisting of knowledge graph structure retrieval, vector semantic retrieval, and graph reordering, it quickly and accurately locates and filters the most relevant decision context information from massive, dispersed knowledge resources, overcoming the low retrieval efficiency and information omissions caused by traditional manual document retrieval. Based on this, it uses a large language model combined with feature vectors and contextual information to reason and generate multiple candidate solutions, automating and parallelizing the solution formulation process that originally relied on manual experience deduction and sequential communication, shortening the solution generation time. Finally, through multi-dimensional quantitative evaluation and global conflict resolution, it can automatically coordinate resource and measure conflicts in multi-event concurrent scenarios, avoiding the problem of exponentially increasing difficulty of manual coordination. This invention solves the problems of long response times and low efficiency in existing highway event handling decisions, improving the efficiency of highway event handling.

[0087] Optionally, the intelligent decision-making method for highways that integrates knowledge base and large model reasoning provided in this embodiment of the invention further includes steps S201-S205 before step S103.

[0088] S201. Collect multi-source knowledge resources required for highway decision-making. These multi-source knowledge resources include structured knowledge, semi-structured knowledge, and unstructured knowledge.

[0089] Structured knowledge includes emergency response plan databases, traffic control standard databases, and organizational and contact information databases; semi-structured knowledge includes historical event handling case databases; and unstructured knowledge includes policy and regulatory documents and technical standards and specifications.

[0090] For example, knowledge resources can be aggregated from multiple business systems and document libraries within a highway operation and management unit through customized data acquisition interfaces or ETL tools. For structured knowledge, such as the plan number, applicable scenarios, response level, and handling process fields in the emergency plan information table, as well as the clause numbers and content in the control and management standard library, and the department names, responsibilities, and contact information in the organizational contact information library, these can be directly imported in batches via database connections. For semi-structured knowledge, namely the historical event handling case library, case summaries, handling measures, and effect evaluations are extracted by parsing its fixed-format fields. For unstructured knowledge, including policy and regulatory documents and technical standards in PDF or Word formats, plain text content is extracted using document parsing tools.

[0091] S202. Perform vector transformation on the multi-source knowledge resources to obtain a unified vector knowledge base.

[0092] For example, embodiments of the present invention can perform unified vectorization processing on various types of collected knowledge resources at the content level. First, for structured and semi-structured knowledge, their field content is converted into natural language description sentences; for unstructured text, the original text is directly retained. Then, a general semantic embedding model suitable for the Chinese domain is used to segment all text content into blocks, dividing it according to paragraphs or fixed lengths. Each text block is independently input into the embedding model to generate a corresponding fixed-dimensional dense vector. These vectors, along with the corresponding original text blocks and metadata tags, are stored in a vector database to form a unified vector knowledge base, and a unique index number is generated for each vector record.

[0093] S203. Based on multi-source knowledge resources, extract entities and relationships between entities; entities include event types, handling measures, equipment resources, organizational structures and personnel roles, and relationships between entities include causal relationships, temporal relationships and hierarchical relationships.

[0094] For example, embodiments of the present invention can extract knowledge elements based on collected multi-source knowledge resources using natural language processing technology. On the one hand, for structured data, field names and values ​​are directly mapped to predefined entity types, such as mapping the plan name to the response measure entity and the lead department to the organization entity. On the other hand, for unstructured text, automated annotation is performed using deep learning-based named entity recognition and relation extraction models to extract entities such as event types (e.g., rear-end collisions, congestion, traffic control), response measures (e.g., road closures, traffic diversions, information boards), equipment resources (e.g., tow trucks, patrol vehicles), organizations (e.g., command centers, road administration teams), and personnel roles (e.g., duty officers, on-site commanders). Simultaneously, through syntactic dependency analysis and semantic role annotation, causal relationships (e.g., accidents causing congestion), temporal relationships (e.g., warnings before responses), and hierarchical relationships (e.g., a general plan includes specific plans) between entities are identified.

[0095] S204. Construct a knowledge graph for the transportation domain based on entities and relationships between entities.

[0096] For example, in this embodiment of the invention, all extracted entities can be treated as nodes in a graph, and the relationships between entities can be treated as edges connecting the nodes, which are then imported into a graph database for storage. A unified attribute structure is defined for each node, including node name, node type, unique identifier ID, source file, and creation time; for each edge, relationship type, weight (such as the degree of relationship closeness), and direction are defined. Simultaneously, indexes are created for both nodes and edges to improve the retrieval efficiency of subsequent graph traversal. Ultimately, a highly connected transportation domain knowledge graph centered on event handling is constructed.

[0097] S205. Based on the unified vector knowledge base and the knowledge graph of the transportation field, establish mapping relationships to obtain the preset knowledge base.

