Highway intelligent operation and maintenance question-answering system based on large language model

Abstract

The invention provides a highway intelligent operation and maintenance question-answering system based on a large language model, and belongs to the technical field of natural language processing. The system takes a large language model as a core reasoning engine and combines a domain knowledge base and an RAG technology to realize accurate question and answer of highway operation and maintenance; the method comprises the following steps: based on original knowledge data cutting, generating a title through a large language model, and customizing a knowledge base; receiving query, analyzing an intention by using a large language model, and matching to generate a function; the query is rewritten by using a large language model, dense and sparse vector query is generated, and a double-layer retrieval mechanism is formed; a two-step recall mode is utilized, coarse-grained recall is firstly carried out, then a recall result is subjected to fine-grained optimization through a screening mechanism, and a reasoning text is generated; and finally, inputting the query and reasoning text into the large language model, and generating an optimal answer through single-round and multi-round questions and answers. According to the method, the professionality and reliability of answers are enhanced, and the technical problem that answers are incomplete and inaccurate in an existing question and answer system is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, information retrieval and large language model, and particularly relates to a highway intelligent operation and maintenance question and answer system based on a large language model. BACKGROUND

[0002] RAG technology (Retrieval-Augmented Generation) is an artificial intelligence technology combining information retrieval technology and large language model (LLMs) text generation. The RAG technology architecture mainly consists of three core modules: index knowledge base, retrieval module (Retriever) and generation module (Generator). For details, refer to the formula:

[0003]

[0004] In the formula: is the complete RAG module, is the generation module; is the retrieval module, wherein and respectively represent the indexer and retriever. The indexer converts the external database into an index knowledge base for efficient query . The retriever retrieves relevant knowledge units from according to the query ; is the user input query; is the external knowledge database; is the index knowledge base.

[0005] In recent years, large language models have developed rapidly and have shown excellent performance in many natural language processing tasks, providing new solutions for highway operation decision-making and knowledge question and answer. However, there is still a significant deviation between the output of general large language models and the needs of vertical fields, especially in highly knowledge-intensive vertical tasks, which often exhibit severe hallucinations and knowledge cutoff problems. RAG technology overcomes the knowledge limitations of large language models, retrieves relevant knowledge units from external index knowledge bases, and fuses them as reasoning prompts (Prompt) input into large language models, giving pre-trained parameterized large language models the ability to dynamically access non-parameter index knowledge bases and enhancing their performance in knowledge-intensive tasks.

[0006] Although RAG technology has shown good performance in vertical field tasks such as medicine, law and education, a large number of industry large language models have emerged, and RAG still has limitations such as missing key information, lack of specific details, and incomplete answers. The exploration of the civil engineering field, especially the highway intelligent operation and maintenance field, is still in the initial exploration stage, mainly in the following aspects:

[0007] 1) Lack of publicly available domain datasets and high-quality non-parametric index knowledge bases, which restricts the generalization ability and reproducibility of RAG in the field;

[0008] 2) The knowledge data in the highway field is mostly standards, specifications, manuals and papers, which generally have high structuralization, modal diversity and high information density. A short text often contains multiple actions, objects and time nodes. If the traditional knowledge base construction method is used, it is easy to cause the loss of main objects and the fragmentation of semantic relationships;

[0009] 3) Each process module of the RAG technology has strong configurability and flexibility, and needs to be customized and optimized for specific tasks, otherwise it is difficult to achieve the expected performance.

[0010] These defects result in low adaptability of RAG technology in the field of highway intelligent operation and maintenance; weak correlation of knowledge base information; poor retrieval accuracy; and shallow and incomplete answers. SUMMARY

[0011] The present application aims to overcome the deficiencies of the prior art in knowledge organization and question answering generation, and provides a highway intelligent operation and maintenance question answering system based on a domain knowledge base and RAG technology. The system introduces a "semantic fingerprint" knowledge base construction method, which effectively improves the semantic correlation of internal information and avoids the problems of knowledge fragmentation and semantic fragmentation. At the same time, by customizing and adjusting each module in the RAG process, the system achieves the expected performance, realizes efficient question answering support for highway intelligent operation and maintenance tasks, knowledge increment updating, and provides a structured and high-quality data basis for further fine-tuning of the field large language model.

[0012] The present application proposes a highway intelligent operation and maintenance question answering system based on a large language model. The system function modules include a self-domain knowledge base, a self-knowledge boundary module, a query enhancement module, a double-layer retrieval module, a screening and reordering module, a generation module, and a knowledge increment and optimization module. Among them, the domain knowledge base, the query enhancement module and the knowledge increment and optimization module are the core innovation points of the system.

[0013] The domain knowledge base construction module provides a knowledge injection base for the highway intelligent operation and maintenance question and answer system by natural language processing techniques and customizing knowledge base construction rules. Compared with traditional methods, the system uses a dual-database parallel approach to enhance the accuracy and comprehensiveness of the search. During the knowledge base construction process, the system performs structured metadata extraction and document structure slicing operations on the pre-processed domain knowledge data, and uses a large language model to generate "semantic fingerprints" in a structured JSON format, presenting the first and second layers of knowledge units. This approach significantly improves the efficiency of structured conversion of raw data, enhances the scalability of data, and improves the system's ability to adapt to complex tasks. At the same time, the generated "semantic fingerprints" highly concentrate semantic information, effectively avoiding the problems of missing subject objects and semantic fragmentation, thereby greatly improving the accuracy of knowledge expression and the overall performance of the system in terms of knowledge search efficiency and coverage.

[0014] The query enhancement module uses various rewriting rules to reconstruct user queries of different depths, effectively improving the relevance and quality of search results while ensuring that the query intent is consistent with the semantic space of the knowledge units. From the perspective of query depth and knowledge dependency, the system designs two types of vertical rewriting strategies: one is the semantic conversion strategy, which is used to achieve equivalent rewriting and semantic expansion of the query intent; the other is the hierarchical task planning strategy, which breaks down complex query tasks into multiple sub-tasks with associated relationships, improving the relevance and coverage of the search. At the same time, from the perspective of horizontal query types, the system considers complete semantic queries as "generalized queries", uses a large language model to generate pseudo answers, and extracts keywords from them to build "keyword queries", achieving horizontal query expansion from semantic driving to keyword driving and enhancing structural diversity. Through query enhancement, the system effectively enriches the granularity and hierarchy of the query without changing the original query semantics, significantly improving the recall rate and overall performance of the search.

