Method and system for constructing operation and maintenance knowledge base based on large model

Abstract

The application discloses a kind of based on big model's operation and maintenance knowledge base construction method and system, belong to artificial intelligence and operation and maintenance management field.The method includes: step S1, gather multi-source operation and maintenance data and carry out auditing, cleaning, utilize big model to extract entity and relationship to construct knowledge base;Step S2, through multi-round dialogue, complete user question context, and reconfigure into normalized query;Step S3, to normalized query parallel execution semantic retrieval and graph retrieval, after the fusion result, input big model generates answer with source reference;Step S4, collect the quality feedback of answer to user, and utilize feedback data to optimize knowledge base or big model.Corresponding system includes five big modules of content management, RAG knowledge base, question supplement, AI accurate reply and feedback learning.The advantage of the application is: realize the automated construction of operation and maintenance knowledge, accurate traceability question and answer and closed loop self-evolution, significantly improve operation and maintenance intelligent level and decision support efficiency.

Description

Technical Field

[0001] This invention relates to a method and system for constructing an operation and maintenance knowledge base based on a large model, belonging to the fields of artificial intelligence and operation and maintenance management technology. Background Technology

[0002] With the widespread application and increasing complexity of intelligent systems, the operation and maintenance (O&M) management of equipment and systems has become a crucial link in ensuring their safe and stable operation. Traditional O&M models heavily rely on expert experience and manual record-keeping. When faced with a large number of diverse and scattered modern intelligent systems, their inefficiencies, difficulties in knowledge transfer, and slow fault response become increasingly apparent. In recent years, knowledge bases and artificial intelligence technologies have been introduced into the O&M field, aiming to systematically manage scattered O&M experience, equipment data, and fault cases, and attempting to assist frontline O&M personnel through intelligent question-and-answer methods.

[0003] In existing technologies, the construction of operations and maintenance (O&M) knowledge bases typically relies on conventional techniques such as data collection, knowledge representation, retrieval, and question answering. At the data level, while some system logs and manuals can be integrated, key data such as real-time analysis, unstructured communications (e.g., instant messaging records), and paper records generated by O&M personnel on-site often struggle to be effectively collected and integrated, resulting in incomplete knowledge sources. Regarding knowledge representation and storage, traditional methods, such as rule-based ontology construction or simple vectorization, struggle to address the complexity, ambiguity, and rich relationships between entities in the O&M domain. Customized modeling is costly and difficult to reuse. At the intelligent application level, although large language models have demonstrated powerful capabilities in general domains, their direct application to highly specialized O&M scenarios suffers from shortcomings such as missing domain knowledge, inaccurate answers, and an inability to effectively trace the source of answers (i.e., the "illusion" problem). Furthermore, existing solutions often focus on static knowledge construction and one-way question answering, lacking a closed-loop mechanism for continuous learning and iteration based on practical feedback from frontline O&M personnel. This leads to lagging knowledge base updates and an inability to evolve in sync with actual O&M scenarios.

[0004] Therefore, how to build an operations and maintenance knowledge base system that can integrate multi-source heterogeneous operations and maintenance data, utilize large models to deeply understand professional knowledge, support accurate traceability and intelligent interaction, and have the ability to continuously evolve based on real feedback has become an urgent technical problem to be solved in the field of intelligent operations and maintenance. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method and system for constructing an operation and maintenance knowledge base based on a large model. The technical solution of this invention is as follows: A method for constructing an operations and maintenance knowledge base based on a large model includes the following steps: Step S1: Collect multimodal raw data from several operation and maintenance business systems, review and clean the raw data, and use a large model to extract entity and relationship information to build an initial operation and maintenance knowledge base; Step S2: Respond to the initial question input by the user through the interactive terminal, complete the question context through multiple rounds of dialogue, and use the large model to reconstruct the dialogue content into a standardized query that conforms to the preset field template; Step S3: Based on the normalized query, perform semantic similarity retrieval and knowledge graph association retrieval in parallel in the operation and maintenance knowledge base, merge the retrieval results and re-rank them; input the ranked knowledge fragments and the normalized query into the large model to generate answers with source citations; Step S4: Collect user evaluations of the answer's quality, and feed back complete interaction data, including queries, knowledge fragments, answers, and evaluations, to update the operation and maintenance knowledge base or optimize the large model.

[0006] Step S1 specifically includes: Step S11, Data Acquisition: Gather text and image data from equipment manuals, brand fault knowledge bases, engineer experience reports, and maintenance work order systems; Step S12, Content Review: The collected data is reviewed in sequence for timeliness, technical correctness, and applicability to operational scenarios; Step S13, Intelligent Cleaning and Segmentation: For the approved data, use a large language model to parse the document structure, split the document into multiple paragraph blocks based on semantic coherence and heading level, and generate summary tags for each paragraph block. Step S14, Multimodal Knowledge Extraction: For text paragraph blocks, use a large model to extract the entities of equipment, faults, causes, solutions and their relationships; for image data, use a visual large model to identify the content and generate descriptive text, and associate it with text entities.

[0007] The intelligent cleaning and splitting in step S13 specifically involves: S131, Document Parsing and Serialization: Parse the document to be processed into a sequence of paragraphs P = [p1, p2, ..., p...] arranged by natural paragraphs. n ], and identify the heading level L of each paragraph pi. i ; S132, Semantic Vectorization: Using a pre-trained large language model, each paragraph pi is encoded into a corresponding semantic vector vi, resulting in a semantic vector sequence V = [v1, v2, ..., v n ]; S133, Similarity Calculation: The semantic relevance between adjacent paragraphs is calculated based on the following cosine similarity formula: ; S134, Boundary Detection: Define a set of split points Cut = {}. Traverse the sequence of paragraphs. If any of the following conditions are met, set a split point after paragraph pi and add i to the set Cut: (a) Semantic similarity sim(v) i ,v i+1 )<θ, where θ is a preset similarity threshold; (b) Heading level L i > L i+1 ; S135, Paragraph Block Generation: Based on the split point set Cut, the original paragraph sequence P is divided into K semantic paragraph blocks B = [b1, b2, ..., b K ] ; S136, Summary Tag Generation: For each semantic paragraph block b k The input is fed into the large language model, which generates a text label Lk that summarizes its semantic content.

[0008] Step S2 specifically includes: Step S21: Receive the initial question through a dialog interface integrated into an instant messaging tool or professional application; Step S22: Based on the preset completion logic, guide the user to fill in the missing information such as the system, device brand, device model, and fault symptoms through a question-and-answer method. Step S23: Input the completed dialogue record into the large language model that has been fine-tuned by instructions. The model is configured to output a standardized query statement that conforms to the structure of "system-brand-model-phenomenon".

[0009] The steps of hybrid retrieval and source-based answer generation in step S3 specifically include: Step S31: Encode the normalized query into a vector and retrieve the K candidate text segments with the highest similarity from the vector database; Step S32: Extract entities from the normalized query, query related entities and relationship paths in the graph database, and obtain related knowledge fragments; Step S33: Combine the retrieval results of steps S31 and S32, use the cross-encoder model to score the relevance of each candidate knowledge fragment to the normalized query, sort them according to the scores, and select the top N fragments. Step S34: The standardized query and the first N knowledge fragments are constructed into prompt information according to a preset template and input into the generative large language model; the prompt information includes instructions requiring the model to generate an answer based on the provided knowledge and clearly mark the source.

