Structured security assessment report intelligent generation system, method and device

Abstract

The invention discloses a structured safety assessment report intelligent generation system, method and device, and relates to the technical field of road traffic safety. Comprising a data input and preprocessing module, a multi-modal semantic mapping and understanding engine, a domain knowledge base module, an LLM-based report generation and suggestion engine, a dynamic compliance verification and alignment module and a structured report output and management module. The problems that a traditional road safety assessment report is low in writing efficiency, poor in normalization, insufficient in data utilization, lack of dynamic compliance verification, insufficient in operable suggestions and the like are solved, the system comprises a data input and preprocessing module, a multi-modal semantic mapping and understanding module and the like, and through the technologies of a Z-Score algorithm, LLM field fine adjustment, a graph neural network, dynamic compliance verification and the like, the real-time performance of the road safety assessment report is improved. Functions of data cleaning, multi-modal semantic fusion, compliance report generation and the like are realized, the report period is automatically shortened in the whole process, the content specialty, compliance and suggestion operability are improved, and the method is suitable for highway traffic safety management.

Description

Technical Field

[0001] This invention relates to the field of highway traffic safety technology, specifically to a structured safety assessment report intelligent generation system, method, and equipment. Background Technology

[0002] Highway safety assessment reports are crucial for highway safety management and decision-making. Traditionally, the writing of highway safety assessment reports relied heavily on manual labor, which presents the following problems:

[0003] 1. Low efficiency and time-consuming: Manually organizing and analyzing data and writing reports according to specifications is a cumbersome and time-consuming process that cannot meet the management needs of rapid response.

[0004] 2. High subjectivity and difficulty in ensuring standardization: The quality of the report is easily affected by the experience level and subjective judgment of the writer. There may be deviations in the understanding and implementation of industry standards such as the "Specification for Safety Evaluation of Highway Projects", resulting in insufficient objectivity and standardization of the report.

[0005] 3. Insufficient data utilization and low level of intelligence: Faced with increasingly complex road detection data and AI analysis results (such as hazard identification information and risk assessment data), it is difficult for humans to efficiently and comprehensively integrate them into reports and transform them into structured content with in-depth insights and clear guidance.

[0006] 4. Lack of dynamic compliance verification: During the report generation process, especially when making safety improvement recommendations, there is often a lack of effective mechanisms to verify in real time whether these recommendations fully comply with the latest design standards, construction specifications and mandatory provisions of traffic management policies, which may lead to the recommendations lacking feasibility or posing compliance risks.

[0007] 5. Insufficient ability to generate actionable recommendations: Improvement recommendations in traditional reports are sometimes too general and lack targeted, specific, prioritized or resource-optimized actionable solutions, such as detailed emergency evacuation route planning or priority lists for safety facility upgrades.

[0008] Although there are some existing text generation tools or attempts to use AI for assisted analysis, in the professional field of highway safety assessment reports, how to deeply integrate the powerful text generation capabilities of large language models with domain knowledge (especially specific national or industry standards) and combine them with dynamic compliance verification mechanisms to ensure the professionalism, authority, and practicality of the generated reports remains a pressing technical challenge.

[0009] Compared to evaluation reports in fields such as building safety and chemical safety, highway traffic safety assessment reports have significant unique characteristics: their assessment objects not only involve static road facilities but also require the integration of dynamic traffic flow data (such as traffic volume, speed, and traffic violations), historical accident data (such as accident types and causal chains), and complex interactions among traffic participants (such as conflicts arising from mixed traffic of motor vehicles and non-motor vehicles); the evaluation criteria must strictly adhere to traffic-specific regulations such as the "Highway Route Design Specifications" and "Road Traffic Signs and Markings," and it is recommended that they possess real-time capabilities (such as emergency traffic management) and scenario adaptability (such as severe weather control). Existing technologies have failed to specifically address these unique data processing, semantic association, and compliance verification needs in the field of traffic safety, resulting in reports lacking specific traffic scenario characteristics.

[0010] In view of this, a structured security assessment report intelligent generation system, method and device are provided to overcome the above problems. Summary of the Invention

[0011] The purpose of this invention is to provide a structured security assessment report intelligent generation system, method, and device to solve the problems mentioned in the background art.

[0012] To address the aforementioned technical issues, this invention provides an intelligent generation system for structured security assessment reports, comprising: a data input and preprocessing module, a multimodal semantic mapping and understanding engine, a domain knowledge base module, an LLM-based report generation and suggestion engine, a dynamic compliance verification and alignment module, and a structured report output and management module.

[0013] The data input and preprocessing module is used to receive and parse the risk data output by the upstream AI analysis engine, and to clean, convert the format and extract key information.

[0014] The multimodal semantic mapping and understanding engine is connected to the data input and preprocessing module, and is used to perform deep semantic understanding and contextual association on the preprocessed multi-source heterogeneous risk data, and transform it into a unified semantic representation that is easy for large language models to process;

[0015] The domain knowledge base module is used to store structured industry standards, specifications, expert experience, and domain ontology knowledge related to highway safety assessment.

[0016] The LLM-based report generation and suggestion engine is connected to the multimodal semantic mapping and understanding engine and the domain knowledge base module, and is used to generate evaluative text and intelligent suggestions by combining the unified semantic representation and domain knowledge;

[0017] The dynamic compliance verification and alignment module is connected to the LLM-based report generation and suggestion engine and associated with the specification library in the domain knowledge base module, and is used to verify and correct the compliance of the generated evaluative text and intelligent suggestions.

[0018] The structured report output and management module is connected to the dynamic compliance verification and alignment module and is used to generate structured reports according to preset templates.

