Iron tower fault operation and maintenance knowledge retrieval method and system based on large model

Abstract

The invention discloses an iron tower fault operation and maintenance knowledge retrieval method and system based on a large model, and belongs to the technical field of text retrieval and knowledge questioning and answering, and the method comprises the steps: classifying iron tower operation and maintenance knowledge according to business scenes, and carrying out father-child segmentation processing on a full-category knowledge text; and preprocessing user question information based on the large language model, and outputting a user core question. And performing mixed retrieval in the hierarchical structure knowledge base based on the user core problem, and outputting a first retrieval result. And checking the first retrieval result according to a preset response rule and outputting a second retrieval result. According to the method, text segmentation is carried out through a parent block-child block segmentation mode, sorting is carried out by extracting a user core problem and retrieving related child blocks, the context of the parent block is associated, and a result is output based on a confidence coefficient verification rule. High-precision knowledge retrieval is achieved, illusion content is avoided, and the fault handling efficiency and knowledge obtaining efficiency of operation and maintenance personnel are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of text retrieval and knowledge question answering, and particularly relates to a tower fault operation and maintenance knowledge retrieval method and system based on a large model. BACKGROUND

[0002] In the field of communication infrastructure, tower operation and maintenance is a key link to ensure the stable operation of mobile networks. The tower operation and maintenance knowledge system is complex, covering historical fault cases, base station power support, dynamic environment monitoring and equipment operation specifications, etc. At present, this knowledge usually exists in a fragmented form, scattered in paper manuals, electronic documents and the personal experience of operation and maintenance personnel, lacking a unified and efficient management and retrieval carrier. In the prior art, the management and retrieval of such knowledge mainly exist in the following schemes: 1. Text retrieval based on keyword matching: This scheme usually adopts a fixed-length blocking method (such as fixed 500 characters / block) to cut the document, and retrieves based on keyword matching or TF-IDF algorithm. However, this method has obvious limitations. First, the fixed blocking strategy cannot balance the context integrity of the knowledge and the accuracy of the retrieval unit, which may lead to missing of key information or introduction of irrelevant content in the retrieval results. Secondly, simple keyword matching cannot effectively understand semantic association, for example, for "RRU offline" and "RRU disconnect", the retrieval system may not be able to effectively associate, resulting in a generally low recall rate (usually less than 70%) of the retrieval, which is difficult to meet the accuracy requirements of tower operation and maintenance. 2. General question answering technology based on large language model: In recent years, general large language models have been directly applied to operation and maintenance knowledge question answering. Although this scheme has certain natural language understanding ability, it has inherent defects in professional field application. The process of generating answers by the model is heavily dependent on its internal parameters, lacking an anchoring mechanism for external authoritative knowledge base, and is prone to produce "hallucinated content" that does not conform to the facts. False or incorrect fault handling schemes may lead to serious equipment operation errors, posing significant safety risks. In addition, this scheme cannot trace the knowledge source of the answer, and cannot accurately locate the knowledge in the sub-field. SUMMARY

[0003] To solve the above problems, the application provides a large model-based tower fault operation and maintenance knowledge retrieval method and system to solve the problems of simple block division, weak context association, difficult balance between retrieval accuracy and integrity, and easy generation of "illusion content" by using large models alone, which leads to operation errors in tower fault operation and maintenance scenarios and cannot meet multi-dimensional operation and maintenance requirements.

[0004] A large model-based tower fault operation and maintenance knowledge retrieval method, comprising: Classifying tower operation and maintenance knowledge according to business scenarios to obtain full-category knowledge text; Performing parent-child segmentation processing on the full-category knowledge text to generate a hierarchical structure knowledge base; Pretreating user question information based on a large language model and outputting a user core question; Performing mixed retrieval in the hierarchical structure knowledge base based on the user core question and outputting a first retrieval result; Verifying the first retrieval result according to a preset response rule and outputting a second retrieval result.

[0005] According to an embodiment of the application, the full-category knowledge text includes base station power support system basic knowledge, dynamic environment monitoring basic knowledge, dynamic environment monitoring device operation specification, and operation and maintenance system application knowledge.

[0006] According to an embodiment of the application, the parent-child segmentation processing of the full-category knowledge text to generate a hierarchical structure knowledge base further comprises: Cutting each category of knowledge text using a parent-child segmentation mode to obtain parent-child hierarchical structure text of the corresponding category, wherein the parent-child hierarchical structure text includes a parent block and a child block, and the parent block contains the context information of the child block; Summarizing each category of parent-child hierarchical structure text and generating a hierarchical structure knowledge base.

[0007] According to an embodiment of the application, the cutting of each category of knowledge text using a parent-child segmentation mode to obtain parent-child hierarchical structure text of the corresponding category further comprises: Divide each category of knowledge text into parent blocks using a paragraph delimiter as a segmentation identifier, and the maximum length of the parent block is less than a first preset threshold; Divide each category of knowledge text into child blocks using a line feed character as a segmentation identifier, and the maximum length of the child block is less than a second preset threshold.

[0008] According to an embodiment of the application, the cutting of each category of knowledge text using a parent-child segmentation mode to obtain parent-child hierarchical structure text of the corresponding category further comprises: When the full-category knowledge text contains a table, an HTML format code is used to retain the table structure.

[0009] According to an embodiment of the present application, the pre-processing of the user question information based on the large language model and outputting the user core question further comprises: Removing adverbs and modal particles in the user question information based on the large language model, and extracting three-element information of fault type, device model and operation demand; Generating the user core question based on the three-element information and outputting.

[0010] According to an embodiment of the present application, the mixed retrieval based on the user core question in the hierarchical structure knowledge base and outputting the first retrieval result further comprises: Calling the hierarchical structure knowledge base and searching a plurality of sub-blocks related to the user core question therefrom; Calculating the similarity of the user core question and the sub-blocks based on a semantic rearrangement model, and sorting the searched plurality of sub-blocks according to the similarity; Contextually associating the sorted plurality of sub-blocks with their parent blocks, and outputting the associated first retrieval result.