[0010] The cross-encoder model in step S33 performs joint encoding by concatenating the normalized query with a candidate knowledge fragment into a whole sequence, uses a self-attention mechanism to capture the deep semantic interaction between the two, and outputs a probability value representing the relevance by the classification layer at the end of the model.

[0011] Step S4 specifically includes: Step S41: Provide a feedback entry point on the answer presentation interface to collect users' binary evaluations of whether the answer solves the problem and whether the cause and solution of the fault are correct; Step S42: The standardized query of this interaction, the final retrieved knowledge fragment, the generated answer and user evaluation are packaged into a training sample and stored in the feedback database. Step S43: Based on the samples in the feedback database, perform supervised fine-tuning on the generative large language model, or dynamically adjust the confidence weights of the referenced knowledge fragments in the operation and maintenance knowledge base.

[0012] The dynamic adjustment of the confidence weight W of the knowledge fragment in step S43 is performed according to the following algorithm: When a knowledge fragment is cited and receives positive feedback, its weight is updated according to the following formula: W new =W old +η×(1-W old ), When a knowledge fragment is cited and receives negative feedback, its weight is updated according to the following formula: W new =W old ×δ, where W old For the weights before adjustment, W new The weights are adjusted, η is the preset learning rate, and δ is the preset decay factor, satisfying 0 < η and δ < 1.

[0013] An operation and maintenance knowledge base construction system for implementing the method includes: The content management module is configured to perform step S1, which is used for data aggregation, review, cleaning and knowledge extraction. The RAG knowledge base module stores the knowledge graph and vector data generated by the content management module and provides a retrieval interface. The question supplementation module is configured to execute step S2 to achieve interactive query understanding and standardization; The AI-powered accurate response module is configured to execute step S3, which is used to perform hybrid retrieval and source-tracing answer generation; The feedback learning module is configured to perform step S4 to process user feedback and drive system optimization.

[0014] The content management module also includes a visual processing submodule, which is used to call the visual big model to process image data, generate text descriptions, and associate them with text knowledge entities.

[0015] The advantages of this invention are: 1. By collecting data from diverse and heterogeneous systems, including equipment manuals, brand fault databases, engineer reports, and maintenance work orders, and performing timeliness, accuracy, and universality checks, this method effectively solves the problem of difficulty in collecting key information such as on-site analysis data and paper records in traditional methods. Furthermore, by utilizing a large model for intelligent cleaning, segmentation, and multimodal knowledge extraction, unstructured text and image data can be transformed into a structured "equipment-fault-cause-solution" knowledge network. This ensures the accuracy, comprehensiveness, and systematic nature of knowledge from the source, laying a solid foundation for high-quality intelligent question answering.

[0016] 2. Addressing the pain points of vague and incomplete descriptions of operational and maintenance issues, this invention proactively completes key contextual information such as "system-brand-model-phenomenon" through multi-round guided dialogue, and then reconstructs it into a standardized query using a large model. This process simulates the scenario of expert questioning, significantly improving the system's accuracy in understanding the user's true intent and overcoming the shortcomings of general-purpose large models in understanding specialized fields.

[0017] 3. A hybrid retrieval strategy combining semantic vector retrieval and knowledge graph-based retrieval is employed, balancing semantic similarity and logical relevance to significantly improve the relevance and coverage of knowledge retrieval. Through cross-encoder re-ranking and source-tracing generation based on retrieval fragments, the final answer is ensured to originate strictly from a trusted knowledge base, with clearly marked citation sources. This generates accurate, verifiable, and traceable operational suggestions, greatly enhancing frontline operations personnel's trust in and reliance on intelligent decision-making support.

[0018] 4. A closed-loop learning mechanism based on user quality feedback has been introduced. The system not only provides answers but also collects user evaluations of the effectiveness of the answers and feeds back complete interaction data. This high-quality feedback data is used to optimize the generative model or adjust the confidence weights of knowledge in the knowledge base, enabling the system to continuously learn from real-world operational scenarios. This achieves dynamic updates to the knowledge base and iterative enhancement of question-answering capabilities, fundamentally solving the problems of traditional static knowledge bases being slow to update and unable to adapt to changes in the field.

[0019] This invention transforms fragmented and implicit operational experience into systematic, reusable, and precisely searchable explicit knowledge, and provides immediate and accurate assistance through natural language interaction. This significantly reduces reliance on a few senior experts, shortens the time for fault diagnosis and resolution, improves the overall skill level and response efficiency of the operations and maintenance team, and provides technical support for the standardization, automation, and intelligentization of operations and maintenance work. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the method of the present invention.

[0021] Figure 2 This is a block diagram of the main structure of the present invention.

[0022] Figure 3 yes Figure 2 A schematic diagram illustrating the working principles of the content management module and the RAG knowledge base module.

[0023] Figure 4 yes Figure 2 A schematic diagram illustrating the working principle of the question supplement module and the AI-based accurate response module. Detailed Implementation

[0024] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0025] See Figures 1 to 4 This invention relates to a method for constructing an operation and maintenance knowledge base based on a large model, comprising the following steps: Step S1: Collect multimodal raw data from several operation and maintenance business systems, review and clean the raw data, and use a large model to extract entity and relationship information to build an initial operation and maintenance knowledge base; Step S2: Respond to the initial question input by the user through the interactive terminal, complete the question context through multiple rounds of dialogue, and use the large model to reconstruct the dialogue content into a standardized query that conforms to the preset field template; Step S3: Based on the normalized query, perform semantic similarity retrieval and knowledge graph association retrieval in parallel in the operation and maintenance knowledge base, merge the retrieval results and re-rank them; input the ranked knowledge fragments and the normalized query into the large model to generate answers with source citations; Step S4: Collect user evaluations of the answer's quality, and feed back complete interaction data, including queries, knowledge fragments, answers, and evaluations, to update the operation and maintenance knowledge base or optimize the large model.

[0026] This invention, through step S1, systematically aggregates data from multiple business systems, including equipment manuals, fault databases, and work orders, and performs a triple review of timeliness, accuracy, and universality. Then, it utilizes a large model for intelligent cleaning and structured extraction. This ensures that the constructed operation and maintenance knowledge base not only has a wide range of data sources but also undergoes rigorous quality control, forming a complete, accurate, and reliable structured knowledge network, laying a solid foundation for subsequent intelligent applications.

[0027] Step S2, through multiple rounds of guided dialogue, proactively completes key contextual information such as "system, brand, model, and phenomenon," and uses a large model to reconstruct it into a standardized query conforming to a preset template. This process simulates the interactive logic of expert questioning, transforming ambiguous natural language intents into precise, machine-processable standardized queries, significantly improving the ability to understand the intent of complex professional issues.