[0019] Furthermore, the data input and preprocessing module preprocesses the data through the following steps:

[0020] Step 1: Overall goals and rules for data preprocessing, including:

[0021] Data cleaning: Removes duplicate records, invalid characters, and obviously erroneous data using preset regular expressions and outlier detection algorithms;

[0022] Format conversion: Convert XML and CSV format data to the system's internal JSON format;

[0023] Key information extraction: Using named entity recognition technology in natural language processing, core fields including hazard type, location, and severity are extracted;

[0024] Step 2: Focus on outlier detection targets, quantify the degree of deviation of a single data point from the overall distribution of the dataset, and identify erroneous or abnormal data that significantly deviates from the normal range;

[0025] Step 3: Introduce data distribution characteristic parameters and calculate the mean of the dataset. and standard deviation ,in:

[0026] mean The calculation formula is:

[0027] ;

[0028] In the formula, For the first in the dataset The value of each data point The number of samples in the dataset;

[0029] Standard deviation The calculation formula is:

[0030] ;

[0031] In the formula, The squared deviation of a single data point from the mean;

[0032] Step 4: Derive the standardized deviation index (Z-Score), the calculation formula is as follows:

[0033] ;

[0034] In the formula, The data point value to be detected. The mean of the dataset. is the standard deviation of the dataset.

[0035] Step 5: Set an outlier detection threshold. Based on statistical principles, when... When this happens, the data point is determined to be an outlier;

[0036] Step 6: Handle outliers, remove data points that are determined to be outliers, and retain the cleaned data for subsequent analysis.

[0037] Furthermore, the multimodal semantic mapping and understanding engine integrates LLM and document structuring tools, and employs attention mechanisms and graph neural networks for semantic fusion of multimodal data. The node feature update formula of the graph neural network is as follows:

[0038] ;

[0039] in:

[0040] For the first The node feature matrix of the layer;

[0041] Add self-loops to adjacency matrix A This indicates the relationship between a node and itself.

[0042] for The degree matrix;

[0043] For the first Layer weight matrix;

[0044] This is the activation function.

[0045] Furthermore, the risk-measure mapping library in the domain knowledge base module uses the Analytic Hierarchy Process (AHP) to determine the priority of governance measures. The AHP includes weight vector calculation and consistency verification, wherein:

[0046] The formula for calculating the weight vector is:

[0047] ;

[0048] In the formula:

[0049] : No. The weight of each risk factor;

[0050] : The first in the judgment matrix The factor and the first The relative importance of each factor;

[0051] The number of risk factors;

[0052] The formula for calculating the consistency index is:

[0053] .

[0054] The formula for calculating the consistency ratio is:

[0055] .

[0056] In the formula:

[0057] To determine the largest eigenvalue of a matrix;

[0058] This is a random, one-time indicator;

[0059] When the value is less than 0.1, the consistency of the judgment matrix is ​​considered acceptable.

[0060] Furthermore, the LLM-based report generation and suggestion engine uses BLEU-4 and ROUGE-L metrics to evaluate the similarity between the generated text and the standard template, wherein:

[0061] The formula for calculating BLEU-4 is:

[0062] ;

[0063] in The formula for calculating the penalty for short sentences is:

[0064] ;

[0065] The formula for calculating accuracy is:

[0066] ;

[0067] In the formula:

[0068] : The number of words in the generated text;

[0069] : Word count of the reference text;

[0070] : The weight, usually ;

[0071] : Matching accuracy;

[0072] : In the generated text Maximum number of occurrences in the reference text;

[0073] The number of reference texts;

[0074] The formula for calculating ROUGE-L is:

[0075] ;

[0076] Recall rate:

[0077] ;

[0078] Accuracy:

[0079] ;

[0080] in:

[0081] : The length of the longest common subsequence between the generated text and the reference text;

[0082] : The number of words involved in the text;

[0083] : The number of words in the generated text;

[0084] Weight parameter, usually set to 1, emphasizing the balance between recall and precision.

[0085] Furthermore, the dynamic compliance verification and alignment module has a built-in rule engine based on mandatory clauses of regulations and standards. It uses a forward chain reasoning algorithm to match the generated suggestions with the normative clauses one by one. Non-compliance items are classified into three levels according to their severity: urgent, important and general. The corrected content needs to be verified again.

[0086] A method for intelligently generating structured security assessment reports, applied to the aforementioned intelligent generation system for structured security assessment reports, includes the following steps:

[0087] Step 1: Risk Data Reception and Preprocessing: Receive and preprocess the structured / semi-structured data output from the upstream AI engine, wherein outlier detection adopts the Z-Score algorithm and the steps described in claim 2;

[0088] Step 2: Multimodal semantic understanding and knowledge representation: Using the Transformer architecture and cross-modal attention mechanism, combined with graph neural networks, the preprocessed data pairs are semantically fused.

[0089] Step 3: LLM and Knowledge Base-Based Report Drafting and Suggestion Generation: The LLM engine interacts with the domain knowledge base to generate a draft report and suggestions, and uses BLEU-4 and ROUGE-L metrics to evaluate and optimize the text;

[0090] Step 4: Dynamic compliance verification and content alignment: The draft report is automatically verified in real time, non-compliance items are identified and graded according to the rule engine, and secondary verification is initiated.

[0091] Step 5: Finalizing and Outputting the Structured Report: Set up a three-level review process. After the review is approved, generate an electronic document according to the template and push it out.

[0092] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent generation method for structured security assessment reports.

[0093] Compared with the prior art, the beneficial effects of the present invention are:

[0094] 1. This invention uses AI technology to automate the entire process from data input to compliance report output, significantly shortening the report writing cycle, freeing professionals from heavy repetitive text work, improving work efficiency, and significantly enhancing report generation efficiency and automation level.

[0095] 2. This invention deeply integrates industry standards and expert knowledge, and utilizes the professional text understanding and generation capabilities of large-scale language models to effectively ensure the professional level of the assessment report content, the accuracy of terminology usage, and the standardization of the overall structure. It reduces the subjectivity and arbitrariness of manual writing and ensures the professionalism and standardization of the report content.