[0011] According to an embodiment of the present application, the verification of the first retrieval result according to the preset response rule and outputting the second retrieval result further comprises: Calculating the confidence degree according to the semantic matching degree of the first retrieval result and the user core question; Comparing the confidence degree with a preset threshold, if the confidence degree < 0.3, suspending the output, if the confidence degree ≥ 0.3, outputting the second retrieval result in descending order of the confidence degree.

[0012] According to an embodiment of the present application, the second retrieval result includes the original context information of the sub-blocks and their associated parent blocks, and the second retrieval result is labeled with its corresponding knowledge classification.

[0013] A large model-based tower fault operation and maintenance knowledge retrieval system, comprising: A text classification module for classifying tower operation and maintenance knowledge according to business scenarios to obtain full-category knowledge text; A knowledge base construction module for parent-child segmentation processing of the full-category knowledge text to generate a hierarchical structure knowledge base; A question processing module for pre-processing user question information based on a large language model and outputting a user core question; A mixed retrieval module for mixed retrieval based on the user core question in the hierarchical structure knowledge base and outputting a first retrieval result; A verification and response module for verifying the first retrieval result according to a preset response rule and outputting a second retrieval result.

[0014] According to an embodiment of the present application, the knowledge base construction module further comprises: a text cutting module, configured to cut the knowledge texts of each category by using a parent-child segmentation mode to obtain parent-child hierarchical structure texts corresponding to the categories, wherein the parent-child hierarchical structure texts comprise parent blocks and child blocks, and the parent blocks contain context information of the child blocks; a knowledge base generation module, configured to aggregate the parent-child hierarchical structure texts of each category and generate hierarchical structure knowledge bases.

[0015] According to an embodiment of the present application, the text cutting module further comprises: a parent block cutting module, configured to divide the knowledge texts of each category into parent blocks by using a paragraph delimiter as a segmentation identifier, and the maximum length of the parent blocks is less than a first preset threshold; a child block cutting module, configured to divide the knowledge texts of each category into child blocks by using a line break as a segmentation identifier, and the maximum length of the child blocks is less than a second preset threshold.

[0016] According to an embodiment of the present application, the text cutting module further comprises: a table processing module, configured to, when the knowledge texts of all categories contain tables, retain the table structure by using HTML format codes.

[0017] According to an embodiment of the present application, the question processing module further comprises: an information extraction module, configured to remove adverbs and modal particles in the user's question information based on a large language model, and extract three-element information of fault types, device models and operation requirements; a core question generation module, configured to generate a user core question based on the three-element information and output the user core question.

[0018] According to an embodiment of the present application, the mixed retrieval module further comprises: a child block matching module, configured to call the hierarchical structure knowledge bases and search a plurality of child blocks related to the user core question from the hierarchical structure knowledge bases; a reordering module, configured to calculate the similarity between the user core question and the child blocks based on a semantic rearrangement model, and reorder the plurality of searched child blocks according to the similarity; a context association module, configured to associate the reordered plurality of child blocks with their parent blocks in context, and output the first retrieval result after association.

[0019] According to an embodiment of the present application, the verification and response module further comprises: a confidence degree calculation module, configured to calculate a confidence degree according to the semantic matching degree between the first retrieval result and the user core question; A judgment module is configured to compare the confidence level with a preset threshold value, if the confidence level is less than 0.3, the output is suspended, if the confidence level is greater than or equal to 0.3, the second search result is output according to the order from high to low of the confidence level, the second search result includes the original context information of the sub-block and the associated parent block of the sub-block, and the second search result is marked with the corresponding knowledge classification.

[0020] An electronic device comprises a processor and a memory, the memory stores a computer program, the computer program is loaded and executed by the processor to realize the above-mentioned large model-based tower fault operation and maintenance knowledge retrieval method.

[0021] A computer readable storage medium stores a computer program, the computer program is loaded and executed by the processor to realize the above-mentioned large model-based tower fault operation and maintenance knowledge retrieval method.

[0022] Compared with the prior art, the present application has the following advantages: 1. The retrieval accuracy and recall rate are significantly improved The present application adopts a "parent-child segmentation" text processing mode, which effectively balances the accuracy and integrity of the retrieval. The sub-block as the smallest retrieval unit ensures accurate matching, and the parent block association mechanism guarantees the integrity of the context. Combined with the semantic rearrangement model, the preliminary results are optimized and sorted, so that the overall retrieval accuracy is greatly improved (can reach more than 90%), effectively solving the problem of low recall rate caused by weak semantic association in traditional keyword matching technology.

[0023] 2. The "illusion" content of the large model is fundamentally eliminated, and the safety is high The present application introduces a response rule checking mechanism based on confidence level calculation, strictly limits the generation ability of the large model within the range of knowledge base anchoring. When the confidence level is lower than the safety threshold, the system refuses to generate content; when the confidence level meets the standard, only the original information in the knowledge base is output. This design completely avoids the risk of false and incorrect information generated by general large models in professional fields from the process, provides reliable protection for high safety requirement scenarios such as tower operation and maintenance, and greatly reduces the possibility of operation errors caused by misleading information.

[0024] 3. The knowledge acquisition efficiency of operation and maintenance personnel is greatly improved, and the training period is shortened The present application greatly reduces the difficulty of knowledge positioning through customized knowledge classification and accurate retrieval result marking. New employees do not need to memorize a large number of document structures, but can quickly obtain accurate and structured knowledge through natural questioning. The training period of new employees is significantly shortened, the knowledge acquisition efficiency of new personnel is improved by 40%, the personnel training cost is effectively reduced, and the fault emergency response efficiency is improved.

[0025] 4. Strong knowledge reusability, high management efficiency The application is based on a hierarchically constructed knowledge base. Knowledge units (sub-blocks) can be flexibly and independently retrieved and called. The same knowledge unit can be efficiently reused in different but related user questioning scenarios, avoiding repeated entry and storage of knowledge, reducing the maintenance workload of the knowledge base, and improving the overall efficiency of knowledge management.

[0026] 5. Traceable results, high decision-making reliability The final output of the application clearly marks the knowledge classification and source to which the retrieval result belongs. Users can easily trace the original context of the answer, enhancing their trust in the content provided by the system, reducing the decision-making hesitation time caused by information uncertainty, and improving the decision-making efficiency and quality of operation and maintenance work. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present disclosure. Those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0028] Figure 1 is a large model-based tower fault operation and maintenance knowledge retrieval method flowchart provided according to an embodiment of the present application.

