LLM-based operation and maintenance knowledge intelligent question and answer method and system

Abstract

The invention discloses an LLM-based intelligent question and answer method and system for operation and maintenance knowledge, relates to the technical field of artificial intelligence, and provides an efficient knowledge management and update mechanism which can continuously absorb latest operation and maintenance knowledge and technical practice, and operation and maintenance personnel can upload new knowledge documents to the system through a data entry unit, so that the operation and maintenance knowledge can be quickly and conveniently answered. After verification of the data checking unit, the knowledge can be automatically partitioned and vectorized and supplemented into the knowledge base, the mechanism not only ensures timeliness and accuracy of the knowledge base, but also promotes accumulation and inheritance of the knowledge, new operation and maintenance personnel can quickly obtain necessary knowledge and skills through the system, and the operation and maintenance efficiency is improved. In addition, the system also supports the processing of multi-round dialogues and context relevance problems, so that the solution of complex problems is smoother and more efficient.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an LLM-based operation and maintenance knowledge intelligent question-answering method and system. Background Art

[0002] With the rapid development of information technology, operation and maintenance work has become more and more complex and arduous. Driven by technologies such as cloud computing, big data and the Internet of Things, the scale and complexity of the system continue to increase, and the difficulty of operation and maintenance management has also increased. The operation and maintenance team needs to process large amounts of data and respond quickly to various emergencies while ensuring the stability and security of the system.

[0003] Then, traditional operation and maintenance methods often rely on the experience and skills of operation and maintenance personnel, and it is difficult to quickly respond to complex operation and maintenance needs. It is not only inefficient but also prone to errors. At the same time, the inheritance and sharing of operation and maintenance knowledge also face many challenges. New employees are slow to get started, and the resignation of old employees leads to knowledge gaps. Faced with massive operation and maintenance data, traditional methods are difficult to quickly and accurately locate problems and provide solutions, affecting the efficiency and accuracy of operation and maintenance.

[0004] Therefore, how to combine the advantages of LLM to intelligently answer operation and maintenance problems and provide intelligent question-answering methods and systems for standardized operating procedures, thereby improving operation and maintenance efficiency and accuracy has become an important issue. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and to provide an intelligent question-and-answer method and system for operation and maintenance knowledge based on LLM. By providing multiple input methods and an effective question checking and processing mechanism, the user's convenience of use and the accuracy of the question-and-answer system are greatly improved. Users can choose the appropriate input method according to their own needs. At the same time, the system can intelligently identify and process repeated, non-compliant or context-related questions to ensure that the returned answers are both accurate and compliant with the specifications. This not only improves the efficiency of operation and maintenance work, but also reduces the potential risks caused by wrong answers.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: On the one hand, an intelligent question-answering method for operation and maintenance knowledge based on LLM, the specific steps of the method are:

[0007] S1. Provide multiple input methods, including text box input, voice-to-text input conversion, and receive questions raised by users;

[0008] S2. Check the submitted questions for duplication, compliance, and contextual relevance, including:

[0009] For repetitive questions, get the results from the cache and return the answers. The questions and answers in the cache are stored in the form of key-value pairs, with the question as the key and the answer as the value.

[0010] For non-compliant issues, we analyze them word by word. For issues that contain sensitive words or violate process specifications, we immediately trigger the feedback mechanism and return reminder information.

[0011] For questions with contextual associations, when users ask questions in multiple rounds of conversations, the logical associations between questions are identified and multiple questions with contextual associations are merged for processing;

[0012] S3. After receiving the query request from the user after the question check, the question input by the user is converted into a vector form by using the vector retrieval technology, and an efficient search is performed in the knowledge base. The operation and maintenance knowledge in the knowledge base is also stored in the form of vectors. By calculating the similarity between the question vector and each knowledge vector in the knowledge base, multiple groups of answers with the most similar vectors are retrieved;

[0013] S4. Summarize the multiple sets of answers returned, evaluate the relevance of the answers to the user's questions, and when there are answers that meet the relevance threshold, refine and summarize the answers and present them to the user;

[0014] When there is no answer that meets the relevance threshold, the external model interface is called to summarize the resources and present them to the user;

[0015] S5. Send the summarized answers to the user in a formatted form.