[0028] In step S3, semantic similarity retrieval and knowledge graph association retrieval are performed in parallel, taking into account both text semantic matching and entity logical association, which greatly improves the comprehensiveness and relevance of knowledge recall. More importantly, the most relevant knowledge fragments are selected through re-ranking, and the generative model is forced to generate answers with source citations based on these fragments.

[0029] In step S4 of this invention, a feedback loop is designed. The system collects direct user evaluations of the answer's validity and feeds back complete interactive data, including the query, knowledge fragment, answer, and evaluation. This high-value data from real-world scenarios is used to update the knowledge base content (e.g., adjust knowledge weights) or optimize the generation model parameters. This enables the system to continuously learn from front-line operational practices, constantly correcting errors and strengthening effective knowledge, achieving dynamic iteration and continuous improvement in the system's intelligence level. This fundamentally solves the problem of traditional systems lagging knowledge updates and being unable to adapt to changes in the field.

[0030] Step S1 specifically includes: Step S11, Data Acquisition: Gather text and image data from equipment manuals, brand fault knowledge bases, engineer experience reports, and maintenance work order systems; Step S12, Content Review: The collected data is reviewed in sequence for timeliness, technical correctness, and applicability to operational scenarios; Step S13, Intelligent Cleaning and Segmentation: For the approved data, use a large language model to parse the document structure, split the document into multiple paragraph blocks based on semantic coherence and heading level, and generate summary tags for each paragraph block. Step S14, Multimodal Knowledge Extraction: For text paragraph blocks, use a large model to extract the entities of equipment, faults, causes, solutions and their relationships; for image data, use a visual large model to identify the content and generate descriptive text, and associate it with text entities.

[0031] The four sub-steps of step S1 above constitute a complete and rigorous process from data aggregation to knowledge structuring. Each step innovatively improves upon specific pain points in traditional operations and maintenance knowledge processing, and its advantages and examples are as follows: Step S11 (Data Acquisition) solves the problem of single data sources and achieves multi-dimensional aggregation of operation and maintenance knowledge. Traditional methods often rely solely on system logs or equipment manuals, neglecting valuable resources such as engineer experience and fault case libraries. This step systematically aggregates four core data categories: static knowledge (manuals), dynamic cases (fault library), expert experience (reporting records), and on-site issues (work orders), forming a data pool covering the entire operation and maintenance lifecycle.

[0032] For example, when dealing with a "precision air conditioner not cooling" fault, the system can simultaneously retrieve: the refrigeration principle diagram in the equipment manual (static knowledge), historical cases of "compressor abnormality of the same model" in the brand fault database (dynamic cases), special experience recorded in the engineer's report of "high humidity in a certain computer room causing condenser to freeze" (expert experience), and "alarm code E03" and on-site temperature described in the current work order (on-site problem), providing comprehensive information support for subsequent analysis.

[0033] Step S12 (Content Review) establishes a professional data quality control system for the operations and maintenance field, ensuring the accuracy and applicability of knowledge from the source. Traditional data cleaning often focuses on format specifications, lacking in-depth review at the domain level. This step introduces a triple filtering standard of "timeliness-accuracy-universality" to ensure that the knowledge entering the database is not only correctly formatted, but also reliable and practically valuable.

[0034] For example, a submitted "Server Memory Replacement Guide" will undergo the following checks: 1) Timeliness review: Determine if it targets an older, discontinued model (if so, it will be flagged or downgraded); 2) Correctness review: Verify that the operational steps (such as "hot-plugging") comply with current safety regulations; 3) Generality review: Assess if it only applies to customized models from a specific manufacturer (if so, its generality tag weight will be reduced). Only knowledge that passes the review is considered high-quality and generalizable.

[0035] Step S13 (Intelligent Cleaning and Segmentation) leverages the semantic understanding capabilities of a large-scale model to achieve intelligent, semantically preserving, granular processing of unstructured documents. Traditional document segmentation methods based on fixed lengths or simple delimiters can compromise semantic integrity. This step utilizes a large language model to understand document structure and semantics, segmenting documents based on semantic coherence and heading hierarchy. This ensures that each paragraph block is an independent semantic unit and automatically generates summary tags, greatly facilitating subsequent retrieval and understanding.

[0036] For example, when faced with a lengthy 100-page "Data Center Infrastructure Operation and Maintenance Manual," traditional methods might mechanically divide it into 500-word segments, disrupting the integrity of the "Uninterruptible Power Supply (UPS) Battery Replacement Procedure." This approach, however, can recognize the chapter structure and intelligently break down "5.3 Replacement Process" under "Chapter 5 Battery Maintenance" and all its subordinate steps (safety preparation, disassembly, installation, testing) into a complete paragraph block, generating tags such as "UPS Battery Standard Replacement Operating Procedure."

[0037] Step S14 (Multimodal Knowledge Extraction) realizes the fusion, understanding and structuring of text and image data, and constructs a machine-readable and reasonable operation and maintenance knowledge graph.

[0038] Traditional methods struggle to process information in images, and textual information is often unstructured. This step uses a large model to extract entities and relationships, transforming text into triples such as "device-fault-cause-solution"; simultaneously, it leverages a visual large model to "understand" the image, generating descriptive text and associating it with relevant textual entities, thereby constructing a visual knowledge network rich in logical relationships.

[0039] For example, from a fault report text, the entities "[Precision Air Conditioner]", "[Alarm E03]", and "[Condenser Blockage]" are extracted, along with the relationships: "[Alarm E03] - Cause - [Condenser Blockage]", and "[Condenser Blockage] - Solution - [Clean the filter and check the air duct]". Simultaneously, the accompanying on-site photos are recognized by the visual big data model, generating a description: "The image shows that the condenser fins are covered with a large amount of dust." The system associates this description with the entity "Condenser Blockage". Ultimately, all this information is integrated into a knowledge graph, supporting subsequent precise queries and root cause reasoning.

[0040] Therefore, step S1, through these four interconnected sub-steps, completes the transformation of operation and maintenance data from "multi-source heterogeneous raw data" to "high-quality, structured, multi-modal fused knowledge", laying a solid data foundation for building an intelligent and reliable operation and maintenance knowledge base.

[0041] The intelligent cleaning and splitting in step S13 specifically involves: S131, Document Parsing and Serialization: Parse the document to be processed into a sequence of paragraphs P = [p1, p2, ..., p...] arranged by natural paragraphs. n ], and identify the heading level L of each paragraph pi. i ; S132, Semantic Vectorization: Using a pre-trained large language model, each paragraph pi is encoded into a corresponding semantic vector vi, resulting in a semantic vector sequence V = [v1, v2, ..., v n ]; S133, Similarity Calculation: The semantic relevance between adjacent paragraphs is calculated based on the following cosine similarity formula: ; S134, Boundary Detection: Define a set of split points Cut = {}. Traverse the sequence of paragraphs. If any of the following conditions are met, set a split point after paragraph pi and add i to the set Cut: (a) Semantic similarity sim(v) i ,v i+1 )<θ, where θ is a preset similarity threshold; (b) Heading level L i > L i+1 ; S135, Paragraph Block Generation: Based on the split point set Cut, the original paragraph sequence P is divided into K semantic paragraph blocks B = [b1, b2, ..., b K ] ; S136, Summary Tag Generation: For each semantic paragraph block b k The input is fed into the large language model, which then generates a text label L that summarizes its semantic content. k .