[0096] 3. This invention, through its built-in dynamic compliance verification and alignment module, can ensure in real time that the generated key content, such as security improvement suggestions, automatically complies with and aligns with the mandatory requirements of current laws and regulations, thereby enhancing the authority and reliability of the report, as well as the feasibility and legality of the suggestions in actual work.

[0097] 4. Based on risk assessment results and a domain knowledge base, this invention can intelligently generate specific, detailed, and prioritized actionable suggestions, including emergency evacuation route planning schemes, priority lists for safety facility renovations, and resource allocation optimization recommendations. This provides stronger support for management departments to make accurate decisions and take efficient actions, and enhances the intelligence and operability of improvement suggestions.

[0098] 5. This invention possesses powerful deep semantic understanding and information integration capabilities. Its multimodal semantic mapping and understanding engine can perform deep semantic understanding and effective integration of complex, multi-source AI analysis results, providing high-quality, high-information-density input for large-scale language models, thereby significantly improving the accuracy, comprehensiveness, and depth of the generated report content.

[0099] 6. This invention can quickly and accurately generate high-quality, standardized safety assessment reports, providing management departments with a powerful technical tool to grasp the road safety status in a timely and comprehensive manner, and to scientifically formulate maintenance plans and management strategies. It helps to achieve intelligent closed-loop management of the entire "perception-assessment-response" chain and improve the overall traffic safety management efficiency.

[0100] 7. The large-scale language model used in this invention can be continuously optimized and fine-tuned for specific corpora and application scenarios in the field of highway safety assessment, making it perform better than general text generation models in professional fields and better adaptable to the needs of specific environments such as rural roads.

[0101] 8. This invention designs a dedicated extraction and association algorithm for traffic flow time-series data and accident source tracing data, which differs from the static parameter processing in the construction field; the domain knowledge base has built-in traffic conflict point identification, line-of-sight verification and other 12 types of traffic safety-specific rules, which are significantly different from equipment corrosion standards in the chemical industry; the output dynamic traffic control schemes, accident prevention measures and other content directly respond to the collaborative safety needs of people, vehicles, roads and environment in traffic scenarios, which are different from general engineering suggestions in other fields, and are distinguished from evaluation report generation methods in other fields through three dimensions. Attached Figure Description

[0102] Figure 1 This is a schematic diagram of a structured security assessment report intelligent generation system, method, and device according to the present invention.

[0103] Figure 2 This is a system block diagram of the generation system in the intelligent generation system, method and device for structured security assessment reports of the present invention. Detailed Implementation

[0104] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0105] See Figure 1 and Figure 2 As shown, an embodiment of a structured security assessment report intelligent generation system, method, and device is provided:

[0106] System Implementation Example:

[0107] A structured security assessment report intelligent generation system includes: a data input and preprocessing module, a multimodal semantic mapping and understanding engine, a domain knowledge base module, an LLM-based report generation and suggestion engine, a dynamic compliance verification and alignment module, and a structured report output and management module;

[0108] The data input and preprocessing module is used to receive and parse the risk data output by the upstream AI analysis engine, and to perform cleaning, format conversion and key information extraction.

[0109] The multimodal semantic mapping and understanding engine is connected to the data input and preprocessing module, and is used to perform deep semantic understanding and context association on the preprocessed multi-source heterogeneous risk data, and transform it into a unified semantic representation that is easy for large language models to process;

[0110] The domain knowledge base module is used to store structured industry standards, specifications, expert experience, and domain ontology knowledge related to highway safety assessment.

[0111] The LLM-based report generation and suggestion engine is connected to the multimodal semantic mapping and understanding engine and the domain knowledge base module, and is used to generate evaluative text and intelligent suggestions by combining the unified semantic representation and domain knowledge;

[0112] The dynamic compliance verification and alignment module is connected to the LLM-based report generation and suggestion engine and associated with the specification library in the domain knowledge base module. It is used to perform compliance verification and correction on the generated evaluation text and intelligent suggestions.

[0113] The structured report output and management module is connected to the dynamic compliance verification and alignment module and is used to generate structured reports according to preset templates.

[0114] Specifically, the data input and preprocessing module uses the Z-Score algorithm to detect outliers. When the Z-Score value of a data point is greater than 3, it is determined to be an outlier. The module also uses named entity recognition technology in natural language processing to extract fields such as hazard type, location, and severity.

[0115] Specifically, the multimodal semantic mapping and understanding engine integrates LLM and document structuring tools, and uses attention mechanisms and graph neural networks to unify and merge different modal data such as text descriptions, numerical indicators, and AI analysis conclusions at the semantic level, constructing a relationship graph between hidden dangers and road attributes, and extracting the causal relationships between core risk elements.

[0116] Specifically, the specification library in the domain knowledge base module uses knowledge graph technology to store regulations, breaking down clauses into nodes and relationships; the risk-measure mapping library constructs association rules based on decision tree algorithm, determines governance measures according to conditions such as hazard type, risk level, and road attributes, and uses AHP hierarchical analysis method to determine priorities; the expert experience and case library uses cosine similarity algorithm to screen cases.

[0117] Specifically, the LLM-based report generation and suggestion engine uses the Qwen series model, which has been fine-tuned with a corpus in the field of highway safety. By adding normative constraint prompts to the input, it ensures that the generated content complies with regulatory standards. It then combines risk assessment results and a domain knowledge base to generate specific emergency evacuation route planning schemes.

[0118] Specifically, the dynamic compliance verification and alignment module has a built-in rule engine based on mandatory clauses of laws and regulations. It uses a forward chain reasoning algorithm to match the generated suggestions with the normative clauses one by one. Non-compliance items are classified into three levels according to their severity: urgent, important and general. The corrected content needs to be verified again.