[0015] The knowledge increment and optimization module aims to address the limitations of the lack of public training data sets and standardized knowledge bases in the field of highway operation and maintenance. This module systematically cleans, denoises, and structures the question and answer data (QA data), continuously enriching the domain knowledge base while effectively expanding the system's self-awareness boundaries and knowledge coverage. More importantly, the generated high-quality structured data can directly serve the subsequent domain fine-tuning of large language models, improving the continuous learning ability and adaptability of the model, and providing stable data support for the long-term evolution, optimization, and iterative upgrade of the system.

[0016] In summary, this invention addresses the shortcomings of current RAG question-answering technologies in areas such as knowledge base construction, query enhancement, and knowledge increment through advanced technologies like domain knowledge base construction, query enhancement, and knowledge increment. Furthermore, by embedding large-scale models of pre-defined tasks into key system components, it significantly reduces human and material costs, providing a low-cost solution for building domain-specific intelligent operation and maintenance question-answering systems. This system significantly improves the efficiency and reliability of intelligent highway operation and maintenance knowledge-based question answering, providing intelligent support for highway maintenance and management, and possesses broad application prospects and profound technical value.

[0017] The specific technical solution is as follows:

[0018] A highway intelligent operation and maintenance question-answering system based on a large language model includes:

[0019] The domain knowledge base construction module is used to retrieve relevant knowledge units through retrieval algorithms, providing the system with accurate, real-time, and domain-specific knowledge injection and information support. This module allows users to update domain knowledge using knowledge base construction rules to meet industry development needs.

[0020] The self-knowledge boundary module is used to receive user queries, analyze the query intent and domain knowledge dependency through an intelligent agent, and select an appropriate path for query generation based on the system capability assessment results.

[0021] The query enhancement module allows for the generation of multiple semantically broad "general queries" and "keyword queries" that focus on key information;

[0022] The two-layer retrieval module, based on two-layer queries, performs coarse-grained candidate knowledge unit retrieval. Each knowledge unit includes structural metadata and segmented document content.

[0023] The filtering and reordering module is responsible for fine-grained filtering of the candidate knowledge unit set, selecting strongly related knowledge units, and reordering them based on the filtering results to form reasoning corpus, thereby achieving retrieval and recall at different granularities.

[0024] The generation module constructs a generative model input based on query and inference corpora, and infers and outputs answer data; at the same time, it evaluates the quality of the answers, generates supplementary queries, and improves the answer data through multiple rounds of dialogue iteration.

[0025] The knowledge increment and optimization module is used to continuously expand the knowledge base and extend its own knowledge boundaries, thereby improving the response speed of knowledge question answering; at the same time, it builds a model fine-tuning dataset based on "query-question answer data" to fill the gaps in the domain dataset.

[0026] Further, the domain knowledge base construction module utilizes customized knowledge base construction rules to construct structured data sets in JSON format through structural metadata extraction, document original structure segmentation, semantic fingerprint label generation, first and second level knowledge unit binding, and index entry extraction, and further constructs a hybrid knowledge base and a text knowledge base, forming a parallel structure of the two bases.

[0027] Further, the self-knowledge boundary module uses a large language model with few-shot chain prompts as an intelligent agent to complete the predetermined task. The intelligent agent sequentially judges the query intent, system knowledge ability, external knowledge dependence, retrieval selection, and generation path matching according to the user query to ensure that the system can handle queries of different depths.

[0028] Further, the query enhancement module uses a large language model to implement query enhancement rules at regular intervals, and implements semantic conversion enhancement or hierarchical task planning according to the query depth, further generates keyword queries, and constitutes a two-layer query to meet the query requirements of different levels and granularities. Among them, hierarchical task planning uses metadata related to the query in the knowledge base as a prompt, refers to the knowledge structure model, uses a large language model to supplement the generation of question and answer drafts, and then generates summary queries and keyword queries.

[0029] Further, the two-layer retrieval module uses dense retrieval algorithms and sparse retrieval algorithms to recall Top-k knowledge units for each query from the hybrid knowledge base and the text knowledge base based on the two-layer query generated by the query enhancement module, and then aggregates the de-duplicated coarse-grained candidate knowledge unit set.

[0030] Further, the screening and reordering module combines the query with the candidate knowledge units one by one to form scoring pairs, performs fine-grained screening of the knowledge units according to semantic relevance, answer contribution, vector score, and keyword score, and reorders the screening scores to form reasoning corpus to support system reasoning generation.

[0031] Further, the generation module uses a large language model with different preset tasks as a generator and a discriminator respectively to generate answers based on the query and the reasoning corpus context; then, the completeness of the answers is evaluated, and the final answers or supplementary queries are output to support multi-round dialogue to enrich the answer content.

[0032] Further, the knowledge increment and optimization module cleans, denoises, and structures the query-answer data pairs into Json data through domain knowledge base construction rules. It can dynamically adjust the self-awareness boundary, update the knowledge base in real time, improve the accuracy, response speed, and resource utilization efficiency of the system, and at the same time, further construct QA data sets and multi-round dialogue data sets to support subsequent large language model fine-tuning work.

[0033] The highway intelligent operation and maintenance question and answer system of the present application combined with the field knowledge base and RAG technology has the following beneficial effects:

[0034] 1. Visual interaction: through the large language model, natural language dialogue is realized, and users can intuitively and conveniently interact with the system, greatly improving the use experience.

[0035] 2. Low-cost knowledge injection: with the help of customized knowledge base and large language model collaborative construction, the applicability and retrieval effect of the system are guaranteed, and the cost of knowledge injection is effectively reduced.

[0036] 3. Efficient retrieval mechanism: a double-layer query strategy combining summary query and keyword query is adopted, combined with double-layer fusion of sparse and dense retrieval algorithms, and two-stage recall of coarse and fine granularity, to realize comprehensive coverage of relevant knowledge units, significantly improving the completeness and accuracy of the answer.