[0042] The above steps represent a leap from "mechanical segmentation" to "semantic understanding": traditional methods, which cut segments by fixed lengths or simple delimiters, often compromise semantic integrity. This method, however, understands semantics through a large model, ensuring that each output knowledge block (semantic paragraph block) is thematically cohesive and logically complete.

[0043] At the same time, it uses cosine similarity to quantify semantic associations while respecting the inherent logic of the document reflected in the heading hierarchy, making the segmentation decision both flexible and rigorous, and greatly improving the accuracy and rationality of the segmentation.

[0044] It not only completes the segmentation, but also automatically generates summary tags for each knowledge block, directly serving efficient retrieval and knowledge management, realizing processing as application.

[0045] The following step-by-step instructions are based on a comprehensive example, the "Chiller Unit Operation Manual," which includes chapters such as "Safety Specifications," "Operating Procedures," and "Troubleshooting."

[0046] Advantages of S131 (Document Parsing and Serialization): This provides a "skeletal map" of the document, giving it structural awareness for subsequent analysis. This step transforms unstructured text into a structured sequence with key metadata (heading levels Li). This is equivalent to establishing a coordinate system for the document, enabling subsequent algorithms to "know" the logical position of each paragraph within its context (such as which chapter or section it belongs to).

[0047] Example: The manual is parsed as a sequence: [p1: “Chapter 1 Safety Specifications” (L1), p2: “1.1 Personal Protective Equipment” (L2), p3: “Insulating gloves must be worn…” (L2), p4: “Chapter 2 Start-up Procedures” (L1), ...].

[0048] The advantages of S132 (semantic vectorization) are: it achieves "deep semantic encoding" of text, transforming language into computable mathematical objects. By pre-training a large model, each paragraph is encoded as a high-dimensional semantic vector vi, and the spatial relationships between these vectors (such as angles and distances) directly and quantitatively reflect the semantic relevance between paragraphs.

[0049] Example: The vectors v3 and v2 of p3 (Personal Protective Equipment Details) and p2 (Personal Protective Equipment Title) are close in direction (high similarity); while the vectors v4 of v3 and p4 (Initiation Process Title) are quite different in direction (low similarity).

[0050] Advantages of S133 (Similarity Calculation): It provides an objective and quantifiable "semantic coherence index," replacing subjective experience-based judgment. Using the cosine similarity formula, it accurately calculates the similarity sim(v) between adjacent paragraph vectors. i , v i+1 This value directly and objectively indicates whether the two texts are semantically closely connected or have undergone a significant shift.

[0051] Example: Calculating sim(v3, v4), which is the similarity between the “Personal Protective Equipment Details” paragraph and the “Chapter 2 Startup Process” title paragraph, may yield a low result (e.g., 0.15), quantitatively confirming the existence of a significant semantic jump here.

[0052] The advantages of S134 (Boundary Detection): It formulates "intelligent segmentation rules" that integrate semantics and structure, achieving accurate and interpretable decisions. This step is the decision-making center of the algorithm. It combines two conditions: (a) semantic similarity threshold and (b) heading hierarchy logic, to scientifically determine the cutting point. Condition (a) captures implicit topic shifts, and condition (b) respects the explicit structure of the document.

[0053] Example: 1. Triggering condition (a): Between p3 and p4, since sim(v3, v4)=0.15 < θ (let θ=0.5), the algorithm determines that this is the semantic boundary between chapters and sets a split point after p3.

[0054] 2. Triggering condition (b): When p is processed 10 : “2.3.1 Start the main pump (L3)” and p 11 The heading "2.4 Operational Monitoring (L2)" changes from L3 to L2 (L2) despite the potentially related semantics. 10 > L 11 This indicates the end of a more granular sub-item of the operation and returns to the parent topic list; therefore, it is also present in p. 10 Then set the split point.

[0055] The advantages of S135 (paragraph block generation): Based on the set of cut points (Cut), the original document is divided into multiple semantic paragraph blocks (B). The paragraphs within each block are highly semantically related, while the boundaries between blocks are relatively clear.

[0056] Example: Ultimately, the manual is divided into something like b1: [p1, p2, p3] (the complete first chapter of the safety specifications), b2: [p4, p5, ..., p 10 [Chapter 2 Startup Procedure to "Starting the Main Pump"], b3: [p11, ...] (Chapter 2 "Operation Monitoring" section) and other independent knowledge blocks.

[0057] The advantages of S136 (Summary Tag Generation): It automatically generates "precise semantic summaries" for knowledge units, greatly improving the efficiency of subsequent retrieval and management. Utilizing the summarizing capabilities of a large language model, it generates a refined tag L_k for each knowledge block b_k. This is equivalent to automatically generating a high-quality keyword index for the knowledge base.

[0058] Example: Tag b1 might generate "Safety Protection Requirements for Chiller Unit Operation," while tag b2 might generate "Standard Start-up Procedures and Main Pump Start-up Steps for Chiller Units." When maintenance personnel search for "Start Main Pump," b2 can be quickly and accurately located through its tag.

[0059] In summary, the core advantage of this series of steps lies in the fact that it solves the problem of intelligent document processing through algorithmic and quantitative methods, achieving three major goals: fidelity (preserving semantic integrity), efficiency (automated processing), and usability (outputting tagged knowledge units). This provides crucial high-quality data preprocessing capabilities for building a high-quality, easily searchable operations and maintenance knowledge base.

[0060] Step S2 specifically includes: Step S21: Receive the initial question through a dialog interface integrated into an instant messaging tool or professional application; Step S22: Based on the preset completion logic, guide the user to fill in the missing information such as the system, device brand, device model, and fault symptoms through a question-and-answer method. Step S23: Input the completed dialogue record into the large language model that has been fine-tuned by instructions. The model is configured to output a standardized query statement that conforms to the structure of "system-brand-model-phenomenon".

[0061] Step S2 (Interactive Query Understanding and Standardization) constructs a bridge from fuzzy natural language descriptions to precise, machine-efficient standardized queries, fundamentally solving the problems of incomplete user query information and arbitrary expression in operation and maintenance scenarios, which lead to misunderstandings and retrieval failures in intelligent systems.

[0062] Step S21 provides a low-barrier, highly convenient interactive entry point that aligns with users' work habits, ensuring the system can seamlessly integrate into actual workflows. Traditional knowledge base systems often require users to log into independent platforms and fill out complex forms, resulting in a high barrier to entry. This step integrates the entry point into instant messaging tools such as WeChat and DingTalk, or maintenance work order apps. Users can directly ask questions through their most familiar chat windows when encountering problems, greatly reducing usage resistance and ensuring the system's accessibility and high-frequency use.

[0063] For example, when maintenance engineer Li was inspecting the data center, he discovered that a precision air conditioner was malfunctioning. Instead of logging into a specific system on his computer, he simply contacted the maintenance assistant in the work WeChat group and typed, "The air conditioner seems to be stopping cooling, please take a look." This step ensured that the problem was detected immediately and conveniently.