[0119] This system consists of the following modules working together to achieve intelligent generation of highway safety assessment reports:

[0120] The data input and preprocessing module is responsible for receiving and parsing various risk data output from upstream AI analysis engines (such as the "Multimodal Perception Data Fusion and Hazard Identification Engine" and the "Risk Intelligent Prediction and Decision Engine"). This includes structured and semi-structured data such as hazard identification information and risk assessment indicators. Input data types include, but are not limited to, identified safety hazard information (such as type, location, image / video evidence links, and severity descriptions), quantified risk assessment results (such as the probability of accidents on specific road sections and data on the distribution of traffic violation hotspots), and dynamic risk prediction information.

[0121] Preprocessing:

[0122] Step 1: Overall Goals and Rules for Data Preprocessing

[0123] The core objective of the data input and preprocessing module is to clean, convert, and extract key information from the risk data (including structured and semi-structured data) output by the upstream AI analysis engine, providing high-quality data for subsequent multimodal semantic understanding and report generation. Specific preprocessing rules are as follows:

[0124] Data cleaning: Removes duplicate records, invalid characters, and obviously erroneous data using preset regular expressions and outlier detection algorithms;

[0125] Format conversion: Convert data in various formats such as XML and CSV to the system's internal JSON format to ensure data structure consistency;

[0126] Key information extraction: Using Named Entity Recognition (NER) technology in Natural Language Processing, core fields such as hazard type, location, and severity are accurately extracted.

[0127] Step 2: Focus on outlier detection targets

[0128] In the data cleaning process, outlier detection is one of the core tasks. The goal is to quantify the degree of deviation of a single data point from the overall distribution of the dataset in order to accurately identify erroneous or abnormal data that deviates significantly from the normal range (such as the severity of erroneous potential hazards, abnormal risk assessment indicators, unreasonable proportion of traffic violations, etc.) and avoid abnormal data from interfering with subsequent analysis.

[0129] Step 3: Introduce data distribution characteristic parameters

[0130] To describe the overall distribution of the dataset, two basic statistics need to be calculated first:

[0131] 1. The mean of the dataset The formula reflects the central trend of the data and is as follows:

[0132] ;

[0133] in, For the first in the dataset The value of a data point (such as the "percentage of speeding behaviors" detected multiple times on a certain road segment). This represents the number of samples in the dataset.

[0134] 2. Standard deviation of the dataset : Reflects the degree of dispersion of the data, and the calculation formula is:

[0135] ;

[0136] in, The standard deviation is the squared deviation of a single data point from the mean. The larger the standard deviation, the more dispersed the data distribution (e.g., the greater the fluctuation in risk scores across different road sections).

[0137] Step 4: Derive the standardized deviation index (Z-Score)

[0138] Single data point The absolute deviation from the mean is However, this deviation is affected by the magnitude of the data itself (e.g., the deviation of a risk score of 90 from the mean and the deviation of hazard location coordinates from the mean cannot be directly compared). To eliminate magnitude interference and achieve a unified measurement of the degree of deviation for different types of data, the deviation needs to be standardized: dividing the absolute deviation by the standard deviation of the dataset yields the standardized deviation index (Z-Score).

[0139] ;

[0140] in, The data point value to be detected. The mean of the dataset. is the standard deviation of the dataset.

[0141] Step 5: Set the outlier detection threshold

[0142] According to statistical principles, if a dataset approximately follows a normal distribution, about 99.7% of the data points will fall within the mean. Within the range (i.e.) Data points outside this range are considered low-probability events (probability < 0.3%), therefore a judgment rule is set: when... At that time, the data point was identified as an outlier.

[0143] Step Six: Real-world Data Application Examples

[0144] Taking the risk data processing of a rural road section from K1+000 to K1+500 as an example, the specific process is as follows:

[0145] 1. Data Collection: Receives various types of data output from the upstream AI engine, including:

[0146] Hazard list (JSON format): [{"Hazard ID":"H001","Type":"Pit","Location":"K1+150","Severity":"Medium","Image Evidence":"img_k1_150_pit.jpg"},{"Hazard ID":"H002","Type":"Tilted Sign","Location":"K1+200","Description":"Speed ​​limit sign tilted inward by about 15 degrees","Video Evidence":"vid_k1_200_sign.mp4"}];

[0147] Risk assessment summary: "The overall risk score for this road section is 75 (high risk), with the main risk contributions coming from poor road surface conditions and inadequate maintenance of traffic facilities. The probability of a traffic safety incident caused by potholes in the next 30 days is predicted to be 0.05%."

[0148] Road attribute data: "Road grade: Class IV highway; Design speed: 40km / h; Pavement type: Asphalt concrete; Location: Mountainous area, near villages";

[0149] Traffic violation hotspots: "Speeding accounted for 20% of violations on the K1+300-K1+400 section" (assuming multiple detection data are 18%, 20%, 21%, 19%, and 50%).

[0150] 2. Format conversion and key information extraction:

[0151] Convert the above data into JSON format;

[0152] NER technology was used to extract key information such as "potholes", "K1+150", "middle", "signboard tilt", and "20%".

[0153] 3. Outlier detection (taking "percentage of speeding behavior" as an example):

[0154] Calculate the mean: ;

[0155] Calculate the standard deviation: ;

[0156] Calculate the Z-Score: Calculated for 50% yields:

[0157] (Threshold not reached);

[0158] If there is an incorrectly entered "100% overspeed percentage", then:

[0159] It was determined to be an outlier.

[0160] 4. Outlier handling: Remove outliers such as "100% overspeed percentage" and retain the cleaned data for subsequent analysis.

[0161] By gradually building a Z-Score outlier detection system:

[0162] First, use the mean and standard deviation Describe the data distribution characteristics (Step 3);

[0163] Then standardized To achieve a unified measurement of the degree of deviation of different types of data (step four);

[0164] Ultimately, it was set based on statistical laws. Outlier criteria (Step 5), and complete the cleaning process by combining the actual data (Step 6).

[0165] It provides a reliable data foundation for subsequent processes such as multimodal semantic fusion and LLM report generation.