[0037] 4. Answer traceability and explainability: based on the reasoning corpus-based question and answer mechanism, the knowledge density, traceability and explainability of the answer are improved, effectively alleviating the "illusion" problem of large language models in the field of highway intelligent operation and maintenance.

[0038] 5. Knowledge increment: the system can dynamically update its knowledge boundary according to the response result, continuously supplement the knowledge base content, and continuously improve the system accuracy and response speed, while reducing energy consumption.

[0039] 6. Data set blank completion: through the organization of "query-answer data", complete QA data set and multi-round dialogue reasoning data set are constructed, providing a solid data foundation for subsequent large language model parameter fine-tuning, and making up for the lack of high-quality data set in the industry. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 It is the overall architecture diagram of the highway intelligent operation and maintenance question and answer system of the present application;

[0041] Figure 2 It is the display diagram of visual interaction of the highway intelligent operation and maintenance question and answer system of the present application;

[0042] Figure 3 It is the result display and other way question and answer result comparison of the present application;

[0043] Figure 4 It is the original data processing flowchart of the field knowledge base construction module of the present application;

[0044] Figure 5 It is the result display diagram of the multi-modal large model converting multi-source heterogeneous knowledge text of the present application;

[0045] Figure 6A "semantic fingerprint" result display diagram is generated for the large language model of the application;

[0046] Figure 7 A self-awareness boundary module visualization display diagram is generated for the application;

[0047] Figure 8 A query enhancement result display diagram is generated for the application;

[0048] Figure 9 A hierarchical task planning strategy flowchart is generated for the application. DETAILED DESCRIPTION

[0049] The specific technical solutions of the application are described in conjunction with the drawings.

[0050] As shown in Figure 1 , the overall architecture of the system of the application is shown, including the relationship between the functional modules and the data flow. The processes implemented by each part are as follows:

[0051] I. Domain knowledge base construction module (see Figure 3 )

[0052] 1. Module function

[0053] Structured data processing (see Figure 5 , Figure 6 ): Structured data processing aims to clean, convert, denoise and cut block the multi-source heterogeneous data with knowledge value, and unify the original data of different formats into structured text format, and further extract knowledge units with clear semantics and clear boundaries, to provide standardized support for subsequent knowledge modeling and retrieval.

[0054] 2. Technical implementation

[0055] The system uses an automatic processing framework based on customized rules to replace the traditional mechanical cutting method relying on Token, significantly improving the integrity and semantic consistency of knowledge units. The construction of the domain knowledge base follows the following technical process:

[0056] (1) Use multi-modal large language models to understand and convert non-text knowledge carriers such as charts, tables, and images to generate processable text information;

[0057] (2) Automatically extract chapter and section titles such as "7 Wave Beam Guardrail" and "7.1 General Provisions" as structured metadata to indicate the hierarchy and organizational relationship of knowledge;

[0058] (3) According to the type of literature, the document structure cutting strategy is adopted: for papers and manual texts, the text is divided according to natural paragraphs; for standard documents such as national standards and industry specifications, the text is divided according to the clause granularity. Each text block is bound with its structure metadata to generate the first level knowledge unit;

[0059] (4) Through a large language model, the semantic understanding and abstract generation of each knowledge unit are carried out, and the "semantic fingerprint" with fingerprint characteristics is extracted to enhance the effect of subsequent semantic retrieval recall;

[0060] (5) The above content is organized in a structured JSON format to uniformly represent the content, structure information and semantic characteristics of each knowledge unit;

[0061] (6) Based on structured data, a domain knowledge base is constructed to form a high-quality, searchable and expandable knowledge index system.

[0062] 3. Example flow:

[0063] Input: (table as follows)

[0064]

[0065] Analysis process:

[0066] Text conversion: The domain knowledge base construction module receives table data presented in the form of an image, uses a multi-modal large language model to identify, clean and understand the semantics of the content, and converts unstructured visual information into a standard text expression: "The measured project: the wave size deviation of the wave beam plate should be controlled within ± 0.16mm, and a thickness gauge or a micrometer should be used for detection, with a sampling ratio of 5%; The wall thickness of the column should be 4.5 ± 0.25mm, and a thickness gauge or a micrometer should be used for detection, with a sampling ratio of 5%."

[0067] First level knowledge unit generation: According to the preset cutting strategy, the text content is divided and processed, and the cutting result is bound with the corresponding structure metadata ("wave beam guardrail / quality acceptance") to form a first level knowledge unit with context level and organizational relationship: "Wave beam guardrail Quality acceptance Measured project: The measured project: The wave size deviation of the wave beam plate should be controlled within ± 0.16mm, and a thickness gauge or a micrometer should be used for detection, with a sampling ratio of 5%; The wall thickness of the column should be 4.5 ± 0.25mm, and a thickness gauge or a micrometer should be used for detection, with a sampling ratio of 5%."

[0068] "Semantic fingerprint" generation: The large language model is called to perform semantic compression and summarization on the first-level knowledge unit, extract the core content, and generate a concise "semantic fingerprint" with search features: "the measured items and specific requirements of the quality acceptance of the wave beam guardrail", and based on this, the second-level knowledge unit is further bound and generated.

[0069] Template installation specifications or allowable deviations: Template elevation controlled within ±10mm; template internal dimensions controlled within ±200mm; height difference between adjacent template surfaces not exceeding 2mm; surface flatness not exceeding 5mm; embedded part centerline position controlled within ±3mm.

[0070] Result generation: The second-level knowledge units are organized into structured JSON data in a unified format, including structural metadata, text content, and semantic fingerprints.

[0071] Output:

[0072] {

[0073] Chapter: "W-beam guardrail"

[0074] Section: Quality Acceptance

[0075] "Semantic Fingerprint": "Measured Items and Specific Requirements for Quality Acceptance of Corrugated Beam Guardrails",

[0076] "Clause": "Measured items: The dimensional deviation of the corrugated beam slab should be controlled within ± 0.16mm, and should be inspected using a thickness gauge or micrometer, with a sampling rate of 5%; the column wall thickness should be 4.5 ± 0.25mm, and should be inspected using a thickness gauge or micrometer, with a sampling rate of 5%."