[0064] Step S22 (Guided Information Completion): Through pre-defined domain logic, the system proactively initiates multi-round dialogues to systematically uncover and complete the key context behind the problem, transforming vague requirements into a complete work order-style description. The system no longer passively accepts incomplete input but, like an experienced colleague, proactively guides the user to complete the information based on the core information framework of the operations and maintenance domain: "System Affiliation - Equipment Brand - Equipment Model - Fault Phenomenon."

[0065] For example, when the system receives a vague question from Engineer Li, it automatically triggers a guided dialogue: The system asked: "Which system does the air conditioner with the problem belong to? Is it the data center cooling system or the office air conditioning system?" Engineer Li replied, "The data center cooling system." The system asked, "Okay, what is the brand of this air conditioner? For example: Gree, Haier, York." Mr. Li replied, "It's from York." The system asked: "What is the specific model of the device? You can find it on the nameplate on the device." After checking, Engineer Li replied, "The model number is YCAG-123." The system asked: "Finally, please describe the 'no cooling' phenomenon in more detail, such as: no airflow from the vent, hot air being blown, alarm codes displayed on the panel, etc." Mr. Li replied, "The airflow from the vent is very small, and the panel displays the code 'AL-01'." Through these rounds of interaction, a vague question that originally only contained the five words "the air conditioner is not cooling" was completed into a full context including "data center cooling system - York - YCAG-123 - low air volume, alarm AL-01" (the following content is only for illustrative purposes).

[0066] The advantage of step S23 (query structure reconstruction) is that by using a large language model that has been fine-tuned by instructions, loose dialogue records are intelligently extracted and formatted into standardized query statements, providing a "structured search key" for subsequent accurate retrieval.

[0067] While the completed chat log is information-rich, it remains in natural language, which is not conducive to the most accurate matching by the knowledge base. This step, through a domain-fine-tuned large model, acts like a skilled technical secretary, organizing the chat log into a standard "query ticket".

[0068] The system takes all the above dialogue records (Q&A) as input and feeds them into a specially trained query reconstruction model. This model strictly follows the template of "system-brand-model-phenomenon" and outputs the following standardized query statement: "System: Data center cooling system, Brand: York, Model: YCAG-123, Fault: Low air volume, Control panel alarm code AL-01." This query is highly structured, eliminates redundancy, and standardizes terminology; it can be used directly for: 1. Accurately match the entity “York YCAG-123” in the knowledge graph.

[0069] 2. In the vector database, perform high similarity matching with documents that record the "AL-01 alarm handling solution".

[0070] 3. As an explicit input for multi-condition retrieval.

[0071] The overall advantage of step S2 lies in its transformation of an unstructured, incomplete social dialogue into a structured, complete, and standardized query. The example of the transformation from "the air conditioner isn't cooling" to the final standard query vividly demonstrates how this process significantly bridges the semantic gap between humans and machines. This lays a decisive foundation for the accurate and comprehensive retrieval in step S3, making it a crucial link in improving accuracy throughout the entire intelligent question-answering process.

[0072] The steps of hybrid retrieval and source-based answer generation in step S3 specifically include: Step S31: Encode the normalized query into a vector and retrieve the K candidate text segments with the highest similarity from the vector database; Step S32: Extract entities from the normalized query, query related entities and relationship paths in the graph database, and obtain related knowledge fragments; Step S33: Combine the retrieval results of steps S31 and S32, use the cross-encoder model to score the relevance of each candidate knowledge fragment to the normalized query, sort them according to the scores, and select the top N fragments. Step S34: The standardized query and the first N knowledge fragments are constructed into prompt information according to a preset template and input into the generative large language model; the prompt information includes instructions requiring the model to generate an answer based on the provided knowledge and clearly mark the source.

[0073] Step S3 (hybrid retrieval and source-based answer generation) constructs a robust answer generation framework that ensures both the breadth and depth of retrieval and the accuracy and credibility of answers. This framework effectively addresses the limitations of traditional single retrieval methods and constrains the generation behavior of large models, making their output verifiable and auditable.

[0074] The following example of troubleshooting a simple maintenance issue illustrates the synergistic effect of each step: Overall Example: Troubleshooting the "Data Center Precision Air Conditioning Alarm E03" Issue Assume that after step S2, the system generates a normalized query: System: Data center cooling system; Brand: Emerson; Model: Liebert PEX4; Fault symptom: Frequent compressor start-stop, control panel displays alarm code E03; Advantages of step S31 (semantic vector retrieval): Achieve broad recall based on deep semantic similarity to capture knowledge that is semantically related but has different expressions.

[0075] Example: The system encodes the query as a vector and retrieves the following candidate fragments from the vector database: 1. Segment A (General Knowledge): "Common reasons for frequent start-stop of air conditioner compressors: insufficient refrigerant, unstable power supply voltage, and poor heat dissipation."

[0076] 2. Excerpt B (Forum Experience): "An engineer reported that his Liebert air conditioner was reporting an E03 error, and the problem was resolved after cleaning the condenser."

[0077] 3. Fragment C (similar case): "The PEX4 air conditioner in a computer room issued an E03 alarm, and the inspection revealed that the high-voltage switch was faulty."

[0078] Advantages of step S32 (knowledge graph retrieval): Achieve precise association and reasoning based on entity logical relationships, and find solutions for direct matching and deep association.

[0079] Example: The system extracts the entities "Emerson", "Liebert PEX4", and "Alarm Code E03" from the query and queries them in the graph database: Locate the entity "Liebert PEX4" and associate it with the "Standard Fault Code Table".

[0080] Following the relationship of "Fault Code: E03", we found its official explanation: "E03: High pressure protection, possible causes: condenser blockage, cooling fan failure, excessive refrigerant, high pressure switch malfunction." Then, the associated "Standard Procedure" entity was found, and fragment D (official SOP) was obtained: "Emerson Liebert PEX4 Series E03 Alarm Handling Steps: 1. Inspect and clean the condenser fins; 2. Test the cooling fan operation; 3. Check the refrigerant pressure; 4. Check the high-pressure switch." The advantages of step S33 (fusion and reordering): Through a deep semantic interaction model, heterogeneous search results are sorted in a unified and accurate manner, and the most relevant and informationally valuable knowledge is selected.

[0081] Example: The system merges fragments A, B, C, and D, and uses a cross-encoder model to evaluate their relevance to the original query one by one. The model can understand: Fragment D directly and accurately corresponds to the brand, model, and alarm code in the query, and has the highest relevance.

[0082] Segments B and C mention "cleaning the condenser" and "high-pressure switch failure" respectively. These pieces of information are highly relevant to the official explanations of "condenser blockage" and "high-pressure switch," and are based on specific cases, so their relevance is secondary.

[0083] Although segment A is related to the topic (compressor start-stop), it does not mention the specific brand and model or the E03 code, and it lacks specificity, making it the least relevant.

[0084] Assume the score ranking is: Segment D (0.96) > Segment C (0.88) > Segment B (0.85) > Segment A (0.65). The system selects the first N=3 segments: D, C, B.

