Dynamic self-adaptive management system for road traffic safety planning under driving of artificial intelligence

Abstract

The invention discloses a dynamic adaptive management system for road traffic safety planning driven by artificial intelligence. According to the system, multi-source data are acquired through a data acquisition and preprocessing layer and are cleaned and standardized; the knowledge engineering layer constructs a policy knowledge graph and a rule engine; the large model service layer realizes natural language generation, semantic understanding and multi-modal reasoning; the intelligent generation layer uses a GAN model to generate multi-modal planning content and dynamically adapts the multi-modal planning content; the implementation process management layer monitors planning implementation, middle-stage evaluation and last-stage summarization in real time; and the interaction and output layer provides visual editing and multi-format output. The system integrates the technologies of deep learning, generative adversarial network and the like, realizes intelligent planning generation and dynamic supervision, has the capabilities of multi-modal data processing, dynamic self-adaption and full-process management, can effectively improve the scientificity, high efficiency and adaptability of road traffic safety planning, and provides powerful support for road traffic safety management.

Description

Technical Field

[0001] This invention belongs to the field of road traffic safety planning technology, and in particular relates to a dynamic adaptive management system for road traffic safety planning driven by artificial intelligence. Background Technology

[0002] With the acceleration of intelligent road traffic and the continuous increase in various traffic elements such as people, vehicles, roads, and enterprises, new transportation business models such as autonomous vehicles and unmanned delivery vehicles are booming, posing higher requirements for road traffic safety. Traditional road traffic safety planning methods mainly rely on expert experience and limited data analysis, which have many problems and limitations.

[0003] On the one hand, traditional methods have significant shortcomings in data processing. They can only process structured traffic data (such as traffic flow statistics) and lack the ability to integrate unstructured data (such as video surveillance and sensor time-series data) and multi-dimensional correlated data (human-vehicle-road-environment interaction characteristics). This results in incomplete data dimensions for safety planning, making it difficult to comprehensively and accurately reflect the complexity and dynamism of the traffic system. On the other hand, traditional methods lack dynamic adaptability. Relying on static historical data to generate planning schemes, they cannot perceive the dynamic safety risks brought about by new traffic patterns (such as autonomous vehicle trajectories and unmanned delivery vehicle cluster behavior) in real time. This leads to a disconnect between the planning scheme and the actual traffic conditions, making it difficult to adapt to rapidly changing traffic environments and demands.

[0004] Furthermore, traditional methods are weak in comprehensive safety correlation analysis. They focus only on single factors (such as vehicle speed) or safety indicators of local road sections, lacking a comprehensive correlation analysis of traffic participants (pedestrians / non-motorized vehicles / motorized vehicles), infrastructure (traffic lights / guardrails), and enterprise operational data (logistics routes). This results in a one-sided approach to safety planning and an inability to grasp the overall traffic safety situation. Traditional methods also lag behind in planning implementation supervision. They can only evaluate the effectiveness of planning through post-accident statistics, lacking real-time monitoring and dynamic adjustment mechanisms for the planning implementation process. This leads to safety hazards not being eliminated in a timely manner, and planning iteration cycles lasting months or even years, making it difficult to achieve effective supervision and timely optimization of planning implementation.

[0005] Finally, traditional methods suffer from gaps in end-to-end digital management. Security planning faces technical barriers at each stage, from data collection and solution generation to implementation and supervision, lacking an integrated digital platform. This results in low planning efficiency and difficulty in achieving closed-loop management, failing to fully leverage the advantages of digital technology in improving planning quality and efficiency.

[0006] Therefore, an innovative system is needed to solve the above problems, improve the scientific nature, efficiency and dynamic adaptability of road traffic safety planning, and support full-process digital management to meet the needs of modern road traffic safety management. Summary of the Invention

[0007] This invention proposes an artificial intelligence-driven dynamic adaptive management system for road traffic safety planning to address the problems existing in the prior art.

[0008] To achieve the above objectives, the present invention provides an artificial intelligence-driven dynamic adaptive management system for road traffic safety planning, comprising:

[0009] The data acquisition and preprocessing layer is used to collect multi-source heterogeneous road traffic safety-related data and perform cleaning, transformation and standardization processing.

[0010] The knowledge engineering layer is used to build policy knowledge graphs and rule engines based on road traffic safety-related policies, regulations, and standards, providing policy basis and professional knowledge support.

[0011] The large model service layer is used to perform natural language generation, semantic understanding, and multimodal reasoning on text data, and output structured analysis results.

[0012] The intelligent generation layer is used to generate multimodal content for road traffic safety planning through a generative adversarial network model based on the structured analysis results, and dynamically adjust the planning generation strategy and regulatory parameters according to real-time traffic data, environmental changes and user needs.

[0013] The implementation process management level is used to monitor the implementation process of road traffic safety planning in real time. In the middle stage of planning implementation, it evaluates the achievement of planning goals and the effectiveness of measures. At the end of planning implementation, it comprehensively summarizes and evaluates the overall effect of the plan.

[0014] The interaction and output layer provides a visual editing interface and multi-format output functionality, enabling users to interact with the system and obtain planning results.

[0015] Optionally, the data acquisition and preprocessing layer includes:

[0016] The multimodal data access module is used to access accident record texts from traffic management departments, video images from road monitoring systems, traffic flow data collected by traffic sensors, and map data from geographic information systems.

[0017] The intelligent cleaning module is used to clean, transform, and standardize the collected data, remove duplicate and erroneous information from accident records, denoise, enhance, and extract key frames from video images, and integrate traffic flow data according to a uniform time interval.

