Troubleshooting guidance method, control device, and failure diagnosis apparatus

Abstract

The application provides a fault maintenance guidance method, a control device and a fault diagnosis device, relates to the technical field of fault diagnosis, and the fault maintenance guidance method comprises the following steps: obtaining a fault maintenance consultation request, determining target equipment model information and fault description information; performing matching search in a local professional knowledge base based on the target equipment model information, obtaining target maintenance data associated with the equipment model, and the target maintenance data comprising a maintenance article fragment and a maintenance reference picture; constructing a round of consultation prompt words based on the target maintenance data, inputting the round of consultation prompt words into a large language model, and obtaining an initial maintenance guidance result output by the large language model; and performing content optimization on the initial maintenance guidance result to generate corresponding maintenance guidance information. The application realizes a full-automatic generation link from an equipment model to executable maintenance guidance, and greatly reduces the manual analysis cost of maintenance personnel.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a fault repair guidance method, control device and fault diagnosis equipment. Background Technology

[0002] As the core equipment of digital currencies, blockchain servers, including their computing boards, power supplies, and fans, operate under high load and high temperature environments for extended periods, leading to frequent failures. Traditional blockchain server repair heavily relies on specialized fixtures; different brands and models of servers often require matching official fixtures for testing and repair. For example, Company A's X-type blockchain server requires a dedicated test board, while its B-type blockchain server requires a separate fixture. This "single-brand, single-fixture" model forces repair companies to equip themselves with a large number of expensive and incompatible specialized devices, tying up huge sums of money. Furthermore, when new models emerge, older fixtures often become incompatible, severely limiting repair efficiency and coverage. To address this issue, universal fault diagnosis equipment has emerged in this field. Through adjustable interfaces, standardized testing procedures, and adaptive power modules, it is compatible with various brands and models of blockchain servers and their accessories, enabling one-stop testing and fault location, thereby reducing the repair threshold and equipment costs.

[0003] However, even with general-purpose diagnostic equipment, maintenance personnel still face another major challenge in actual maintenance: obtaining fault analysis and maintenance guidance. Currently, maintenance personnel typically rely on their own experience or consult scattered materials such as paper maintenance manuals, online forum posts, and technical bulletins to diagnose problems. This approach has significant shortcomings: First, the efficiency of retrieving maintenance information is extremely low. Faced with a massive and unstructured maintenance knowledge base, maintenance personnel struggle to quickly locate solutions that accurately match the current equipment model and fault symptoms. Second, maintenance information from different sources may be contradictory or outdated, leading to misjudgments. Third, complex faults often require integrating multiple maintenance articles, reference images, and even videos to form a complete understanding, and manual integration is time-consuming, labor-intensive, and prone to overlooking crucial details. Therefore, how to efficiently and automatically construct accurate and easy-to-understand maintenance guidance information from large amounts of maintenance data has become a key bottleneck in improving the efficiency and quality of blockchain server maintenance. While existing diagnostic equipment can detect abnormal hardware parameters, it cannot automatically generate targeted maintenance steps and illustrated guidance content, forcing maintenance personnel to manually analyze data and failing to fully utilize equipment capabilities. Summary of the Invention

[0004] The main objective of this invention is to provide a fault repair guidance method that aims to significantly reduce the cost of manual analysis and improve the accuracy and efficiency of repair.

[0005] To achieve the above objectives, the present invention provides a fault repair guidance method, applied to fault diagnosis equipment, the fault repair guidance method comprising: Obtain fault repair consultation requests and determine the target equipment model information and fault description information; Based on the target device model information, a matching search is performed in the local professional knowledge base to obtain target repair data associated with the device model. The target repair data includes repair article excerpts and repair reference pictures. Based on the target maintenance data, a round of consultation prompts is constructed, and the round of consultation prompts is input into the large language model to obtain the initial maintenance guidance results output by the large language model. The initial maintenance guidance results are optimized to generate corresponding maintenance guidance information.