[0016] Furthermore, the intelligent question-answering method further includes knowledge base construction and updating, that is, inputting and uploading operation and maintenance knowledge document data, the operation and maintenance knowledge document data includes text and pictures, and the text information in the pictures is extracted and converted by optical character recognition (OCR) technology to be converted into editable text data, and the input operation and maintenance knowledge document data is checked for repeatability and compliance, and the newly added operation and maintenance knowledge document data that passes the check and verification is divided into blocks, that is, the number of blocks after division is n, and the i-th block is represented by B i , embed the divided content into the vector and encoder model, and represent the vector of the i-th block after model processing as V i , construct an index by calculating the similarity between vectors, that is, Where V i ·V j Represents vector V i and V j The dot product of |V i | and |V j | respectively represent vector Vi and V j Model, create LLM prompts according to the knowledge content, add the processed knowledge to the knowledge base, and realize the update of the knowledge base.

[0017] Furthermore, the specific steps of checking for repetitive questions and returning answers in S2 are:

[0018] The question raised by the user is matched with the key in the cache, where the key-value pair storage structure in the cache is represented as: Cache = {(Q1, A1), (Q2, A2), ..., (Q n , A n )} Among them, Q i Representative question, A i Represents the answer;

[0019] If the match is successful, the corresponding value is directly returned as the answer;

[0020] If the match fails, the compliance check and contextual relevance problem check process continues.

[0021] Furthermore, when S3 performs vectorization processing on the user's question, the user's question is cleaned and stop words are removed. The original user's question is represented as Q, the cleaned question is Q', and the stop word set is ST. Then Q'=Q-{w|w∈ST}, the cleaned question Q' is vectorized, and the vector is represented as V Q , by comprehensively analyzing the vocabulary, grammatical structure, and semantic information in the question to generate vectors, the vector representation set corresponding to the vocabulary is obtained as {v w |w∈Q′}, the importance weight of the word w in the operation and maintenance field is I w , the frequency of occurrence is f w , then the weight of word w in the vector is W w =I w ×f w , through V Q =∑ w∈Q′ W w ×v w The final vector V is obtained by Q .

[0022] Furthermore, when S3 performs index search, the knowledge vector set in the knowledge base is K = {V k |k=1,2,…,m}, the problem vector is V Q , calculate the similarity score between the question vector and each knowledge vector in the knowledge base. The calculation process is: Score(V Q , V k )=α×Similarity(V Q , Vk )+β×Weight(V k ), where α and β are weight parameters, Similarity(V Q , V k ) is the similarity calculated based on the vector, Weight(V k ) is the knowledge vector V k The knowledge vectors are sorted according to the similarity score, and the text blocks corresponding to the vectors with high similarity scores are selected as the retrieval results. At the same time, a comprehensive evaluation is performed based on the weight and timeliness of the knowledge in the knowledge base. For the newly updated operation and maintenance knowledge, a high weight γ is given, and the longer the update time is, the lower the weight is, that is, the update time of knowledge k is t k , then its timeliness weight is Where t is the current time, λ is the attenuation coefficient, and the final comprehensive weight is W total (V k )=γ×T k ×Weight(V k ) and adjust the search results according to the comprehensive weight.

[0023] Furthermore, the S4 determines the relevance score by the degree of semantic association, evaluates the relevance of the answer to the user question, defines the answer as A, and the user question as Q, and calculates the degree of matching between the answer and the question keywords, that is, Where count(w, A) represents the number of times keyword w appears in answer A, w is the importance weight of keyword w, |Q| is the number of keywords in question Q, and the answers that meet the relevance threshold θ are refined and summarized and presented to the user. When there is no answer that meets the relevance threshold θ, the answer set E = {A e |e=1,2,…,p}, integrate and summarize to get the final answer A′ and present it to the user.

