A multi-scene emergency disposal decision system in transportation safety management

Abstract

The application relates to a multi-scene emergency disposal decision system in traffic safety management. The system comprises a dynamic decision library module, an accident perception positioning module, an intelligent decision matching module and a multi-department linkage scheduling module. The system can analyze historical accident cases through structuring, build a knowledge graph and a simulation model, generate a dynamic fusion decision knowledge base, obtain multi-source heterogeneous perception data in real time, perform abnormal analysis and directional analysis, generate structured accident confirmation information, retrieve matching cases and strategy packages in the knowledge base, generate and optimize disposal schemes, distribute the deconstructed disposal tasks to emergency departments, monitor the execution process in real time, generate a disposal complete data package and update the knowledge base. The method can significantly improve the efficiency and accuracy of emergency disposal in traffic safety management by building a dynamic fusion decision knowledge base and combining real-time data-driven simulation optimization, and can enhance multi-scene adaptability and multi-department collaboration ability.

Description

Technical Field

[0001] This invention belongs to the field of transportation safety management technology, and in particular relates to a multi-scenario emergency response decision-making system for transportation safety management. Background Technology

[0002] In the field of transportation safety management technology, the efficiency and accuracy of emergency response to road transport accidents are directly related to the safety of people's lives, the control of property losses, and the speed of restoring traffic order. However, current emergency response technologies and systems have several shortcomings that urgently need to be addressed: First, existing emergency systems mostly focus on single accident types such as ordinary vehicle collisions, lacking targeted designs for special scenarios such as hazardous chemical transport leaks, brake failure of vehicles on long downhill slopes, and rollovers on mountainous roads, making it difficult to provide customized response solutions and limiting the scope of scenarios covered. Second, in terms of accident response, existing emergency systems rely excessively on manual reporting of accident locations and types, which not only leads to large positioning errors but also causes delays in information transmission, resulting in lagging dispatch of rescue forces and seriously affecting the efficiency of emergency response. Third, the formulation of response plans relies heavily on the experience and judgment of on-site personnel, lacking the scientific support of "historical case reference + data simulation verification," which easily leads to poor adaptability of plans to actual accident scenarios and an inability to effectively reduce the harm of accidents. Fourth, for scenarios requiring multi-departmental collaboration, such as hazardous chemical leaks, existing emergency systems lack automated cross-departmental information push and resource dispatch mechanisms, resulting in lagging coordination among emergency rescue stations, 122 command centers, and other relevant departments, and weak secondary risk prevention and control capabilities. Summary of the Invention

[0003] Based on this, it is necessary to provide a multi-scenario emergency response decision-making system for transportation safety management to address the aforementioned technical issues. This system aims to improve the efficiency and accuracy of emergency response, enhance adaptability to multiple scenarios, improve the speed of emergency response, increase the scientific nature of solutions, and enhance multi-departmental collaboration capabilities.

[0004] Firstly, this application provides a multi-scenario emergency response decision-making system for transportation safety management, including:

[0005] The dynamic decision-making database module is used to obtain structured feature vectors by performing structured analysis on historical accident cases; construct a historical case knowledge graph based on the structured feature vectors; establish parameterized simulation models based on multiple typical accident scenarios, and generate optimized disposal strategy packages through multi-objective optimization calculations using the parameterized simulation models; and fuse and associate the historical case knowledge graph with the optimized disposal strategy packages to obtain a dynamic fusion decision-making knowledge base.

[0006] The accident perception and localization module is used to acquire vehicle-mounted data streams, roadside data streams, and environmental data streams in real time to obtain multi-source heterogeneous perception data streams. Based on high-frequency feature templates extracted from the dynamic fusion decision knowledge base, it performs anomaly analysis on the multi-source heterogeneous perception data streams to generate initial accident hypotheses. Based on the initial accident hypotheses, it retrieves corresponding typical evidence patterns from the dynamic fusion decision knowledge base to perform targeted analysis on the multi-source heterogeneous perception data streams to obtain targeted analysis results. Based on the targeted analysis results, it generates structured accident confirmation information containing accident type, location, and initial condition labels.

[0007] The intelligent decision matching module is used to retrieve matching historical cases, actual handling plans, and matching optimized handling strategy packages based on the initial condition tags in the structured accident confirmation information in the dynamic fusion decision knowledge base; it integrates the recommended actions in the actual handling plan and the matching optimized handling strategy package to generate a preliminary handling plan draft; and it uses real-time environmental data to drive a preset digital twin simulation model to deduce and optimize the preliminary handling plan draft to generate the final recommended handling plan.

[0008] The multi-departmental collaborative scheduling module is used to deconstruct the final recommended disposal plan into atomic task sequences; distribute the atomic task sequences to the corresponding emergency execution departments, monitor multi-dimensional data in real time during task execution, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain analysis results; and update the dynamic fusion decision knowledge base in a targeted manner based on the analysis results.

[0009] In one embodiment, the dynamic decision base module includes a graph construction subunit and a knowledge base construction subunit. The graph construction subunit is used for:

[0010] Based on historical road transport accident records, each accident is analyzed in a structured manner to extract accident scenario type, spatiotemporal coordinates, environmental conditions, sequence of response actions, list of resources used, and final response effect indicators, resulting in structured data for each accident. Based on the structured data of each accident, the causal dependencies between features, the correspondence between scenarios and response actions, and the correlation between resource consumption and effect are analyzed to obtain feature interrelationships. The structured data of each accident and the feature interrelationships are integrated to obtain a structured feature vector.

[0011] The structured feature vectors and their interrelationships are stored in a pre-defined graph database. Synonymous relationships between entities are identified through a semantic mapping symmetric discriminant function, and implicit logical chains are obtained by combining a causal scoring function, thus constructing a historical case knowledge graph.

[0012] The knowledge base construction subunit is used for:

[0013] For multiple typical accident scenarios, a parametric simulation model is constructed. The input of the parametric simulation model is the initial parameters of the accident and the environmental parameters, and the output is the accident development trend and the prediction data of the effects of different handling actions.

[0014] The non-dominated sorting genetic algorithm II is used to perform multi-objective optimization calculations on the preset combination of accident initial conditions and response actions. The optimization objectives are set as rescue efficiency, loss control and resource consumption. The termination criteria are that the consolidation ratio is greater than the preset ratio or the optimization performance indicator shows an inflection point. Pareto optimal optimization response strategy package is generated.

[0015] By linking the historical case knowledge graph with the optimization strategy package, the corresponding optimization strategy references are labeled for the cases in the historical case knowledge graph, and similar historical cases are linked for the optimization strategy package, thus establishing a multi-dimensional index structure and obtaining a dynamic fusion decision knowledge base.

[0016] In one embodiment, the accident perception and positioning module includes a real-time data acquisition subunit and a real-time data analysis subunit. The real-time data acquisition subunit is used to sequentially perform noise reduction and format standardization processing on the real-time acquired vehicle data stream, roadside data stream and environmental data stream to obtain a multi-source heterogeneous perception data stream.

[0017] The real-time data analysis subunit is used for:

[0018] The frequency of abnormal features is counted from the historical case knowledge graph of the dynamic fusion decision knowledge base, and high-frequency feature combinations are extracted through cluster analysis to form high-frequency feature templates;

[0019] The multi-source heterogeneous sensing data stream is input into a machine learning classifier built based on high-frequency feature templates to perform anomaly detection and generate one or more initial accident hypotheses with confidence.

[0020] For each initial hypothesis of an accident, key feature combinations corresponding to the accident type are retrieved from the dynamic fusion decision knowledge base as typical evidence patterns. Multi-source data cross-validation is performed in combination with environmental data streams to obtain targeted analysis results.

[0021] When the matching degree between the directional analysis results and the typical evidence pattern reaches a preset threshold, the occurrence of the accident is confirmed. Combining the directional analysis results and the typical evidence pattern, structured accident confirmation information containing accident type, precise spatiotemporal location, and initial condition labels is generated.

[0022] In one embodiment, the intelligent decision matching module includes a preliminary solution generation subunit and a solution update subunit. The preliminary solution generation subunit is used for:

[0023] Using initial condition tags as the core retrieval terms, the cosine similarity algorithm is used to perform semantic matching on multiple historical cases in the dynamic fusion decision knowledge base. Historical cases with similarity higher than the preset similarity threshold are used as matching historical cases, and the actual handling solutions corresponding to the matching historical cases and matching optimization handling strategy packages adapted to the current accident scenario are obtained.

[0024] The weighted fusion algorithm is used to analyze the success factors and shortcomings of actual disposal solutions, and the case analysis results are obtained. Combined with the recommended actions in the matching and optimization disposal strategy package, the solutions are integrated and matched through conflict resolution rules to generate a preliminary disposal plan draft that includes a resource scheduling list, action sequence and spatial control scope.

[0025] The scheme updates the sub-unit for:

[0026] The preliminary draft of the disposal plan and the real-time environmental data stream are input into a preset digital twin simulation model to simulate the execution process and obtain predictive indicators including rescue time, resource utilization rate and risk control effect. The comprehensive utility value is then calculated based on the predictive indicators.