[0098] For example, embodiments of the present invention can establish a bidirectional mapping relationship between a unified vector knowledge base and a transportation domain knowledge graph. For instance, in the metadata corresponding to each vector record in the unified vector knowledge base, a field is added to store the node ID or relation ID corresponding to that knowledge in the graph; simultaneously, in the attributes of each node in the knowledge graph, the vector index number corresponding to the node content in the unified vector knowledge base is recorded. Through this bidirectional ID association mechanism, when performing vector semantic retrieval, the corresponding node and its surrounding structured information in the graph can be quickly located based on the returned vector record; conversely, when performing graph structure retrieval, the vector representation of the corresponding node can be quickly retrieved for semantic similarity calculation.

[0099] Thus, by constructing a pre-defined knowledge base that integrates vector knowledge base and knowledge graph, this invention enables multi-source heterogeneous knowledge resources to be uniformly represented and interconnected, providing structured and semantic dual support for three-level hierarchical retrieval, significantly improving the relevance and coverage of knowledge retrieval, and shortening information acquisition time from minutes to seconds.

[0100] Optionally, the intelligent decision-making method for highways that integrates knowledge base and large model reasoning provided in this embodiment of the invention further includes steps S301-S306 before step S104.

[0101] S301, collect historical event handling reports, emergency drill records, traffic control cases, and various emergency plan texts for highways.

[0102] In some embodiments, the historical archive system and business record platform of highway operation and management units acquire four types of original data through data interfaces or batch import methods. Historical event handling reports record the time, location, process, handling measures taken, actual execution time of each measure, resources used, and final handling effects of various past events, typically stored in a mixed format of structured forms and text descriptions. Emergency drill records document the drill process simulating various event scenarios, problems exposed during the drills, and improvement suggestions, stored in the form of drill summary reports. Traffic control case records document the implementation of daily traffic control measures, including the reasons for control, control methods (such as entrance closures, mainline diversion, speed limits, etc.), control periods, and control effect evaluations. Various emergency plan texts include general emergency plans and specific emergency plans (such as severe weather emergency plans, hazardous chemical accident emergency plans, tunnel emergency plans, etc.).

[0103] S302. Generate a fine-tuning corpus based on historical event handling reports, emergency drill records, traffic control cases, and various emergency plan texts.

[0104] For example, embodiments of the present invention can perform text extraction, cleaning, and formatting on various types of raw data collected. For Word and PDF documents, document parsing tools are used to extract plain text content, removing noise information such as headers, footers, watermarks, chart annotations, and non-text annotations. For structured form data, the content of each field is converted into coherent natural sentence descriptions (e.g., converting "Event type: rear-end collision, affected lane: lane 2, response measure: lane 2 closed" into the text "This event is a rear-end collision, affecting lane 2, and the response measure is to close the lane"). All the cleaned and converted text is segmented according to event units or contingency plan units, with each record as an independent corpus entry, and attached with structured tags (such as source type, event type, and good / bad marks for response results).

[0105] S303. Extract event summaries, environmental and resource constraints, disposal target preferences, and disposal schemes from the fine-tuning corpus to construct an instruction fine-tuning dataset.

[0106] The input of each fine-tuning sample in the instruction fine-tuning dataset includes event perception information, decision context information, and target weight vector. The target weight vector is used to identify the dominant preferences among priority of handling timeliness, minimization of road traffic impact, and minimum handling resource consumption. The output is the handling plan that conforms to each dominant preference.

[0107] For example, embodiments of the present invention can extract the event summary (event overview description), environmental and resource constraints (weather, traffic flow, and available resources at the time of the event), disposal target preference (the dominant optimization direction of this disposal), and the final disposal plan adopted for each record from the fine-tuning corpus, and assemble them into training samples according to the data format of the instruction fine-tuning. The input part of each sample includes three items: the first item is event perception information (i.e., the event attribute description after feature vector parsing); the second item is decision context information (i.e., relevant knowledge fragments retrieved from the knowledge base); the third item is the target weight vector (composed of three values, corresponding to the weights of the three dimensions of "priority of disposal timeliness", "minimum impact on road traffic", and "lowest disposal resource consumption", the sum of which is 1, where the highest weight represents the dominant preference direction of the sample). The output part of the sample is the text of the disposal plan that is finally actually adopted or recommended by experts under the guidance of the dominant preference.

[0108] S304. Using the instruction fine-tuning dataset, supervised fine-tuning is performed using a pre-trained general large language model as the base model to obtain the initial decision model.