[0085] The advantage of step S34 (source generation) is that it forces the generation process to be strictly faithful to the provided authoritative knowledge source and outputs it in an explicit citation manner, ensuring the accuracy and credibility of the answer.

[0086] Example: The system constructs the following prompt message to input a generative large model: Please answer the question about Emerson Liebert PEX4 air conditioner alarm E03 and frequent compressor start-stop based on the information provided below.

[0087] Knowledge: [1] Official handling steps: Emerson Liebert PEX4 series E03 alarm handling steps: 1. Check and clean the condenser fins; 2. Test the operation of the cooling fan; 3. Check the refrigerant pressure; 4. Check the high pressure switch.

[0088] [2] Related case: The PEX4 air conditioner in a computer room had an E03 alarm. The inspection revealed that the high-voltage switch was faulty.

[0089] [3] Engineer's experience: Some engineers have reported that when Liebert air conditioners report E03, the problem is solved after cleaning the condenser.

[0090] Please generate your answer strictly based on the above knowledge and indicate the source of the reference, such as [1].

[0091] The large model generates the final answer based on this: "According to official information [1], the E03 alarm you encountered on your Emerson Liebert PEX4 air conditioner is a 'high pressure protection'. It is recommended to troubleshoot in the following steps: 1. Check and clean the condenser first: this is the most common cause [1] [3]=3. Dirty fins can cause poor heat dissipation and trigger high pressure protection. 2. Check the cooling fan: ensure that the fan is running normally to ensure heat dissipation of the condenser [1]. 3. Check the refrigerant pressure: excessive pressure may also trigger protection [1]. 4. Check the high pressure switch: a faulty component can falsely report E03 [1] [2]; please troubleshoot in this order, which can usually solve the problem." This design effectively tames and guides the powerful generation capabilities of large models towards professional and reliable output, which is the technical guarantee for the present invention to achieve accurate, reliable, and traceable intelligent operation and maintenance Q&A.

[0092] The cross-encoder model in step S33 jointly encodes the normalized query and a candidate knowledge fragment into a whole sequence, using a self-attention mechanism to capture the deep semantic interaction between them, and outputs a probability value representing the relevance by the classification layer at the end of the model. Through deep and fine-grained semantic interaction matching, it fundamentally solves the problems of coarse semantic understanding and inability to perceive complex semantic relationships between queries and documents in traditional retrieval systems (such as independent vector retrieval), thus achieving a crucial leap from "semantic similarity" to "semantic relevance."

[0093] For example, given the standardized query "alarm code AL-01" and the candidate fragment "check fan operating status and capacitor," a dual encoder might only learn that "alarm" and "check" have a certain degree of similarity. However, a cross encoder, during joint encoding, can dynamically learn the strong association between the specific entity "AL-01" and "fan" and "capacitor," and can even capture the implicit logical relationship between "code" and "status" (the code indicates the status), thus determining that "this document specifically explains the check items for the AL-01 alarm," resulting in a very high relevance score. The cross encoder can effectively suppress the former (outputting low probability) while greatly improving the latter (outputting high probability) because it understands that although the former is topic-related, it does not specifically answer the core of the current query (specific brand, specific model, specific code).

[0094] After hybrid retrieval, the result set may contain duplicate, similar, or unevenly related fragments. The reordering step of the cross-encoder acts as a crucial quality filter and information concentrator. It uses precise scoring to select the few most relevant, informative, and likely-to-be-used fragments for constructing the answer, and then passes them to the subsequent generative model.

[0095] Assuming the graph retrieval returns the official Standard Operating Procedure (SOP), and the vector retrieval returns engineer experience and a general explanation, the cross-encoder, through deep interaction, might give a score: SOP (0.98) > Engineer Experience (0.85) > General Explanation (0.70) > Another general maintenance guide (0.20). The system selects the top three. Thus, the knowledge fragment ultimately input into the generative model (step S34) is a high-quality, highly relevant, and noise-free collection, greatly reducing the risk of the generative model generating irrelevant or erroneous content ("illusions") and improving the accuracy and conciseness of the generated answers.

[0096] Step S4 specifically includes: Step S41: Provide a feedback entry point on the answer presentation interface to collect users' binary evaluations of whether the answer solves the problem and whether the cause and solution of the fault are correct; Step S42: The standardized query of this interaction, the final retrieved knowledge fragment, the generated answer and user evaluation are packaged into a training sample and stored in the feedback database. Step S43: Based on the samples in the feedback database, perform supervised fine-tuning on the generative large language model, or dynamically adjust the confidence weights of the referenced knowledge fragments in the operation and maintenance knowledge base.

[0097] The advantage of step S4 (feedback-driven knowledge iteration optimization) is that it constructs a "learning-optimization" closed loop based on real user feedback, enabling the entire system to upgrade from a static, one-time knowledge base to a dynamic, intelligent system that can continuously learn and evolve from practical experience.

[0098] For example: the optimization process for handling the "server hard drive alarm" problem. Suppose the system receives a query: "A certain brand of server hard drive indicator light is constantly on, alarm", and generates an answer through steps S1-S3, which quotes knowledge fragment X: "A constantly on hard drive indicator light usually indicates that the hard drive is about to fail, and it is recommended to back up the data immediately and prepare to replace the hard drive." Advantages of step S41 (feedback collection): It provides a simple, low-barrier feedback channel to ensure that a large number of high-quality real-world verification signals can be collected efficiently.

[0099] Example and Explanation: After seeing the answer, the operations engineer followed the instructions (backing up and replacing the hard drive), but the problem persisted (the indicator light remained on). He directly clicked the feedback option on the answer screen: [Answer did not solve the problem], [Indicated cause and solution: Incorrect]. This binary evaluation design is simple and direct, adding almost no burden to the user, yet it captures the most crucial "effectiveness" and "correctness" signals, providing a clear direction for system optimization.

[0100] The advantage of step S42 (data encapsulation) is that it fully records the decision-making chain and encapsulates an interaction into a context-rich "causal" sample that can be used for machine learning.

[0101] Example and Explanation: The system automatically packages the complete data from this interaction into: 1) Standardized query (including server model and phenomenon); 2) The cited knowledge fragment X; 3) The generated answer; 4) The user's negative evaluation. These four elements together constitute a high-quality "negative sample." It tells the system: "When encountering this type of query, citing fragment X and providing this answer is deemed invalid by the user." This structured record is the cornerstone for subsequent precise optimization.

[0102] The advantages of step S43 (closed-loop optimization) are: it provides two complementary optimization paths, enabling dual iteration of the system's "knowledge content" and "expressive ability".

[0103] Path 1: Adjust the confidence weight of knowledge fragments Example and Explanation: According to the algorithm in step S43, the system applies the negative feedback weight decay formula to knowledge fragment X. Assuming its original weight is 0.8 and the decay factor δ = 0.6, the new weight becomes 0.8 × 0.6 = 0.48. This means that the reliability of fragment X is significantly reduced in the system. Result: In the future, when similar queries occur again, due to the reduced weight of fragment X, its priority in the retrieval and sorting stages will decrease, and the system will be more inclined to refer to other alternative solutions with higher weights (e.g., "check hard drive backplane connection" or "update RAID card firmware"), thereby avoiding repeated errors.