[0166] Multimodal semantic mapping and understanding engine: Connected to the data input and preprocessing module, it integrates advanced LLM (such as Q wen series models optimized for specific tasks) and document structuring tools (such as Dolphin) to perform deep semantic understanding and context association on preprocessed multi-source heterogeneous risk data, transforming it into a unified semantic representation that is easy for large language models to process.

[0167] Working Mechanism: An attention mechanism and graph neural networks (GNNs) are employed to unify and integrate different modalities of data, such as text descriptions, numerical indicators, and AI analysis conclusions, at the semantic level. For example, a relationship graph between hazards and road attributes is constructed using GNNs to extract causal relationships between core risk elements; leveraging the natural language understanding capabilities of LLMs, the text descriptions of image / video evidence are semantically aligned with hazard information, ultimately forming a comprehensive scene knowledge representation containing information such as "potholes at K1+150 were caused by rainfall in the mountainous area, exacerbating the driving risk on roads near nearby villages."

[0168] In multi-model semantic mapping, a graph of the relationship between hazards and road attributes is constructed using a graph neural network (GNN). The node feature update formula is as follows: Graph Neural Network (GNN) node embedding formula:

[0169] BNN layer update formula (taking GCN as an example);

[0170] ;

[0171] in:

[0172] : No. Layer node feature matrix (such as semantic vectors of nodes like "road potholes" and "mountain road sections").

[0173] Adjacency matrix A plus self-loops This indicates the relationship between a node and itself.

[0174] : The degree matrix (diagonal matrix) (used for normalization).

[0175] :No. The layer's weight matrix (learned through training to transform features).

[0176] : Activation function (such as ReLU, referring to nonlinear transformation).

[0177] Domain Knowledge Base Module: Stores structured industry standards, specifications, expert experience, and domain ontology knowledge related to highway safety assessment, providing decision support for LLM.

[0178] Standards Library: Utilizing knowledge graph technology to store regulations such as the "Highway Project Safety Evaluation Standard," clauses are broken down into nodes and relationships, supporting rapid retrieval and logical reasoning. For example, when searching for the clause "sign installation angle," relevant design standards and acceptance specifications can be linked.

[0179] Risk-Measure Mapping Library: This library constructs association rules based on decision tree algorithms, determining remediation measures according to factors such as hazard type, risk level, and road attributes. For example, "potholes on a Class IV mountain highway with medium risk" corresponds to the measure "repairing with hot patching within 15 days, followed by a smoothness test." Priority logic is determined using the Analytic Hierarchy Process (AHP), and applicable conditions are judged through semantic matching algorithms.

[0180] In determining the priority of the risk-measure mapping library, the Analytic Hierarchy Process (AHP) is used for logical derivation, and its core calculation process is as follows:

[0181] Weight vector calculation:

[0182] formula:

[0183] ;

[0184] in:

[0185] : No. The weight of each risk factor;

[0186] : The first in the judgment matrix The factor and the first The relative importance of each factor, such as the severity of the hazard and the cost of repair;

[0187] The number of risk factors;

[0188] Consistency check:

[0189] formula:

[0190] Consistency Indicators:

[0191] .

[0192] Consistency ratio:

[0193] .

[0194] in:

[0195] : Determine the largest eigenvalue of a matrix;

[0196] Random, one-time indicator, based on Retrieve from table, such as hour ;

[0197] When the value is less than 0.1, the consistency of the judgment matrix is ​​considered acceptable.

[0198] Expert experience and case library: The cosine similarity algorithm is used to screen cases. When the system deals with the "signboard tilt" hazard, it automatically matches the treatment solutions for similar road sections and similar tilt angles in historical cases to help LLM generate more accurate suggestions.

[0199] The LLM-based report generation and suggestion engine is the core intelligent processing unit of the system. It adopts the Qwen series model, which has been fine-tuned with corpus data in the highway safety field, and interacts closely with the output of the multimodal semantic mapping and understanding engine and the domain knowledge base module.

[0200] Evaluation Text Generation: By adding regulatory constraint prompts to the LLM input, such as "Describe the safety status of this road section according to the requirements of the 'Highway Project Safety Evaluation Specification'", the generated descriptive text, analytical content, and evaluation conclusions are ensured to comply with regulatory standards. For example, in the hazard investigation and analysis section, based on the requirement in the specification library that "potholes deeper than 3cm require special explanation", a detailed description is given: "The pothole depth at K1+150 is approximately 5cm, which exceeds the safety threshold stipulated in the 'Highway Maintenance Technical Specification' and seriously affects driving smoothness and safety."

[0201] Intelligent suggestion generation: Taking emergency evacuation route planning as an example, LLM combines risk assessment results (such as "high-risk road section with an accident probability of 0.05%), obtains rules from the risk-measure mapping library, and then calls geographic information system (GIS) data to generate a specific plan such as "It is recommended to set up temporary traffic control points at K1+000 and K1+500 to guide vehicles to detour via XX village road, which is only 2 kilometers away from the main road and has good road conditions".

[0202] Dynamic compliance verification and alignment module: It is closely connected to the output of the LLM report generation and suggestion engine and associated with the specification library in the domain knowledge base. It has a built-in rule engine (such as Drools) based on the mandatory clauses of standards such as the "Specification for Safety Evaluation of Highway Projects".

[0203] Verification logic: A forward chain reasoning algorithm is used to match the recommendations generated by LLM with the regulatory clauses one by one. For example, for the recommendation to "add warning signs", not only is it checked whether the placement location meets the visibility requirements, but also natural language processing technology is used to analyze whether the sign content contains the necessary information required by regulations (such as hazard type and distance indication).

[0204] Non-compliance handling: Non-compliance is categorized into three levels based on severity (urgent, important, and minor). Urgent non-compliance (such as serious violations of the repair deadline) is immediately reported to the LLM engine for correction; important non-compliance (such as non-standard logo graphics) requires supplementary explanation before output; minor non-compliance (such as imprecise wording) is marked with a warning in the report. The corrected content must be validated again to ensure 100% compliance.