[0024] On the other hand, an intelligent question-answering system for operation and maintenance knowledge based on LLM, the system comprises: a questioning module, a checking module, a question retrieval module, a summarizing module, a feedback module, and a knowledge base maintenance module;

[0025] The question module is used to obtain questions submitted by users;

[0026] The checking module is used to check the repetitiveness, compliance and contextual relevance of the submitted questions, obtain the results from the cache and return the answers for the repetitive questions, return reminder information for the non-compliant questions, and merge multiple questions for the questions with contextual relevance;

[0027] The question retrieval module accepts the user's query request and retrieves multiple groups of similar vector answers from the knowledge base according to the input;

[0028] The summarizing module summarizes the answers returned by the search, and for answers that do not meet the relevance threshold, attempts to obtain other answers by calling the external model interface and then summarizes them;

[0029] The feedback module replies to the user in a formatted language according to the answer of the summary module;

[0030] The knowledge base maintenance module supports the entry of operation and maintenance knowledge document data in multiple formats, performs duplication and compliance checks on the operation and maintenance knowledge document data, and adds the verified knowledge to the knowledge base, thereby updating and expanding the knowledge base in real time.

[0031] Furthermore, the question retrieval module is composed of a preprocessing unit, a retrieval unit and a response unit, wherein:

[0032] The preprocessing unit vectorizes the user's question using a transformer encoder model;

[0033] The retrieval unit performs an index search on the question vector, retrieves the corresponding text block from the knowledge base, and retrieves the results whose relevance meets the threshold range;

[0034] The response unit performs splicing based on the user's question, prompts and query results that meet the relevance threshold, and connects multiple answers.

[0035] Furthermore, the knowledge base maintenance module is composed of a data entry unit, a data checking unit and a knowledge base supplement unit, wherein:

[0036] The data entry unit is used to enter and upload a variety of operation and maintenance knowledge document data, and the operation and maintenance knowledge document data includes text and pictures;

[0037] The data checking unit is used to check and verify the repeatability and compliance of the operation and maintenance knowledge document data, and when the data verification passes, confirm that the operation and maintenance knowledge document data is supplemented into the knowledge base;

[0038] The knowledge base supplement unit is used to divide the newly added operation and maintenance knowledge document data into blocks, embed the blocks into vectors and transformer encoder models, and build vector indexes. According to the characteristics of the knowledge content and the user's question type, LLM prompts are created and supplemented to the knowledge base.

[0039] Compared with the prior art, the LLM-based intelligent question-answering method and system for operation and maintenance knowledge has the following beneficial effects:

[0040] 1. The present invention provides an efficient knowledge management and updating mechanism, which can continuously absorb the latest operation and maintenance knowledge and technical practices. Through the data entry unit, the operation and maintenance personnel can upload new knowledge documents to the system. After verification by the data inspection unit, the knowledge will be automatically divided into blocks, vectorized, and added to the knowledge base. This mechanism not only ensures the timeliness and accuracy of the knowledge base, but also promotes the accumulation and inheritance of knowledge. Newly employed operation and maintenance personnel can quickly acquire necessary knowledge and skills through the system, and the experience of senior operation and maintenance personnel can also be effectively preserved and shared. In addition, the system also supports multi-round dialogues and the processing of context-related problems, making the solution of complex problems smoother and more efficient. In this way, the overall technical level of the operation and maintenance team is improved, enabling enterprises to better cope with the ever-changing IT environment.

[0041] 2. The present invention integrates a large language model LLM to build an intelligent operation and maintenance knowledge base, so that operation and maintenance personnel can quickly obtain standardized operation instructions and solutions. By providing multiple input methods and an effective problem inspection and processing mechanism, the user's convenience and the accuracy of the question and answer system are greatly improved. With intelligent operation and maintenance guidance, operation and maintenance personnel can handle various complex problems, thereby significantly improving the intelligence level of operation and maintenance work.

[0042] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a flowchart of an intelligent question-answering method for operation and maintenance knowledge based on LLM;

[0045] Figure 2 This is a module composition diagram of an LLM-based operation and maintenance knowledge intelligent question answering system;

[0046] Figure 3 This is a unit composition diagram of the knowledge base maintenance module of an LLM-based operation and maintenance knowledge intelligent question-answering system;

[0047] Figure 4This is a unit composition diagram of the question retrieval module of an operation and maintenance knowledge intelligent question answering system based on LLM. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Embodiment 1

[0050] This example demonstrates how the LLM-based intelligent question-and-answer system for operation and maintenance knowledge can efficiently handle routine operation and maintenance problems in daily operation and maintenance scenarios of enterprises. Through a detailed operation process, including question submission, multi-faceted inspection, vector retrieval, answer summary and feedback, it demonstrates the significant advantages of the system in improving operation and maintenance efficiency, enhancing personnel skills, and ensuring the effective inheritance and application of operation and maintenance knowledge.