[0018] Optionally, the knowledge engineering layer includes:

[0019] The policy knowledge graph module is used to collect and organize policies, regulations, and standards related to road traffic safety through natural language processing technology, and to build a policy knowledge graph that graphically displays the relationships, scope of application, and implementation requirements between policies.

[0020] The rules engine module is used to build a knowledge base that includes traffic rules, safe driving knowledge, and accident case analysis. By analyzing and summarizing historical accident cases, it extracts key factors and patterns of traffic safety.

[0021] Optionally, the large model service layer includes:

[0022] The natural language generation module is used to generate planning reports based on a pre-trained large language model through fine-tuning and prompting engineering techniques.

[0023] The semantic understanding module is used to perform semantic understanding on the collected text data through deep learning algorithms, and to extract key information, intent and sentiment.

[0024] The multimodal reasoning module is used to fuse multimodal data including text, images, videos, and numerical values, and to perform comprehensive analysis using artificial intelligence reasoning algorithms to output analysis results.

[0025] Optionally, the intelligent generation layer includes:

[0026] The multimodal generation module is used to generate multimodal content for road traffic safety planning based on the analysis results provided by the large model service layer through a generative adversarial network model. This includes visualization of planning schemes, voice explanations, and virtual reality scene simulations.

[0027] The dynamic adaptation engine module is used to dynamically adjust the planning generation strategy and regulatory parameters based on real-time traffic data, environmental changes and user needs, and optimize the generated planning scheme based on reinforcement learning algorithms.

[0028] Optionally, the implementation process management layer includes:

[0029] The dynamic monitoring subsystem is used to monitor the implementation process of road traffic safety planning in real time through Internet of Things technology and sensor networks. It monitors the construction progress of traffic facilities, the installation status of traffic signs and the operation status of traffic lights, and issues early warnings when problems are found in the implementation process.

[0030] The mid-term evaluation subsystem is used to evaluate the achievement of planning goals and the effectiveness of measures by collecting and analyzing relevant data during the mid-term phase of planning implementation. Based on the evaluation results, it proposes adjustment suggestions through artificial intelligence algorithms.

[0031] The final summary subsystem is used to comprehensively summarize and evaluate the overall effectiveness of the plan at the end of its implementation and to feed the evaluation results back to the knowledge base.

[0032] Optionally, the interaction and output layer includes:

[0033] A visual editor is used to provide a visual interactive interface, through which users can query, edit, modify and optimize road traffic safety planning content;

[0034] The multi-format output module is used to output the system's final planning results and regulatory reports. The formats of the final planning results and regulatory reports include electronic sand table, PDF document, PPT presentation and video report.

[0035] Optionally, the expression for the comprehensive analysis is:

[0036] ;

[0037] In the formula, This represents the result of comprehensive reasoning. Represents a multimodal inference function. Represents a multi-source data set. The parameters of the multimodal inference model are: ϕ(x) represents the feature extraction function; ⨁ represents the multimodal fusion operator. The prediction function receives the fused features (i.e. The results are used as input to generate a comprehensive reasoning result. .

[0038] Optionally, the calculation expression of the multimodal generation module is:

[0039] ;

[0040] In the formula, Represents a generator. Indicates the discriminator, Data distribution representing historical planning schemes, This represents the prior normal distribution of random noise. Represents the value function. Represents the mathematical expectation. This represents a random variable.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] This invention achieves a deep integration of artificial intelligence technology and road traffic safety planning, resulting in significant technical benefits. The system utilizes natural language generation, semantic understanding, and multimodal reasoning technologies to intelligently generate and optimize plans, improving the efficiency and quality of planning text generation. Simultaneously, the system can process and analyze data from multiple modalities, including text, images, videos, and numerical data, fully utilizing information from various data sources to make planning more scientific and rational. Furthermore, the system possesses dynamic adaptive capabilities, enabling timely adjustments to planning schemes and regulatory strategies based on real-time traffic data, environmental changes, and user needs, ensuring the effectiveness and adaptability of the plans. The system also achieves end-to-end management from plan generation and implementation to supervision and evaluation, improving the efficiency and effectiveness of plan execution. The interaction and output layer provides an intuitive and user-friendly visual interface and multiple output formats to meet the needs of different users, enhancing user experience and practicality. Through data-driven plan generation and optimization, the system can accurately identify key traffic safety risks and highly effective improvement measures, ensuring that limited resources are invested in the most needed areas, thus improving resource utilization efficiency. In summary, this invention significantly enhances the scientific rigor, efficiency, and adaptability of road traffic safety planning, providing strong technical support for road traffic safety management. Attached Figure Description

[0043] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0044] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0045] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0047] Example 1

[0048] like Figure 1 As shown, this embodiment provides an artificial intelligence-driven dynamic adaptive management system for road traffic safety planning, including:

[0049] The data acquisition and preprocessing layer 101 is used to collect multi-source heterogeneous road traffic safety-related data and perform cleaning, transformation and standardization processing.

[0050] Knowledge Engineering Layer 102 is used to construct a policy knowledge graph and rule engine based on road traffic safety-related policies, regulations, and standards, providing policy basis and professional knowledge support;

[0051] The large model service layer 103 is used to perform natural language generation, semantic understanding and multimodal reasoning on text data, and output structured analysis results;

[0052] The intelligent generation layer 104 is used to generate multimodal content of road traffic safety planning through a generative adversarial network model based on the structured analysis results, and dynamically adjust the planning generation strategy and regulatory parameters according to real-time traffic data, environmental changes and user needs.