[0077] }

[0078] II. Self-Knowledge Boundaries Module (see...) Figure 7 )

[0079] 1. Module Functions

[0080] This module is responsible for identifying and deeply analyzing the intent of user-input queries, accurately determining the degree to which the query relies on domain knowledge. Simultaneously, the module assesses the system's current knowledge coverage and generation capabilities, dynamically determining whether external knowledge retrieval is necessary to improve answer quality, thus achieving intelligent collaboration and efficient integration of generation and retrieval.

[0081] 2. Technical Implementation

[0082] Model selection: The core capabilities of the module are provided by a large language model deployed locally, such as the Qwen2.5 series. The model acts as an agent and undertakes key tasks such as intent recognition, query type determination, and response strategy decision-making.

[0083] Model prompting: The model behavior is set using the few-shot prompting method. Through carefully designed prompt words, the model completes sub-tasks such as intent classification, knowledge demand judgment, and answer generation path selection. This knowledge Q&A system divides queries into the following three categories based on the degree of dependence on domain knowledge:

[0084] (1) Conceptual / Explanatory Questions: This type of query focuses on understanding and explaining basic concepts and usually relies on built-in general knowledge to generate accurate answers with low dependence on external domain knowledge. For example: "What is a rigid guardrail?"

[0085] (2) Operational / Regulatory Questions: This type of query involves understanding and applying operational methods, construction processes, and related technical specifications, relying on high-quality, structured domain knowledge, and requiring knowledge base retrieval and knowledge injection to achieve response. For example: "What are the weather impacts and restrictions on marking line construction?"

[0086] (3) Scheme Questions: This type of query focuses on complete problem solutions, usually covering multiple stages of problem handling (such as pre-diagnosis, mid-treatment, and post-evaluation), with high dependence on systematic domain knowledge, requiring hierarchical task planning strategies to gradually analyze and generate responses. For example: "How should a damaged wave-shaped guardrail be repaired?"

[0087] 3. Analysis process

[0088] Input: What are the weather impacts and restrictions on marking line construction?

[0089] Analysis process:

[0090] Query reception: The self-knowledge boundary module receives input and calls the large language model for preprocessing, identifying key entities (such as weather, marking lines), actions (construction), etc.

[0091] Semantic analysis and intent recognition: Call the large language model for semantic analysis, intent recognition, and path selection. The specific analysis process is as follows:

[0092] • Object is: "marking line construction" → specific operational behavior;

[0093] • Question is: "impact and restriction" → focus on construction conditions and influencing factors;

[0094] Implicit motivation: Wondering if weather factors affect construction, under what conditions construction cannot be carried out (such as rain, low temperature), that is, understanding the relevant "construction limitation standards".

[0095] More biased (normative / operational)

[0096] Focus on the "restrictions and effects" of external conditions (weather) on construction operations;

[0097] In fact, it is discussing the normative conditions of construction (such as temperature, humidity, weather requirements);

[0098] Similar questions will list clear restriction conditions in construction technical specifications and operation procedures, such as: "Do not construct when the air temperature is below 10℃", "The road surface must be dry", etc;

[0099] Answering such questions requires citing specific technical specifications and operation procedures as a basis, and it is recommended to inject knowledge by searching related knowledge units to enhance the accuracy and professionalism of the answer.

[0100] Output:

[0101] Three, query enhancement module (see Figure 8 )

[0102] 1. Module function

[0103] Relying on the powerful text generation capability of large language models, multi-dimensional and multi-level double-layer queries (including summary queries and keyword queries) are generated to obtain content-rich and complete answer data.

[0104] 2. Technical implementation

[0105] In the process of generating double-layer queries, the system uses semantic conversion strategy and hierarchical task planning strategy based on the large language model of preset prompts to expand and reconstruct the original user query, forming a more expressive and valuable summary query. At the same time, with the help of the summary query pseudo-answer generated by the large language model, the system intelligently identifies and extracts core terms and keywords through semantic analysis and information extraction algorithms, and then constructs keyword queries. Its processing formula is as follows:

[0106]

[0107] In the formula: represents the user's input query; and represent the summary query and the keyword query, respectively; and represent the summary query and the keyword query generation functions, respectively, which are completed by the large language model of preset tasks.

[0108] Semantic transformation strategy: By defining query enhancement rules, the original query is reconstructed using methods such as multi-angle semantic rewriting, semantic equivalent replacement, content distillation and format conversion, generating a query form with equivalent semantic expression, thereby improving its semantic coverage and versatility.

[0109] Hierarchical task planning strategy (see) Figure 9 Guided by the metadata prompts, the large language model first generates a preliminary task response draft, then mines the implicit task chain and key steps from it, and further refines it into a well-structured, hierarchical, and decomposable summary query.

[0110] 3. Analysis Process

[0111] Input: What impact and limitations does weather have on pavement marking construction?

[0112] Analysis process:

[0113] Query Receiving: The query enhancement module receives the query (What impact and limitations does the weather have on road marking construction?) and tag (2) from the previous module, and expands multiple general queries through a proxy and semantic conversion strategy.

[0114] Semantic transformation: After receiving a query, the large language model expands multiple general queries through the following steps: intent understanding (the impact and limitations of weather on "road marking construction," focusing on "influencing factors" and "construction conditions"), semantic decomposition ("weather" → natural conditions, such as rainfall, humidity, temperature, wind, etc.; "impact and limitations" → focusing on the limitations of weather on construction in terms of "whether construction can proceed," "construction quality," and "construction safety"; "road marking construction" → refers to the process of applying road markings, involving adhesion, drying speed, construction equipment, etc.), and expression rewriting (concretizing "impact and limitations" into "unsuitable conditions" and "meteorological factors affecting quality or progress"; changing the sentence tone from interrogative to a structure guided by conditions and results).