[0051] In the specific implementation, first, in the daily operation and maintenance work environment of the enterprise, the operation and maintenance personnel are equipped with computer terminals installed with the operation and maintenance knowledge intelligent question and answer system based on LLM. The information technology infrastructure of the enterprise includes multiple key components: server clusters, which run multiple operating system instances and provide computing resources for various business applications; network equipment, including routers, switches, etc., responsible for data transmission and exchange, to ensure the internal and external network communication of the enterprise; database systems, which store massive business data, such as customer information, transaction records, etc.; and enterprise resource planning (ERP) systems for enterprise resource management and customer relationship management (CRM) systems for customer relationship maintenance. Before the system is activated, the preliminary construction of the knowledge base has been completed, which covers a wealth of operation and maintenance knowledge, including common server failure types and solutions, such as the diagnosis and repair steps of server hardware failures (such as hard disk damage, memory failure, etc.); the troubleshooting process of network failures (such as network connection interruption, slow network speed, etc.); the processing strategy of database problems (such as slow data query, data loss, etc.); and solutions to common problems of various business applications. At the same time, the knowledge base maintenance module is in normal operation, able to monitor and obtain the latest operation and maintenance knowledge in real time, and update and supplement the knowledge base in a timely manner.

[0052] Through simulation, the operation and maintenance personnel found that the ERP system was running slowly during the daily inspection, which affected the business operation efficiency. Therefore, the operation and maintenance personnel chose to enter the question "ERP system is running slowly, how to find out the reason?" in the text box in the user interface of the intelligent question-answering system. After the question module receives the question entered by the user, it starts a series of processing processes. The inspection module immediately conducts a comprehensive inspection of the submitted question. First, a repeatability check is performed, and the question is compared with the historical questions in the cache one by one. The key-value pair storage structure in the cache is represented as: Cache = {(Q1, A1), (Q2, A2), …, (Q n ,A n )}, where Q i Representative question, A i Represents the answer. Through efficient string matching, the similarity between the new question and the question in the cache is calculated. If the similarity calculation result is high, it is considered to be a repeated question and the corresponding value is directly returned as the answer. If no matching repeated question is found, the compliance check and context-related question check process will continue to be executed. The compliance check module analyzes the question word by word and checks whether the question contains sensitive words or violates the operation and maintenance process specifications based on the compliance rule library. If sensitive words or violations of specifications are detected, the feedback mechanism will be triggered immediately to return detailed reminder information to the operation and maintenance personnel, informing them of the reasons for the non-compliance of the problem and providing modification suggestions. In the context-related check scenario of multiple rounds of dialogue, this module will identify the logical associations between questions, merge multiple questions with contextual relationships, and form a more complete and accurate comprehensive problem description so that subsequent modules can better understand user needs.

[0053] After the question check and processing, the question retrieval module starts working. The preprocessing unit first cleans the user's question and removes stop words. The user's original question is defined as Q, the cleaned question is Q', and the stop word set is ST. Then Q' = Q-{w|w∈ST}, remove common stop words, and vectorize the cleaned question Q'. During the vectorization process, the transformer encoder model is used to comprehensively analyze the vocabulary, grammatical structure, and semantic information in the question. The model decomposes Q' into vocabulary units, and the vector representation set corresponding to the vocabulary is obtained as {V w |w∈Q′}. For each word w, its importance weight in the operation and maintenance field is I w , the frequency of occurrence is f w , then the weight of word w in the vector is W w =I w ×f w , and finally through V Q =∑ w∈Q′ W w ×v w, we get the vector V representing the problem Q , the retrieval unit receives the question vector V Q After that, an index search is performed in the knowledge base, and the knowledge vector set in the knowledge base is K = {V k |k=1,2,…,m}, calculate the similarity score between the question vector and each knowledge vector in the knowledge base. The calculation process is Score(V Q , V k )=α×Similarity(V Q , V k )+β×Weight(V k ), where α and β are weight parameters, Similarity(V Q , V k ) is the similarity calculated based on the vector, Weight(V k ) is the knowledge vector V k The system will sort the knowledge vectors according to the similarity score and select the text blocks corresponding to the vectors with similarity scores higher than the preset threshold θ as the retrieval results. The system will make a comprehensive evaluation based on the weight and timeliness of the knowledge in the knowledge base. For the recently updated operation and maintenance knowledge, a higher weight γ (γ>1) will be given. The longer the update time is, the lower the weight will be. That is, the update time of knowledge k is t k , then its timeliness weight is (where t is the current time and λ is the attenuation coefficient), the final comprehensive weight is W total (V k )=γ×T k ×Weight(V k ), further adjust the search results according to the comprehensive weight to ensure that the returned answers are both relevant and timely.