[0027] Based on predictive indicators and comprehensive utility values, sensitivity analysis is conducted on the resource allocation parameters, action timing parameters, and spatial control parameters in the preliminary draft disposal plan to determine key control parameters. Multiple rounds of gradient fine-tuning are performed on the key control parameters according to a preset step size. After each round of fine-tuning, the adjusted plan is re-input into a digital twin simulation model for simulation, and the comprehensive utility value for the corresponding round is calculated. By comparing the comprehensive utility values ​​of each round, the parameter combination corresponding to the highest comprehensive utility value is selected to generate the final recommended disposal plan.

[0028] In one embodiment, the multi-department coordinated scheduling module includes a task execution subunit and a task feedback update subunit. The task execution subunit is used for:

[0029] Based on the division of responsibilities among functional departments and the dependency relationship between tasks, the final recommended disposal plan is broken down into independent task units with implementing entities, resource requirements, start and end times, and acceptance standards, forming an atomic task sequence;

[0030] The atomic task sequence is pushed to the command system of the corresponding emergency execution department through the emergency response platform, and the task commitment receipt returned by the command system is received and verified.

[0031] The task feedback update subunit is used for:

[0032] Real-time collection of multi-dimensional information, including task execution progress, resource availability, changes in the on-site environment, accident situation evolution data, and actual execution details of response actions, yields actual execution records. This data is then integrated with structured accident confirmation information, the final recommended response plan, and the actual execution records to generate a complete response data package.

[0033] The actual effect indicators, including casualty control rate, property loss reduction rate, rescue response time and traffic restoration efficiency, are calculated from the complete data package of the disposal. The actual effect indicators are compared with the predicted effect indicators of the final recommended disposal plan to calculate the deviation rate. The causes of the deviation are analyzed in combination with the actual execution records to obtain the analysis results.

[0034] The entire data set of disposal cases is updated into the historical case knowledge graph as a new case. The parameterized simulation model is adjusted based on the analysis results. The optimized disposal strategy package corresponding to the new case is generated based on the updated parameterized simulation model. The new case is then stored in the historical case knowledge graph to obtain the updated dynamic fusion decision knowledge base.

[0035] In one embodiment, the formula for calculating the overall utility value is as follows:

[0036]

[0037] in, This is the overall utility value, ranging from [0, 10]. These correspond to the rescue efficiency target, loss control target, and resource consumption target, respectively. For the first The dynamic weight of the project target, and , This refers to the timeline following the accident; Norm For the first Project Target Indicators The normalized value; For the first Project target and the first The coupling coefficient of the project target, with a value range of [-0.5, 0.5].

[0038] Secondly, this application also provides a multi-scenario emergency response decision-making method in transportation safety management, including:

[0039] Step S1: Obtain structured feature vectors by performing structured analysis on historical accident cases; construct a historical case knowledge graph based on the structured feature vectors; establish a parametric simulation model based on multiple typical accident scenarios, and generate an optimized handling strategy package through multi-objective optimization calculation using the parametric simulation model; integrate and associate the historical case knowledge graph with the optimized handling strategy package to obtain a dynamic fusion decision knowledge base.

[0040] Step S2: Real-time acquisition of vehicle data stream, roadside data stream, and environmental data stream to obtain a multi-source heterogeneous sensing data stream; based on high-frequency feature templates extracted from the dynamic fusion decision knowledge base, anomaly analysis is performed on the multi-source heterogeneous sensing data stream to generate an initial accident hypothesis; based on the initial accident hypothesis, corresponding typical evidence patterns are retrieved from the dynamic fusion decision knowledge base to perform targeted analysis on the multi-source heterogeneous sensing data stream to obtain targeted analysis results; based on the targeted analysis results, structured accident confirmation information containing accident type, location, and initial condition labels is generated.

[0041] Step S3: Based on the initial condition tags in the structured accident confirmation information, search the dynamic fusion decision knowledge base to obtain matching historical cases, actual handling plans, and matching optimized handling strategy packages; integrate the recommended actions in the actual handling plans and matching optimized handling strategy packages to generate a preliminary handling plan draft; use real-time environmental data to drive the preset digital twin simulation model to deduce and optimize the preliminary handling plan draft to generate the final recommended handling plan.

[0042] Step S4: Deconstruct the final recommended disposal plan into a task sequence; distribute the atom task sequence to the corresponding emergency execution department, monitor the multi-dimensional data during the task execution process in real time, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain the analysis results; based on the analysis results, update the dynamic fusion decision knowledge base in a targeted manner.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0044] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0045] The aforementioned multi-scenario emergency response decision-making system for transportation safety management achieves intelligent handling of road transport accidents throughout the entire process, from perception and identification, plan generation, collaborative scheduling to knowledge base optimization, through the coordinated operation of a dynamic decision base module, an accident perception and location module, an intelligent decision matching module, and a multi-departmental linkage and dispatch module. Specifically, the dynamic decision base module constructs a dynamic fusion decision knowledge base by combining historical case analysis with simulation models, providing scientific data support for decision-making and solving the problem of traditional solutions lacking experience and simulation verification; the accident perception and location module integrates multi-source data streams and achieves accurate accident identification and location based on knowledge base feature templates and evidence patterns, overcoming the shortcomings of manual reporting, such as delays and large errors; the intelligent decision matching module integrates historical plans and simulation strategies and optimizes them through deduction to generate customized handling plans, improving plan adaptability and avoiding improper handling caused by reliance on experience; the multi-departmental linkage and dispatch module realizes task distribution, process monitoring, and closed-loop updates of the knowledge base, strengthening cross-departmental collaboration efficiency and solving the pain points of delayed linkage and inability to continuously optimize the system.

[0046] Compared with traditional emergency response systems, this system significantly improves the speed of accident response, the accuracy of response plans, and the efficiency of inter-departmental coordination through technologies such as multi-source data fusion, intelligent algorithm matching, multi-departmental collaboration, and dynamic optimization. It effectively reduces accident losses and secondary risks, and enhances the intelligence and reliability of transportation safety management. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart of a multi-scenario emergency response decision-making method in transportation safety management is provided as an exemplary embodiment of the present invention;

[0049] Figure 2 A flowchart of a method for obtaining structured accident confirmation information is provided as an exemplary embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of a multi-scenario emergency response decision-making system in transportation safety management, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] In one embodiment, such as Figure 1 As shown, a multi-scenario emergency response decision-making system 100 for transportation safety management is provided. This embodiment illustrates the system's application to a terminal, but it is understood that the system can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the system includes:

[0053] The dynamic decision base module 101 is used to obtain structured feature vectors by performing structured analysis on historical accident cases; construct a historical case knowledge graph based on the structured feature vectors; establish a parametric simulation model based on multiple typical accident scenarios, and generate an optimized disposal strategy package by performing multi-objective optimization calculations through the parametric simulation model; and integrate and associate the historical case knowledge graph with the optimized disposal strategy package to obtain a dynamic fusion decision knowledge base.

[0054] Specifically, existing historical accident data stored by multiple departments is mostly in unstructured or semi-structured form, scattered across different information systems such as traffic management, emergency rescue, and insurance, and cannot be directly used for intelligent decision-making. Therefore, a standardized parsing process can be used to extract key features such as accident scenario type, spatiotemporal coordinates, environmental conditions, sequence of response actions, list of resources used, and final response effect indicators to form a structured feature vector, providing standardized entity data for subsequent knowledge graph construction. Based on this structured feature vector, the entity-relationship storage advantages of graph databases can be utilized. Elements such as accidents, response plans, resources, and environment can be treated as entities, and the relationships between elements (such as a specific response action corresponding to an accident, or an environmental condition affecting the response effect) can be treated as edges, forming a visualized relational network to construct a historical case knowledge graph. Compared to traditional relational databases, this approach can more efficiently support multi-dimensional feature-based relational queries and similar case matching, providing historical experience support for subsequent decision-making.

[0055] Specifically, historical cases cannot cover all possible initial conditions and environmental combinations of accidents. Therefore, based on the physical mechanisms of various scenarios, such as computational fluid dynamics of hazardous chemical leaks and multibody dynamics of vehicle rollovers, initial accident parameters and environmental parameters can be used as inputs to establish parametric simulation models. Numerical calculations can simulate the development trend of accidents and the effects of different response actions, ensuring that the model output data accurately reflects the actual accident evolution. Furthermore, multi-objective optimization calculations through parametric simulation models can comprehensively consider core objectives such as rescue efficiency, loss control, and resource consumption, avoiding imbalances caused by a single objective. Optimization algorithms can select the set of response strategies that perform best under multiple objectives, generating an optimized response strategy package that provides theoretically optimal solutions for decision-making. Integrating and associating historical case knowledge graphs with optimized response strategy packages allows each historical case to correspond to a suitable optimized strategy reference, and each optimized strategy to link similar historical cases, forming a unified, dynamic, fusion-based decision knowledge base. This ensures that subsequent decisions can draw on past practical experience while incorporating the scientific basis of simulation simulations, improving decision reliability.