[0115] Keyword query generation: For general queries, the core intent and high-frequency words are used to generate pseudo-answers (e.g., rainy weather makes roads slippery, preventing paint from adhering effectively and affecting the adhesion and lifespan of road markings; low temperatures prolong paint drying time, affecting construction progress and road marking durability; high humidity may cause blistering or peeling of road markings. In addition, strong winds are detrimental to the accuracy of spraying operations, easily causing contamination or deviation). From this, we identify terms that express an influencing mechanism (e.g., "adhesion," "wind speed limitation") and operational terms (e.g., "paint drying time," "temperature requirements").

[0116] Output:

[0117] Summary query: 1. What impact and limitations does weather have on pavement marking construction?

[0118] 2. Under what weather conditions is it unsuitable to carry out road marking construction?

[0119] Keyword search: 1. Road surface drying, adhesion, paint drying time, wind speed limit, temperature requirements

[0120] 2. Rainfall restrictions, maximum wind speed, minimum construction temperature, humidity control, and road surface dryness.

[0121] IV. Two-layer search module

[0122] 1. Module Functions

[0123] The two-layer retrieval module uses a dense retrieval algorithm (cosine similarity) and a sparse retrieval algorithm (BM25) to retrieve Top-k knowledge units from the indexed knowledge base, and a coarse-grained set of candidate knowledge units is retrieved. The processing formula is as follows:

[0124]

[0125] In the formula: and This indicates the results obtained from a two-level search; and These respectively represent retrieving knowledge units from a hybrid knowledge base and a text knowledge base; and These represent candidate knowledge units for hybrid knowledge bases and text knowledge bases, respectively. and These represent the cosine similarity algorithm and the BM25 algorithm, respectively. This represents the set of candidate knowledge units obtained by merging and deduplicating the top-k most relevant knowledge units.

[0126] 2. Technical Implementation

[0127] Cosine similarity:

[0128]

[0129] In the formula: Represents vector

[0130] BM25 Algorithm:

[0131]

[0132] In the formula: Indicates the first One query item; Indicates the first Inverse document frequency of each query item; This represents the knowledge unit to be retrieved; and represents the length of the knowledge unit and the average field length; control the default value of the non-linear term frequency normalization is 1.2; control the penalty strength of document length on the score; and represents the total number of knowledge units in the index and the number of knowledge units containing query terms .

[0133] Five, screening and reordering module

[0134] 1, module function

[0135] Use the scoring module and the target reordering model to perform fine-grained identification and screening on the recalled candidate knowledge unit set, extract knowledge units highly related to the user query, and realize two-stage accurate recall.

[0136] 2, technical implementation

[0137] Data preparation: Pair the user query with the candidate knowledge unit one by one to construct query-knowledge unit (Q-K) data pairs.

[0138] Scoring module: Comprehensive consideration of multi-dimensional indicators between candidate knowledge units and user queries, including content relevance, answer contribution, vector similarity, and keyword matching score, to generate a comprehensive score. The system sets a predetermined threshold to filter out strongly related knowledge units.

[0139] Content relevance and answer contribution: Use the zero-shot scoring strategy to input each Q-K pair into a large language model, and guide the model to conduct comprehensive evaluation from two dimensions through structured prompts: one is content relevance, that is, the consistency of query and document in semantic theme; the second is answer contribution, that is, the support degree of the document to the query answer. The prompt words explicitly guide the model to focus on key factors such as "whether it contains the core object in the query", "whether it has semantic focus", "whether it provides effective information", etc., and output two floating-point scores between 0 and 1, representing semantic matching degree and answer value respectively. The specific rules are as follows:

[0140]

[0141] Vector similarity score: Based on text semantic embedding (embedding) calculation, use a pre-trained text embedding model to convert the query and document title into fixed-dimensional vector representations respectively. Then calculate the cosine similarity between the query vector and each knowledge unit vector as a measure of semantic relevance. The cosine similarity value ranges from [-1,1], and is mapped to the [0,1] interval through normalization processing.

[0142] Key word matching score: The coverage of key words is measured by counting the frequency of key words in the document body. Specifically, the query sentence and the document content are segmented by using a Chinese segmentation tool, and meaningless stop words and punctuation marks are removed. Then, it is counted whether each word in the query appears in the document segmentation list, and the proportion of the number of key words appearing in the total number of query key words is calculated. This proportion is the key word matching score, and the value range is [0, 1].

[0143] The whole scoring process is as follows:

[0144]

[0145] In the formula: represents the comprehensive score of the candidate knowledge unit by the scoring module; , , , respectively represent the content relevance score, the answer contribution score, the vector similarity score and the key word matching score; … represent the respective weights; represents the set of strongly related knowledge units obtained by fine-grained screening, wherein represents a threshold value.

[0146] Re-ranking module: According to the comprehensive score, the knowledge units of fine-grained secondary recall are sorted in descending order to generate high-quality reasoning corpus. Its processing formula is as follows:

[0147]

[0148] In the formula: represents the reasoning corpus obtained by descending sorting and splicing according to the comprehensive score .

[0149] 3. Analysis process

[0150] Input:

[0151] {

[0152] "Q": "What are the impacts and restrictions of weather on marking line construction?",

[0153] "K": "Marking line - construction details - traffic marking line construction weather requirements - traffic marking line should be constructed in the daytime. In rainy, snowy, sandstorm, strong wind, and low temperature weather, the construction should be suspended. Rain, snow, and other bad weather will affect the adhesion between the road surface and the paint, and sandstorms and strong winds will affect the operation of traffic marking line construction. For traffic marking paint, primer, and protruding road sign adhesive materials, the air temperature during construction should also meet the corresponding use requirements."

[0154] }

[0155] Analysis process:

[0156] Score one, content relevance:

[0157] · The query topic is "weather" and "marking line construction";

[0158] · The document topic is clearly focused on "weather requirements for traffic marking line construction";

[0159] · Weather conditions, construction behavior, and material specifications are highly aligned with the query;

[0160] · Score 1.0.

[0161] Score two, answer contribution:

[0162] · The query asks about the "impact and restrictions" of weather on construction;

[0163] · The document specifically states that:

[0164] Impact: Bad weather will affect the adhesion between the road surface and the paint, and wind and sand will affect the operation;

[0165] Restrictions: Construction should be suspended in rainy, snowy weather, and material use should meet temperature requirements;

[0166] · The answer is accurate, detailed, and comprehensive;

[0167] · Score 0.9.