[0054] The summary module processes multiple sets of answers returned by the retrieval, evaluates the relevance of the answers to the user questions, defines the answer as A, and the user question as Q. By calculating the degree of matching between the answer and the question keywords, that is, Where count(w, A) represents the number of times keyword w appears in answer A, wis the importance weight of keyword w, |Q| is the number of keywords in question Q, and when the relevance score exceeds the preset threshold θ, it is considered to meet the relevance requirement; for answers that meet the relevance threshold, the summary module will refine and summarize, remove redundant information, and generate concise and clear answers with prominent emphasis. For example, in this embodiment, the key troubleshooting steps such as "first check the server CPU and memory usage, if the usage is too high, analyze the process occupying resources; then check whether the database query statement is optimized and check the execution plan; finally check whether the network connection is normal and whether there is packet loss" are summarized. The feedback module displays the summarized answers in a clear and formatted text form on the terminal screen of the operation and maintenance personnel. The display format includes point-by-point listing of steps and highlighting of key information so that the operation and maintenance personnel can quickly understand and follow. The operation and maintenance personnel will gradually perform troubleshooting operations according to the answers provided by the system, check the monitoring indicators of the server, and determine that it is a background process related to the ERP system by further analyzing the process information. Then check the database query statement according to the answer prompt. In the whole process, the operation and maintenance personnel not only solved the current problem, but also learned the standard processing procedures and methods for such problems through the detailed troubleshooting steps and solutions provided by the system. This helps improve the skills of operation and maintenance personnel, and also promotes the inheritance and sharing of enterprise operation and maintenance knowledge. When encountering similar problems in the future, operation and maintenance personnel can handle them more skillfully, or share this experience with other colleagues, further enriching the enterprise's operation and maintenance knowledge base.

[0055] In summary, this example fully demonstrates the effectiveness and advantages of the LLM-based intelligent question-and-answer system for operation and maintenance knowledge in dealing with routine operation and maintenance problems. Through the collaborative work of multiple modules, from question submission to final answer feedback, each link reflects the intelligence and efficiency of the system. The system can not only quickly and accurately provide problem solutions for operation and maintenance personnel, reduce problem troubleshooting time, and improve operation and maintenance efficiency, but also play a positive role in knowledge inheritance and personnel skills improvement.

[0056] Embodiment 2

[0057] This embodiment shows how the LLM-based operation and maintenance knowledge intelligent question-answering system plays a key role in emergency situations when business is interrupted due to sudden complex failures in the core business system of an enterprise. It elaborates on the entire process from question submission, special situation handling, in-depth retrieval and analysis to final problem solving, highlighting the system's ability to respond quickly in emergency situations, its ability to effectively handle complex problems, and its important role in ensuring the continuity of enterprise business.