[0053] The implementation process management layer 105 is used to monitor the implementation process of road traffic safety planning in real time. In the mid-term stage of planning implementation, it evaluates the achievement of planning goals and the effectiveness of measures. In the final stage of planning implementation, it comprehensively summarizes and evaluates the overall effect of the plan.

[0054] The interaction and output layer 106 provides a visual editing interface and multi-format output functions, supporting user interaction with the system and obtaining planning results.

[0055] Furthermore, the data acquisition and preprocessing layer 101 includes:

[0056] The multimodal data access module 1011 can access road traffic safety-related data from various sources and in different formats, such as accident record texts from traffic management departments, video images from road monitoring systems, traffic flow data collected by traffic sensors, and map data from Geographic Information Systems (GIS). This extensive data access provides a comprehensive information foundation for subsequent analysis and planning.

[0057] The intelligent cleaning module 1012 cleans, transforms, and standardizes the collected data. For example, it removes duplicate and erroneous information from accident records, denoises, enhances, and extracts keyframes from video images, and integrates traffic flow data according to a uniform time interval to ensure data quality and consistency and improve the accuracy of subsequent processing.

[0058] Furthermore, the knowledge engineering layer 102 includes:

[0059] The policy knowledge graph module 1021 collects and organizes textual materials such as policies, regulations, standards, and norms related to road traffic safety, and constructs a policy knowledge graph using natural language processing technology. This graph graphically displays information such as the relationships between policies, their scope of application, and implementation requirements, providing policy basis and reference for planning.

[0060] The rule engine module 1022 establishes a knowledge base containing traffic rules, safe driving knowledge, and accident case analysis. Through the analysis and summarization of numerous historical accident cases, it extracts key factors and patterns in traffic safety, providing professional knowledge support for planning and supervision.

[0061] Furthermore, the large model service layer 103 includes:

[0062] The natural language generation module 1031 uses a large model to generate natural language texts such as planning reports and recommendations. Based on a pre-trained large language model (such as GPT-4, ERNIE-Bot, or Mistral), it uses fine-tuning and prompt engineering techniques to automatically generate texts such as planning reports, policy recommendations, and draft plans, thereby improving the efficiency of planning text generation.

[0063] The semantic understanding module 1032 uses deep learning algorithms to perform semantic understanding on the collected text data, extracting key information, intent, and sentiment. For example, it analyzes public feedback text on traffic safety to understand the public's needs and opinions on traffic facilities, traffic rules, etc., providing a reference for planning adjustments.

[0064] The multimodal reasoning module 1033 integrates data from multiple modalities, including text, images, videos, and numerical data, and performs comprehensive analysis using artificial intelligence reasoning algorithms. For example, by combining images of a traffic accident scene with accident report text, it infers the causes of the accident, potential risks, and possible preventative measures, providing more comprehensive and in-depth analysis results for planning generation.

[0065] Furthermore, the intelligent generation layer 104 includes:

[0066] The multimodal generation module 1041, based on the analysis results provided by the core AI processing layer, utilizes a Generative Adversarial Network (GAN) model to generate multimodal content for road traffic safety planning, such as visualization of planning schemes, voice explanations, and virtual reality (VR) scene simulations, making the planning content more intuitive and easy to understand. Legal elements are integrated into a traffic-related regulations database, and regulations are parsed using an NLP model to generate compliance constraints. The regulations database server deploys an Elasticsearch cluster, supporting millisecond-level regulations retrieval (single-node configuration, 32-core CPU, 128GB memory, 1TB SSD). Road features are integrated with high-precision map data and historical accident heatmaps. A safety-oriented loss function is used, introducing safety indicator weighting terms (such as the accident rate reduction rate and compliance score) on top of the original GAN ​​loss function. Multi-task learning synchronously generates planning text, visualization schemes, and VR simulations, ensuring cross-modal consistency.

[0067] The multimodal generation module 1041 in the intelligent generation layer 104 refers to the generative model of the Generative Adversarial Network (GAN), and its formula and meaning are as follows:

[0068]

[0069] Represents a generator. Indicates the discriminator, Data distribution representing historical planning schemes, This represents the prior normal distribution of random noise. Represents the value function. It represents the mathematical expectation.

[0070] The dynamic adaptation engine module 1042 dynamically adjusts the planning generation strategy and regulatory parameters based on real-time traffic data, environmental changes, and user needs. It employs reinforcement learning algorithms to optimize the generated planning schemes. By continuously simulating and evaluating the effects of different schemes, the optimal planning scheme is selected, improving the scientific validity and feasibility of the planning.

[0071] Furthermore, the implementation process management layer 105 includes:

[0072] The dynamic monitoring subsystem 1051 utilizes Internet of Things (IoT) technology and sensor networks to monitor the implementation process of road traffic safety planning in real time. For example, it monitors the construction progress of traffic facilities, the installation status of traffic signs, and the operational status of traffic lights, promptly identifying problems during implementation and issuing early warnings.

[0073] The mid-term evaluation subsystem 1052, during the mid-term phase of planning implementation, collects and analyzes relevant data to assess the achievement of planning objectives and the effectiveness of measures. Based on the evaluation results, it uses artificial intelligence algorithms to propose adjustment suggestions to ensure that the plan can adapt to changes in actual conditions.

[0074] The final summary subsystem 1053 comprehensively summarizes and evaluates the overall effectiveness of the plan at the end of its implementation. It analyzes lessons learned during the implementation process and transmits feedback information to the intelligent generation layer 104, providing a reference for subsequent planning.