[0205] Structured report output and management module: Processes the final report content after verification and alignment according to the preset template of "Standard Format of Rural Road Safety Evaluation Report in XX Province" which conforms to the requirements of "Specifications for Safety Evaluation of Highway Projects".

[0206] Template Design and Updates: The template adopts an XML structured design, which facilitates flexible adjustment of chapter structure and style; the update mechanism is based on monitoring of regulatory changes, and automatically updates the format and expression requirements of relevant content in the template when the "Technical Specifications for Highway Maintenance" is revised.

[0207] Version Management: Adopting the Git version control concept, each reported modification generates a new version number (such as V1.0, V1.1), recording the modification time, modifier, and modified content; users can view historical versions at any time through the SaaS platform and compare the differences between different versions.

[0208] Method Implementation Example: A method for intelligently generating structured security assessment reports, applied to the aforementioned intelligent generation system for structured security assessment reports, includes the following steps:

[0209] Step 1: Risk Data Reception and Preprocessing: Receive and preprocess structured / semi-structured data output from the upstream AI engine;

[0210] Step 2: Multimodal semantic understanding and knowledge representation: Using the Transformer architecture and cross-modal attention mechanism, deep semantic analysis and cross-modal information fusion are performed on the preprocessed data;

[0211] Step 3: LLM and Knowledge Base-Based Report Drafting and Suggestion Generation: The LLM engine interacts with the domain knowledge base to query relevant knowledge, generate draft reports and suggestions, and uses BLEU-4 and ROUGE-L metrics to evaluate the similarity between the generated text and the standard template and optimize it.

[0212] Step 4: Dynamic compliance verification and content alignment: The draft report is automatically verified in real time. The rule engine in the specification library makes judgments and adopts a graded processing strategy for non-compliance items of different severity, and initiates secondary verification.

[0213] Step 5: Finalizing and Outputting the Structured Report: Set up a three-level review process. After approval, generate an electronic document file according to the template and push it to the user.

[0214] Taking the assessment of a rural road as an example, the specific method is as follows:

[0215] Step 1: Risk Data Reception and Preprocessing: The system receives structured / semi-structured data from the upstream AI engine through the data input and preprocessing module. The initial processing standards are: data integrity rate ≥ 95%, error rate < 5%, and key information extraction accuracy rate ≥ 90%. The processing includes: using data cleaning tools (such as OpenRefine) to remove duplicate and erroneous information; converting the data to JSON using a format conversion script; and using Python's spaCy library to extract key information such as the location and severity of potential hazards.

[0216] Step 2: Multimodal Semantic Understanding and Knowledge Representation: The multimodal semantic mapping and understanding engine adopts the Transformer architecture and cross-modal attention mechanism to perform deep semantic analysis and cross-modal information fusion on the preprocessed data. For example, the feature vector of the potential hazard image is associated with the text description through the attention mechanism to generate a structured knowledge representation containing image semantics, which serves as the input of LLM.

[0217] Step 3: Report Drafting and Suggestion Generation Based on LLM and Knowledge Base: The LLM engine interacts with the domain knowledge base via HTTP API to query relevant knowledge. For quality assessment, BLEU-4 and ROUGE-L metrics are used to evaluate the similarity between the generated text and the standard template. During optimization, for content with low scores, reinforcement learning algorithms are used to adjust the LLM generation parameters until the requirements are met.

[0218] In the evaluation of the similarity between the report text and the standard module, the BLEU-4 index is used for quantitative calculation, and its formula is as follows:

[0219] ;

[0220] Brevity penalty (BP) short sentence punishment:

[0221] ;

[0222] Accuracy:

[0223] ;

[0224] in:

[0225] : The number of words in the generated text;

[0226] : Word count of the reference text;

[0227] : The weight, usually ;

[0228] : Matching accuracy;

[0229] : In the generated text Maximum number of occurrences in the reference text;

[0230] The number of reference texts;

[0231] In addition to BLEU-4, the ROUGE-L metric is used to further evaluate the semantic relevance of the text. The calculation logic is as follows:

[0232] ROUGE-L indicator formula:

[0233] ;

[0234] Recall rate:

[0235] ;

[0236] Accuracy:

[0237] ;

[0238] in:

[0239] : The length of the longest common subsequence between the generated text and the reference text;

[0240] : The number of words involved in the text;

[0241] : The number of words in the generated text;

[0242] Weight parameter, usually set to 1, emphasizing the balance between recall and precision.

[0243] Step 4: Dynamic Compliance Verification and Content Alignment: The compliance module performs real-time automatic verification of the draft report, judging based on the rule engine in the specification library. For non-compliance items of different severity, a tiered processing strategy is adopted: urgent items trigger an immediate correction process; important items generate a list of correction suggestions; general items are logged and the user is notified. After automatic alignment, a second verification is initiated to ensure that all content complies with the specifications.

[0244] Step 5: Finalizing and Outputting the Structured Report: Before delivering the structured report to the user, the output and management module sets up a three-level review process (initial review, expert review, and final review). Different roles have different permissions (e.g., the initial reviewers can only view the report, while experts can modify some content). After the review is approved, a PDF file is generated according to the template, pushed to the user through the SAAS platform, and a notification email containing the report link and version number is sent.

[0245] Equipment Example:

[0246] A computer device includes a memory, a processor, and a computer program, the computer program being stored in the memory and capable of running on the processor, wherein the processor, when executing the computer program, implements the intelligent generation method for structured security assessment reports.

[0247] Hardware architecture:

[0248] Server / Cloud Deployment: A distributed architecture is adopted, with the data processing module deployed on a cluster of computing nodes, the LLM computing engine using a GPU-accelerated cluster, and the knowledge base storage server using a distributed file system (such as Ceph) to ensure high availability and scalability of the system.