[0058] The core business system of the enterprise suddenly encountered a serious failure, causing the business to be interrupted. This emergency situation posed a huge threat to the normal operation of the enterprise. The operation and maintenance team immediately activated the emergency response mechanism. At this time, the mobile device was connected to the LLM-based operation and maintenance knowledge intelligent question and answer system. The system is configured with a knowledge base, which contains system architecture knowledge, that is, the connection mode and dependency relationship of each component of the system; comprehensive fault handling cases, the phenomenon, cause and solution of various historical faults; and detailed emergency operation procedures, which clearly define the emergency measures and operation steps to be taken in different fault scenarios. The system supports voice input, which can facilitate the operation and maintenance personnel to quickly submit questions in an emergency. It also has multi-language processing capabilities to meet the language habits of different operation and maintenance personnel and the diverse needs of multinational companies. The operation and maintenance personnel input the question through voice: "The core business system has a large area of ​​service unavailable, and the error prompt is 'database connection timeout'. What should I do?" After the voice input is converted into text, the question module quickly receives the question and simultaneously starts various processing flows. The inspection module quickly starts working, and the repeatability inspection module compares the question with the historical questions in the cache and finds that the question is not exactly the same as the previous one. However, semantic analysis identifies that there is a certain correlation, so the relevant context information is retained for subsequent comprehensive consideration. After the compliance check ensures that the question does not involve any security or process violations, the context relevance check plays an important role, merging the current question with previous related questions to form a more logically complete and information-rich problem description, laying the foundation for subsequent accurate retrieval of answers.

[0059] After the question retrieval module receives the processed questions, the preprocessing unit begins to vectorize the merged questions. The preprocessing unit corrects the text and strengthens the extraction and analysis of keywords. Through the transformer encoder model, the vocabulary, grammatical structure, and semantic information in the question are deeply understood and encoded. The model assigns a specific vector representation to each word and weights the word according to its importance in the operation and maintenance field. w (its value is determined based on the operation and maintenance knowledge system and historical data statistics) and the frequency of occurrence f w (obtained through statistics of the number of times words appear in the knowledge base), calculate the weight W of the word in the vector w =I w ×f w , and finally through the weighted summation method V Q =∑ w∈Q′ W w ×v w , where Q′ is the question text after cleaning and error correction, v w is the vector representation of the word w, and we get the vector V that accurately represents the problem Q The retrieval unit is based on the question vector VQ Perform a comprehensive search in the knowledge base, and the knowledge vector set in the knowledge base is K = {V k |k=1,2,…,m},calculate the similarity score between the question vector and each knowledge vector in the knowledge base, calculated as Score(V Q , V k )=α×Similarity(V Q , V k )+β×Weight(V k ), where α and β are weight parameters, Similarity(V Q , V k ) is the similarity calculated based on the vector space model, Weight(V k ) is the knowledge vector V k When considering the timeliness factor, let the current time be t and the update time of knowledge k be t k , then its timeliness weight is (λ is the attenuation coefficient, which is determined according to the knowledge update frequency and importance). A higher weight γ (γ>1) is given to the recently updated fault handling knowledge. The final comprehensive weight is W total (V k )=γ×T k ×Weight(V k ), sort the knowledge vectors according to the similarity score and comprehensive weight, and select the text blocks corresponding to the knowledge vectors with the highest relevance as the retrieval results. The summary module processes the retrieved answers. Due to the complexity and urgency of the questions, when the answers retrieved initially do not fully meet the relevance threshold θ, the summary module quickly calls the external model interface (with the database fault diagnosis model) to obtain more resources and information.

[0060] Integrate and summarize all the answers obtained (including answers returned by internal retrieval and external models), and calculate the relevance score between the answers and the questions Where A is the answer, Q is the question, count(w, A) indicates the number of times keyword w appears in answer A, I w is the importance weight of keyword w, |Q| is the number of keywords in question Q, and the most relevant and valuable answer information is screened out. The answer is refined and summarized, redundant and irrelevant information is removed, and a clear, accurate, and operational final answer is generated. The feedback module presents the summarized answer to the operation and maintenance personnel in two ways: voice broadcast and text display. The voice broadcast function enables the operation and maintenance personnel to obtain key information in time during the operation, and the text display makes it easy to carefully view the detailed steps.

[0061] After the fault is resolved, the system summarizes the entire emergency response process, documents the details of the fault, the resolution process, and the lessons learned, and adds new knowledge and experience to the knowledge base through the knowledge base maintenance module, realizing real-time updating of the knowledge base. This not only enriches the content of the knowledge base, but also provides a more comprehensive and effective solution reference for similar faults that may occur in the future, further enhancing the system's ability to deal with complex faults.