[0075] Furthermore, the interaction and output layer 106 includes:

[0076] The visual editor 1061 provides an intuitive and easy-to-use visual interactive interface, facilitating users to query, edit, modify, and optimize road traffic safety planning content. Users can interact with the system through charts, maps, 3D models, and other methods, improving operational convenience and user experience.

[0077] The multi-format output module 1062 outputs the system's final planning results and supervision reports in various formats, such as electronic sand table, PDF document, PPT presentation, video report, etc., to meet the needs of different users in different scenarios and facilitate the reporting, dissemination and archiving of plans.

[0078] Specifically, the large model service layer 103 is the server that carries the large road traffic safety planning model. The large road traffic safety planning model consists of a hierarchical processing architecture composed of three core modules. The following is an explanation of the model formulas and symbols.

[0079]

[0080] in, This represents a natural language generation function; This represents a semantic understanding function; Represents a multimodal inference function; This represents a large-scale model for road traffic safety planning.

[0081] The first step is for the natural language generation module 1031 to generate a planning report R by taking the requirements document D as input.

[0082]

[0083] in, Generate model parameters for natural language processing; Represents a word sequence; This indicates that... Maximize all possible values;

[0084] The second step involves the semantic understanding module 1032 performing a triple analysis on the text data T.

[0085]

[0086] Where K represents the key information vector, , I represents the feature dimension; I represents the intent label. , E represents the number of intention categories; E represents the emotional polarity. -1 represents negative, and 1 represents positive; This represents a semantic understanding function, where T represents text data; For semantic understanding model parameters;

[0087] The third step involves the multimodal reasoning module 1033 fusing multi-source data X={T,I,V,N} for comprehensive reasoning.

[0088]

[0089] In the formula, This represents the result of comprehensive reasoning. Represents a multimodal inference function. Represents a multi-source data set. The parameters of the multimodal inference model are: ϕ(x) represents the feature extraction function; ⨁ represents the multimodal fusion operator. The prediction function receives the fused features (i.e. The results are used as input to generate a comprehensive reasoning result. Typically, it consists of one or more fully connected layers.

[0090] The multimodal generation module 1041 in the intelligent generation layer 104 refers to the generative model of the Generative Adversarial Network (GAN), and its formula and meaning are as follows:

[0091]

[0092] in, Represents a generator. Indicates the discriminator, Data distribution representing historical planning schemes, This represents the prior normal distribution of random noise. Represents the value function. It represents the mathematical expectation.

[0093] Specifically, we will take the formulation of the road traffic safety plan for Province S over the next five years as an example to explain in detail.

[0094] Step 1, Data Acquisition and Preprocessing Layer 101;

[0095] The multimodal access module of the data acquisition and preprocessing layer 101 is used to access road traffic safety-related data from various sources and in different formats. Text data is accessed via an ETL (Extract-Transform-Load) pipeline built using Apache NiFi (Apache data stream processing tool), supporting PDF / Word / Excel parsing. Video data is accessed via an FFmpeg (Fast Forward MPEG, a multimedia processing tool) transcoding cluster, slicing surveillance videos into 10-second keyframes. Numerical data is analyzed using Apache Flink (Apache stream processing framework) for real-time traffic statistics (5-second granularity). The intelligent cleaning module 1012 is used to clean, transform, and standardize the acquired data. Text cleaning uses regular expressions to filter illegal characters and BERT-base (Bidirectional Encoder Representations from Transformers-base) to detect semantic repetition. Video cleaning uses YOLOv7 (You Only Look Once version)... 7. The target detection model v7 identifies invalid frames (such as all-black / all-white images), automatically removes noise, and standardizes data by using a unified timestamp format to complete spatial coordinate transformation.

[0096] The second step is Knowledge Engineering Layer 102;

[0097] The policy knowledge graph module 1021 of the knowledge engineering layer 102 collects and organizes textual materials such as policies, regulations, standards and norms related to road traffic safety. It uses natural language processing technology to construct a policy knowledge graph, uses spaCy (Industrial-strength Natural Language Processing) for entity recognition (policy name, implementation time, responsible department), stores the graph through Neo4j (Graph Database), and constructs policy relationships (such as "State Council Document No. 12 of 2023" → "Revision of the Road Traffic Safety Law"). The rule engine module 1022 establishes a knowledge base containing traffic rules, safe driving knowledge, accident case analysis, etc. It builds a historical accident case database with 10,000+ cases labeled, extracts accident causal chains (such as "speeding → loss of control → collision"), and the safety rule base converts the "Highway Route Design Specifications" into executable rules (such as "warning signs must be set for curves with a radius of less than 50m").

[0098] The third step is to serve the large model layer 103.

[0099] The natural language generation module 1031 of the large model service layer 103 is used to generate natural language text such as planning reports and suggestions using the large model. Based on the pre-trained large language model, through fine-tuning and prompting engineering techniques, it realizes the automatic generation of texts such as planning reports, policy suggestions, and draft plans. Pre-training continues on the S Province accident report dataset. The fine-tuning strategy uses LoRA (Low-Rank Adaptation) technology, freezing 90% of the parameters and training only the adapter layer. Prompt engineering is used to design a three-stage prompt template (requirements analysis → plan generation → compliance check). The semantic understanding module 1032 uses deep learning algorithms to perform semantic understanding on the collected text data, extracting key information, intent, and sentiment. Sentiment analysis is used to construct a S Province dialect sentiment dictionary (e.g., "stuck" → negative emotion). Intent recognition provides a classifier to support 8 types of user intents (planning consultation, data query, complaints and suggestions, etc.). The multimodal reasoning module 1033 is used to fuse data from multiple modalities such as text, images, videos, and numerical values, and uses artificial intelligence reasoning algorithms for comprehensive analysis. Cross-modal attention adopts the UniT architecture (Universal... Transformer (a general-purpose transformer architecture), shared encoder, feature fusion inputs text TF-IDF (Term Frequency-Inverse Document Frequency) vectors (200-dimensional), video C3D (Convolutional 3D Network) features (4096-dimensional), and numerical vectors (32-dimensional).