[0056] The accident perception and localization module 102 is used to acquire vehicle data streams, roadside data streams, and environmental data streams in real time to obtain multi-source heterogeneous perception data streams. Based on high-frequency feature templates extracted from the dynamic fusion decision knowledge base, it performs anomaly analysis on the multi-source heterogeneous perception data streams to generate initial accident hypotheses. Based on the initial accident hypotheses, it retrieves corresponding typical evidence patterns from the dynamic fusion decision knowledge base to perform directional analysis on the multi-source heterogeneous perception data streams to obtain directional analysis results. Based on the directional analysis results, it generates structured accident confirmation information containing accident type, location, and initial condition labels.

[0057] Specifically, vehicle-mounted data streams, roadside data streams, and environmental data streams suffer from inconsistent formats and noise interference. Therefore, preprocessing is crucial. This includes denoising to filter sensor errors and environmental interference signals, standardizing data formats and units, and aligning timestamps to ensure data synchronization and the accuracy and effectiveness of subsequent analysis. For example, since anomalous features of similar accidents in historical cases are frequently repeated, these features can be statistically extracted and combined to form templates. This allows for rapid initial screening of potential accidents, reducing false negatives and false negatives. For instance, comparing multi-source heterogeneous sensing data streams with templates and using anomaly detection algorithms to generate confident initial accident hypotheses enables rapid accident identification. Because different accident types possess unique key evidence features, typical evidence patterns can be retrieved based on the initial accident hypothesis for targeted analysis. This effectively addresses the uncertainty of the initial hypothesis, allowing for in-depth analysis of core data dimensions. Furthermore, cross-validation with environmental parameters can eliminate interference factors and improve identification accuracy. Finally, structured accident confirmation information containing accident type, location, and initial condition labels can be generated, providing clear and standardized input to the intelligent decision-making matching module.

[0058] The intelligent decision matching module 103 is used to retrieve matching historical cases, actual handling plans, and matching optimized handling strategy packages based on the initial condition tags in the structured accident confirmation information in the dynamic fusion decision knowledge base; it integrates the recommended actions in the actual handling plan and the matching optimized handling strategy package to generate a preliminary handling plan draft; and it uses real-time environmental data to drive a preset digital twin simulation model to deduce and optimize the preliminary handling plan draft to generate the final recommended handling plan.

[0059] Specifically, initial condition tags can accurately characterize the core features of an accident, such as accident type, environmental parameters, and initial state. Therefore, these can be used as core search terms to quickly locate relevant matching historical cases and optimized response strategy packages in the dynamic fusion decision knowledge base, ensuring the targeting and efficiency of the search. The response plans for historical cases are practically feasible, while the optimized response strategy packages are theoretically optimal. Subsequently, by analyzing the success factors and shortcomings of historical plans and combining them with the recommended actions of the optimized strategies, a preliminary plan can be formed through integration and matching. This ensures that the plan conforms to the actual execution scenario and is scientifically reasonable. The plan can include resource scheduling lists, action sequences, and spatial control scope to ensure operability at the execution level. However, this preliminary plan does not fully consider the dynamic changes in the real-time environment. Therefore, digital twin simulation technology can be used to recreate the accident scene, inputting the preliminary plan and real-time environmental data into the model to simulate the entire execution process, outputting predictive indicators such as rescue time, resource utilization, and risk control effectiveness. The effectiveness of the plan can be evaluated by calculating a comprehensive utility value. In addition, the parameters in the preliminary draft of the disposal plan can be fine-tuned and repeatedly simulated. By comparing the comprehensive utility values ​​after multiple rounds of fine-tuning, the parameter combination with the best comprehensive performance can be selected to generate the final recommended disposal plan, ensuring that the plan can be adapted to the current real-time scenario to the greatest extent, reduce accident losses, and improve disposal efficiency.

[0060] The multi-departmental collaborative scheduling module 104 is used to deconstruct the final recommended disposal plan into atomic task sequences; distribute the atomic task sequences to the corresponding emergency execution departments, monitor multi-dimensional data in real time during task execution, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain analysis results; and update the dynamic fusion decision knowledge base in a targeted manner based on the analysis results.

[0061] Specifically, the final recommended response plan cannot be directly implemented by each emergency response department. Therefore, it can be broken down into independent atomic tasks according to functional divisions (firefighting, traffic police, medical, environmental protection, etc.) and task dependencies (e.g., traffic control before personnel evacuation). Each task should clearly define the implementing entity, resource requirements, start and end times, and acceptance criteria, forming an atomic task sequence to ensure that each department clearly understands its responsibilities and implementation requirements. The atomic task sequence can then be distributed to the corresponding emergency response departments, while simultaneously receiving task commitment confirmations from these departments. Unconfirmed or unexecuted tasks can be reassigned to avoid delays caused by poor information transmission.

[0062] Furthermore, data on task execution progress, resource availability, changes in the on-site environment, and the evolution of the accident situation can be collected through GPS positioning, on-site sensors, and departmental reports. This data, along with structured accident confirmation information, the final recommended handling plan, and actual execution records, forms a complete handling data package. Comparing and analyzing the actual and predicted performance indicators within this data package identifies deviations and problems in plan execution. By calculating the deviation rate through item-by-item comparisons, the causes of deviations, such as resource scheduling delays or environmental changes exceeding predictions, can be traced using actual execution records, leading to analytical results and clarifying directions for system improvement. Based on these results, the dynamic fusion decision-making knowledge base can be updated in a targeted manner. For example, the handling data package can be added as a new case to the historical case knowledge graph, adaptively adjusting the parametric simulation model to generate and store optimized handling strategy packages corresponding to the new cases. This enables dynamic iteration of the knowledge base, ensuring continuous improvement in decision-making quality and handling efficiency when dealing with similar accidents in the future.

[0063] The system comprises four main modules: The Dynamic Decision Base Module 101, through structured analysis of historical accident cases, construction of a historical case knowledge graph, establishment of parameterized simulation models, and generation of optimized handling strategy packages, forms a dynamic fusion decision knowledge base, providing rich case support and optimized strategy references for subsequent emergency decision-making. The Accident Perception and Location Module 102, by acquiring multi-source heterogeneous perception data streams and analyzing them in conjunction with high-frequency feature templates and typical evidence patterns in the dynamic fusion decision knowledge base, generates structured accident confirmation information, achieving rapid accident identification, accurate location, and clarification of key information. The Intelligent Decision Matching Module 103, based on the structured accident confirmation information retrieval knowledge base, integrates historical solutions and optimized strategies, and generates a final recommended handling plan through simulation and optimization, drawing on historical experience while adapting to real-time scenarios, thus improving the scientific validity and feasibility of the handling plan. The Multi-Department Collaborative Scheduling Module 104 breaks down handling plans, distributes atomic tasks, monitors the execution process, and updates the knowledge base, not only achieving efficient multi-department collaborative handling but also continuously optimizing the system's decision-making capabilities through closed-loop feedback.

[0064] In one embodiment, the dynamic decision base module includes a graph construction subunit and a knowledge base construction subunit. The graph construction subunit is used for:

[0065] Based on historical road transport accident records, each accident is analyzed in a structured manner to extract accident scenario type, spatiotemporal coordinates, environmental conditions, sequence of response actions, list of resources used, and final response effect indicators, resulting in structured data for each accident. Based on the structured data of each accident, the causal dependencies between features, the correspondence between scenarios and response actions, and the correlation between resource consumption and effect are analyzed to obtain feature interrelationships. The structured data of each accident and the feature interrelationships are integrated to obtain a structured feature vector.

[0066] The structured feature vectors and their interrelationships are stored in a pre-defined graph database. Synonymous relationships between entities are identified through a semantic mapping symmetric discriminant function, and implicit logical chains are obtained by combining a causal scoring function, thus constructing a historical case knowledge graph.

[0067] The knowledge base construction subunit is used for:

[0068] For multiple typical accident scenarios, a parametric simulation model is constructed. The input of the parametric simulation model is the initial parameters of the accident and the environmental parameters, and the output is the accident development trend and the prediction data of the effects of different handling actions.

[0069] The non-dominated sorting genetic algorithm II is used to perform multi-objective optimization calculations on the preset combination of accident initial conditions and response actions. The optimization objectives are set as rescue efficiency, loss control and resource consumption. The termination criteria are that the consolidation ratio is greater than the preset ratio or the optimization performance indicator shows an inflection point. Pareto optimal optimization response strategy package is generated.

[0070] By linking the historical case knowledge graph with the optimization strategy package, the corresponding optimization strategy references are labeled for the cases in the historical case knowledge graph, and similar historical cases are linked for the optimization strategy package, thus establishing a multi-dimensional index structure and obtaining a dynamic fusion decision knowledge base.