[0104] Path 2: Supervised fine-tuning of generative large language models Example and Explanation: The system periodically uses accumulated feedback data (such as the negative samples mentioned above, as well as numerous positive samples) as training data to fine-tune the large language model for generating answers. After learning from these samples, the model gradually understands that: "When the query involves 'the hard drive light of a certain brand of server is always on,' it is not appropriate to hastily generate a suggestion of 'replacing the hard drive immediately,' but rather to be more cautious or suggest combining other inspection steps." Result: Even if the same fragment X is retrieved, the fine-tuned model may add more rigorous statements such as "but connection or firmware issues should be ruled out first" when generating answers, thereby improving the accuracy and reliability of the generated answers.

[0105] The overall advantages of step S4 are clearly demonstrated through the above examples: 1. Effective knowledge (receiving positive feedback) is reinforced by increasing its weight, while ineffective or outdated knowledge (receiving negative feedback) is quickly demoted or marginalized, thus automatically maintaining a high level of usability in the knowledge base.

[0106] 2. Under the "tutelage" of a large amount of real feedback, the generative model has learned how to better combine and express knowledge, reducing arbitrariness and errors, and making its answers more professional and comprehensive.

[0107] 3. It is no longer a tool that provides information in a one-way manner, but a learning partner that can continuously improve itself based on user feedback and grow together with users. This enhances the long-term value and vitality of the system.

[0108] The dynamic adjustment of the confidence weight W of the knowledge fragment in step S43 is performed according to the following algorithm: When a knowledge fragment is cited and receives positive feedback, its weight is updated according to the following formula: W new =W old +η×(1-W old ), When a knowledge fragment is cited and receives negative feedback, its weight is updated according to the following formula: W new =W old ×δ, where W old For the weights before adjustment, W new The weights are adjusted, η is the preset learning rate, and δ is the preset decay factor, satisfying 0 < η and δ < 1.

[0109] The dynamic adjustment of the confidence weight W for the knowledge fragment in step S43 transforms the abstract concept of "knowledge reliability" into a concrete, calculable weight value W. This weight is no longer static but changes dynamically with each use and verification result.

[0110] The adjusted weight W can directly or indirectly affect subsequent steps. For example, in the reordering stage of step S33 or the retrieval stage of S31, knowledge fragments with high weights can receive bonus points, making them more likely to be prioritized for retrieval and adoption; conversely, fragments with low weights will be marginalized. This enables a positive cycle of high-quality knowledge within the system.

[0111] Suppose that in the operations and maintenance knowledge base, there are two knowledge fragments about solving the problem of "frequent server restarts": Fragment P: Weight W P = 0.6. The message reads: "Check if the CPU cooling fan is dusty and clean it." Fragment Q: Weight W Q = 0.8. The content is: "For a certain model of A100 server, frequent restarts may be due to the 'Power Limit' setting in the BIOS being too low. It is necessary to enter the BIOS and adjust the 'PL Limit' parameter."

[0112] Scene evolution: 1. Initial state: Because fragment Q is more specific and professional, its initial weight (0.8) is higher than that of fragment P (0.6). When a user queries "A100 server frequently restarts", the system is more likely to retrieve and use fragment Q to generate an answer.

[0113] 2. Positive Feedback Reinforcement: The user adopts the suggestion in fragment Q, successfully solves the problem, and gives a positive evaluation. The system updates the weights of fragment Q. Assuming the learning rate η = 0.1, then: W Qnew =0.8+0.1×(1-0.8)=0.8+0.02=0.82; The weight of fragment Q was slightly increased, consolidating its status as "high-value knowledge".

[0114] 3. Negative Feedback and Correction: Another user reported frequent restarts on an "old desktop computer," suggesting the system might also be referencing fragment P. However, the problem persisted after cleaning the fan, leading to a negative review. The system updated the weight of fragment P. Assuming the attenuation factor delta = 0.7, then: W Pnew =0.6×0.7=0.42, the weight of segment P drops significantly due to an invalid feedback.

[0115] 4. System Evolution Results: After multiple rounds of interaction, fragment Q, due to its continued effectiveness, may steadily increase its weight to above 0.9, becoming high-confidence core knowledge; while fragment P, either due to limited applicability or being too general, will see its weight drop to very low (e.g., 0.3) after receiving several negative feedbacks, gradually becoming marginalized by the system. When new users ask questions, the system will prioritize and recommend high-weight knowledge like fragment Q, which has been "tested in practice."

[0116] The fundamental advantage of the weight dynamic adjustment algorithm in step S43 lies in its transformation of human feedback into an algorithmic driving force for the self-optimization of the knowledge base. Through a mathematical design that combines gradual reinforcement with rapid decay, it achieves automation, data-driven continuous iteration, and quality improvement of the knowledge base.

[0117] This invention also relates to an operation and maintenance knowledge base construction system for implementing the method, comprising: Content management module 1 is configured to perform step S1 for data aggregation, review, cleaning and knowledge extraction; RAG knowledge base module 2 stores the knowledge graph and vector data generated by the content management module and provides a retrieval interface; Question supplement module 3 is configured to execute step S2 to achieve interactive query understanding and standardization; AI-powered precise response module 4 is configured to execute step S3, which is used to perform hybrid retrieval and source-tracing answer generation; Feedback learning module 5 is configured to perform step S4 to process user feedback and drive system optimization.

[0118] The content management module also includes a visual processing submodule, which is used to call the visual big model to process image data, generate text descriptions, and associate them with text knowledge entities.

[0119] Through the collaborative work of five core modules, a data-driven, intelligent interactive, and closed-loop evolutionary intelligent operation and maintenance decision support platform has been built. Its core advantage lies in its ability to systematically transform fragmented operation and maintenance data, implicit expert experience, and real-time user feedback into searchable, reasonable, and evolving explicit knowledge through large-scale model technology, and to provide accurate and reliable intelligent question-and-answer services.

[0120] The working principle of this system is explained step by step as follows: Knowledge construction and storage (corresponding to step S1): Content Management Module 1 begins its work, aggregating multimodal data from various business systems (such as the equipment database and work order system). It performs review and cleaning, and uses a large model (including a visual processing submodule) to extract entities, relationships, and image descriptions, outputting structured knowledge.

[0121] This knowledge is fed into RAG Knowledge Base Module 2, where it is constructed into a knowledge graph that facilitates relational queries and a vector database that facilitates semantic retrieval, forming the "brain" of the system.

[0122] Intelligent question answering service (corresponding to steps S2 and S3): When maintenance personnel encounter problems and ask questions through WeChat, the app, or other entry points, the question supplement module 3 is activated.

[0123] It completes information through multi-turn dialogues and uses a large model to reconstruct the dialogues into standardized queries (such as "system-brand-model-phenomenon").

[0124] The query was simultaneously sent to the AI-powered precise response module 4. This module initiated a "hybrid search" request to the RAG knowledge base module 2: performing a semantic search in the vector database on one hand, and a relational query in the graph database on the other.

[0125] The AI-powered precise response module 4 integrates and refines the search results, and finally submits the most relevant knowledge fragments along with the query to the generative big data model to generate an answer with clear citation sources, which is then returned to the user.