[0100] Step 4, intelligent generation layer 104;

[0101] The multimodal generation module 1041 of the intelligent generation layer 104 is used to generate multimodal content for road traffic safety planning based on the analysis results provided by the core artificial intelligence processing layer. This content includes visualization of planning schemes, voice explanations, and virtual reality (VR) scene simulations. The T5-large model (Text-To-Text Transfer Transformer-large) is used to generate planning text and compliance descriptions simultaneously. Stable Diffusion XL is used to generate risk distribution maps based on heatmap visualizations. The engine constructs a 3D road network model and accesses real-time traffic flow data. The dynamic adaptation engine module 1042 is used to dynamically adjust the planning generation strategy and regulatory parameters based on real-time traffic data, environmental changes, and user needs. The reinforcement learning environment defines a reward function (accident rate reduction × 0.7 + budget execution rate × 0.3) and uses the PPO algorithm (Proximal Policy Optimization) to update the policy network every 24 hours.

[0102] This system adopts a hierarchical processing architecture, with the core formula being mul = {G, U, M}. Here, G (natural language generation function) is responsible for transforming structured data and analysis results into human-readable, standardized planning text; U (semantic understanding function) is responsible for performing deep semantic analysis on textual data to extract key information, intent, and sentiment; and M (multimodal reasoning function) is responsible for integrating heterogeneous data such as text, images, videos, and numerical data for comprehensive judgment and reasoning.

[0103] In the first phase, a preliminary planning report was generated.

[0104] Input (Requirements Document D);

[0105] The large model system accesses and parses multiple internal and public documents from Province S, forming a set of requirement documents D.

[0106] The Comprehensive Transportation Development Plan of Province S;

[0107] "White Paper on Road Traffic Accident Statistics and Analysis in S Province, 2023";

[0108] "Construction Requirements for Intelligent Transportation Systems in the S Province Expressway Network";

[0109] Traffic safety pain point reports submitted by various cities and prefectures (e.g., "Request for Instructions on Frequent Accidents on a Certain Section of XX Expressway" and "Report on the Gap in Safety Protection Facilities for Rural Roads").

[0110] Draft provincial-level fiscal budget for transportation (2025-2027);

[0111] In the processing, G represents the natural language generation module 1031 learning the internal logic and planning language style of the above document D based on the parameter θg, and generating a preliminary planning draft R by maximizing the probability sequence.

[0112] R = G(D; θg) = argmax P(w|D; θg)

[0113] Output (Preliminary Planning Draft R) to generate the "S Province Road Traffic Safety Plan (2025-2030) (Draft)", the core contents of which include engineering measures, management measures, technological applications, budget estimates, etc.

[0114] The second stage involves multidimensional semantic understanding and feedback analysis.

[0115] Input (text data T), and input the preliminary planning draft R and the subsequent multi-source text data into the semantic understanding module 1032.

[0116] Draft R full text;

[0117] Text of the expert review meeting minutes;

[0118] Text of online comments and written opinions submitted to the public hearing;

[0119] Media reports and commentaries on the draft;

[0120] Processing procedure;

[0121] U indicates that the semantic understanding module 1032 performs triple analysis on the text set T and outputs a structured vector.

[0122] K, I, E = U(T; θu);

[0123] Output analysis results, key information vector (K∈R) n );

[0124] K = [0.92, 0.85, 0.78, 0.63, 0.41] (n=5);

[0125] K1=0.92, and the precise management of accident hotspots was identified as the highest priority keyword.

[0126] With K²=0.85, the coverage rate of smart devices is a key focus.

[0127] K3=0.78, and the rationality of the fund allocation has been discussed many times.

[0128] K4=0.63, the impact of the construction period on traffic is a secondary concern.

[0129] K5=0.41, indicating a low level of consensus on the application of new technologies (such as autonomous driving testing).

[0130] Intent label (I∈{0,1}ᵐ);

[0131] I = [1, 0, 1, 0, 1, 0] (m=6);

[0132] [Engineering renovation, regulation revision, technology upgrade, education and publicity, budget approval, cross-departmental collaboration];

[0133] Analysis shows that the strongest consensus currently is on engineering renovation, technological upgrading, and budget approval.

[0134] Emotional polarity (E∈[-1,1]), E = 0.76.

[0135] The overall sentiment is strongly positive, indicating that experts and the public highly approve of the direction of using high-tech means to manage traffic safety, but there are disagreements on the details of the draft.

[0136] The third stage involves multimodal data fusion and intelligent reasoning.

[0137] The input is multi-source data X = {T, L, V, N}. To optimize the draft, the model incorporates more modal data for analysis. The multimodal inference function M extracts features from each modality of data through the feature extraction function φ(x), and then performs comprehensive inference through the fusion operator.

[0138] Feature extraction;

[0139] The image L was used to extract a certain ramp on the XX Expressway as a spatial blind spot.

[0140] Analysis of video V revealed that 78% of the accidents at that location were related to heavy trucks losing control while crossing lanes and going downhill.

[0141] The numerical value N shows that the accident rate on this road section is 35% higher at night than during the day, and this is highly correlated with the fog meteorological data.