[0109] For example, embodiments of the present invention can use a pre-trained general-purpose large language model (such as open-source base models like Qwen, Baichuan, or the ChatGLM series) as the initial base model. The pre-trained weights of the base model are loaded, and domain adaptation is performed using either low-rank adaptation or full-parameter fine-tuning. Specifically, during training, the input parts of the training samples in the instruction fine-tuning dataset are concatenated to form the model's input sequence, and the output parts serve as supervision labels. The difference between the model-generated text and the labels is calculated using the cross-entropy loss function, and the model parameters are iteratively updated using the backpropagation algorithm. During training, the loss value and generation quality index on the validation set are monitored. Training is terminated when the validation set loss tends to plateau or no longer decreases for several consecutive epochs, and the optimal checkpoint is saved as the initial decision model. This model already possesses the ability to generate reasonable handling solutions based on input features and context in highway event decision-making scenarios.

[0110] S305. Collect multi-dimensional preference ranking data of traffic management experts on the solutions generated by the initial decision-making model under typical event scenarios.

[0111] In some embodiments, multi-dimensional preference ranking data includes preference ranking data for each dimension: processing timeliness, traffic impact, and resource consumption.

[0112] For example, embodiments of the present invention can design typical highway incident scenarios (covering combinations of different types, severity levels, and weather conditions), and use an initial decision model to generate multiple candidate handling solutions for each scenario. Multiple candidate solutions under the same scenario are then combined in pairs or triplets, and senior experts in the field of traffic management are invited to rank the solutions based on their preferences from the dimensions of handling timeliness, traffic impact, and resource consumption. That is, for each dimension, experts independently determine which solution performs better and provide a ranking result. For example, in the dimension of handling timeliness, if experts believe that solution A has the shortest completion time, solution B is next, and solution C has the longest, then the ranking is A > B > C. All expert ranking judgments for all scenarios and all dimensions are collected, and consistency checks are performed. After removing abnormal data with low consistency, a high-quality expert preference ranking dataset is formed.

[0113] S306. Based on multi-dimensional preference ranking data, the initial decision model is optimized using a preference alignment algorithm to obtain a large language model.

[0114] For example, embodiments of the present invention can further optimize the initial decision-making model by utilizing multi-dimensional preference ranking data from experts and employing preference alignment algorithms or PPO algorithms in reinforcement learning from human feedback (RLHF). Taking the DPO algorithm as an example, its specific operation is as follows: the expert preference ranking data is transformed into pairwise preference pairs (selection pairs vs. rejection pairs), and a preference loss function is constructed. This function measures the degree to which the probability of the model generating a selected solution is higher than the probability of a rejected solution. By minimizing this loss function, a finite number of gradient updates are performed in the parameter space of the initial decision-making model, causing the model's output distribution to gradually shift towards the direction of expert preferences. The resulting large language model after optimization not only possesses domain knowledge of highway incident handling, but also generates solutions that better align with the judgment preferences of human experts in multiple dimensions such as handling timeliness, traffic impact, and resource consumption, thereby improving the professionalism and acceptability of the decision-making solutions.

[0115] Thus, this invention optimizes the large language model by fine-tuning instructions and aligning with expert preferences, making the model-generated solutions highly consistent with the realities of traffic management in terms of professionalism and multi-objective balance, reducing the time spent on manual solution deduction and communication, and fundamentally improving the quality and feasibility of decision-making.

[0116] Optionally, the intelligent decision-making method for highways that integrates knowledge base and large model reasoning provided in this embodiment of the invention further includes steps S401-S405 after step S105.

[0117] S401. During the implementation of the collaborative response plan, continuously collect actual execution data.

[0118] In some embodiments, actual execution data includes the actual execution time of each response measure, the actual resource consumption, the actual traffic flow recovery time, and whether secondary events occur.

[0119] For example, after the collaborative response plan is issued and implemented, this embodiment of the invention can connect to the business execution system of the highway operation management department (such as the dispatch and command system, resource management system, traffic flow monitoring system, etc.) to collect the actual execution status of various response measures in real time or near real time. The specific data items collected include: the actual start time and actual completion time of each measure (obtained through reporting from on-site personnel's mobile terminals, vehicle GPS trajectories, or changes in equipment status); the actual quantity of various resources (manpower, vehicles, materials) consumed; the actual time when the traffic flow in the affected area recovers from an abnormal state to a normal traffic state, and the traffic flow recovery time is calculated accordingly; and whether secondary events (such as secondary accidents, new congestion points, etc.) were caused due to the ineffective implementation of the original plan or changes in external conditions during the entire response process.

[0120] S402. Compare the actual execution data with the expected data in the collaborative disposal plan item by item, and calculate the execution deviation of each disposal measure.