[0029] Figure 2 is a method flowchart for parent-child segmentation processing of full-category knowledge text and generating a hierarchical structure knowledge base according to an embodiment of the present application.

[0030] Figure 3 is a method flowchart for cutting various types of knowledge text using a parent-child segmentation mode according to an embodiment of the present application.

[0031] Figure 4 is a method flowchart for pre-processing user questioning information based on a large language model according to an embodiment of the present application.

[0032] Figure 5 is a method flowchart for mixed retrieval in a hierarchical structure knowledge base based on user core questions and outputting a first retrieval result according to an embodiment of the present application.

[0033] Figure 6 is a method flowchart for checking the first retrieval result according to a preset response rule and outputting a second retrieval result according to an embodiment of the present application.

[0034] Figure 7It is a large model-based tower fault operation and maintenance knowledge retrieval system structure diagram provided according to an embodiment of the application.

[0035] Figure 8 It is a knowledge base construction module structure diagram provided according to an embodiment of the application.

[0036] Figure 9 It is a structure diagram of a text cutting module provided according to an embodiment of the application.

[0037] Figure 10 It is a structure diagram of a question processing module provided according to an embodiment of the application.

[0038] Figure 11 It is a structure diagram of a mixed retrieval module provided according to an embodiment of the application.

[0039] Figure 12 It is a structure diagram of a verification and response module provided according to an embodiment of the application.

[0040] Figure 13 It is a computer device structure schematic diagram provided according to an embodiment of the application.

[0041] Reference signs: 01-text classification module; 02-knowledge base construction module; 03-question processing module; 04-mixed retrieval module; 05-verification and response module; 0211-parent block cutting module; 0212-sub-block cutting module; 0213-table processing module; 031-information extraction module; 032-core question generation module; 041-sub-block matching module; 042-reordering module; 043-context association module; 051-confidence calculation module; 052-judgment module. DETAILED DESCRIPTION

[0042] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0043] Embodiment 1 Additional aspects and advantages of the embodiments of the application will be in part apparent and in part pointed out hereinafter in the description of the embodiments of the application. The aspects and advantages of the embodiments of the application will be more clearly understood from the following description with reference to the drawings. Figures 1-6The embodiment of the application provides a tower fault operation and maintenance knowledge retrieval method based on a large model, comprising: S1: classifying tower operation and maintenance knowledge according to business scenarios to obtain full-category knowledge text.

[0044] S2: performing parent-child segmentation processing on the full-category knowledge text to generate a hierarchical structure knowledge base.

[0045] S3: preprocessing user question information based on a large language model and outputting a user core question.

[0046] S4: performing mixed retrieval in the hierarchical structure knowledge base based on the user core question and outputting a first retrieval result.

[0047] S5: verifying the first retrieval result according to a preset response rule and outputting a second retrieval result.

[0048] Specifically, in the step S1, the tower operation and maintenance knowledge is classified according to business scenarios to obtain full-category knowledge text, and the full-category knowledge text includes base station power supporting system basic knowledge, dynamic environment monitoring basic knowledge, dynamic environment monitoring device operation specification and operation and maintenance system application knowledge.

[0049] In a specific embodiment of the application, the tower operation and maintenance knowledge is divided into four categories according to business scenarios, ensuring that the classification covers the whole operation and maintenance process, and fundamentally solving the problems of tower operation and maintenance knowledge fragmentation and lack of systematic management by standardizing and classifying the knowledge according to scenarios, laying a foundation for subsequent hierarchical processing and accurate retrieval. The full-category knowledge text specifically includes the following four dimensions: 1) base station power supporting system basic knowledge, which relates to the core infrastructure for providing stable power support for communication equipment. Including but not limited to battery charging and discharging specification, generator start-stop process. 2) Dynamic environment monitoring basic knowledge, which focuses on the data system and alarm mechanism for real-time monitoring of dynamic environment. Including but not limited to: monitoring system parameter configuration, alarm signal interpretation. 3) Dynamic environment monitoring device operation specification, which directly guides the operation and maintenance personnel to intervene and handle faults on specific devices, has clear steps and operability, including but not limited to sensor calibration steps, monitoring terminal fault troubleshooting. 4) Operation and maintenance system application knowledge, which relates to the digital management, process collaboration and decision support of operation and maintenance work. Including but not limited to fault logbook entry, system report generation. The application classifies the tower operation and maintenance knowledge according to the above four dimensions, realizing full-scene and full-process knowledge coverage of the tower operation and maintenance “from basic theory to on-site operation, from data monitoring to system management”.

[0050] Specifically, the step S2 performs parent-child segmentation processing on the full-category knowledge text to generate a hierarchical structure knowledge base, which further comprises: S21: adopt the parent-child segmentation mode to cut each category of knowledge text, to obtain the parent-child hierarchical structure text of the corresponding category, wherein the parent-child hierarchical structure text includes a parent block and a child block, the parent block contains the context information of the child block, and further includes: S211: divide each category of knowledge text into a parent block with a paragraph delimiter as a segmentation identifier, and the maximum length of the parent block is less than a first preset threshold.

[0051] S212: divide each category of knowledge text into a child block with a line break as a segmentation identifier, and the maximum length of the child block is less than a second preset threshold.

[0052] S213: when the full category knowledge text contains a table, adopt an HTML format code to retain the table structure.

[0053] In an embodiment of the present application, the text is cut by introducing a two-level structure of “parent block-child block” to ensure the balance between context association and retrieval accuracy, so that when retrieving, both a small knowledge unit (child block) and the necessary background information (parent block) maintaining the semantic of the unit can be located, and the contradiction between retrieval accuracy and semantic integrity that cannot be considered by the traditional single block strategy is solved. Specifically, a paragraph delimiter (such as “\n\n”) is used as a segmentation identifier, which is used as the division boundary of the parent block to segment the text, forming a parent block with a maximum length of 1800 characters, which can carry the complete context of the knowledge unit, for example, a complete fault case containing fault phenomenon description, pre-checking steps and detailed processing flow is divided into the same parent block. A line break (such as “\n”) is used as a segmentation identifier, which is used as the division boundary of the child block to segment the text, forming a child block with a maximum length of 250 characters, which is used as the smallest retrieval unit to ensure retrieval accuracy. The line break is usually used to divide the points, steps or list items in the text, which is used as the division identifier of the child block, and can further refine the knowledge in the parent block into the smallest independent unit that can be retrieved. For example, the core processing steps 1-2 in the above fault case parent block are divided into independent child blocks.