[0142] Multimodal fusion reasoning: [Text T, "Add surveillance"]⨁[Image L, blind spot coordinates]⨁[Value N, Nighttime accident rate ↑35%] => Decision: Deploy radar-visual integrated machines with infrared night vision capabilities at the blind spot coordinates; [Value N, High truck traffic]⨁[Video V, Truck behavior analysis]⨁[Text T, "Fatigue driving"] => Decision: Add a fatigue driving warning system at the highway exit and link it with the law enforcement system. Output an optimized planning scheme, generating the "S Province Road Traffic Safety Plan (2025-2030)", the key contents of which are as follows:

[0143] 1. Planning goals (by 2030);

[0144] The accident prevention and control goals are to strive to reduce the road traffic accident fatality rate per 10,000 vehicles in the province by X (sensitive data, omitted) or more, resolutely curb traffic accidents with 5 or more deaths, and eliminate major traffic accidents with 10 or more deaths. The infrastructure construction goals are to achieve a coverage rate of X% (sensitive data, omitted) for corrugated guardrails or other safety protection facilities on rural roads with potential hazards throughout the province; and to ensure that the average spacing between video surveillance and radar sensing equipment on highways is no greater than X kilometers, and no greater than X kilometers on key road sections. The capacity building goals are to establish a province-wide integrated "cloud-edge-terminal" vehicle-road collaborative safety early warning system; and to achieve a network access rate of X% (sensitive data, omitted) for dynamic monitoring of key vehicles (passenger vehicles, dangerous goods vehicles, freight vehicles, and motorcycles), and a violation processing rate of X% (sensitive data, omitted) or more.

[0145] 2. Main tasks and measures;

[0146] 2.1 Improve the modern governance system for road traffic safety;

[0147] Strengthen legislation and standards development, promote the revision of the "S Province Road Traffic Safety Regulations," and improve local standards for traffic safety management of expressways and rural roads. Strictly implement the main responsibility of transportation enterprises and establish a cross-departmental joint punishment mechanism. Deepen the innovation of regulatory mechanisms, fully promote the application of the "one road, multiple parties" linkage mechanism, and improve the collaborative governance model for traffic safety management of expressways, ordinary national and provincial highways, and rural roads. Establish a provincial-level road traffic safety risk assessment and early warning release platform.

[0148] 2.2 Improve the ability to govern road traffic safety according to law;

[0149] To enhance law enforcement and control efficiency, we will leverage big data, AI video analytics, and other technologies to accurately investigate and punish serious traffic violations that easily lead to accidents, such as overloading, speeding, fatigue driving, illegal passenger transport in trucks, and drunk driving. We will improve the accountability system, strictly investigate and punish dereliction of duty, malfeasance, and illegal activities related to road traffic accidents in accordance with laws and regulations, and strengthen the supervision and accountability for rectification of problems exposed by accidents.

[0150] 2.3 Create a safe and orderly road traffic environment;

[0151] We will continue to address road safety hazards. For highways, we will focus on upgrading and renovating accident-prone sections, including optimizing alignment, adding emergency escape lanes, and improving signage, markings, and intelligent sensing equipment. For rural roads, we will implement safety and life protection projects for rural roads in hilly areas, prioritizing the treatment of high-risk sections such as sharp bends, steep slopes, and roads near water or cliffs. For urban roads, we will optimize management measures such as mandatory right turns for large vehicles and improve safety facilities around schools and hospitals. We will enhance the sophistication of traffic organization, promoting the application of flexible control strategies such as dynamic lane management and time-based traffic restrictions at complex urban intersections and highway interchanges.

[0152] 2.4 Strengthen safety supervision of vehicles and drivers;

[0153] Promote the technological upgrading of key vehicles, and comprehensively promote the installation and application of intelligent video monitoring alarms, collision avoidance and other active safety technologies in passenger vehicles and hazardous material transport vehicles. Encourage the gradual installation and application of advanced driver assistance systems (ADAS) in heavy-duty trucks. Strengthen the full-lifecycle management of drivers, strictly supervise the training and examination of motor vehicle drivers, and establish a professional entry and exit mechanism for passenger and freight transport drivers. Utilize VR and AR technologies to enhance drivers' emergency response capabilities training.

[0154] 2.5 Enhance the level of technological innovation and application in road traffic safety;

[0155] Construct a provincial-level smart big data platform for road traffic safety, integrating data from transportation, public security, meteorology, and health departments to achieve full-chain management, including precise risk profiling, intelligent hazard identification, situation prediction and early warning, and performance evaluation and feedback. Deepen the application of vehicle-road cooperative technology, conducting demonstrations on sections such as the ring expressway to provide drivers with information services such as intersection collision warnings, wrong-way overtaking warnings, and severe weather warnings. Develop and apply new equipment and technologies, promoting the use of drones in rapid accident response on highways and patrol and inspection of rural roads. Explore the use of high-precision maps and BeiDou positioning technology to provide precise safety services for key vehicles.

[0156] 2.6 Improve the road traffic safety emergency rescue system;

[0157] Enhance rapid response and rescue capabilities, optimize the layout of emergency rescue stations on highways and in remote areas, and establish a "golden one-hour" rescue circle. Strengthen collaboration among fire, health, and traffic police departments, improve emergency rescue plans, and conduct regular drills. Deepen cooperation among police, medical, and fire departments, improve the rescue and treatment mechanism for traffic accident victims, establish a "green channel" for the treatment of critically injured patients, and reduce the mortality and disability rates from accidents.