[0168] Score three, vector similarity score

[0169] · Query segmentation: ['weather','marking line', 'construction', 'impact','restriction']

[0170] · Knowledge unit keyword segmentation: ['marking line', 'construction', 'details', 'traffic','marking line', 'construction', 'weather', 'bad weather','strong wind','requirements', …]

[0171] · Hit keywords: ['weather','marking line', 'construction', 'impact']

[0172] • MB25 computes score result 0.4444;

[0173] • Score: 0.4444.

[0174] Score four, keyword matching score

[0175] • Query embedding result: [-0.02960000000894, -0.0318000018596,…

[0176] • Knowledge unit embedding result: [0.00510000018402, -0.025100000202655,…

[0177] • Cosine similarity computes score result 0.7273;

[0178] • Score: 0.7273.

[0179] Output:

[0180] 0.4 * 1.0 + 0.25 * 0.9 + 0.15 * 0.4444 + 0.2 * 0.7273 = 0.837

[0181] Six, generation module (see Figure 2 )

[0182] 1. Module function

[0183] The generation module is the core component of the system, responsible for generating natural language answer data according to user queries and reasoning corpus. This module not only undertakes the answer generation task, but also supports answer integrity discrimination, supplementary query generation, and multi-round dialogue function.

[0184] 2. Technical implementation

[0185] Generate answer data: Use a large language model (such as Qwen2.5) deployed locally to generate preliminary answer content based on user queries and related reasoning corpus. The processing formula is:

[0186]

[0187] In the formula: denotes the answer data output by the generated answer unit; denotes the dialogue generation function, implemented by a large language model;

[0188] Answer integrity discrimination: Perform post-processing operations such as semantic analysis, format specification, and redundant information filtering on the model output results, and comprehensively judge the completeness and accuracy of the answer content.

[0189] Supplementary query generation: For the information missing or further expandable in the model generation, intelligent supplementary suggestions and supplementary queries are proposed to guide the user to improve information acquisition. Its processing formula is:

[0190]

[0191] In the formula: represents the output response of the judgment model unit; represents the answer completeness judgment function, which is realized by a large language model. The function output 1 represents completeness, and output 0 represents incompleteness, which can be supplemented; represents the supplementary query;

[0192] Multi-round dialogue mechanism: Combined with the model response and supplementary suggestions, the user can choose whether to initiate a supplementary query, and the system can continuously track the user's context to realize multi-round interaction and reasoning generation based on historical context. Its processing formula is:

[0193]

[0194] In the formula: represents the answer data of the system output in the round; represents the supplementary query in the round; represents the dialogue state up to the round; represents the reasoning corpus in the round.

[0195] 3. Analysis process

[0196] Input:

[0197] {

[0198] "Q": "What impact and restrictions does the weather have on marking line construction?",

[0199] "K": "Marking line - construction details - traffic marking line construction weather requirements - traffic marking line should be constructed during the day. In rainy, snowy, sandstorm, strong wind, and low-temperature weather below the material's specified construction temperature, construction should be suspended. Rain, snow, and other severe weather can affect the adhesion between the road surface and the paint, and sandstorms and strong winds can affect the traffic marking line construction operation. For traffic marking paint, primer, and raised road sign adhesive materials, the air temperature during construction should also meet the corresponding usage requirements."

[0200] }

[0201] Analysis process:

[0202] Prompt Splicing: The system receives structured inputs, including a user query (Q) and a knowledge unit (K). Then, according to the set prompt template, the query is spliced with the corresponding corpus into a context input that is logically clear and semantically focused. This input is used to guide the language model to accurately understand the user's intent and generate answers around the key content.

[0203] First Round Model Answer: The language model generates a preliminary answer based on the constructed context Prompt: "Weather conditions have a significant impact on traffic marking construction. Adverse weather such as rain, snow, dust storms, and strong winds can deteriorate the construction environment, affect the adhesion of paint to the road surface, and reduce construction quality. In addition, construction should be suspended when the temperature is lower than the specified material usage to avoid damage to material performance."

[0204] Supplementary Suggestions: The system analyzes the model's answer and identifies possible information gaps or content that can be further refined. Combining the original corpus and generated content, it automatically proposes supplementary suggestions and potential supplementary queries to enrich the answer content or improve the completeness of the expression. For example: Supplementary suggestions and queries: Does the marking paint need to explain the impact of temperature on paint construction?

[0205] User Adjustment: The system presents the supplementary suggestions to the user, who decides whether to further query or adjust the current question. For example, the user chooses to continue asking: "Does strong wind weather have an impact on paint construction?"

[0206] Multi-round dialogue iteration: The system records the user's new input and takes the previous round of answer response, supplementary query, and retrieval context as new input to generate an enhanced answer. Each iteration progresses on the existing context, ensuring semantic coherence and continuous content completion.

[0207] Output:

[0208] Weather conditions have a significant impact on traffic marking construction. Adverse weather such as rain, snow, dust storms, and strong winds can deteriorate the construction environment, affect the adhesion of paint to the road surface, and reduce construction quality. In addition, construction should be suspended when the temperature is lower than the specified material usage to avoid damage to material performance.

[0209] Environmental temperature is a key factor in determining construction effectiveness. Generally, when the environmental temperature is below 5°C, the quality of marking construction is unstable, prone to peeling and other problems, and the low melting temperature will seriously affect the adhesion of paint to the road surface. If you do not use a primer, it is recommended that the environmental temperature be above 10°C before construction. During the paint construction process, the temperature of the paint applied to the road surface should not be lower than 180°C to ensure sufficient adhesion. Especially in early winter, it is advisable to schedule the construction time between 10 am and 3 pm to avoid the low-temperature period in the early morning and late afternoon, ensuring construction quality.

[0210] Seven, Knowledge Increment Module

[0211] 1. Module Function

[0212] The Knowledge Increment Module is responsible for implementing the dynamic updating and adaptive expansion of the knowledge base in the question and answer system. It also creates high-quality data sets to support subsequent large language model fine-tuning and other work, and fills the gaps in industry data.