[0071] Specifically, the core function of the knowledge graph construction subunit is to transform scattered historical accident data into a structured and interconnected knowledge graph, providing experience support for decision-making. For example, a comprehensive structured analysis can be performed on historical road transport accident records. This record data can be obtained from the information systems of multiple related departments such as traffic management, emergency rescue, and insurance, including complete information on various typical accidents within a preset historical time period. The analysis process can employ standardized data field extraction rules to ensure consistency in the extraction criteria for each type of feature, extracting accident scenario type, spatiotemporal coordinates, environmental conditions, response action sequence, resource list, and final response effect indicators, forming structured data for a single accident. Specifically, the accident scenario type is classified and coded according to eight preset typical scenarios; the spatiotemporal coordinates are obtained by extracting the longitude, latitude, and timestamp of the accident; environmental conditions include quantitative indicators such as meteorological parameters, road condition parameters, and terrain parameters; the response action sequence can be obtained by sorting out the key response steps executed in chronological order; the resource list clarifies the type and quantity of rescue teams, equipment, and materials; and the final response effect indicators can be represented by quantifiable data such as the number of casualties, the amount of property damage, and the traffic recovery time.

[0072] Furthermore, based on this, association rule mining algorithms can be used to analyze the structured data of single accidents from multiple incidents. Causal dependencies can be identified by calculating the support and confidence between features. For example, when the co-occurrence support of "heavy rain" and "chain-reaction rear-end collisions" is higher than a set threshold, a causal relationship is determined. The correspondence between scenarios and actions can be determined by statistically analyzing frequently occurring combinations of response actions in similar scenarios. Furthermore, correlation analysis can be used to establish a mapping relationship between resource consumption and effectiveness indicators, such as quantifying the negative correlation between the number of rescue equipment deployed and the accident response time, thereby obtaining the interrelationships between features. Integrating the structured data of single accidents with these interrelationships results in a structured feature vector for each accident, containing feature attributes and association weights. This ensures that the vector comprehensively represents the core attributes and inherent connections of the accident.

[0073] Specifically, structured feature vectors and their interrelationships can be stored in a pre-defined graph database. For example, during storage, elements such as accidents, environments, resources, and response actions within the structured feature vectors are treated as independent entities. Each entity is assigned a unique identifier, and the interrelationships between features are represented as edges connecting entities. Each edge is labeled with its relationship type and corresponding weight value. To avoid entity redundancy in the knowledge graph, a semantic mapping symmetric discriminant function can be used to identify synonymous relationships between entities. The core principle of this function is similarity calculation based on entity attribute features, expressed as follows: ,in Representing entities and Synonym similarity, and Representing entities respectively and The system uses attribute sets to calculate the ratio of the intersection to the union of attribute sets of two entities. If the similarity exceeds a set threshold, the two entities are considered synonyms and merged. Simultaneously, a causal scoring function is used to obtain the implicit logical chain. This function quantifies the confidence level of a causal relationship between two entities, and its expression is... ,in This represents the causal confidence level between the cause entity (cause) and the effect entity (effect). This represents the probability that the result entity, effect, occurs given the existence of the cause entity. This represents the probability of the result entity "effect" occurring alone. By traversing entity pairs in the knowledge graph, calculating causal confidence, and filtering out entity pairs with confidence scores higher than a set threshold, an implicit causal logic chain is constructed. This enriches the association dimension of the knowledge graph and ultimately forms a complete and closely related historical case knowledge graph.

[0074] Specifically, the core function of the knowledge base construction sub-unit is to build parameterized simulation models and generate optimized disposal strategy packages. These are then integrated with historical case knowledge graphs to form a dynamic fusion decision-making knowledge base, providing theoretical support for decision-making. For example, for multiple preset typical accident scenarios, such as eight scenarios including hazardous chemical leaks, long downhill brake failure, and rollovers on mountain roads, parameterized simulation models can be built based on the physical characteristics and evolution patterns of each scenario, using a combination of physical mechanisms and data-driven approaches. For instance, in the hazardous chemical leak scenario, a diffusion simulation model can be built based on computational fluid dynamics principles, combined with the physicochemical properties of the leaked substance. Similarly, in the long downhill brake failure scenario, a vehicle trajectory simulation model can be built based on vehicle dynamics principles, considering parameters such as slope and road surface friction coefficient. The inputs to this parametric simulation model are explicitly defined as initial accident parameters and environmental parameters. Initial accident parameters may include vehicle status at the time of the accident and the initial severity of the accident, while environmental parameters may include wind speed, wind direction, terrain slope, and road conditions. The model's outputs are accident development trend data and predicted data on the effectiveness of different response actions. Accident development trend data includes, for example, the curve of the hazardous chemical diffusion range over time, while predicted data on the effectiveness of different response actions includes, for example, the success rate of personnel evacuation under different evacuation radii. By constructing this model, it is possible to ensure that dynamically adapted prediction results are output based on the initial conditions and environmental parameters of different accidents.

[0075] Subsequently, a non-dominated sorting genetic algorithm II can be used to perform multi-objective optimization calculations on the preset combination of initial accident conditions and response actions. This algorithm can effectively solve multi-objective optimization problems and find Pareto optimal solutions among multiple conflicting objectives. The optimization objectives can be rescue efficiency, loss control, and resource consumption. Rescue efficiency is quantified by the arrival time of the rescue team and the completion time of the accident response; loss control is quantified by the number of casualties, the amount of property damage, and the degree of environmental damage; and resource consumption is quantified by the total amount of rescue teams, equipment, and materials invested. The algorithm execution process includes steps such as population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation. During population initialization, each individual corresponds to a combination of initial accident conditions and response actions, and the initial population is constructed through random generation. Non-dominated sorting stratifies individuals in the population according to their dominance relationship, where dominance means that an individual's performance on all objectives is no worse than another individual's and it is better on at least one objective. Crowding calculation quantifies the distance between individuals within the same stratum, avoiding the concentration of optimal solutions in local areas. The population is continuously iterated and evolved through selection, crossover, and mutation operations.

[0076] Furthermore, the algorithm's termination criteria can be set as either a consolidation ratio greater than a preset ratio or the optimization performance indicator showing an inflection point. The consolidation ratio characterizes the convergence degree of the Pareto optimal solution in the population, calculated based on the ratio of the objective function value of the best individual in the population to the initial best individual. When the consolidation ratio exceeds the preset ratio, it indicates that the population has converged. The optimization performance indicator evaluates the quality of the Pareto optimal solution set. When the indicator curve shows an inflection point, it indicates that continued iteration cannot significantly improve the solution set quality, at which point the algorithm can terminate and output a Pareto optimal optimization strategy package. Each strategy package contains specific resource scheduling schemes, traffic control schemes, evacuation schemes, etc.

[0077] Specifically, a multi-source data association mechanism can be employed to extract the core features of cases from the historical case knowledge graph and match them with the applicable conditions of the optimized handling strategy package. Each historical case is labeled with a corresponding optimization strategy reference, and each optimized handling strategy package is linked to similar historical cases, achieving mutual support between empirical data and theoretical optimization data. Furthermore, a multi-dimensional index structure can be established, including indexes for accident types, spatial locations, environmental parameters, and resource requirements. The accident type index is categorized into eight typical scenarios, and the spatial location index is built based on geographic coordinates, supporting rapid retrieval by region. This multi-dimensional index structure ensures that the required historical cases and optimized handling strategy packages can be quickly located from different query perspectives, ultimately forming a dynamically integrated decision-making knowledge base that is highly efficient in retrieval and closely interconnected.

[0078] In one embodiment, the accident perception and localization module includes a real-time data acquisition subunit and a real-time data analysis subunit. The real-time data acquisition subunit is used to sequentially perform noise reduction and format standardization processing on the real-time acquired vehicle data stream, roadside data stream, and environmental data stream to obtain a multi-source heterogeneous perception data stream, as illustrated below. Figure 2 As shown, the real-time data analysis subunit is used to obtain structured accident confirmation information through the following steps:

[0079] S201: Statistically count the frequency of abnormal features from the historical case knowledge graph of the dynamic fusion decision knowledge base, and extract high-frequency feature combinations through cluster analysis to form high-frequency feature templates;

[0080] S202: Input the multi-source heterogeneous sensing data stream into a machine learning classifier built based on high-frequency feature templates to perform anomaly detection and generate one or more initial accident hypotheses with confidence.

[0081] S203: For each initial hypothesis of an accident, the key feature combination of the corresponding accident type is retrieved from the dynamic fusion decision knowledge base as a typical evidence pattern, and multi-source data cross-validation is performed in combination with environmental data stream to obtain targeted analysis results;

[0082] S204: When the matching degree between the directional analysis results and the typical evidence pattern reaches a preset threshold, the accident is confirmed to have occurred. Combining the directional analysis results and the typical evidence pattern, structured accident confirmation information containing accident type, precise spatiotemporal location, and initial condition labels is generated.