[0054] In an embodiment of the present application, when the full category knowledge text contains a table (such as a device parameter table), an HTML format code is used to retain the table structure, avoiding format disorder caused by block division, and ensuring the integrity of the table after retrieval.

[0055] In an embodiment of the present application, a three-level title symbol “###” is used as the division boundary of the parent block, and a semicolon “;” is used as the division boundary of the child block to segment the document. This embodiment is applicable to structured documents divided by titles, i.e. the use scenario of the original document of the knowledge base mainly in the form of title levels (such as word outline format).

[0056] S22: aggregate the parent-child hierarchical structure texts of each category and generate a hierarchical structure knowledge base.

[0057] In a specific embodiment of the present application, the parent-child hierarchical structure texts formed by step S21 are aggregated to generate a hierarchical structure knowledge base. The hierarchical structure knowledge base is not simply a pile of texts, but a digital knowledge system embedded with rich metadata and multi-level association relationships. In the retrieval stage, high-precision matching can be achieved through sub-blocks. In generating the response, the context information in the parent block can be quickly associated and supplemented through the parent-child relationship index, thereby significantly improving the accuracy and completeness of the final output answer and effectively avoiding misunderstandings or incorrect answers due to fragmented information. By systematically organizing the dispersed knowledge fragments that have been classified and parent-child segmented into a unified, structured data set that can be efficiently retrieved and managed, an optimized data foundation is established for subsequent hybrid retrieval.

[0058] Specifically, step S3 further includes: S31: based on the large language model, remove adverbs and modal particles from the user's question information, and extract three element information of fault type, device model and operation demand.

[0059] S32: generate a user core question based on the three element information and output.

[0060] In a specific embodiment of the present application, the Qwen3-32B large model is called to extract the core information of the user's question through the standardized Prompt. First, the modifying words (such as "please ask" and "help check") and the mood words (such as "oh" and "huh") in the user's question information are removed, and the repetitive expressions are removed. Based on the purified text, the model extracts structured information according to the preset knowledge graph in the tower operation and maintenance field, extracts three key element information including fault type (such as RRU offline), device model (such as RRU5235E), and operation demand (such as troubleshooting steps), and reorganizes the three extracted element information according to a preset, simple, and unambiguous syntax structure to generate the final user core question. At the same time, the model is prohibited to generate any speculative answers or explanations, and only outputs the refined core question. This constraint mechanism fundamentally cuts off the possibility of "hallucination" content generated in the preprocessing stage, ensures that the input source of the search is pure and objective, and strictly limits the role of the large language model to "question interpreter" rather than "answer generator". For example, the user asks "please ask how to troubleshoot RRU5235E offline?", and the output of the refined result is "RRU5235E offline troubleshooting steps". The present application uses the natural language understanding ability of the large language model to convert the user's unstructured question that may contain redundant information into a standardized and high-precision query instruction suitable for knowledge base retrieval, effectively solving the problem of search failure caused by differences in question expression. By designing specific Prompt to constrain the model behavior, it is ensured that it strictly outputs according to this three-element framework, avoiding the generation of irrelevant or over-interpreted content, thereby improving the hit rate and accuracy of subsequent knowledge base retrieval.

[0061] Specifically, step S4 further comprises: S41: calling the hierarchical structure knowledge base and searching a plurality of sub-blocks related to the user core question from the hierarchical structure knowledge base.

[0062] S42: calculating the similarity between the user core question and the sub-blocks based on the semantic rearrangement model, and sorting the searched plurality of sub-blocks according to the similarity.

[0063] S43: associating the sorted plurality of sub-blocks with their parent blocks in context, and outputting the associated first search result.

[0064] In a specific embodiment of the present application, based on the user core problem (such as "RRU 5235E offline troubleshooting steps") refined in step S3, the retrieval interface of the hierarchical knowledge base is called, and the preliminary retrieval is performed by preferentially matching the sub-blocks in the hierarchical database, and the retrieval range is limited to the sub-block (i.e. the smallest retrieval unit), so that a batch of candidate sub-blocks with the highest similarity to the user core problem (for example, 50 or 100 sub-blocks with the highest similarity are returned) are quickly retrieved, ensuring that all possible relevant knowledge units are included in the candidate set and avoiding missed detection. Since the preliminary retrieval result may contain some noise that is semantically similar but actually irrelevant, for example, the sub-blocks of "RRU 5235E offline troubleshooting" and "RRU 5235E installation steps" are retrieved at the same time, in order to solve this problem, the present application introduces a Rerank semantic rearrangement model to perform fine sorting on the multiple sub-blocks retrieved in the preliminary retrieval. This model can more deeply understand the semantic correlation between the user core problem and each sub-block, and calculate the similarity between the user core problem and the sub-block, and then sort the multiple sub-blocks according to the similarity. Although the sorted sub-blocks are accurate, they may lack necessary background information, for example, a sub-block content is "check the power switch", without context, the user may not be clear whether it is for "offline" fault checking or for "new device" checking. Therefore, the present application uses the parent-child relationship index constructed by the knowledge base to automatically match the parent block to which the sorted sub-block belongs, to supplement the context information, and finally outputs a structured list (i.e. the first retrieval result). Each item in the list corresponds to a sub-block and the content of the associated parent block, for example, the sub-block is "RRU power-off check", and the associated parent block is "RRU offline pre-check process". Through a multi-stage hybrid retrieval strategy, the present application ensures that the final result not only accurately hits the user's intention, but also has complete context information, realizes the best combination of sub-blocks providing accurate answer points and parent blocks providing background support, effectively avoids misunderstanding caused by fragmented information, and finally outputs the first retrieval result which is significantly better than the traditional single retrieval method in accuracy, relevance and completeness, laying a solid data foundation for generating reliable answers.