[0062] In summary, this embodiment vividly demonstrates the excellent performance of the LLM-based operation and maintenance knowledge intelligent question and answer system in emergency situations when enterprises face complex failures in core business systems. The system, with its powerful functions and efficient processing procedures, plays an irreplaceable role in emergency response, problem handling, answer provision, and knowledge updating. Through a series of advanced technologies and mechanisms such as voice input, multi-language support, contextual association processing, deep vector retrieval, and external model calls, it achieves rapid diagnosis and effective resolution of complex failures, effectively ensures the continuity of enterprise business, and reduces economic losses and reputation risks caused by failures.

[0063] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An intelligent question-answering method for operation and maintenance knowledge based on LLM, characterized in that: The specific steps of this method are: S1. Provide multiple input methods, including text box input, voice-to-text input conversion, and receive questions raised by users; S2. Check the submitted questions for duplication, compliance, and contextual relevance, including: For repetitive questions, get the results from the cache and return the answers. The questions and answers in the cache are stored in the form of key-value pairs, with the question as the key and the answer as the value. For non-compliant issues, we analyze them word by word. For issues that contain sensitive words or violate process specifications, we immediately trigger the feedback mechanism and return reminder information. For questions with contextual associations, when users ask questions in multiple rounds of conversations, the logical associations between questions are identified and multiple questions with contextual associations are merged for processing; S3. After receiving the query request from the user after the question check, the question input by the user is converted into a vector form by using the vector retrieval technology, and an efficient search is performed in the knowledge base. The operation and maintenance knowledge in the knowledge base is also stored in the form of vectors. By calculating the similarity between the question vector and each knowledge vector in the knowledge base, multiple groups of answers with the most similar vectors are retrieved; S4. Summarize the multiple sets of answers returned, evaluate the relevance of the answers to the user's questions, and when there are answers that meet the relevance threshold, refine and summarize the answers and present them to the user; When there is no answer that meets the relevance threshold, the external model interface is called to summarize the resources and present them to the user; S5. Send the summarized answers to the user in a formatted form.

2. According to claim 1, an intelligent question-answering method for operation and maintenance knowledge based on LLM is characterized in that: The intelligent question-answering method also includes knowledge base construction and updating, that is, inputting and uploading operation and maintenance knowledge document data, the operation and maintenance knowledge document data includes text and pictures, and the text information in the pictures is extracted and converted by optical character recognition (OCR) technology to be converted into editable text data, and the input operation and maintenance knowledge document data is checked for repeatability and compliance, and the newly added operation and maintenance knowledge document data that passes the inspection and verification is divided into blocks, that is, the number of blocks after division is n, and the i-th block is represented by B i , embed the divided content into the vector and encoder model, and represent the vector of the i-th block after model processing as V i , construct an index by calculating the similarity between vectors, that is, Where V i ·V j Represents vector V i and V i The dot product of |V i | and |V j | respectively represent vector V i and V j Model, create LLM prompts according to the knowledge content, add the processed knowledge to the knowledge base, and realize the update of the knowledge base.

3. The intelligent question-answering method for operation and maintenance knowledge based on LLM according to claim 1 is characterized in that: The specific steps of checking for repetitive questions and returning answers in S2 are: The question raised by the user is matched with the key in the cache, where the key-value pair storage structure in the cache is represented as: Cache = {(Q1, A1), (Q2, A2), ..., (Q n ,A n )} Among them, Q i Representative question, A i Represents the answer; If the match is successful, the corresponding value is directly returned as the answer; If the match fails, the compliance check and contextual relevance problem check process continues.

4. The intelligent question-answering method for operation and maintenance knowledge based on LLM according to claim 1 is characterized in that: When S3 performs vectorization processing on user questions, the user's questions are cleaned and stop words are removed. The original user question is represented as Q, the cleaned question is Q', and the stop word set is ST. Then Q'=Q-{w|w∈ST}, and the cleaned question Q' is vectorized, and the vector is represented as V Q , by comprehensively analyzing the vocabulary, grammatical structure, and semantic information in the question to generate vectors, the vector representation set corresponding to the vocabulary is obtained as {v w |w∈Q′}, the importance weight of the word w in the operation and maintenance field is I w , the frequency of occurrence is f w , then the weight of word w in the vector is W w =I w ×f w , through V Q =∑ w∈Q′ W w ×v w The final vector V is obtained by Q .