[0083] Specifically, the core function of the real-time data acquisition subunit is to preprocess multi-source raw data streams, eliminating data noise and format differences to form a unified and standardized multi-source heterogeneous sensing data stream. For example, real-time vehicle-mounted data streams can be acquired through the vehicle's BeiDou positioning unit and multi-dimensional sensor unit. The BeiDou positioning unit can collect the vehicle's geographical coordinates and timestamp data in real time, while the multi-dimensional sensor unit can include pressure sensors, gas sensors, tilt sensors, etc., to collect status data such as vehicle tank pressure, surrounding gas composition and concentration, and vehicle attitude angle. Roadside data streams can be acquired through roadside monitoring equipment such as surveillance cameras, millimeter-wave radar, and microwave detectors deployed along the road, including data such as traffic flow parameters, video frames of abnormal vehicle behavior, and road conditions. Environmental data streams can be extracted from real-time data released by meteorological and environmental monitoring departments, including environmental parameters such as wind speed, wind direction, temperature, humidity, precipitation, and topography. Since the above three types of data streams come from different devices and systems, there are problems such as inconsistent data formats, noise interference, and time asynchrony. Therefore, they can be preprocessed sequentially. One approach is to use a moving average filtering algorithm for noise reduction. This algorithm replaces abnormal fluctuations in data by calculating the average value of data within a certain window, thus eliminating random noise caused by sensor errors and environmental interference. The core logic of this algorithm is... ,in For the i-th data point after denoising, To adjust the sliding window size, For the k-th raw data point within the window, the smoothness and real-time performance of the data are balanced by appropriately setting the window size. Subsequently, heterogeneous data output from different devices can be converted into a unified data format and unit according to preset data specifications. For example, different units of gas concentration can be uniformly converted to mass concentration units, and geographic coordinates can be uniformly converted to latitude and longitude formats. Simultaneously, time synchronization and alignment can be performed based on the timestamp information in each data stream, ensuring that vehicle-mounted, roadside, and environmental data collected at the same time form a temporal correspondence. Ultimately, a multi-source heterogeneous sensing data stream with unified structure, time synchronization, and noise removal is obtained.

[0084] Specifically, the core function of the real-time data analysis subunit is to perform hierarchical analysis on preprocessed multi-source heterogeneous sensing data streams based on prior knowledge from a dynamic fusion decision knowledge base. This analysis ranges from initial anomaly screening to precise confirmation, generating structured accident confirmation information. Illustratively, firstly, abnormal feature data from all accident cases can be extracted from the historical case knowledge graph of the dynamic fusion decision knowledge base. This data is then categorized and statistically analyzed by accident type, calculating the frequency and co-occurrence probability of each abnormal feature within each accident category. These abnormal features can include quantifiable characteristics such as sudden changes in sensor data, parameter exceeding limits, and abnormal states. Examples include gas sensor readings exceeding safety thresholds, tilt sensor readings changing beyond a set angle, and abnormal drops in braking system pressure data. Subsequently, a density clustering algorithm can be used to cluster the statistically analyzed abnormal features. This algorithm does not require a pre-defined number of clusters and can automatically identify high-density regions in the feature space, grouping frequently co-occurring abnormal features into one category. The core feature combinations of each cluster are then extracted to form high-frequency feature templates corresponding to various accident scenarios. Each high-frequency feature template can contain key information such as feature terms, feature threshold ranges, and correlation weights between features. For example, a high-frequency feature template for a hazardous chemical leak scenario includes the gas concentration exceeding the standard threshold, the tank pressure drop, the corresponding meteorological parameter ranges, and the correlation weights of each feature.

[0085] Specifically, multi-source heterogeneous sensing data streams can be converted into feature vectors and input into a machine learning classifier built based on high-frequency feature templates for anomaly detection. This machine learning classifier can be a random forest classifier, which has advantages such as resistance to overfitting and strong ability to handle high-dimensional data. During its construction, anomaly feature data and normal operation data from historical case knowledge graphs can be used as training samples, with feature items from the high-frequency feature templates as input features and "whether an accident occurred" and "accident type" as output labels. The decision rules of the classifier are determined through training. Subsequently, the classifier can analyze the input feature vectors one by one, calculating the degree of matching between the feature vector and each high-frequency feature template, and outputting the probability value of each possible accident type as a confidence level. When the confidence level is higher than a set preliminary screening threshold, a corresponding initial accident hypothesis can be generated. If the confidence levels of multiple accident types all meet the requirements, multiple initial accident hypotheses with confidence ranking are generated.

[0086] For each initial hypothesis of an accident, key feature combinations corresponding to the accident type can be retrieved from the dynamic fusion decision knowledge base as typical evidence patterns. These typical evidence patterns represent the core feature set that distinguishes a particular type of accident from others. For example, the typical evidence pattern for a long downhill brake failure scenario includes key parameter combinations such as the slope parameter range, vehicle speed change trend, brake system pressure data characteristics, and road surface friction coefficient. Based on these typical evidence patterns, it is possible to focus on the data dimensions related to the core features in the multi-source heterogeneous sensing data stream for in-depth extraction and analysis. For example, for the initial hypothesis of brake failure, the focus is on extracting slope data, vehicle speed sequences, and brake pressure data. Simultaneously, cross-validation of multi-source data can be performed by combining relevant parameters from the environmental data stream. For instance, the extracted slope data can be compared with terrain parameters in the environmental data stream, and the vehicle speed change trend can be compared with speed data collected by roadside radar. Through mutual validation of different data sources, the error interference of single data points can be eliminated, resulting in targeted analysis results.

[0087] Specifically, a cosine similarity algorithm can be used to calculate the matching degree between the directional analysis results and typical evidence patterns. When the matching degree reaches a preset threshold, the occurrence of an accident is confirmed. Subsequently, relevant data can be integrated to generate structured accident confirmation information. That is, by comparing the core features of the directional analysis results with those of typical evidence patterns, the accident type can be determined. For example, if the gas concentration characteristics highly match the typical evidence pattern of a hazardous chemical leak, it is determined to be a hazardous chemical leak accident. Furthermore, the geographic location data collected by the Beidou positioning unit can be fused with the auxiliary positioning data of roadside monitoring equipment to obtain a precise spatiotemporal location, ensuring that the positioning accuracy meets the actual handling requirements. In addition, environmental parameters, vehicle initial state parameters, and key scene parameters at the time of the accident can be extracted and matched with the initial condition classification system in the dynamic fusion decision knowledge base to generate initial condition labels. These labels accurately represent the core initial features of the accident. Finally, by integrating the accident type, precise spatiotemporal location, and initial condition labels, structured and standardized structured accident confirmation information can be formed.

[0088] In one embodiment, the intelligent decision matching module includes a preliminary solution generation subunit and a solution update subunit. The preliminary solution generation subunit is used for:

[0089] Using initial condition tags as the core retrieval terms, the cosine similarity algorithm is used to perform semantic matching on multiple historical cases in the dynamic fusion decision knowledge base. Historical cases with similarity higher than the preset similarity threshold are used as matching historical cases, and the actual handling solutions corresponding to the matching historical cases and matching optimization handling strategy packages adapted to the current accident scenario are obtained.

[0090] The weighted fusion algorithm is used to analyze the success factors and shortcomings of actual disposal solutions, and the case analysis results are obtained. Combined with the recommended actions in the matching and optimization disposal strategy package, the solutions are integrated and matched through conflict resolution rules to generate a preliminary disposal plan draft that includes a resource scheduling list, action sequence and spatial control scope.

[0091] The scheme updates the sub-unit for:

[0092] The preliminary draft of the disposal plan and the real-time environmental data stream are input into a preset digital twin simulation model to simulate the execution process and obtain predictive indicators including rescue time, resource utilization rate and risk control effect. The comprehensive utility value is then calculated based on the predictive indicators.

[0093] Based on predictive indicators and comprehensive utility values, sensitivity analysis is conducted on the resource allocation parameters, action timing parameters, and spatial control parameters in the preliminary draft disposal plan to determine key control parameters. Multiple rounds of gradient fine-tuning are performed on the key control parameters according to a preset step size. After each round of fine-tuning, the adjusted plan is re-input into a digital twin simulation model for simulation, and the comprehensive utility value for the corresponding round is calculated. By comparing the comprehensive utility values ​​of each round, the parameter combination corresponding to the highest comprehensive utility value is selected to generate the final recommended disposal plan.

[0094] Specifically, the core function of the preliminary solution generation subunit is to retrieve and adapt historical experience and theoretical optimization strategies based on the core features of the accident, and then form a preliminary draft of a feasible handling plan through fusion and adaptation. For example, the initial condition tags in the structured accident confirmation information can be used as the core search terms, and the initial condition tags can be transformed into search feature vectors containing dimensions such as accident type, environmental parameters, and initial state. Subsequently, a cosine similarity algorithm can be used to semantically match multiple historical cases stored in the dynamic fusion decision knowledge base, thereby effectively measuring the similarity between two feature vectors without being affected by data dimensions. By calculating the similarity between the current accident and each historical case, historical cases with similarity higher than a preset similarity threshold can be identified as matching historical cases. Simultaneously, the actual handling plans corresponding to these matching historical cases, as well as the matching optimization handling strategy package adapted to the current accident scenario in the dynamic fusion decision knowledge base, are obtained, ensuring that the search results include both empirically verified solutions and theoretically optimal strategy references.