[0121] For example, embodiments of the present invention can compare the actual execution data with the preset expected data in the collaborative handling plan (including the expected start time, expected duration, expected resource consumption, expected traffic flow recovery time, and expected probability of secondary events for each measure) item by item, and calculate the deviation value. The deviation is calculated as follows: time deviation = actual execution time - planned execution time (positive value is a delay); resource consumption deviation = actual consumption - planned consumption (positive value is an overspending); traffic flow recovery time deviation = actual recovery time - planned recovery time. Regarding whether a secondary event occurs, if an unexpected secondary event actually occurs, it is marked as a major deviation. The present invention presets allowable deviation thresholds for each deviation type (for example, a time deviation of no more than 5 minutes is normal, and a resource consumption deviation of no more than 20% is normal).

[0122] S403. If the execution deviation exceeds the preset deviation threshold, the affected measures affected by the execution deviation are identified based on the causal and temporal relationships stored in the preset knowledge base, and the identification results are obtained.

[0123] For example, when the execution deviation of a certain measure exceeds a preset deviation threshold, this embodiment of the invention can, based on the causal relationships (e.g., if the clearing operation on a certain road segment is delayed, it will directly cause subsequent traffic diversion measures to fail to start on time) and temporal relationships (e.g., information release must be completed 5 minutes before traffic control begins, otherwise the warning effect will be lost), stored in a preset knowledge base, start from the deviating measure and traverse downstream along the relationship chain to find all subsequent measures that are logically dependent on the deviating measure and have not yet started or are being implemented. These measures are marked as "affected measures". The identification results include the name of the affected measure, the planned time, the type of dependency relationship with the original deviating measure, and the expected severity of the impact.

[0124] S404. Based on the identification results, dynamically adjust the start time of the affected measures and the resource allocation scheme.

[0125] For example, embodiments of the present invention can recalculate the optimal start execution time and resource allocation scheme for affected measures and their dependencies. The adjustment algorithm uses the original plan as a baseline, the current actual completion time of the deviating measures as the new anchor point, and extends it sequentially along the time-series dependency chain, while simultaneously checking whether resources can be reallocated (e.g., allocating unused resources originally allocated to completed measures to delayed measures in advance). The goal of the adjustment is to minimize the impact on all subsequent measures while satisfying all time-series dependencies and resource constraints, and to make the overall scheme completion time as close as possible to the original plan. The adjustment process also needs to consider changes in current real-time weather and traffic flow data. If conditions change significantly, the parameters of relevant measures are simultaneously corrected in the adjustment scheme (e.g., adding anti-skid measures due to increased rainfall).

[0126] S405. Based on the dynamically adjusted start execution time and resource allocation scheme of the affected measures, regenerate the collaborative handling scheme.

[0127] For example, embodiments of the present invention can merge the start execution times and resource allocation schemes of all affected measures after dynamic adjustment, as well as the original plans of unaffected measures, to form an updated and complete collaborative handling plan. The new plan clearly indicates the adjustments that differ from the original plan and their reasons. The regenerated plan is pushed to on-site command personnel and various implementing departments in the form of a message, and the task dashboard of the dispatch system is updated simultaneously. At the same time, the subsequent execution data of the new plan continues to be collected, forming a closed loop of cyclical monitoring—deviation detection—dynamic adjustment—plan update, until all events are handled and traffic returns to normal.

[0128] Thus, by continuously collecting actual data during the execution of the solution and comparing it with the expected data item by item, the present invention automatically identifies the affected measures by utilizing the causal and temporal relationships in the knowledge graph when the deviation exceeds the threshold, and dynamically adjusts the time and resource allocation before regenerating the solution. This enables the collaborative handling solution to have closed-loop adaptive capabilities, effectively avoiding the problem of overall failure of static solutions due to execution deviations, and significantly improving the robustness of solution execution and the final handling effect in multi-event concurrent scenarios.

[0129] Optionally, the intelligent decision-making method for highways that integrates knowledge base and large model reasoning provided in this embodiment of the invention further includes steps S501-S506 before step S104.

[0130] S501. Based on the event type and impact range length in the feature vector of each event, and the emergency plan level in the decision context information, determine the urgency level and confidence score of each event.

[0131] In some embodiments, the confidence score is determined by weighting the overall confidence score in the complete perceived record and the average similarity score of the knowledge fragments returned by the three-level hierarchical retrieval.