[0065] In a specific embodiment of the present application, the rearrangement model can also select a hybrid algorithm of "BM25 algorithm + semantic similarity weighting" to reduce the dependence on large model computing power.

[0066] Specifically, step S5 further comprises: S51: Calculate the confidence degree according to the semantic matching degree between the first retrieval result and the user core problem.

[0067] S52: compare the confidence with a preset threshold, if the confidence < 0.3, suspend output, if the confidence ≥ 0.3, output the second search results in order from high to low according to the confidence, wherein the second search results include the original context information of the sub-block and its associated parent block, and the second search results are labeled with the corresponding knowledge classification to which they belong.

[0068] In a specific embodiment of the present application, a confidence-driven verification mechanism is introduced to quantitatively evaluate the reliability of the first search results, and strict output control is performed based on the evaluation results, thereby avoiding the risk of large models producing "illusion content", ensuring the rigor and reliability of the output. The risk of providing false or uncertain information in high safety requirement scenarios such as tower operation is fundamentally avoided. Specifically, first, the semantic matching degree between the first search results and the user's core problem is calculated, and this matching degree is quantified as a specific confidence score. The calculation process preferably uses the cosine similarity algorithm to objectively measure the semantic relevance between the user's core problem and the vector of the search result sub-block by comparing the cosine value of the angle between the vectors in high-dimensional space, the higher the score, the higher the matching degree of the search result with the user's problem, and the stronger the reliability. The present application sets a strictly verified confidence threshold (i.e. a preset threshold), preferably 0.3, which is used to distinguish between "effective knowledge" and "uncertain information". If the confidence < 0.3, it indicates that there is a lack of highly relevant knowledge in the existing knowledge base to the user's core problem, at which point the system will suspend output of any specific content or any speculative content derived from the knowledge base, and output a preset safety prompt message, for example: "Sorry, according to the existing knowledge base content, accurate answers cannot be provided at the moment." When the confidence ≥ 0.3, it indicates that the first search results have high reference value, at which point the system will strictly output these results in order from high to low according to the confidence, forming the second search results. The output text is limited to the original context of the sub-block and its associated parent block in the knowledge base, while retaining the format of the original knowledge, such as tables, step numbers, etc., to ensure accurate presentation of technical details. Any form of instruction fine-tuning or instruction optimization of the large model is prohibited, as well as any form of summarization, summary or creative rewriting of the original content. And each result is labeled with the corresponding knowledge classification (such as "base station power supporting system") to which it belongs, facilitating users to quickly screen the applicable scenarios of knowledge, and enhancing the traceability and credibility of the results.

[0069] In a specific embodiment of the present application, the question of "RRU5235E offline" is taken as an example for illustrative description: when the user inputs "I am a new operation and maintenance, and want to know how to troubleshoot RRU5235E offline, what checks do I need to do first?", the system first extracts the user's core question through the Qwen3-32B large model, outputs the user's core question of "RRU5235E offline troubleshooting pre-check steps", and based on the retrieval and rearrangement mechanism, the sub-block "RRU offline pre-check: 1. Check if the power switch is closed; 2. Check if the network cable connection is loose" is matched first, the confidence of the matched sub-block and the user's core question is 0.92, and the parent block information "RRU5235E series offline fault handling process: phenomenon description: device indicator light is off... Pre-check steps are used to exclude basic hardware problems..." is associated. Since the confidence is ≥ 0.3, the retrieval results are displayed in order from high to low according to the confidence, and the classification "dynamic environment monitoring equipment operation" is marked.

[0070] Embodiment 2 Based on the above method, the present application further provides a large model-based tower fault operation and maintenance knowledge retrieval system, as shown in Figures 7-12 , which comprises: A text classification module 01 for classifying tower operation and maintenance knowledge according to business scenarios to obtain full-category knowledge text, which includes base station power supporting system basic knowledge, dynamic environment monitoring basic knowledge, dynamic environment monitoring equipment operation specification and operation and maintenance system application knowledge.

[0071] A knowledge base construction module 02 for parent-child segmentation processing of full-category knowledge text to generate a hierarchical structure knowledge base.

[0072] A problem processing module 03 for pre-processing user question information based on a large language model and outputting a user core question.

[0073] A hybrid retrieval module 04 for hybrid retrieval in the hierarchical structure knowledge base based on the user's core question and outputting a first retrieval result.

[0074] A verification and response module 05 for verifying the first retrieval result according to a preset response rule and outputting a second retrieval result.

[0075] In a specific embodiment of the present application, the present application divides the tower operation and maintenance knowledge into four categories by the text classification module 01 according to the business scenarios, ensures that the classification covers the whole operation and maintenance process, classifies the knowledge by standardization and scene, fundamentally solves the problem of fragmentation and lack of systematic management of tower operation and maintenance knowledge, and lays a foundation for subsequent hierarchical processing and accurate retrieval. The full-category knowledge text specifically includes the following four dimensions: 1) base station power supporting system basic knowledge, this kind of knowledge involves the core infrastructure that provides stable power support for communication equipment. Including but not limited to battery charging and discharging specifications, generator start-stop process. 2) Dynamic environment monitoring basic knowledge, this kind of knowledge focuses on the data system and alarm mechanism for real-time monitoring of dynamic environment. Including but not limited to: monitoring system parameter configuration, alarm signal interpretation. 3) Dynamic environment monitoring device operation specification, this kind of knowledge directly guides the on-site intervention and fault handling of the specific device by the operation and maintenance personnel, has clear steps and operability, including but not limited to sensor calibration steps, monitoring terminal troubleshooting. 4) Operation and maintenance system application knowledge, this kind of knowledge is related to the digital management, process collaboration and decision support of operation and maintenance work. Including but not limited to fault log entry, system report generation. The present application classifies the tower operation and maintenance knowledge according to the above four dimensions, realizes the full-scene and full-process knowledge coverage of the tower operation and maintenance "from basic theory to on-site operation, from data monitoring to system management".

[0076] Specifically, the knowledge base construction module 02 further includes: The text cutting module 021 is configured to cut the knowledge texts of each category by adopting a parent-child segmentation mode to obtain parent-child hierarchical structure texts of the corresponding category, wherein the parent-child hierarchical structure texts include parent blocks and child blocks, and the parent blocks contain context information of the child blocks.