[0249] Processors: The data processing nodes use Intel Xeon Platinum series CPUs, which support multi-threaded parallel processing; the LLM inference nodes are equipped with NVIDIA A100 GPUs, which utilize CUDA acceleration technology to improve model computation efficiency.

[0250] Storage: High-speed DDR4 memory with a capacity of 128GB or more ensures fast read and write speeds for multimodal data processing; the database uses a MySQL cluster (master-slave architecture) to store structured data and MongoDB to store unstructured expert cases and multimedia evidence, improving query performance through data sharding technology.

[0251] Input / Output Interfaces: The data interface uses the gRPC framework to implement REST API, supporting high-concurrency data transmission; the report output interface integrates the Apache PDFBox and Docx4j libraries to generate PDF / Word documents and connects to the SaaS platform API via the OAuth2.0 protocol.

[0252] Software components:

[0253] Operating System: CentOS 8 is used for compute nodes, with kernel parameters optimized to adapt to high-concurrency data processing; Ubuntu 20.04 is used for GPU nodes, with CUDA and cuDNN drivers installed to ensure the efficient operation of LLM services.

[0254] Core software:

[0255] LLM Service: The Qwen model service is built on the HuggingFaceTransformers library, optimized on the highway safety corpus through fine-tuning technology, and deployed on the TensorFlowServing framework to support hot model updates.

[0256] Compliance verification engine: Based on the Drools rule engine, it converts the standard clauses into Drools rule files (.drl) and uses KIEWorkbench for rule management and version control.

[0257] Report generation engine: Employs Apache POI and Freemarker template engines to automatically populate and format report content, and supports custom template syntax extensions.

[0258] Efficiency improvement: Through hardware acceleration and distributed processing, the cycle time is shortened compared to manual writing.

[0259] Compliance assurance: After the dynamic compliance verification module was launched, the compliance rate of reports improved compared to manual processing, significantly reducing legal risks and rework costs.

[0260] Enhanced operability: The generated reports show that the adoption rate of emergency evacuation route planning schemes has increased, and the priority list for safety facility renovation has improved resource allocation efficiency, effectively guiding highway maintenance and management decisions.

[0261] This invention aims to solve the aforementioned problems in existing methods of generating highway safety assessment reports, specifically to achieve the following technical objectives:

[0262] (1) To achieve automated and intelligent generation of assessment reports:

[0263] To address the issues of low efficiency and high subjectivity in manual writing, a technical solution is provided that can automatically transform risk data (such as hazard identification results and risk assessment indicators) output by upstream AI analysis engines into structured and professional assessment reports.

[0264] (2) Ensure the professionalism and standardization of the report content:

[0265] To address the issues of inconsistent report quality and difficulty in ensuring standardization, this approach deeply integrates industry standards such as the "Specification for Safety Evaluation of Highway Projects" with expert knowledge, ensuring that the generated reports are accurate, professional, and compliant with industry standards.

[0266] (3) Integrate a dynamic compliance verification mechanism:

[0267] To address the potential conflict between improvement recommendations and existing standards, a dynamic compliance verification function is embedded in the report generation process to ensure that key content such as the output security improvement recommendations is automatically aligned with the mandatory provisions of relevant laws and standards.

[0268] (4) Enhance the intelligence and operability of improvement suggestions:

[0269] It addresses the problem of traditional suggestions being vague and lacking specific guidance, enabling the intelligent generation of specific and actionable suggestions such as emergency evacuation route planning, priority lists for safety facility upgrades, and resource allocation optimization recommendations.

[0270] (5) Provide efficient decision support tools:

[0271] This provides highway management departments with a tool that can quickly generate high-quality safety assessment reports, assisting them in making scientific decisions and supporting the construction of a full-chain intelligent governance system encompassing "perception-assessment-response".

Claims

1. A structured security assessment report intelligent generation system, characterized in that, Comprise: a data input and preprocessing module, a multi-modal semantic mapping and understanding engine, a domain knowledge base module, an LLM-based report generation and suggestion engine, a dynamic compliance verification and alignment module, and a structured report output and management module; the data input and preprocessing module is used for receiving and analyzing risk data output by an upstream AI analysis engine, and performing cleaning, format conversion and key information extraction; the multi-modal semantic mapping and understanding engine is connected with the data input and preprocessing module, and is used for deep semantic understanding and context association of the preprocessed multi-source heterogeneous risk data, and converting the multi-source heterogeneous risk data into a unified semantic representation easy for a large language model to process; the domain knowledge base module is used for storing structured highway safety evaluation related industry standards, specifications, expert experience and domain ontology knowledge; the LLM-based report generation and suggestion engine is connected with the multi-modal semantic mapping and understanding engine and the domain knowledge base module, and is used for generating evaluation text and intelligent suggestions in combination with the unified semantic representation and domain knowledge; the dynamic compliance verification and alignment module is connected with the LLM-based report generation and suggestion engine, and is associated with the specification library in the domain knowledge base module, and is used for compliance verification and correction of the generated evaluation text and intelligent suggestions; the structured report output and management module is connected with the dynamic compliance verification and alignment module, and is used for generating a structured report according to a preset template.

2. The structured security assessment report intelligent generation system of claim 1, wherein, The data input and preprocessing module uses Z-Score algorithm for anomaly value detection, and determines that a data point is an anomaly value when the Z-Score value of the data point is greater than 3; the naming entity recognition technology in natural language processing is used to extract hazard types, locations and severity; LSTM algorithm is used for anomaly detection of traffic flow time series data, and congestion risk correlation anomaly is determined when the sudden drop amplitude of vehicle speed in peak period exceeds 20%; The naming entity recognition technology is used to additionally extract accident traceability fields, including collision form, vehicle type involved in the accident, accident period and road alignment factors, and time and space correlation labeling is performed on the hazard data.