5. The intelligent question-answering method for operation and maintenance knowledge based on LLM according to claim 1 is characterized in that: When S3 performs index search, the knowledge vector set in the knowledge base is K={V k |k=1,2,…,m}, the problem vector is V Q , calculate the similarity score between the question vector and each knowledge vector in the knowledge base. The calculation process is: Score(V Q ,V k )=α×Similarity(V Q ,V k )+β×Weight(V k ), where α and β are weight parameters, Similarity(V Q ,V k ) is the similarity calculated based on the vector, Weight(V k ) is the knowledge vector V k The knowledge vectors are sorted according to the similarity score, and the text blocks corresponding to the vectors with high similarity scores are selected as the retrieval results. At the same time, a comprehensive evaluation is performed based on the weight and timeliness of the knowledge in the knowledge base. For the newly updated operation and maintenance knowledge, a high weight γ is given, and the longer the update time is, the lower the weight is, that is, the update time of knowledge k is t k , then its timeliness weight is Where t is the current time, λ is the attenuation coefficient, and the final comprehensive weight is W total (V k )=γ×T k ×Weight(V k ) and adjust the search results according to the comprehensive weight.

6. The LLM-based intelligent question-answering method for operation and maintenance knowledge according to claim 1, characterized in that: The S4 determines the relevance score by the degree of semantic association, evaluates the relevance of the answer to the user question, defines the answer as A, and the user question as Q, and calculates the degree of matching between the answer and the question keywords, that is, Where count(w,A) represents the number of times keyword w appears in answer A. w is the importance weight of keyword w, |Q| is the number of keywords in question Q, and the answers that meet the relevance threshold θ are refined and summarized and presented to the user. When there is no answer that meets the relevance threshold θ, the answer set E = {A e |e=1,2,…,p}, integrate and summarize to get the final answer A′ and present it to the user.

7. An intelligent question-answering system for operation and maintenance knowledge based on LLM, characterized in that: The system consists of: question module, inspection module, question retrieval module, summary module, feedback module, knowledge base maintenance module; The question module is used to obtain questions submitted by users; The checking module is used to check the repetitiveness, compliance and contextual relevance of the submitted questions, obtain the results from the cache and return the answers for the repetitive questions, return reminder information for the non-compliant questions, and merge multiple questions for the questions with contextual relevance; The question retrieval module accepts the user's query request and retrieves multiple groups of similar vector answers from the knowledge base according to the input; The summarizing module summarizes the answers returned by the search, and for answers that do not meet the relevance threshold, attempts to obtain other answers by calling the external model interface and then summarizes them; The feedback module replies to the user in a formatted language according to the answer of the summary module; The knowledge base maintenance module supports the entry of operation and maintenance knowledge document data in multiple formats, performs duplication and compliance checks on the operation and maintenance knowledge document data, and adds the verified knowledge to the knowledge base, thereby updating and expanding the knowledge base in real time.

8. The intelligent question-answering system for operation and maintenance knowledge based on LLM according to claim 7 is characterized in that: The question retrieval module is composed of a preprocessing unit, a retrieval unit and a response unit, wherein: The preprocessing unit vectorizes the user's question using a transformer encoder model; The retrieval unit performs an index search on the question vector, retrieves the corresponding text block from the knowledge base, and retrieves the results whose relevance meets the threshold range; The response unit performs splicing based on the user's question, prompts and query results that meet the relevance threshold, and connects multiple answers.

9. The LLM-based operation and maintenance knowledge intelligent question-answering system according to claim 7, characterized in that: The knowledge base maintenance module is composed of a data entry unit, a data checking unit and a knowledge base supplement unit, wherein: The data entry unit is used to enter and upload a variety of operation and maintenance knowledge document data, and the operation and maintenance knowledge document data includes text and pictures; The data checking unit is used to check and verify the repeatability and compliance of the operation and maintenance knowledge document data, and when the data verification passes, confirm that the operation and maintenance knowledge document data is supplemented into the knowledge base; The knowledge base supplement unit is used to divide the newly added operation and maintenance knowledge document data into blocks, embed the blocks into vectors and transformer encoder models, and build vector indexes. According to the characteristics of the knowledge content and the user's question type, LLM prompts are created and supplemented to the knowledge base.

Images