[0077] Further, the text cutting module 021 further includes: The parent block cutting module 0211 is configured to divide the knowledge texts of each category into parent blocks by taking a paragraph delimiter as a segmentation identifier, and the maximum length of the parent blocks is less than a first preset threshold.

[0078] The child block cutting module 0212 is configured to divide the knowledge texts of each category into child blocks by taking a line feed character as a segmentation identifier, and the maximum length of the child blocks is less than a second preset threshold.

[0079] The table processing module 0213 is configured to, when the full-category knowledge texts contain tables, retain the table structure by adopting HTML format codes.

[0080] In an embodiment of the present application, the text is cut by the text cutting module 021 to ensure the balance between context association and retrieval accuracy, so that when retrieving, both a small knowledge unit (sub-block) and the necessary background information (parent block) to maintain the semantics of the unit can be located, solving the contradiction between retrieval accuracy and semantic integrity that cannot be considered by traditional single blocking strategy. Specifically: the text is segmented by the parent block cutting module 0211 to form parent blocks with a maximum length of 1800 characters, which can carry the complete context of the knowledge unit, for example: a complete fault case containing fault phenomenon description, pre-checking steps and detailed processing flow is divided into the same parent block. The text is segmented by the sub-block cutting module 0212 to form sub-blocks with a maximum length of 250 characters, which are used as the smallest retrieval unit to ensure retrieval accuracy.

[0081] In an embodiment of the present application, when the full-category knowledge text contains a table (such as a device parameter table), the table processing module 0213 is used to retain the table structure to avoid format disorder caused by blocking and ensure the integrity of the table after retrieval.

[0082] The knowledge base generation module 022 is used to collect and generate a hierarchical structure knowledge base from the parent-child hierarchical structure text of each category.

[0083] In an embodiment of the present application, the parent-child hierarchical structure text is collected by the knowledge base generation module 022 to generate a hierarchical structure knowledge base. The hierarchical structure knowledge base is not simply a pile of text, but a digital knowledge system embedded with rich metadata and multi-level association relationships. In the retrieval stage, high-precision matching can be achieved through sub-blocks, and in the generation of answers, the context information in the parent block can be quickly associated and supplemented through the parent-child relationship index, thereby significantly improving the accuracy and integrity of the final output answer and effectively avoiding misunderstandings or mistakes caused by fragmented information. By systematically organizing the dispersed knowledge fragments that have been classified and parent-child segmented into a unified, structured data set that can be efficiently retrieved and managed, an optimized data foundation is established for subsequent hybrid retrieval.

[0084] Specifically, the question processing module 03 further includes: The information extraction module 031 is used to remove adverbs and adverbs in user question information based on a large language model, and extract three-element information of fault type, device model and operation demand.

[0085] The core question generation module 032 is used to generate a user core question based on the three-element information and output.

[0086] In a specific embodiment of the present application, the Qwen3-32B large model is called by the question processing module 03, the core information of the user's question is extracted through the standardized Prompt, the information extraction module 031 is first used to remove the adjectives (such as "please" and "please help check") and the adverbs (such as "oh" and "huh") in the user's question information, and the repeated expressions, based on the purified text, the model extracts structured information according to the preset knowledge graph of the tower operation field, extracts three key element information including fault type (such as RRU offline), device model (such as RRU5235E), and operation demand (such as troubleshooting steps), and then uses the core question generation module 032 to reorganize the extracted three element information according to a preset, simple and unambiguous syntax structure to generate the final user core question, while prohibiting the model from generating any speculative answers or explanations, and only outputting the refined core question. This constraint mechanism fundamentally cuts off the possibility of generating "illusion" content in the preprocessing stage, ensuring that the input source of the search is pure and objective, and strictly limiting the role of the large language model to "question interpreter" rather than "answer generator". For example, the user's question "please ask how to troubleshoot RRU5235E offline?", the output of the refined result is "RRU5235E offline troubleshooting steps". The present application uses the natural language understanding ability of the large language model to convert the user's unstructured question that may contain redundant information into a standardized and high-precision query instruction suitable for knowledge base retrieval, effectively solving the problem of search failure caused by differences in question expression, and by designing specific prompt words (Prompt) to constrain the model behavior, ensuring that it strictly outputs according to this three-element framework, avoiding the generation of irrelevant or over-interpreted content, thereby improving the hit rate and accuracy of subsequent knowledge base retrieval.

[0087] Specifically, the mixed search module 04 further includes: The sub-block matching module 041 is configured to call the hierarchical structure knowledge base and search a plurality of sub-blocks related to the user core question from the hierarchical structure knowledge base.

[0088] The reordering module 042 is configured to calculate the similarity between the user core question and the sub-blocks based on a semantic rearrangement model, and sort the searched plurality of sub-blocks according to the similarity.

[0089] The context association module 043 is configured to associate the sorted plurality of sub-blocks with their parent blocks in context, and output the associated first search result.

[0090] In a specific embodiment of the present application, based on the refined user core question (such as "RRU 5235E offline troubleshooting steps"), the sub-block matching module 041 is called to perform a preliminary search by preferentially matching sub-blocks in the hierarchical structure database, limiting the search range to sub-blocks (i.e., the smallest search unit), quickly retrieving a batch of candidate sub-blocks with the highest similarity to the user core question (for example, returning the top 50 or 100 sub-blocks with the highest similarity), ensuring that all possible relevant knowledge units are included in the candidate set and avoiding missed searches. Since the preliminary search results may contain some noise that is semantically similar but actually irrelevant, for example, sub-blocks of "RRU 5235E offline troubleshooting" and "RRU 5235E installation steps" are retrieved at the same time, to solve this problem, the reordering module 042 is used to perform fine sorting on the multiple sub-blocks retrieved in the preliminary search. This module can more deeply understand the semantic association between the user core question and each sub-block, and calculate the similarity between the user core question and the sub-block, and then sort the multiple sub-blocks according to the similarity. Although the sorted sub-blocks are accurate, they may lack necessary background information, for example, a sub-block content is "check the power switch", without context, the user may not be clear whether this is a check for "offline" failure or a check for "new equipment", therefore, the present application uses the context association module 043 to automatically match the sorted sub-blocks with their parent blocks to supplement the context information, and finally outputs a structured list (i.e., the first search result). Each item in the list corresponds to a sub-block and the content of the parent block associated with it, for example, the sub-block is "RRU power failure check", and the associated parent block is "RRU offline pre-check process". Through a multi-stage hybrid search strategy, the present application ensures that the final result not only accurately hits the user's intention, but also has complete context information, achieving the best combination of sub-blocks providing accurate answer points and parent blocks providing background support, effectively avoiding misunderstandings caused by fragmented information, and finally outputting the first search result that significantly outperforms traditional single search methods in accuracy, relevance and completeness, laying a solid data foundation for generating reliable responses.