3. The structured security assessment report intelligent generation system of claim 2, wherein, The data input and preprocessing module pre-processes data through the following steps: Step one: overall target and rules of data preprocessing, including: data cleaning: removing duplicate records, invalid characters and obvious error data through preset regular expressions and anomaly value detection algorithms; format conversion: uniformly converting XML and CSV format data into internal JSON format; key information extraction: using the naming entity recognition technology in natural language processing to extract core fields including hazard types, locations and severity; Step two: focusing on anomaly value detection target, quantifying the deviation degree of a single data point from the overall distribution of the data set, and identifying error or abnormal data obviously deviating from the normal range; Step three: Introduce data distribution characteristic parameters, calculate the mean of the data set and standard deviation wherein: Mean The formula for the calculation is: ； In the formula, For the first in the dataset The value of each data point The number of samples in the dataset; Standard deviation The formula for calculating the standard deviation is: ； wherein is the squared deviation of the individual data point from the mean; Step four: deriving the standardized deviation degree index (Z-Score), and the calculation formula is: ； wherein is the data point value to be detected, is the data set mean, is the data set standard deviation. Step five: set the threshold of abnormal value judgment, according to the statistical law, when the data point is determined as an abnormal value; Step six: processing of abnormal values, removing the data points determined to be abnormal, and retaining the cleaned data for subsequent analysis.

4. The structured security assessment report intelligent generation system of claim 1, wherein, The multi-modal semantic mapping and understanding engine integrates LLM and file structured tools, adopts attention mechanism and graph neural network, unifies and fuses different modal data such as text description, numerical indicators and AI analysis conclusions at the semantic level, constructs a relationship graph of hidden dangers and road attributes, and extracts the causal relationship between core risk elements; The multi-modal semantic mapping and understanding engine also includes a three-element dynamic relationship graph of traffic participants-road facilities-environmental conditions: For intersection scenes, the semantic chain of signal timing, pedestrian crossing demand and vehicle conflict risk is associated; For adverse weather scenes, environmental data such as rainfall and light are semantically bound with road surface skid resistance coefficient and vehicle braking distance to extract environment-facility-risk causal relationship; For key road sections such as sharp bends and long downhills, preset scene-based semantic understanding templates are used to forcibly associate linear parameters and safety facility effectiveness evaluation.

5. The structured security assessment report intelligent generation system of claim 4, wherein, The multi-modal semantic mapping and understanding engine integrates LLM and file structured tools, adopts attention mechanism and graph neural network for semantic fusion of multi-modal data, and the node feature update formula of the graph neural network is: ； Wherein: is the first layer of node feature matrices; adding a loop to the adjacency matrix A denotes the association of a node with itself; is a degree matrix; For the first weight matrix of the layer; is the activation function.

6. The structured security assessment report intelligent generation system of claim 1, wherein, The risk-measure mapping library in the domain knowledge base module adopts AHP hierarchical analysis method to determine the priority of governance measures, and the AHP hierarchical analysis method includes weight vector calculation and consistency check, wherein: The weight vector calculation formula is: ； In the formula: : No. The weight of each risk factor; : determining the relative importance of the i-th factor and the j-th factor in the matrix ; and ; and : number of risk factors; The consistency index calculation formula is: 。 The consistency ratio calculation formula is: 。 In the formula: To determine the largest eigenvalue of the matrix; is a random one-time indicator; When the value is less than 0.1, it is considered that the consistency of the judgment matrix is acceptable.

7. The structured security assessment report intelligent generation system of claim 1, wherein, The report generation and suggestion engine based on LLM adopts BLEU-4 and ROUGE-L indexes to evaluate the similarity between the generated text and the standard template, wherein: The calculation formula of BLEU-4 is: ； wherein The formula for the penalty for short sentences is: ； The formula for calculating accuracy is: ； In the formula: : number of words in the generated text; : Number of words in the reference text; : the weights of the features, typically ; : matching accuracy; : generate in text maximum number of occurrences in reference text; : number of references; The calculation formula of ROUGE-L is: ; Recall rate: ； Precision: ； Wherein: : generating a length of a longest common subsequence of the text and the reference text; : number of words in the text; : number of words of generated text; : weight parameter, usually takes 1, emphasizes the balance between recall rate and precision rate.

8. The structured security assessment report intelligent generation system of claim 1, wherein, The dynamic compliance checking and alignment module has a rule engine based on mandatory provisions of regulatory standards, and uses forward chain reasoning algorithm to match the generated suggestions with the provisions item by item, and the non-compliant items are classified into three levels of emergency, important and general according to the severity, and the revised content needs to be checked again. 9.A method for intelligent generation of structured security assessment report, applied to the system for intelligent generation of structured security assessment report according to any one of claims 1-6, characterized in that, The method comprises the following steps: Step 1: risk data receiving and preprocessing: receiving and preprocessing the structured / semi-structured data output by the upstream AI engine, wherein the Z-Score algorithm and the step of claim 2 are used for abnormal value detection; Step 2: multi-modal semantic understanding and knowledge representation: using the Transformer architecture and cross-modal attention mechanism, combining graph neural network for semantic fusion of preprocessed data; Step 3: report drafting and suggestion generation based on LLM and knowledge base: LLM engine interacts with domain knowledge base to generate report draft and suggestion, and uses BLEU-4 and ROUGE-L indexes to evaluate and optimize the text; Step 4: dynamic compliance checking and content alignment: automatically checking the report draft in real time, judging the non-compliant items and classifying them according to the rule engine, and starting secondary checking; Step 5: structured report finalization and output: setting a three-level review process, generating an electronic document according to the template and pushing it after passing the review.

10. A computer device, comprising: A computer program product comprising a storage medium to store the program code of a computer program, the program code dynamicall y executable by a processor to cause the processor to carry out the method according to claim 7. A computer program comprising program code to cause a processor to carry out the method according to claim 7, when the program code is executed by the processor. A computer program product comprising a storage medium to store the program code of a computer program, the program code dynamicall y

Images