[0095] Furthermore, a weighted fusion algorithm can be used to conduct in-depth analysis of the actual handling plans for matching historical cases. Weights are assigned to each execution stage of the actual handling plan, with the weights determined based on the handling effect indicators of the historical plans; stages with better handling effects have higher weights. Through weighted calculation, the successful execution elements and implementation shortcomings of the actual handling plans can be quantitatively analyzed, yielding case analysis results. Successful execution elements include suitable resource types and reasonable action sequences, while implementation shortcomings include resource scheduling delays and unreasonable control scope. In conjunction with the recommended actions in the matching and optimization handling strategy package, potential resource configuration conflicts and action sequence contradictions between the two types of plans can be addressed through preset conflict resolution rules for plan integration and matching. These conflict resolution rules are based on a priority order of "safety first, efficiency first, and resource optimization." For example, when the resource scheduling of a historical plan conflicts with the recommended resources in the optimization strategy package, the configuration method that meets the core safety objectives and has better resource consumption is prioritized. Ultimately, a preliminary draft of the response plan can be formed, which includes a resource allocation list, an action sequence plan, and a spatial control scope definition. The resource allocation list clarifies the types and quantities of rescue teams, equipment, and supplies, the action sequence plan arranges key response steps in chronological order, and the spatial control scope defines the specific areas for traffic control and personnel evacuation.

[0096] Specifically, the core function of the solution update sub-unit is to improve the adaptability and execution effectiveness of the preliminary response plan draft through digital twin simulation and iterative optimization, generating the final recommended response plan. For example, the preliminary response plan draft and the real-time updated environmental data stream can be input into a preset digital twin simulation model. This model constructs a 1:1 virtual simulation scene based on the geographical information, traffic flow data, and environmental parameters of the accident site, accurately reproducing the accident evolution pattern and the plan execution process. Subsequently, the preliminary response plan draft is input into the model in a parameterized form, including the time nodes of resource scheduling, the spatial coordinates of action execution, and the boundary parameters of the control area. The real-time environmental data stream can include dynamic parameters such as current wind speed, wind direction, traffic flow changes, and terrain details. Through this model, the entire process of plan execution can be simulated and deduced, outputting multiple predictive indicators, including rescue time, resource utilization rate, and risk control effect. Among them, rescue time quantifies the total time from resource deployment to accident handling completion, resource utilization rate quantifies the actual efficiency of resource use, and risk control effect quantifies the degree of accident loss reduction and the probability of secondary accidents. Furthermore, based on these predictive indicators, a pre-defined formula can be used to calculate the overall utility value, thereby comprehensively evaluating the overall implementation effect of the plan. The formula for calculating the overall utility value is as follows:

[0097]

[0098] in, The comprehensive utility value ranges from [0, 10], with a higher value indicating a better performance of the solution. These correspond to the rescue efficiency target, loss control target, and resource consumption target, respectively. For the first The dynamic weight of the project target, and , This is a timeline following an accident, adjusted in real-time based on the accident's spread rate and environmental complexity. For example, initial rescue efficiency is crucial for minimizing losses. The weight given to [the specific event / event] is higher, while post-accident loss control becomes the core issue. Norm has a higher weight; For the first Project Target Indicators The normalized value of the rescue efficiency index The rescue response time is normalized by reciprocal; the shorter the response time, the closer the normalized value is to 1, which is a key indicator for loss control. Loss reduction rate normalization is used; the higher the reduction rate, the closer the normalized value is to 1. (Resource consumption indicator) Resource utilization rate is normalized; the higher the utilization rate, the closer the normalized value is to 1. For the first Project target and the first The coupling coefficient of the project objectives ranges from -0.5 to 0.5. A positive number indicates that there is a synergistic effect between the two objectives, such as improving rescue efficiency can promote the optimization of loss control. A negative number indicates that there is a trade-off effect between the two objectives, such as moderately increasing resource consumption may improve rescue efficiency. This coefficient is trained based on historical case data using the gradient descent method to ensure that it can accurately reflect the mutual influence between different objectives.

[0099] Specifically, based on predictive indicators and comprehensive utility values, the control variable method can be used to conduct sensitivity analysis on resource allocation parameters, action timing parameters, and spatial control parameters in the preliminary draft response plan. For example, while keeping other parameters constant, the values ​​of individual parameters are adjusted one by one, and the magnitude of change in the comprehensive utility value is observed. Parameters whose magnitude of change exceeds a preset threshold are identified as key control parameters. These parameters have a decisive impact on the effectiveness of the plan's execution, such as the departure time of the rescue team and the radius of the evacuation area. Furthermore, multiple rounds of gradient fine-tuning can be performed on the key control parameters according to a preset step size. The preset step size can be determined based on the quantitative attributes of the parameters to ensure that the fine-tuning range covers the potential optimal range without missing the optimal solution due to excessively large step sizes. After each round of fine-tuning, the adjusted plan can be re-input into the digital twin simulation model for simulation, and the comprehensive utility value for the corresponding round can be calculated. By recording the values ​​of the key control parameters and the corresponding comprehensive utility values ​​for each round, a parameter-utility value mapping relationship is formed. By comparing the overall utility values ​​of each round, the parameter combination with the highest overall utility value can be selected. Furthermore, by combining the feasibility verification of the solution, it can be confirmed that the solution under this parameter combination does not have execution conflicts or resource constraints. Finally, a final recommended solution with quantitative predictive performance indicators is generated.

[0100] In one embodiment, the multi-department coordinated scheduling module includes a task execution subunit and a task feedback update subunit. The task execution subunit is used for:

[0101] Based on the division of responsibilities among functional departments and the dependency relationship between tasks, the final recommended disposal plan is broken down into independent task units with implementing entities, resource requirements, start and end times, and acceptance standards, forming an atomic task sequence;

[0102] The atomic task sequence is pushed to the command system of the corresponding emergency execution department through the emergency response platform, and the task commitment receipt returned by the command system is received and verified.

[0103] The task feedback update subunit is used for:

[0104] Real-time collection of multi-dimensional information, including task execution progress, resource availability, changes in the on-site environment, accident situation evolution data, and actual execution details of response actions, yields actual execution records. This data is then integrated with structured accident confirmation information, the final recommended response plan, and the actual execution records to generate a complete response data package.

[0105] The actual effect indicators, including casualty control rate, property loss reduction rate, rescue response time and traffic restoration efficiency, are calculated from the complete data package of the disposal. The actual effect indicators are compared with the predicted effect indicators of the final recommended disposal plan to calculate the deviation rate. The causes of the deviation are analyzed in combination with the actual execution records to obtain the analysis results.

[0106] The entire data set of disposal cases is updated into the historical case knowledge graph as a new case. The parameterized simulation model is adjusted based on the analysis results. The optimized disposal strategy package corresponding to the new case is generated based on the updated parameterized simulation model. The new case is then stored in the historical case knowledge graph to obtain the updated dynamic fusion decision knowledge base.

[0107] Specifically, the core function of the task execution sub-unit is to transform the abstract final recommended disposal plan into specific tasks that each department can directly execute, and to complete the task distribution and confirmation, thus ensuring the implementation of the plan. For example, the final recommended disposal plan can be systematically decomposed according to the division of responsibilities among functional departments and task dependencies. The division of responsibilities among functional departments can be determined based on the core responsibilities and capability boundaries of each department, including emergency rescue stations, 122 command centers, medical emergency departments, environmental protection departments, traffic control departments, etc., ensuring that each task is assigned to a department with the corresponding disposal qualifications. Task dependencies can be determined by sorting out the logical order of disposal actions in the plan. For example, traffic control tasks must be completed first, followed by personnel evacuation tasks, to avoid disposal conflicts caused by reversing the task order. Furthermore, a task decomposition algorithm can be used in the decomposition process to first identify the core disposal actions in the final recommended disposal plan, and then decompose each action into independent task units containing the executing entity, resource requirements, start and end times, and acceptance criteria, forming an atomic task sequence. The implementation entity clearly identifies the specific responsible department, and the resource requirements detail the types of rescue teams, equipment models, types and quantities of materials needed to complete the task. The start and end times can be set according to the timeliness requirements of the rescue and the dependency relationship of the task to ensure that the critical task is completed within the golden handling time. The acceptance criteria can adopt quantifiable indicators. For example, the acceptance criteria for traffic control tasks is "the traffic density of the designated road section drops below the preset threshold", and the acceptance criteria for personnel evacuation tasks is "the personnel evacuation completion rate of the target area reaches 100%".