[0091] Specifically, the verification and response module 05 further comprises: The confidence calculation module 051 is configured to calculate the confidence degree according to the semantic matching degree between the first search result and the user core question.

[0092] The judgment module 052 is configured to compare the confidence degree with a preset threshold value, if the confidence degree < 0.3, the output is suspended, if the confidence degree ≥ 0.3, the second search result is output in descending order of confidence degree, the second search result includes the original context information of the sub-block and the associated parent block, and the second search result is marked with the corresponding knowledge classification.

[0093] In a specific embodiment of the present application, the reliability of the first search result is quantitatively evaluated by the verification and response module 05, and strict output control is performed based on the evaluation result, thereby avoiding the risk of the large model producing "illusion content", ensuring the rigor and reliability of the output. The risk of providing false or uncertain information in high safety requirement scenarios such as tower operation is fundamentally avoided. Specifically, first, the semantic matching degree between the first search result and the user's core problem is calculated by using the confidence calculation module 051, and the matching degree is quantified as a specific confidence score. The calculation process preferably uses the cosine similarity algorithm to objectively measure the semantic relevance between the two by comparing the cosine value of the angle between the vector of the user's core problem and the vector of the search result sub-block in high-dimensional space. The higher the score, the higher the matching degree of the search result with the user's problem, and the stronger the reliability. Then the calculated confidence is compared with the preset threshold by using the judgment module 052. The present application sets a strictly verified confidence threshold (i.e. the preset threshold), which is preferably 0.3. The preset threshold is used to distinguish between "effective knowledge" and "uncertain information". If the confidence is <0.3, it indicates that there is a lack of highly relevant knowledge to the user's core problem in the existing knowledge base. At this time, the system will suspend output of any specific content or any speculative content derived from the knowledge base, and output a preset safety prompt message, for example: "Sorry, according to the existing knowledge base content, accurate answers cannot be provided at the moment." When the confidence is ≥0.3, it indicates that the first search result has high reference value. At this time, the system will strictly output these results in order of confidence from high to low, forming the second search result. The output text is limited to the original context of the sub-block and its associated parent block in the knowledge base, while retaining the format of the original knowledge, such as tables, step numbers, etc., to ensure accurate presentation of technical details. Any form of instruction fine-tuning or instruction optimization of the large model is prohibited, as well as any form of summarization, summary or creative rewriting of the original content. And each result is clearly labeled with its corresponding knowledge classification (such as "base station power supporting system"), which facilitates users to quickly filter the applicable scenarios of knowledge and enhances the traceability and credibility of the results.

[0094] Embodiment 3 As Figure 13 shown, the embodiment of the present application also provides an electronic device, which includes a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the above-mentioned large model-based tower fault operation knowledge retrieval method. The device in the present application can be a server, PC, PAD, mobile phone, etc.

[0095] Further, the embodiment of the present application also provides a computer readable storage medium, and the computer readable storage medium stores a computer program, the computer program is loaded and executed by a processor to realize the above-mentioned large model-based tower fault operation and maintenance knowledge retrieval method.

[0096] To sum up, the large model-based tower fault operation and maintenance knowledge retrieval method and system provided by the present application has the following advantages: 1. The retrieval precision and recall rate are significantly improved The present application adopts a "parent-child segmentation" text processing mode, effectively balancing the precision and integrity of retrieval. The sub-blocks ensure accurate matching as the smallest retrieval unit, and the parent block association mechanism guarantees the integrity of the context. Combined with the semantic rearrangement model, the preliminary results are optimized and sorted, so that the overall retrieval accuracy is greatly improved (can reach more than 90%), effectively solving the problem of low recall rate caused by weak semantic association in traditional keyword matching technology.

[0097] 2. The "illusion" content of the large model is fundamentally eliminated, and the safety is high The present application introduces a response rule verification mechanism based on confidence calculation, strictly limits the generation ability of the large model within the range of knowledge base anchoring. When the confidence is lower than the safety threshold, the system refuses to generate content; when the confidence meets the standard, only the original information in the knowledge base is output. This design completely avoids the risk of generating fictitious and false information in professional fields by general large models from the process, providing reliable protection for high-safety-required scenarios such as tower operation and maintenance, and greatly reducing the possibility of operation errors caused by misleading information.

[0098] 3. The knowledge acquisition efficiency of operation and maintenance personnel is greatly improved, and the training period is shortened The present application greatly reduces the difficulty of knowledge positioning through customized knowledge classification and accurate retrieval result labeling. New employees do not need to memorize a large number of document structures, but can quickly obtain accurate and structured knowledge through natural questioning. The training period of new employees is significantly shortened, the knowledge acquisition efficiency of new personnel is improved by 40%, the personnel training cost is effectively reduced, and the fault emergency response efficiency is improved.

[0099] 4. Strong knowledge reusability and high management efficiency Based on the hierarchical knowledge base, the knowledge unit (sub-block) can be flexibly and independently retrieved and called, and the same knowledge unit can be efficiently reused in different but related user questioning scenarios, avoiding repeated entry and storage of knowledge, reducing the maintenance workload of the knowledge base, and improving the overall efficiency of knowledge management.

[0100] 5. The results are traceable, and the decision-making credibility is high The final output search result of the application is clearly marked with the knowledge classification and source thereof, the user can conveniently trace the original context of the answer, the trust of the user to the content provided by the system is enhanced, the decision hesitation time caused by information uncertainty is reduced, and the decision efficiency and quality of operation and maintenance work are improved.