[0132] For example, embodiments of the present invention can extract the event type code (such as "traffic accident", "vehicle breakdown", "fire" etc.) and the length of the impact range (in meters) from the feature vector of each event, and simultaneously extract the matched emergency response plan level (such as Level 1 response, Level 2 response, Level 3 response) from the decision context information. These three parameters are input into a preset urgency determination matrix, which predefines urgency levels according to different types, impact ranges, and plan levels. The levels are divided into three grades: "Level I (Extremely Urgent)", "Level II (Emergency)", and "Level III (General)". For example, if the event type is "fire" or "hazardous chemical leak", it is directly determined to be Level I regardless of the impact range; if it is a "traffic accident" and the impact range is >500 meters and the plan is Level 1 response, it is determined to be Level I; if the impact range is ≤200 meters and the plan is Level 3 response, it is determined to be Level III. The confidence score is calculated as follows: the overall confidence score of the complete perception record (derived from multi-source cross-validation, with a value range of 0 to 1) is weighted and summed with the average similarity score of all knowledge fragments returned by vector semantic retrieval (with a value range of 0 to 1). The weights can be set as 0.6 for perception confidence and 0.4 for knowledge similarity. The final score is between 0 and 1. The higher the score, the higher the reliability of the perception and knowledge matching of the current event.

[0133] S502. Select the target reasoning pattern from multiple candidate reasoning patterns based on the urgency level and confidence score.

[0134] In some embodiments, candidate reasoning modes include lightweight reasoning mode, fast reasoning mode, standard reasoning mode, and deep reasoning mode.

[0135] For example, this invention selects one of four candidate reasoning modes based on the urgency level and confidence score, according to a preset mode selection rule. The candidate reasoning modes include a lightweight reasoning mode, a fast reasoning mode, a standard reasoning mode, and a deep reasoning mode. The specific rules for mode selection are as follows: If the urgency level is Level III (general) and the confidence score is ≥0.9, the lightweight reasoning mode is selected because the event is not urgent and the information is highly reliable, requiring no complex reasoning; if the urgency level is Level II (urgent) or the confidence score is between 0.6 and 0.9, the fast reasoning mode is selected, requiring compressed response time while ensuring a certain level of reasoning quality; if the urgency level is Level II and the confidence score is <0.6, or the urgency level is Level I but the confidence score is ≥0.8, the standard reasoning mode is selected, requiring complete reasoning even when the information is not entirely reliable; if the urgency level is Level I (extremely urgent) and the confidence score is <0.8, the deep reasoning mode is selected because the event is extremely serious and the information reliability is low, requiring the most comprehensive reasoning combined with human verification. If a rule conflict occurs (such as simultaneously satisfying the conditions of multiple modes), the heavier reasoning mode is selected according to the principle of "choosing the higher priority," that is, the deep reasoning mode has the highest priority, and so on down.

[0136] S503. If the target reasoning mode is the lightweight reasoning mode, then the lightweight instance of the large language model is called to perform analysis and reasoning. The lightweight instance skips the target recognition and constraint recognition steps and directly generates candidate disposal solutions based on the emergency plan in the decision context information.

[0137] For example, when the lightweight inference mode is selected, a pre-deployed lightweight instance of the large language model is invoked (this instance is a small-parameter version after knowledge distillation or pruning, and its inference speed is significantly faster than the standard instance). During the inference process, this instance skips the event target identification step (i.e., it no longer re-analyzes the event type and handling target) and the constraint identification step (i.e., it no longer re-parses the environment, resources, and normative constraints). It directly extracts the "emergency plan" content from the "decision context information" part of the inference prompt sequence, and uses a retrieval-enhanced template filling method to extract the standard handling process corresponding to the event type and level from the plan and fill it as a candidate handling solution.

[0138] S504. If the target reasoning mode is fast reasoning mode, then call the standard instance of the large language model to perform analysis and reasoning, and force the output of candidate disposal solutions within the preset time threshold.

[0139] For example, when the fast inference mode is selected, a standard instance of the large language model (i.e., the full-parameter version) is invoked, but a mandatory time threshold (e.g., 3 seconds) is set. When the timer reaches the preset time threshold, the model's generation process is forcibly truncated, and the longest complete sentence generated so far is taken as the candidate solution output. To compensate for any content loss that may be caused by truncation, compliance checks and content completion processing are automatically added after the truncated output, extracting and inserting any missing necessary elements from the contingency plan to ensure the integrity of the output solution.

[0140] S505. If the target reasoning mode is the standard reasoning mode, then call the standard instance of the large language model to fully execute the analysis and reasoning process and generate multiple candidate solutions for each event.

[0141] For example, when the standard inference mode is selected, the standard instance of the large language model is invoked to fully execute the entire inference process from steps S1041 to S1045. This includes parsing feature vectors, matching hierarchical standards, generating a complete inference prompt sequence (containing perceptual information and all knowledge types in the decision context), setting multiple sets of differentiated target weight parameters to guide the model to generate multiple solutions in parallel, and performing compliance verification and content completion on each solution. No mandatory truncation time is set during the inference process; the model uses randomized sampling generation (e.g., setting the temperature parameter to 0.7) to maintain solution diversity until a natural end-of-process token (EOS Token) is generated. The entire process typically takes 5-15 seconds, outputting multiple candidate solutions for each event.