[0213] 2. Technical Implementation

[0214] By collecting query-answer data in real user interactions, combining semantic fingerprints and structured conversion, automatically completing knowledge fusion and boundary supplementation, and supporting the continuous evolution of the system and the generation of large language model fine-tuning data.

[0215] Data Extraction and Cleaning: Identify valid query-answer pairs through pre-set rules, perform cleaning and standardization.

[0216] Semantic Compression and Semantic Fingerprint Generation: Use large language models for semantic understanding, extract core information to form semantic fingerprints, and reduce redundancy.

[0217] Knowledge Fusion and Boundary Adjustment Mechanism: Inject new summaries into new enhanced knowledge bases, and reconstruct knowledge boundaries based on new content to improve system adaptability.

[0218] Structured Data Set Construction: Automatically generate standardized single-turn question answering and multi-turn dialogue data for training data accumulation and model fine-tuning.

[0219] 3. Analysis Process

[0220] Input:

[0221] {

[0222] "Q": "What are the impacts and restrictions of weather on marking line construction?",

[0223] "A": "Weather conditions have a significant impact on traffic marking line construction. Adverse weather such as rain, snow, sandstorms, and strong winds can deteriorate the construction environment, affect the adhesion of paint and pavement, and reduce construction quality. In addition, construction should be suspended when the temperature is below the specified material usage, to avoid damage to material performance.

[0224] Ambient temperature is one of the key factors determining the construction effect. Generally, when the ambient temperature is below 5℃, the construction quality of road markings is unstable, and problems such as peeling and flaking are prone to occur. Furthermore, a low melting temperature will severely affect the adhesion between the paint and the road surface. If no primer is used, it is recommended to carry out construction only when the ambient temperature reaches above 10℃. During paint application, the temperature of the paint applied to the road surface should not be lower than 180℃ at the moment of application to ensure sufficient adhesion. Especially in early winter, it is advisable to schedule construction between 10 am and 3 pm to avoid the low temperatures of early morning and evening, thus ensuring construction quality.

[0225] }

[0226] The parsing process involves rule-driven data extraction and cleaning, knowledge segmentation, semantic compression and semantic fingerprint generation, and the construction of a structured dataset, outputting structured data to support further work.

[0227] Output:

[0228] {

[0229] Chapter: "Markings",

[0230] "SemanticFingerprint": "Impact of Weather Conditions on Traffic Marking Construction"

[0231] "Clause": "Atmospheric conditions have a significant impact on traffic marking construction. Severe weather conditions such as rain, snow, sandstorms, and strong winds can deteriorate the construction environment, affecting the adhesion between the paint and the road surface and reducing construction quality. Furthermore, construction should be suspended when the temperature is below the material's performance specifications to avoid damage to the material."

[0232] }

[0233] {

[0234] Chapter: "Markings",

[0235] "Semantic Fingerprint": "The Influence of Temperature on Road Marking Paint Application"

[0236] "Clause": "Generally, when the ambient temperature is below 5℃, the quality of pavement marking construction is unstable, easily leading to problems such as peeling and detachment. Furthermore, a low melting temperature severely affects the adhesion between the paint and the road surface. If no primer is used, it is recommended to proceed with construction only when the ambient temperature reaches above 10℃. During paint application, the temperature of the paint applied to the road surface should not be lower than 180℃ to ensure adequate adhesion. Especially in early winter, it is advisable to schedule construction between 10 AM and 3 PM to avoid the low temperatures of early morning and evening, thus ensuring construction quality."

[0237] }。

Claims

1. A highway intelligent operation and maintenance question and answer system based on a large language model, characterized in that, Comprise the following steps: S1, domain knowledge base construction module: use multi-modal large model to text, standardize, structure the domain multi-source heterogeneous knowledge; and cut, organize it into knowledge units with semantic association, and then build an index knowledge base that can be efficiently retrieved; S2, self-knowledge boundary module: receive user input query, based on pre-trained large language model and its built-in knowledge boundary mechanism, carry out deep semantic analysis to identify whether the user query is beyond the model knowledge range, and judge whether it needs to rely on the domain knowledge base for external enhancement, automatically select the generation strategy and function; S3, query enhancement module: receive query, use pre-set, task-adapted large language model to reconstruct and expand user query from multiple angles, generate "summary query" with extensive semantics and "keyword query" focusing on key information; S4, double-layer retrieval module: receive the summary query and keyword query in S3, respectively use deep semantic retrieval model based on dense retrieval and keyword retrieval model based on sparse retrieval to perform coarse-grained retrieval on the knowledge base, and recall a set of highly relevant candidate knowledge units; S5, screening and reordering module: based on the set of sub-knowledge units recalled by S4, make a "query-knowledge unit" scoring dataset; use a scoring model with pre-set scoring rules to perform fine-grained screening of strongly relevant knowledge units; on this basis, use a target reordering model to sort the screened knowledge units by relevance, and finally generate high-quality inference corpus for large language model inference; S6, generation module: according to the user query and inference corpus, drive the large language model to generate an answer data that meets the user's needs, is semantically coherent and accurate in information, and at the same time, use the set answer evaluation model to check the conformity of the generated answer, if it meets the user's purpose, then output the answer; if not, backtrack to step S3, iterate and optimize the generation process until a satisfactory answer is obtained; S7, knowledge increment and optimization module: used for structured cleaning, anonymization processing and quality evaluation of user "query-answer data", construction of incremental knowledge base and continuous optimization of model, and auxiliary formation of high-quality supervised fine-tuning dataset.

2. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The domain knowledge preprocessing specifically comprises: Standardize the text data of national standards, local specifications, industry operation manuals and domain papers, and at the same time, convert non-text content such as images and tables into structured natural language text using a multi-modal large language model; Perform cleaning operations to remove special symbols, chapter numbers and unstructured noise information, and extract chapter and section titles as hierarchical structure metadata; According to the inherent structure of the document, the text is logically blocked, and the hierarchical structure metadata is bound to form a first-level knowledge unit with structured metadata; The large language model of the preset task is used to perform deep semantic analysis and information extraction on the first-level knowledge unit, highly condense the core semantics, key entities and their mutual relationships, generate a "semantic fingerprint" title with "identity" information, and use the "semantic fingerprint" title for condensed representation of semantics; then, the "semantic fingerprint" title is associated and bound with the corresponding first-level knowledge unit again to form a second-level knowledge unit with more rich semantic features, thereby providing data support for constructing a domain knowledge base.