[0158] 2.7 Enhance road traffic safety and civility;

[0159] Innovative methods of publicity and education will be adopted, targeting key groups such as the elderly, children, and drivers in rural areas, and utilizing new media platforms to conduct precise and targeted traffic safety publicity and education. Typical accident cases will be publicized, and in-depth investigation results of typical traffic accidents and the accountability of relevant responsible parties will be regularly released to the public to strengthen cautionary education.

[0160] 3. Key projects;

[0161] Construction project of "Smart and Safe Corridor" for expressways;

[0162] Rural road safety demonstration project with "thousands of lights and ten thousand belts";

[0163] Construction project of provincial-level road traffic safety big data center and risk monitoring and early warning platform;

[0164] Mandatory installation and application upgrade project for intelligent monitoring and alarm equipment for key vehicles;

[0165] Construction project of traffic safety publicity and education base and integrated media center;

[0166] Through a hierarchical large-scale model architecture of "natural language generation - semantic understanding - multimodal reasoning," the massive, scattered, and heterogeneous traffic data of Province S was successfully transformed into a highly scientific, accurate, and operable road traffic safety plan. This process not only improved the efficiency and intelligence of planning but also ensured, through data-driven insights, that limited resources were invested in the areas with the highest risk and greatest benefits, providing a clear technical path and decision support for road traffic safety governance in Province S over the next five years.

[0167] Step 5, Implementation process management 105;

[0168] 1. Dynamic monitoring subsystem 1051;

[0169] The operational objective of the dynamic monitoring subsystem 1051 is to perceive the physical progress and status of the road traffic safety planning implementation phase in real time, compare it with the digital plan, and promptly identify problems. In the specific implementation process, project management software (such as Primavera P6, MS Project) is used to automatically obtain Gantt charts, critical path node completion status, and resource input.

[0170] 2. Mid-term evaluation subsystem 1052;

[0171] The operational objective of the mid-term evaluation subsystem 1052 is to objectively evaluate the effectiveness of the plan at key milestones (such as six months or one year) to provide a scientific basis for subsequent adjustments.

[0172] The specific operational process includes multi-source data collection and analysis, comparative verification, and generation of a mid-term evaluation report. The first step, multi-source data collection and analysis, automatically connects to the traffic management platform to calculate core planning indicators for this phase, such as the percentage decrease in accident rate, changes in congestion index, and the detection rate of key violations. The semantic understanding module 1032 is used to crawl and analyze citizens' evaluations of traffic changes from social media and government hotlines (e.g., "Traffic efficiency improved after the B intersection renovation" is positive; "The addition of guardrails at the entrance of School C caused congestion" is negative). The second step, comparative verification, automatically compares actual data with the expected targets set in the plan, generating a gap analysis report. Based on the analysis results, the system uses the multimodal reasoning module 1033 to propose data-driven adjustment suggestions. The third step, generating a mid-term evaluation report, automatically generates a structured report that clearly displays "goals-achievements-gaps-recommendations," providing decision-makers with a reference for the next steps.

[0173] 3. Final Summary Subsystem 1053;

[0174] The operational objective of the final summary subsystem 1053 is to conduct a comprehensive review at the end of the planning cycle (e.g., 5 years), extract lessons learned, and empower the next round of planning. The specific operational process includes comprehensive effect evaluation, knowledge distillation, feedback, and iteration. The first step, comprehensive effect evaluation, involves summarizing all data from the entire implementation cycle and conducting a comprehensive analysis both longitudinally (compared to the initial planning stage) and horizontally (compared to other similar cities), calculating the overall return on investment (ROI). The second step, knowledge distillation, involves the system automatically generating "knowledge cards," for example, "In projects related to 'improving safety around schools,' the investment benefit ratio of 'adding pedestrian overpasses' is 3.5 times that of 'only painting zebra crossings.'" This identifies which measures are "best practices" and which are "ineffective measures." The third step, feedback and iteration, automatically feeds these structured "knowledge cards" and summary reports back to the knowledge base and model training module of the intelligent generation layer 104.

[0175] Step 6, Interaction and Output Layer 106;

[0176] This layer is responsible for delivering all the system's results to users in different roles in the most user-friendly and convenient way.

[0177] 1. Visual Editor 1061;

[0178] The goal of the Visual Editor 1061 is to lower the barrier to entry for professional planning systems, allowing managers to participate in planning fine-tuning based on real-time data. The specific operation process is as follows: First, a "single-map" operation allows users to operate on a 3D digital sandbox integrating CIM, BIM, GIS, and real-time data. Drag-and-drop simulation allows users to drag and drop models of "tidal flow lanes," "smart streetlights," or "secondary pedestrian safety islands" from the sidebar and place them at any intersection on the map. Second, real-time simulation and deduction: based on traffic flow models, the system simulates in real-time the impact of the facility's implementation on traffic flow, queue length, and accident risk values ​​in the surrounding road network, presenting the results in a data panel and color changes (e.g., from red to green). Third, scheme comparison: users can create multiple alternative schemes (e.g., Scheme A: build an overpass; Scheme B: dig an underpass), the system simulates in parallel, generates a comparison report to assist decision-making, facilitates collaborative editing, and supports online annotations and comments from multiple departments to form the final decision.

[0179] 2. Multi-format output module 1062;

[0180] The multi-format output module 1062 addresses the reporting, approval, and execution needs of different roles. The specific workflow includes one-click generation, public announcement, cloud synchronization and archiving. For pre-output configuration, users select the target audience and purpose from a drop-down menu. The first step, one-click generation, is for reporting to the government. Selecting "PPT Presentation" automatically generates a richly illustrated slideshow highlighting key indicators, performance comparisons, and future plans. The second step, public announcement, involves selecting "Video Report." The system automatically synthesizes planning animations, simulation results, and real-world videos, adding narration to create an easy-to-understand promotional video, which is then published on the official website for emergency command or scheme demonstration. Selecting "Interactive Electronic Sand Table" generates an independently runnable 3D scene file, allowing users to immerse themselves in the sand table and demonstrate their plans. Cloud synchronization and archiving automatically synchronizes all output files to the cloud knowledge base, adding timestamps and version numbers to form a complete planning archive for easy auditing and traceability.