[0108] Subsequently, the atomic task sequence can be pushed to the command system of the corresponding emergency execution department through the emergency response platform. This platform has cross-departmental interface adaptation capabilities and can convert the atomic task sequence into the corresponding format for transmission based on the existing interface protocols of each department's command system (such as message queues, database interfaces, etc.), ensuring that the task package can be correctly parsed by each department's system. After the push is completed, the system can activate the receipt reception mechanism, set a preset waiting time, and receive task commitment receipts returned by each department's command system. Furthermore, the validity of the receipts can be verified to confirm whether the receipts clearly respond to the intention to execute the task, whether resource supplementation requests are made, and whether the task start and end times are adjusted, etc. If no receipt is received within the preset waiting time, or if the receipt shows that the department cannot execute the task, the system can automatically initiate the task reallocation process, adjusting the task execution subject based on the capabilities and locations of the remaining available departments, ensuring that all atomic tasks are undertaken by the corresponding departments.

[0109] Specifically, the core function of the task feedback update subunit is to monitor the entire task execution process in real time, collect and analyze data, and update the dynamic fusion decision knowledge base based on the analysis results, forming a closed-loop optimization mechanism for the system. For example, a multi-dimensional information real-time collection channel can be established to obtain comprehensive data during task execution through various technical means. For instance, task execution progress can be obtained through the task status (e.g., preparing, executing, completed, blocked) reported regularly by the command systems of various departments, with the reporting frequency dynamically adjusted according to the urgency of the task; resource availability can be tracked in real time through GPS / BeiDou positioning modules on rescue vehicles and equipment to track the location and arrival status of resources; changes in the on-site environment can be collected through sensors deployed at the accident site (e.g., gas sensors, meteorological sensors, video surveillance equipment), including data on changes in the concentration of leaked substances, wind speed and direction adjustments, and changes in road conditions; accident situation evolution data can be updated in real time by integrating information reported by roadside monitoring equipment and on-site personnel to update the scope and severity of the accident's impact; and the actual execution details of the handling actions can be collected through the execution record systems of various departments, including specific operating steps, execution time, and participating personnel. By integrating the above multi-dimensional information, an actual execution record can be formed. Furthermore, structured accident confirmation information and the final recommended handling plan can be merged, and the data can be standardized and labeled to generate a complete handling data package covering the entire accident process, providing complete data support for subsequent effect evaluation and knowledge base updates.

[0110] Furthermore, key data can be extracted from the entire dataset to calculate actual performance indicators, such as casualty control rate, property damage reduction rate, rescue response time, and traffic restoration efficiency. The formula for calculating the casualty control rate is as follows: ,in For the casualty control rate, The predicted number of casualties in the final recommended response plan, The actual number of casualties represents the effectiveness of the control scheme in reducing casualties. The value ranges from [0,1], with values ​​closer to 1 indicating better control. The property loss reduction rate can be calculated as (predicted property loss amount - actual property loss amount) / predicted property loss amount. The rescue response time is the time interval from task allocation to the arrival of the first rescue team at the accident site, while the traffic restoration efficiency is the normalized reciprocal of the traffic interruption time. By comparing these actual effectiveness indicators with the predicted effectiveness indicators in the final recommended disposal plan, the deviation rate can be calculated. This indicator quantifies the degree of deviation between the actual and predicted effects. Further analysis of the causes of deviation can be conducted using actual execution records. If the deviation stems from resource scheduling delays, it can be traced back to inaccurate resource positioning or traffic congestion. If the deviation stems from the disposal effect not meeting expectations, it can be analyzed whether the plan parameters do not match the actual scenario or whether the execution operation is flawed. The final result includes quantitative deviation data and causal analysis.

[0111] Specifically, the complete case study data can be used as a new case, supplementing the historical case knowledge graph according to its entity-relationship structure. The entities in the new case include the scene characteristics of the accident, the handling tasks, participating departments, resource allocation, and actual effects. Relationships between entities include accident-task, task-department, and resource-effect relationships, enriching the coverage and association dimensions of the knowledge graph. Based on the analysis results, the parametric simulation model can be adjusted. If the actual accident evolution trend deviates from the model prediction—for example, the actual spread of hazardous chemicals exceeds the predicted range—the corresponding spread coefficient and meteorological impact weights in the model can be adjusted. If the actual effect of the handling actions does not match the model simulation results, the effect quantification function of the handling actions in the model can be optimized. Through the adjusted parametric simulation model, multi-objective optimization calculations can be performed on the accident scenario corresponding to the new case, generating a dedicated optimized handling strategy package for the new case. This strategy package integrates the successful experience and improvement directions of the current handling, and after establishing an association annotation with the new case, it can be synchronously updated to the historical case knowledge graph. By integrating the updated historical case knowledge graph with all optimized handling strategy packages, the multi-dimensional index structure can be optimized to ensure retrieval efficiency. Ultimately, an updated dynamic fusion decision knowledge base can be obtained, enabling continuous improvement of the system's decision-making capabilities.

[0112] Based on the same inventive concept, this application also provides a method for implementing the multi-scenario emergency response decision-making system in transportation safety management as described above. The solution provided by this method is similar to the implementation scheme described in the above system. Therefore, the specific limitations of one or more embodiments of the multi-scenario emergency response decision-making method in transportation safety management provided below can be found in the limitations of the multi-scenario emergency response decision-making system in transportation safety management described above, and will not be repeated here.

[0113] In one exemplary embodiment, such as Figure 3 As shown, a multi-scenario emergency response decision-making method for transportation safety management is provided, including:

[0114] Step S301: By performing structured analysis on historical accident cases, structured feature vectors are obtained; based on the structured feature vectors, a historical case knowledge graph is constructed; a parametric simulation model is established based on multiple typical accident scenarios, and multi-objective optimization calculations are performed through the parametric simulation model to generate an optimized handling strategy package; the historical case knowledge graph and the optimized handling strategy package are fused and associated to obtain a dynamic fusion decision knowledge base;

[0115] Step S302: Real-time acquisition of vehicle data stream, roadside data stream, and environmental data stream to obtain a multi-source heterogeneous sensing data stream; based on high-frequency feature templates extracted from the dynamic fusion decision knowledge base, anomaly analysis is performed on the multi-source heterogeneous sensing data stream to generate an initial accident hypothesis; based on the initial accident hypothesis, corresponding typical evidence patterns are retrieved from the dynamic fusion decision knowledge base to perform targeted analysis on the multi-source heterogeneous sensing data stream to obtain targeted analysis results; based on the targeted analysis results, structured accident confirmation information containing accident type, location, and initial condition labels is generated.

[0116] Step S303: Based on the initial condition tags in the structured accident confirmation information, a search is performed in the dynamic fusion decision knowledge base to obtain matching historical cases, actual handling plans, and matching optimized handling strategy packages; the recommended actions in the actual handling plans and matching optimized handling strategy packages are integrated to generate a preliminary handling plan draft; the preliminary handling plan draft is deduced and optimized using a preset digital twin simulation model driven by real-time environmental data to generate the final recommended handling plan.

[0117] Step S304: Deconstruct the final recommended disposal plan into a task sequence; distribute the atom task sequence to the corresponding emergency execution department, monitor the multi-dimensional data during the task execution process in real time, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain the analysis results; based on the analysis results, update the dynamic fusion decision knowledge base in a targeted manner.

[0118] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a multi-scenario emergency response decision-making system for transportation safety management according to this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0119] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a multi-scenario emergency response decision-making system for transportation safety management as described in this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0120] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A multi-scenario emergency response decision-making system for transportation safety management, characterized in that, The system includes: The dynamic decision-making database module is used to obtain structured feature vectors by performing structured analysis on historical accident cases; construct a historical case knowledge graph based on the structured feature vectors; establish a parameterized simulation model based on multiple typical accident scenarios, and generate an optimized handling strategy package by performing multi-objective optimization calculations through the parameterized simulation model; and fuse and associate the historical case knowledge graph with the optimized handling strategy package to obtain a dynamic fused decision-making knowledge base. The accident perception and localization module is used to acquire vehicle data streams, roadside data streams, and environmental data streams in real time to obtain a multi-source heterogeneous perception data stream. Based on high-frequency feature templates extracted from the dynamic fusion decision knowledge base, anomaly analysis is performed on the multi-source heterogeneous perception data stream to generate an initial accident hypothesis. Based on the initial accident hypothesis, corresponding typical evidence patterns are retrieved from the dynamic fusion decision knowledge base to perform targeted analysis on the multi-source heterogeneous perception data stream to obtain targeted analysis results. Based on the targeted analysis results, structured accident confirmation information containing accident type, location, and initial condition labels is generated. The intelligent decision matching module is used to search the dynamic fusion decision knowledge base based on the initial condition tags in the structured accident confirmation information to obtain matching historical cases, actual handling plans, and matching optimized handling strategy packages; it integrates the actual handling plans with the recommended actions in the matching optimized handling strategy packages to generate a preliminary handling plan draft; and it uses real-time environmental data to drive a preset digital twin simulation model to deduce and optimize the preliminary handling plan draft to generate a final recommended handling plan. A multi-departmental collaborative scheduling module is used to decompose the final recommended disposal plan into a task sequence; distribute the atom task sequence to the corresponding emergency execution department, monitor multi-dimensional data during task execution in real time, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain analysis results; and update the dynamic fusion decision knowledge base in a targeted manner based on the analysis results.