3. The highway intelligent operation and maintenance question and answer system based on a large language model according to claims 1 and 2, characterized in that, The constructed domain knowledge base further includes: The semantic fingerprint title is converted into a dense vector representation by an embedding model, and the hierarchical structure metadata of the second-level knowledge unit, the semantic fingerprint title and the original text block are separated to form independent triple index entries; Based on the dense vector of the semantic fingerprint title and the corresponding first-level knowledge unit text, a key-value pair dictionary is constructed and stored in the hybrid knowledge base, wherein the semantic fingerprint title vector serves as the retrieval key and the knowledge unit text serves as the corresponding value; Based on the separated hierarchical structure metadata, semantic fingerprint title and original text block, a text knowledge base is stored to support information retrieval and matching of different granularities.

4. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The self-knowledge boundary module uses a large language model of a few-shot prompt word limited task as an intelligent agent; After receiving a user query, the module transfers it to the intelligent agent for deep semantic analysis, intent recognition and answer generation path matching, wherein the answer generation path includes a large language model direct generation function with a preset specific background, role and target, and a retrieval enhanced generation function based on the domain knowledge base.

5. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The query enhancement module relies on the text generation capability of the large language model to generate more expandable queries as summarized queries according to the user query, and to generate keyword queries by simulating the generation of pseudo answers and intelligently extracting core keywords therefrom to improve the recall rate and precision of the retrieval; the formula for generating the two types of queries is: ； In the formula: represents a query input by a user; and respectively represent a summary query and a keyword query; and respectively represent a summary query and a keyword query generation function, completed by a large language model of a preset task.

6. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 5, characterized in that, The query enhancement method includes: Semantic conversion strategy: performing multi-angle semantic rewriting, semantic equivalent replacement, content distillation and format conversion strategies by the large language model to generate diversified summarized queries and keyword queries; Hierarchical task planning strategy: regarding the user complex query as a high-level task, extracting the knowledge base metadata related to the query as the semantic context; then, using the preset knowledge structure model and the large language model to perform structured understanding and semantic completion on the task, and based thereon, refining, organizing and supplementing the high-level task to gradually decompose and generate middle-level and bottom-level sub-tasks with clear targets, thereby constructing a complete task chain from user intent recognition to specific behavior execution; finally, based on the task chain, using the large language model to generate corresponding summarized queries and keyword queries.

7. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The double-layer retrieval module uses the summarized queries and keyword queries to recall Top-k knowledge units as a candidate knowledge unit set from the hybrid knowledge base and the text knowledge base by using the dense retrieval algorithm (cosine similarity) and the sparse retrieval algorithm (BM25), and the formula for the retrieval process is: ； ； ； In the formula: and represent the results obtained by double-layer retrieval; and respectively represent the retrieval of knowledge units from the mixed knowledge base and the text knowledge base; and respectively represent the candidate knowledge units of the mixed knowledge base and the text knowledge base; and respectively represent the cosine similarity algorithm and the BM25 algorithm; represent the candidate knowledge unit set obtained by selecting the top-k most relevant knowledge units, merging, and deduplicating.

8. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The screening reordering module is composed of a scoring module and a target reordering module, aiming to identify and extract strongly related knowledge units from the recalled candidate knowledge unit set in a fine-grained manner; wherein the scoring module comprehensively considers multiple dimensions such as content relevance between candidate knowledge units and user queries, answer contribution, vector similarity score, and keyword matching score, generates a comprehensive score, and sets a predetermined threshold to filter out strongly related knowledge units; the target reordering module sorts the strongly related knowledge units in descending order according to the comprehensive score, and finally generates high-quality reasoning corpus, the specific formula is: ； ； ； In the formula: represents the comprehensive score of the candidate knowledge unit by the scoring module; , , , respectively represent the content relevance score, the answer contribution score, the vector similarity score, and the keyword matching score; … respectively represent the respective weights; represents the strongly related knowledge unit set obtained through fine-grained screening, wherein represents the threshold value; represents the reasoning corpus obtained by sorting and splicing in descending order according to the comprehensive score .

9. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The generation module is composed of three parts: answer generation unit, answer evaluation unit and multi-round dialogue unit: The answer generation unit is used to receive input composed of reasoning text and user query, and output corresponding answer data through a large language model generation function with pre-set tasks and knowledge injection, the specific formula is: ； In the formula: represents the answer data generated by the answer unit output; represents a dialogue generation function, implemented by a large language model; The answer evaluation unit uses a large language model with pre-set tasks to evaluate the completeness of the generated answer data for the user query, and can give supplementary query suggestions or output complete answer data according to the evaluation results, the specific formula is: ； In the formula: represents the output response of the judgment model unit; represents an answer completeness judgment function, realized by a large language model, and the function outputs 1 to represent completeness and 0 to represent incompleteness, which can be supplemented; represents a supplementary query; The multi-round dialogue unit receives the feedback of the judgment model unit and the user's adjusted query, and repeatedly executes the retrieval enhancement and answer generation process, realizes the iterative output of satisfactory answers, realizes closed-loop multi-round question and answer and result tracking, the specific formula is: ； wherein: represents the answer data output by the system for the th turn; represents a follow-up query for the th turn; represents the dialogue state as of the th turn; represents the reasoning corpus for the th turn.

10. The highway intelligent operation and maintenance question and answer system based on a large language model according to claim 1, characterized in that, The knowledge increment module continuously optimizes the question and answer system and makes fine-tuning data set according to "user feedback" and "query-answer data"; specifically, through pre-defined data processing rules, "query-answer data" is dynamically collected, cleaned and formatted, and its semantic fingerprint is supplemented to the hybrid knowledge base, realizing real-time updating of the knowledge base and dynamic adjustment of the adaptive knowledge boundary; at the same time, this module is also responsible for generating structured question and answer data sets and multi-round dialogue data sets to support continuous fine-tuning of system core model parameters and performance improvement.

Images