[0101] Although the application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for knowledge retrieval of tower fault operation and maintenance based on a large model, characterized in that, include: The tower operation and maintenance knowledge is categorized according to business scenarios to obtain a complete set of knowledge texts. The knowledge texts of all categories are segmented into parent and child segments to generate a hierarchical knowledge base; The system preprocesses user query information based on a large language model and outputs the user's core question. Based on the user's core question, a mixed search is performed in the hierarchical knowledge base, and the first search result is output. The first search result is verified according to the preset response rules, and the second search result is output.

2. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 1, characterized in that, The comprehensive knowledge texts include basic knowledge of base station power support systems, basic knowledge of environmental monitoring, operating procedures for environmental monitoring equipment, and application knowledge of operation and maintenance systems.

3. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 1, characterized in that, The step of performing parent-child segmentation on the full-category knowledge text to generate a hierarchical knowledge base further includes: The parent-child segmentation pattern is used to cut the knowledge text of each category to obtain the corresponding parent-child hierarchical structure text, wherein the parent-child hierarchical structure text includes a parent block and a child block, and the parent block contains the context information of the child block; The hierarchical parent-child structure texts of various categories are summarized and a hierarchical knowledge base is generated.

4. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 3, characterized in that, The step of segmenting knowledge text into categories using a parent-child segmentation model to obtain corresponding parent-child hierarchical text structures further includes: Each category of knowledge text is divided into parent blocks using paragraph separators as segmentation identifiers, and the maximum length of the parent block is less than a first preset threshold. The knowledge text of each category is divided into sub-blocks using line breaks as segment identifiers, and the maximum length of each sub-block is less than a second preset threshold.

5. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 3, characterized in that, The method of segmenting knowledge text into categories using a parent-child segmentation pattern to obtain corresponding parent-child hierarchical text structures also includes: When the knowledge text of the entire category contains tables, the table structure is preserved using HTML format code.

6. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 1, characterized in that, The preprocessing of user query information based on a large language model and the output of the user's core question further include: Based on the large language model, modifiers and modifiers in user query information are removed, and the three key elements of fault type, equipment model and operation requirements are extracted. Based on the three key information elements, the user's core questions are generated and output.

7. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 1, characterized in that, The step of performing a mixed search in the hierarchical knowledge base based on the user's core question and outputting the first search result further includes: Invoke the hierarchical knowledge base and search it for multiple sub-blocks related to the user's core question; The similarity between the user's core question and the sub-blocks is calculated based on the semantic reordering model, and the searched sub-blocks are sorted according to the similarity. The sorted sub-blocks are context-associated with their parent blocks, and the first search result after association is output.

8. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 1, characterized in that, The step of verifying the first search result according to the preset response rules and outputting the second search result further includes: Confidence is calculated based on the semantic matching degree between the first search result and the user's core question; The confidence level is compared with a preset threshold. If the confidence level is <0.3, the output is paused. If the confidence level is ≥0.3, the second search results are output in descending order of confidence level.

9. The method for knowledge retrieval of tower fault operation and maintenance based on a large model according to claim 8, characterized in that, The second search result includes the original context information of the sub-block and its associated parent block, and the second search result is labeled with its corresponding knowledge category.

10. A knowledge retrieval system for tower fault operation and maintenance based on a large model, characterized in that, include: The text classification module is used to classify tower operation and maintenance knowledge according to business scenarios to obtain all categories of knowledge text; The knowledge base construction module is used to perform parent-child segmentation on the full-category knowledge text to generate a hierarchical knowledge base. The question processing module is used to preprocess user question information based on a large language model and output the user's core question. The hybrid retrieval module is used to perform hybrid retrieval in the hierarchical knowledge base based on the user's core question and output the first retrieval result. The verification and response module is used to verify the first search result according to the preset response rules and output the second search result.

11. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 10, characterized in that, The knowledge base construction module also includes: The text segmentation module is used to segment knowledge text of various categories using a parent-child segmentation mode to obtain parent-child hierarchical structure text of the corresponding category. The parent-child hierarchical structure text includes a parent block and a child block, and the parent block contains the context information of the child block. The knowledge base generation module is used to summarize parent-child hierarchical texts of various categories and generate a hierarchical knowledge base.

12. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 11, characterized in that, The text segmentation module also includes: The parent block segmentation module is used to divide knowledge text of each category into parent blocks using paragraph separators as segmentation identifiers, and the maximum length of the parent block is less than a first preset threshold. The sub-block segmentation module is used to divide knowledge text of each category into sub-blocks using newline characters as segmentation identifiers, and the maximum length of the sub-block is less than a second preset threshold.

13. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 12, characterized in that, The text segmentation module also includes: The table processing module is used to preserve the table structure using HTML format code when the full-category knowledge text contains tables.

14. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 10, characterized in that, The problem handling module also includes: The information extraction module is used to remove modifiers and interjections from user query information based on a large language model, and extract the three key elements of fault type, equipment model and operation requirements. The core question generation module is used to generate and output the user's core question based on the three key information elements.

15. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 10, characterized in that, The hybrid retrieval module also includes: The sub-block matching module is used to call the hierarchical knowledge base and search for multiple sub-blocks related to the user's core question. The reordering module is used to calculate the similarity between the user's core question and the sub-blocks based on the semantic reordering model, and to sort the searched sub-blocks according to the similarity. The context association module is used to associate multiple sorted sub-blocks with their parent blocks and output the first search result after association.

16. The knowledge retrieval system for tower fault operation and maintenance based on a large model according to claim 10, characterized in that, The verification and response module also includes: The confidence calculation module is used to calculate the confidence score based on the semantic matching degree between the first search result and the user's core question. The judgment module is used to compare the confidence level with a preset threshold. If the confidence level is <0.3, the output is paused. If the confidence level is ≥0.3, the second search results are output in descending order of confidence level. The second search results include the original context information of the sub-block and its associated parent block, and the second search results are labeled with their corresponding knowledge category.

17. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the knowledge retrieval method for tower fault operation and maintenance based on a large model as described in any one of claims 1 to 9.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the knowledge retrieval method for tower fault operation and maintenance based on a large model as described in any one of claims 1 to 9.

Images