[0142] S506. If the target reasoning mode is deep reasoning mode, then call the standard instance of the large language model to fully execute the analysis and reasoning process, generate multiple candidate solutions for each event, and push the candidate solutions to the manual review end for confirmation before outputting the solutions.

[0143] For example, when the deep reasoning mode is selected, the standard instance of the large language model is called and the entire reasoning process from S1041 to S1045 is executed, which is completely consistent with the standard reasoning mode in terms of model calling and generation process. However, before the solution output, a manual review step is added: the generated multiple candidate disposal solutions, along with their evaluation scores for each dimension, comprehensive scores, links to the knowledge sources on which they are based, and reasoning paths, are pushed to the manual review terminal of traffic management experts (such as the large screen system of the dispatch center or a mobile approval terminal) in the form of a visual report. The review terminal interface provides functions such as solution comparison, confirmation of each measure, modification and editing, and rejection and regeneration. Wait for the manual review terminal to return a confirmation signal or modification comments. If the review is passed, the solution is output as a candidate solution and enters step S105; if the review is rejected and modification comments are attached, the modification comments are used as supplementary constraints to call the model again for secondary reasoning until manual confirmation is passed.

[0144] Thus, by dynamically selecting lightweight, fast, standard, or deep inference modes based on urgency and confidence scores, this invention ensures that the investment of large model inference resources is precisely matched with the actual needs of the event, avoiding excessive computing resources and redundant overhead caused by low-urgency events. At the same time, it ensures that high-urgency, low-confidence events receive the dual guarantee of the most rigorous inference and human confirmation, significantly improving the overall inference efficiency and the reliability of decision-making in key scenarios.

[0145] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0146] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention; the electronic device 600 includes: a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601; when the processor 601 executes the computer program 603, it implements the steps in the above method embodiments; or, when the processor 601 executes the computer program 603, it implements the functions of each module / unit in the above device embodiments; For example, computer program 603 can be divided into one or more modules / units, one or more modules / units are stored in memory 602 and executed by processor 601 to complete the present invention; one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which are used to describe the execution process of computer program 603 in electronic device 600.

[0147] The processor 601 may be a central processing unit, or it may be other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.; the general-purpose processor may be a microprocessor or any conventional processor, etc. The memory 602 can be an internal storage unit of the electronic device 600, such as a hard disk or memory of the electronic device 600; the memory 602 can also be an external storage device of the electronic device 600, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the electronic device 600; furthermore, the memory 602 can include both internal storage units and external storage devices of the electronic device 600. The memory 602 is used to store computer programs and other programs and data required by the terminal; the memory 602 can also be used to temporarily store data that has been output or will be output.

[0148] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A highway intelligent decision-making method integrating knowledge base and large model reasoning, characterized in that, include: Acquire multi-source sensing data when highway incidents occur; Based on the multi-source sensing data, feature extraction is performed to obtain the feature vector of each event; Based on the feature vector of each event and the preset knowledge base, a three-level hierarchical retrieval is performed: knowledge graph structure retrieval, vector semantic retrieval, and graph reordering, to obtain the decision context information of each event. Based on the feature vector and decision context information of each event, the large language model is used for analysis and reasoning to generate multiple candidate solutions for each event. Based on multiple candidate solutions for each event, a multi-dimensional quantitative evaluation and global conflict resolution are conducted to generate a collaborative solution for highway events.

2. The intelligent decision-making method for highways integrating knowledge base and large-scale model reasoning according to claim 1, characterized in that, The multi-source sensing data includes video surveillance data, radar detection data, traffic flow data, event alarm data, and meteorological data; The feature extraction based on the multi-source sensing data to obtain the feature vector for each event includes: The video surveillance data is subjected to target detection to identify vehicles, pedestrians, debris, and maintenance and construction areas in the video surveillance data, and a first event detection result is generated; the first event detection result includes target type, target location, target quantity, detection time, and first confidence level; The radar detection data is used to track targets, extract the target vehicle's position, speed, acceleration, and trajectory, determine whether there is abnormal driving behavior, and generate a second event detection result; the second event detection result includes an abnormal target vehicle identifier, position, speed, acceleration, abnormal behavior type, and second confidence level; The event alarm data is parsed to extract the alarm event type, alarm location, and alarm time, generating a third event detection result; the third event detection result includes the alarm event type, alarm location, alarm time, alarm source, and third confidence level; The detection results of the first event, the second event, and the third event are cross-validated to generate a complete perception record; the complete perception record includes the event type, event location, affected lane, affected range, detection source, and overall confidence level. The traffic flow data is analyzed to extract lane-level traffic flow, time occupancy, space occupancy, and average vehicle speed upstream and downstream of the event location. The traffic flow abrupt change coefficient is calculated, which is used to characterize the impact of the event on the traffic flow. The meteorological data is analyzed to extract weather phenomena, visibility values, road surface temperature, road surface humidity, and wind force at the location of the event. Each meteorological parameter is then classified, quantified, and coded to generate meteorological impact factors. These meteorological impact factors are used to characterize the degree to which meteorological conditions constrain event handling. The complete perception record, the traffic flow mutation coefficient, the meteorological influence factor, and the GPS station location information are fused to generate a feature vector for each event.

3. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 1, characterized in that, The process involves a three-tiered retrieval system based on the feature vector of each event and a pre-defined knowledge base. This system includes knowledge graph structure retrieval, vector semantic retrieval, and graph reordering to obtain the decision context information for each event, including: Based on the event type and event location in the feature vector of each event, a breadth-first traversal is performed in the preset knowledge base, starting from each event node, to obtain knowledge nodes and reasoning paths that have a direct causal relationship, temporal relationship or hierarchical relationship with each event; Using the feature vector of each event as the query vector, semantic similarity retrieval is performed in the preset knowledge base. The similarity score between the feature vector and each knowledge vector in the preset knowledge base is calculated, and the similarity scores are sorted from high to low to obtain the knowledge fragments that are semantically related to each event. Using the structured information of the knowledge graph in the preset knowledge base, the semantically related knowledge fragments of each event are reordered to generate a reordered knowledge set; The reordered knowledge set is weighted and fused according to knowledge type, similarity score, and source reliability to construct decision context information for each event; the knowledge types include emergency plans, historical cases, regulatory clauses, resource information, and contact information.

4. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 3, characterized in that, Before obtaining the decision context information for each event by performing a three-level hierarchical retrieval based on the feature vector of each event and a preset knowledge base—knowledge graph structure retrieval, vector semantic retrieval, and graph reordering—the process also includes: The system collects multi-source knowledge resources required for highway decision-making, including structured knowledge, semi-structured knowledge, and unstructured knowledge. The structured knowledge includes an emergency response plan database, a traffic control standard database, and an organizational and contact information database. The semi-structured knowledge includes a historical event handling case database. The unstructured knowledge includes policy and regulatory documents and technical standards and specifications. The multi-source knowledge resources are transformed into vectors to obtain a unified vector knowledge base; Based on the multi-source knowledge resources, entities and relationships between entities are extracted; the entities include event types, handling measures, equipment resources, organizational structures and personnel roles, and the relationships between entities include causal relationships, temporal relationships and hierarchical relationships. Based on the entities and the relationships between them, a knowledge graph for the transportation domain is constructed; Based on the unified vector knowledge base and the transportation domain knowledge graph, a mapping relationship is established to obtain the preset knowledge base.

5. The intelligent decision-making method for highways integrating knowledge base and large-scale model reasoning according to claim 1, characterized in that, Based on the feature vector and decision context information of each event, a large model is used for analysis and reasoning to generate multiple candidate handling solutions for each event, including: Based on the feature vector of each event, the event type, number of affected lanes, length of affected area, traffic flow mutation coefficient, and meteorological influence factor are obtained through analysis. Based on the event type, number of affected lanes, length of affected area, traffic flow mutation coefficient and meteorological influence factor, a preset highway event classification standard is matched to determine the severity level, response level and resource allocation constraints of each event. Based on the complete perception record of each event, traffic flow mutation coefficient and meteorological impact factors, as well as emergency plans, historical cases, control standards, available resources and contact information in the decision context information, a reasoning prompt sequence is generated; Based on the inference prompt sequence and the large language model, by setting multiple sets of differentiated target weight parameters, the large model is guided to generate multiple initial disposal schemes in parallel along different optimization directions; the optimization directions include prioritizing disposal timeliness, minimizing the impact on road traffic, and minimizing disposal resource consumption; Based on the normative clauses and emergency plans in the decision-making context information, the compliance of each initial handling plan is verified and the content is supplemented to obtain multiple candidate handling plans for each event.

6. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 1, characterized in that, Before generating multiple candidate solutions for each event through analysis and reasoning using a large model based on the feature vector and decision context information of each event, the process also includes: Collect historical incident handling reports, emergency drill records, traffic control cases, and various emergency plan texts for highways; Based on the aforementioned historical event handling reports, emergency drill records, traffic control cases, and various emergency plan texts, a fine-tuning corpus is generated. Event summaries, environmental and resource constraints, disposal target preferences, and disposal schemes are extracted from the fine-tuning corpus to construct an instruction fine-tuning dataset. The input of each fine-tuning sample in the instruction fine-tuning dataset includes event perception information, decision context information, and target weight vector. The target weight vector is used to identify the dominant preferences among disposal timeliness priority, road traffic impact minimum, and disposal resource consumption minimum. The output is a disposal scheme that conforms to each dominant preference. Using the instruction fine-tuning dataset, supervised fine-tuning is performed using a pre-trained general large language model as the base model to obtain the initial decision model; Traffic management experts are collected to rank the initial decision-making model's generated solutions in typical event scenarios using multi-dimensional preference ranking data. This multi-dimensional preference ranking data includes preference ranking data for each dimension: handling timeliness, traffic impact, and resource consumption. Based on the multi-dimensional preference ranking data, the initial decision model is optimized using a preference alignment algorithm to obtain the large language model.

7. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 1, characterized in that, The process involves multi-dimensional quantitative evaluation and global conflict resolution of multiple candidate handling schemes for each event to generate a collaborative handling scheme for highway events, including: For each event, multiple candidate solutions are quantitatively evaluated from six dimensions: timeliness, safety, resource consumption, traffic impact, feasibility, and historical effects, to obtain the evaluation score of each candidate solution in each evaluation dimension. Based on the evaluation scores of each candidate solution in each evaluation dimension and the preset weight coefficients, the comprehensive score of each candidate solution is calculated. The candidate solutions are sorted from highest to lowest based on the comprehensive scores, and the top-ranked solution is selected as the preferred solution for each event. Conflict detection is performed on the preferred solutions for each event to obtain identification results. The identification results include whether there is a resource conflict where the same resource is requested by multiple events at the same time, an information conflict where the same information board is required to publish different information at the same time, and a spatiotemporal conflict where the control areas of different events overlap or the control times intersect. Based on the identification results and the priority of each event, conflict resolution is carried out to obtain a collaborative handling plan for highway events.

8. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 1, characterized in that, After generating a collaborative handling plan for highway events by performing multi-dimensional quantitative evaluation and global conflict resolution on multiple candidate handling schemes for each event, the plan further includes: During the execution of the collaborative response plan, actual execution data is continuously collected. The actual execution data includes the actual execution time of each response measure, the actual resource consumption, the actual traffic flow recovery time, and whether secondary events occur. The actual execution data is compared item by item with the expected data in the collaborative handling plan to calculate the execution deviation of each handling measure; If the execution deviation exceeds a preset deviation threshold, then based on the causal and temporal relationships stored in the preset knowledge base, the affected measures affected by the execution deviation are identified, and the identification result is obtained. Based on the identification results, the start time and resource allocation scheme of the affected measures are dynamically adjusted; Based on the dynamically adjusted start time and resource allocation scheme of the affected measures, a collaborative handling scheme is regenerated.

9. The intelligent decision-making method for highways integrating knowledge base and large model reasoning according to claim 1, characterized in that, Before generating multiple candidate solutions for each event by analyzing and reasoning through a large language model based on the feature vector and decision context information of each event, the process also includes: Based on the event type and impact range length in the feature vector of each event, as well as the emergency plan level in the decision context information, the urgency level and confidence score of each event are determined; the confidence score is determined by weighting the comprehensive confidence in the complete perception record and the average similarity score of the knowledge fragments returned by the three-level hierarchical retrieval. Based on the urgency level and the confidence score, a target reasoning pattern is selected from multiple candidate reasoning patterns; the candidate reasoning patterns include a lightweight reasoning pattern, a fast reasoning pattern, a standard reasoning pattern, and a deep reasoning pattern. If the target reasoning mode is the lightweight reasoning mode, then the lightweight instance of the large language model is called to perform analysis and reasoning. The lightweight instance skips the target recognition and constraint recognition steps and directly generates candidate disposal solutions based on the emergency plan in the decision context information. If the target reasoning mode is fast reasoning mode, then the standard instance of the large language model is called to perform analysis and reasoning, and candidate solutions are forcibly output within the preset time threshold. If the target reasoning mode is the standard reasoning mode, then the standard instance of the large language model is called to fully execute the analysis and reasoning process and generate multiple candidate solutions for each event. If the target reasoning mode is deep reasoning mode, the standard instance of the large language model is called to fully execute the analysis and reasoning process, generate multiple candidate solutions for each event, and push the candidate solutions to the manual review end for confirmation before the solution is output.

10. A highway intelligent decision-making system integrating knowledge base and large-scale model reasoning, characterized in that, The intelligent decision-making system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 9.

Images