[0181] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic self-adaptive management system for road traffic safety planning driven by artificial intelligence, characterized in that, Comprise: a data collection and preprocessing layer (101) for collecting multi-source heterogeneous road traffic safety related data and performing cleaning, conversion and standardization processing; a knowledge engineering layer (102) for constructing a policy knowledge graph and a rule engine according to road traffic safety related policies, regulations and standard specifications, and providing policy basis and professional knowledge support; a large model service layer (103) for natural language generation, semantic understanding and multi-modal reasoning on text data, and outputting structured analysis results; an intelligent generation layer (104) for generating multi-modal content of road traffic safety planning through a generative adversarial network model according to the structured analysis results, and dynamically adjusting planning generation strategies and supervision parameters according to real-time traffic data, environmental changes and user needs; an implementation process management layer (105) for real-time monitoring of the implementation process of road traffic safety planning, evaluating the achievement of planning goals and the effectiveness of measures in the middle stage of planning implementation, and comprehensively summarizing and evaluating the overall effect of the planning in the end stage of planning implementation; an interaction and output layer (106) for providing a visual editing interface and multi-format output function, supporting user interaction with the system and obtaining planning results.

2. The system of claim 1, wherein, The data collection and preprocessing layer (101) comprises: a multi-modal data access module (1011) for accessing accident record texts of traffic management departments, video images of road monitoring systems, traffic flow data collected by traffic sensors, and geographic information system map data; an intelligent cleaning module (1012) for cleaning, converting and standardizing the collected data, removing duplicate and incorrect information in accident records, denoising, enhancing and key frame extraction of video images, and integrating traffic flow data according to uniform time intervals.

3. The system of claim 1, wherein, The knowledge engineering layer (102) comprises: a policy knowledge graph module (1021) for collecting and organizing road traffic safety related policies, regulations and standard specifications through natural language processing technology, constructing a policy knowledge graph, and displaying the correlation, applicable scope and implementation requirements between policies in a graphical manner; a rule engine module (1022) for establishing a knowledge base containing traffic rules, safe driving knowledge and accident case analysis, and extracting key factors and rules of traffic safety through analysis and summary of historical accident cases.

4. The system of claim 1, wherein, The large model service layer (103) comprises: a natural language generation module (1031) for generating planning reports through fine-tuning and prompt engineering technology based on pre-trained large language models; a semantic understanding module (1032) for performing semantic understanding on collected text data through deep learning algorithms, extracting key information, intent and sentiment orientation; a multi-modal reasoning module (1033) for fusing multi-modal data including text, image, video and numerical value, and utilizing artificial intelligence reasoning algorithms for comprehensive analysis and output of analysis results.

5. The system of claim 1, wherein, The intelligent generation layer (104) comprises: The multi-modal generation module (1041) is configured to generate multi-modal content of the road traffic safety planning by a generative adversarial network model according to the analysis result provided by the large model service layer (103), wherein the multi-modal content includes visual display of the planning scheme, voice explanation and virtual reality scene simulation. The dynamic adaptation engine module (1042) is configured to dynamically adjust a planning generation strategy and a supervision parameter according to real-time traffic data, environmental changes and user demand, and optimize the generated planning scheme based on a reinforcement learning algorithm.

6. The system of claim 1, wherein, The implementation process management layer (105) includes: The dynamic monitoring subsystem (1051) is configured to perform real-time monitoring on the implementation process of the road traffic safety planning by an Internet of Things technology and a sensor network, monitor construction progress of traffic facilities, installation of traffic signs and operation state of traffic signal lights, and issue a warning when a problem in the implementation process is found. The mid-term evaluation subsystem (1052) is configured to perform evaluation on achievement of a planning target and effectiveness of a measure in a mid-term stage of the planning implementation by collecting and analyzing relevant data, and propose adjustment suggestions by an artificial intelligence algorithm according to an evaluation result. The final summary subsystem (1053) is configured to comprehensively summarize and evaluate overall effects of the planning in a final stage of the planning implementation, and feed back an evaluation result to the knowledge base.

7. The system of claim 1, wherein, The interaction and output layer (106) includes: The visual editor (1061) is configured to provide a visual interactive interface, and a user queries, edits, modifies and optimizes road traffic safety planning content through the visual interactive interface. The multi-format output module (1062) is configured to output a final planning result and a supervision report of the system, and formats of the final planning result and the supervision report include an electronic sand table, a PDF document, a PPT presentation and a video report.

8. The system of claim 4, wherein, The expression of the comprehensive analysis is: ； In the formula, represents the comprehensive reasoning result, represents a multi-modal reasoning function, represents a multi-source data set, a multi-modal reasoning model parameter, and represents a feature extraction function. ⨁ denotes a multi-modal fusion operator, represents that the prediction function receives the fused features as input, represents the generated comprehensive inference result.

9. The system of claim 5, wherein, The calculation expression of the multi-modal generation module (1041) is: ； wherein denotes a generator, denotes a discriminator, denotes a data distribution of historical planning scenarios, denotes a prior normal distribution of random noise, denotes a value function, denotes a mathematical expectation, denotes a random variable.

Images