2. The system according to claim 1, characterized in that, The dynamic decision base module includes a graph construction subunit and a knowledge base construction subunit. The graph construction subunit is used for: Based on the recorded data of historical road transport accidents in the aforementioned historical accident cases, each accident is analyzed in a structured manner to extract the accident scene type, spatiotemporal coordinates, environmental conditions, sequence of response actions, list of resources used, and final response effect indicators, thus obtaining single-accident structured data. Based on the single-accident structured data of each accident, the causal dependencies between features, the correspondence between scenes and response actions, and the correlation between resource consumption and effects are analyzed to obtain feature interrelationships. The single-accident structured data and the feature interrelationships are integrated to obtain the structured feature vector. The structured feature vectors and their interrelationships are stored in a preset graph database. Synonymous relationships between entities are identified through a semantic mapping symmetric discriminant function, and implicit logical chains are obtained by combining a causal scoring function, thereby constructing the historical case knowledge graph. The knowledge base construction subunit is used for: For the preset multiple typical accident scenarios, the parametric simulation model is constructed. The input of the parametric simulation model is the initial parameters of the accident and the environmental parameters, and the output is the accident development trend and the prediction data of the effects of different handling actions. The non-dominated sorting genetic algorithm II is used to perform multi-objective optimization calculations on the preset combination of accident initial conditions and disposal actions. The optimization objectives are set as rescue efficiency, loss control and resource consumption. The termination criteria are that the consolidation ratio is greater than the preset ratio or the optimization performance indicator shows an inflection point. The Pareto optimal optimized disposal strategy package is generated. The historical case knowledge graph is associated with the optimization strategy package. The cases in the historical case knowledge graph are labeled with corresponding optimization strategy references, and similar historical cases are linked to the optimization strategy package to establish a multi-dimensional index structure, thereby obtaining the dynamic fusion decision knowledge base.

3. The system according to claim 1, characterized in that, The accident perception and positioning module includes a real-time data acquisition subunit and a real-time data analysis subunit. The real-time data acquisition subunit is used to perform noise reduction and format standardization processing on the real-time acquired vehicle data stream, roadside data stream and environmental data stream in sequence to obtain the multi-source heterogeneous perception data stream. The real-time data analysis subunit is used for: The frequency of abnormal features is counted from the historical case knowledge graph of the dynamic fusion decision knowledge base, and high-frequency feature combinations are extracted through cluster analysis to form the high-frequency feature template. The multi-source heterogeneous sensing data stream is input into a machine learning classifier constructed based on the high-frequency feature template to perform anomaly detection and generate one or more initial hypotheses of the accident with confidence. For each initial hypothesis of the accident, the key feature combination of the corresponding accident type is retrieved from the dynamic fusion decision knowledge base as the typical evidence pattern, and multi-source data cross-validation is performed in combination with the environmental data stream to obtain the targeted analysis results. When the matching degree between the directional analysis result and the typical evidence pattern reaches a preset threshold, the occurrence of the accident is confirmed. Combining the directional analysis result and the typical evidence pattern, the structured accident confirmation information containing the accident type, the precise spatiotemporal location, and the initial condition label is generated.

4. The system according to claim 1, characterized in that, The intelligent decision matching module includes a preliminary scheme generation subunit and a scheme update subunit. The preliminary scheme generation subunit is used for: Using the initial condition tags as the core retrieval items, the cosine similarity algorithm is used to perform semantic matching on multiple historical cases in the dynamic fusion decision knowledge base. Historical cases with similarity higher than a preset similarity threshold are used as the matching historical cases, and the actual handling plan corresponding to the matching historical cases and the matching optimization handling strategy package adapted to the current accident scenario are obtained. The success factors and shortcomings of the actual disposal plan are analyzed using a weighted fusion algorithm to obtain case analysis results. Combined with the recommended actions in the matching and optimization disposal strategy package, the plan is integrated and matched through conflict resolution rules to generate the preliminary disposal plan draft, which includes a resource scheduling list, action sequence and spatial control scope. The scheme updates the subunit for: The preliminary draft of the disposal plan and the real-time acquired environmental data stream are input into the preset digital twin simulation model to perform execution process simulation, and predictive indicators including rescue time, resource utilization rate and risk control effect are obtained. The comprehensive utility value is calculated based on the predictive indicators. Based on the predicted indicators and the comprehensive utility value, a sensitivity analysis is conducted on the resource allocation parameters, action timing parameters and spatial control parameters in the preliminary disposal plan draft to determine the key control parameters. The key control parameters are fine-tuned in multiple rounds according to a preset step size. After each round of fine-tuning, the adjusted scheme is re-input into the digital twin simulation model to perform deduction, calculate the comprehensive utility value of the corresponding round, compare the comprehensive utility values ​​of each round, select the parameter combination corresponding to the highest comprehensive utility value, and generate the final recommended treatment scheme.

5. The system according to claim 1, characterized in that, The multi-department coordinated scheduling module includes a task execution subunit and a task feedback update subunit. The task execution subunit is used for: Based on the division of responsibilities among functional departments and the dependency relationship between tasks, the final recommended disposal plan is broken down into independent task units with execution entities, resource requirements, start and end times, and acceptance standards, forming the atomic task sequence; The atomic task sequence is pushed to the command system of the corresponding emergency execution department through the emergency response platform, and the task commitment receipt returned by the command system is received and verified. The task feedback update subunit is used for: Real-time collection of multi-dimensional information including task execution progress, resource availability, changes in the on-site environment, accident situation evolution data, and actual execution details of response actions, to obtain actual execution records, and integration of the structured accident confirmation information, the final recommended response plan, and the actual execution records to generate the complete response data package; The actual effect indicators, including casualty control rate, property loss reduction rate, rescue response time and traffic restoration efficiency, are calculated from the complete set of disposal data. The actual effect indicators are compared with the predicted effect indicators of the final recommended disposal plan to calculate the deviation rate. The causes of the deviation are analyzed in combination with the actual execution records to obtain the analysis results. The complete set of disposal data is updated to the historical case knowledge graph as a new case. The parameterized simulation model is adjusted according to the analysis results. An optimized disposal strategy package corresponding to the new case is generated based on the updated parameterized simulation model. The new case is then stored in the historical case knowledge graph to obtain an updated dynamic fusion decision knowledge base.

6. The system according to claim 4, characterized in that, The formula for calculating the comprehensive utility value is as follows: in, The comprehensive utility value ranges from [0, 10]. These correspond to the rescue efficiency target, loss control target, and resource consumption target, respectively. For the first The dynamic weight of the project target, and , This refers to the timeline following the accident; Norm For the first Project Target Indicators The normalized value; For the first Project target and the first The coupling coefficient of the project target, with a value range of [-0.5, 0.5].

7. A multi-scenario emergency response decision-making method in transportation safety management, characterized in that, The method includes: Step S1: Obtain structured feature vectors by performing structured analysis on historical accident cases; construct a historical case knowledge graph based on the structured feature vectors; establish a parameterized simulation model based on multiple typical accident scenarios, and generate an optimized handling strategy package by performing multi-objective optimization calculations through the parameterized simulation model; fuse and associate the historical case knowledge graph with the optimized handling strategy package to obtain a dynamic fusion decision knowledge base; Step S2: Real-time acquisition of vehicle data stream, roadside data stream, and environmental data stream to obtain a multi-source heterogeneous sensing data stream; based on the high-frequency feature templates extracted from the dynamic fusion decision knowledge base, anomaly analysis is performed on the multi-source heterogeneous sensing data stream to generate an initial accident hypothesis; based on the initial accident hypothesis, corresponding typical evidence patterns are retrieved from the dynamic fusion decision knowledge base to perform targeted analysis on the multi-source heterogeneous sensing data stream to obtain targeted analysis results; based on the targeted analysis results, structured accident confirmation information containing accident type, location, and initial condition labels is generated; Step S3: Based on the initial condition tags in the structured accident confirmation information, search the dynamic fusion decision knowledge base to obtain matching historical cases, actual handling plans, and matching optimized handling strategy packages; fuse the actual handling plans with the recommended actions in the matching optimized handling strategy packages to generate a preliminary handling plan draft; use real-time environmental data to drive a preset digital twin simulation model to deduce and optimize the preliminary handling plan draft to generate a final recommended handling plan; Step S4: Deconstruct the final recommended disposal plan into a task sequence; distribute the atom task sequence to the corresponding emergency execution department, monitor multi-dimensional data during task execution in real time, and generate a complete disposal data package; compare and analyze the actual effect indicators in the complete disposal data package with the predicted effect indicators of the final recommended disposal plan to obtain the analysis results; based on the analysis results, update the dynamic fusion decision knowledge base in a targeted manner.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the system according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the system according to any one of claims 1 to 6.

Images