[0126] Feedback and evolution (corresponding to step S4): Users can submit simple feedback on the answer page (such as "Is it resolved?" or "Is it correct?").

[0127] The feedback learning module 5 captures this feedback and encapsulates the complete data of this interaction (query, knowledge fragments used, generated answers, user reviews) into a training sample.

[0128] This module performs optimizations based on these samples: first, it dynamically adjusts the confidence weights of referenced knowledge fragments in RAG knowledge base module 2 (e.g., increasing the weight for positive feedback and decreasing it for negative feedback); second, it periodically uses accumulated samples to perform supervised fine-tuning of the generative model in AI accurate response module 4. The optimization results directly affect the next round of knowledge retrieval and answer generation, thus achieving closed-loop evolution of the system.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing an operation and maintenance knowledge base based on a large model, characterized in that, Includes the following steps: Step S1: Collect multimodal raw data from several operation and maintenance business systems, review and clean the raw data, and use a large model to extract entity and relationship information to build an initial operation and maintenance knowledge base; Step S2: Respond to the initial question input by the user through the interactive terminal, complete the question context through multiple rounds of dialogue, and use the large model to reconstruct the dialogue content into a standardized query that conforms to the preset field template; Step S3: Based on the normalized query, perform semantic similarity retrieval and knowledge graph association retrieval in parallel in the operation and maintenance knowledge base, merge the retrieval results and re-rank them; input the ranked knowledge fragments and the normalized query into the large model to generate answers with source citations; Step S4: Collect user evaluations of the answer's quality, and feed back complete interaction data, including queries, knowledge fragments, answers, and evaluations, to update the operation and maintenance knowledge base or optimize the large model.

2. The method for constructing an operation and maintenance knowledge base based on a large model according to claim 1, characterized in that, Step S1 specifically includes: Step S11, Data Acquisition: Gather text and image data from equipment manuals, brand fault knowledge bases, engineer experience reports, and maintenance work order systems; Step S12, Content Review: The collected data is reviewed in sequence for timeliness, technical correctness, and applicability to operational scenarios; Step S13, Intelligent Cleaning and Segmentation: For the approved data, use a large language model to parse the document structure, split the document into multiple paragraph blocks based on semantic coherence and heading level, and generate summary tags for each paragraph block. Step S14, Multimodal Knowledge Extraction: For text paragraph blocks, use a large model to extract the entities of equipment, faults, causes, solutions and their relationships; for image data, use a visual large model to identify the content and generate descriptive text, and associate it with text entities.

3. The method according to claim 2, characterized in that, The intelligent cleaning and splitting in step S13 specifically involves: S131, Document Parsing and Serialization: Parse the document to be processed into a sequence of paragraphs P = [p1, p2, ..., p...] arranged by natural paragraphs. n ], and identify the heading level L of each paragraph pi. i ; S132, Semantic Vectorization: Using a pre-trained large language model, each paragraph pi is encoded into a corresponding semantic vector vi, resulting in a semantic vector sequence V = [v1, v2, ..., v n ]; S133, Similarity Calculation: The semantic relevance between adjacent paragraphs is calculated based on the following cosine similarity formula: ; S134, Boundary Detection: Define a set of split points Cut = {}. Traverse the sequence of paragraphs. If any of the following conditions are met, set a split point after paragraph pi and add i to the set Cut: (a) Semantic similarity sim(v) i ,v i+1 )<θ, where θ is a preset similarity threshold; (b) Heading level L i > L i+1 ; S135, Paragraph Block Generation: Based on the split point set Cut, the original paragraph sequence P is divided into K semantic paragraph blocks B = [b1, b2, ..., b...]. K ] ; S136, Summary Tag Generation: Generate each semantic paragraph block b k The input is fed into the large language model, which generates a text label Lk that summarizes its semantic content.

4. The method according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Receive the initial question through a dialog interface integrated into an instant messaging tool or professional application; Step S22: Based on the preset completion logic, guide the user to fill in the missing information such as the system, device brand, device model, and fault symptoms through a question-and-answer method. Step S23: Input the completed dialogue record into the large language model that has been fine-tuned by instructions. The model is configured to output a standardized query statement that conforms to the structure of "system-brand-model-phenomenon".

5. The method according to claim 1, characterized in that, The steps of hybrid retrieval and source-based answer generation in step S3 specifically include: Step S31: Encode the normalized query into a vector and retrieve the K candidate text segments with the highest similarity from the vector database; Step S32: Extract entities from the normalized query, query related entities and relationship paths in the graph database, and obtain related knowledge fragments; Step S33: Combine the retrieval results of steps S31 and S32, use the cross-encoder model to score the relevance of each candidate knowledge fragment to the normalized query, sort them according to the scores, and select the top N fragments. Step S34: The standardized query and the first N knowledge fragments are constructed into prompt information according to a preset template and input into the generative large language model; the prompt information includes instructions requiring the model to generate an answer based on the provided knowledge and clearly mark the source.

6. The method according to claim 5, characterized in that, The cross-encoder model in step S33 jointly encodes the normalized query and a candidate knowledge fragment into a whole sequence, uses a self-attention mechanism to capture the deep semantic interaction between the two, and outputs a probability value representing the relevance by the classification layer at the end of the model.

7. The method according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Provide a feedback entry point on the answer presentation interface to collect users' binary evaluations of whether the answer solves the problem and whether the cause and solution of the fault are correct; Step S42: The standardized query of this interaction, the final retrieved knowledge fragment, the generated answer and user evaluation are packaged into a training sample and stored in the feedback database. Step S43: Based on the samples in the feedback database, perform supervised fine-tuning on the generative large language model, or dynamically adjust the confidence weights of the referenced knowledge fragments in the operation and maintenance knowledge base.

8. The method according to claim 7, characterized in that, The dynamic adjustment of the confidence weight W of the knowledge fragment in step S43 is performed according to the following algorithm: When a knowledge fragment is cited and receives positive feedback, its weight is updated according to the following formula: IN new =W old +η×(1-W old ), When a knowledge fragment is cited and receives negative feedback, its weight is updated according to the following formula: W new =W old ×δ, where W old For the weights before adjustment, W new The weights are adjusted, η is the preset learning rate, and δ is the preset decay factor, satisfying 0 < η and δ < 1.

9. A system for constructing an operation and maintenance knowledge base to implement the method of any one of claims 1 to 8, characterized in that, include: The content management module is configured to perform step S1, which is used for data aggregation, review, cleaning and knowledge extraction. The RAG knowledge base module stores the knowledge graph and vector data generated by the content management module and provides a retrieval interface. The question supplementation module is configured to execute step S2 to achieve interactive query understanding and standardization; The AI-powered accurate response module is configured to execute step S3, which is used to perform hybrid retrieval and source-tracing answer generation; The feedback learning module is configured to perform step S4 to process user feedback and drive system optimization.

10. The system according to claim 9, characterized in that, The content management module also includes a visual processing submodule, which is used to call the visual big model to process image data, generate text descriptions, and associate them with text knowledge entities.

Images