[0006] Optionally, constructing a round of consultation prompts based on the target maintenance data specifically involves: Determine the preset paragraph titles contained in the repair article fragment; Based on the type and arrangement order of the preset paragraph titles, determine the corresponding multiple analysis instruction text segments, and concatenate the multiple analysis instruction text segments in the recognition order to generate a maintenance article analysis and processing instruction segment; Obtain preset constraint requirements information, and determine the word limit, maintenance reference image processing requirements, content exclusion rules, and data source limitation clauses contained in the constraint requirements information; The first round of consultation prompts are generated by following the preset constraint order of the character limit, the requirements for processing the repair reference images, the content exclusion rules, and the data source limitation clauses.

[0007] Optionally, optimizing the initial repair guidance results to generate corresponding repair guidance information includes: The initial maintenance guidance results are semantically segmented and identified, and the title area, fault phenomenon description area, fault analysis area and solution area are identified based on a preset paragraph semantic tag set. Extract the title text from the identified title region to generate the corresponding title field; Extract the fault phenomenon description text from the fault phenomenon description area to generate the corresponding fault phenomenon description segment; Extract the fault analysis text from the fault analysis area to generate the corresponding fault analysis segment; Extract the solution text from the solution region to generate the corresponding solution segment; The title field, the fault phenomenon description section, the fault analysis section, and the solution section are reorganized according to a preset order to generate corresponding maintenance guidance information.

[0008] Optionally, optimizing the initial repair guidance results to generate corresponding repair guidance information includes: Extract the title field, fault phenomenon description section, fault analysis section, and solution section from the initial maintenance guidance results; Determine the original text location information of each repair reference image in the target repair data within the repair article fragment; The repair reference image is inserted into the corresponding text content position in the initial repair guidance result according to the original text position information to generate the repair guidance information.

[0009] Optionally, the step of performing a matching search in a local professional knowledge base based on the target equipment model information to obtain target maintenance data associated with the equipment model includes: The target device model information is broken down into three levels of keyword units: brand terms, series terms, and model serial number; Using the brand name as the first matching level, retrieve all device models belonging to that brand from the local professional knowledge base; Using the series terms as the second matching level, a first subset of candidate models belonging to the corresponding series is selected from the set of all device models; Using the model number as the third matching level, a unique target model is matched in the first candidate model subset; Extract all repair article fragment indexes and all repair reference image indexes corresponding to the unique target model in the local professional knowledge base, concatenate the original text content corresponding to all the repair article fragment indexes, and associate the image files corresponding to all the repair reference image indexes to generate the target repair data.

[0010] Optionally, the fault repair guidance method further includes: Obtain a secondary fault repair consultation request, and determine the first equipment model information and secondary consultation information in the secondary fault repair consultation request; Based on the first device model information, a corresponding consultation database is determined, which includes a local professional knowledge base and an external knowledge base; Retrieve maintenance knowledge entries from the consultation database that match the first equipment model information and the secondary consultation information; Construct a second-round consultation prompt phrase corresponding to the maintenance knowledge item. The second-round consultation prompt phrase includes a system role setting section and a secondary output format constraint section.

[0011] Optionally, determining the corresponding consultation database based on the first device model information includes: Determine whether at least one maintenance knowledge entry exists in the local professional knowledge base that matches the model information of the first device; If at least one maintenance knowledge item exists, the consultation database is determined to be a local professional knowledge base; If no maintenance knowledge entry exists, the consultation database is determined to be an external knowledge base.

[0012] Optionally, determining the corresponding consultation database based on the first device model information includes: Determine the consistency between the first device model information and the target model information; If the consistency is consistent, the consultation database is determined to be a local professional knowledge base; If the consistency is inconsistent, the consultation database is determined to be an external knowledge base.

[0013] In addition, to achieve the above objectives, the present invention also provides a control device, the control device comprising: a memory, a processor, and a fault repair guidance program stored in the memory and executable on the processor, the fault repair guidance program being configured to implement the fault repair guidance method as described above.

[0014] In addition, to achieve the above objectives, the present invention also provides a fault diagnosis device, including the control device described above.

[0015] This invention provides a fault repair guidance method for fault diagnosis equipment. The method obtains a fault repair consultation request, determines the target equipment model information and fault description information, then performs a matching search in a local professional knowledge base based on the target equipment model information to obtain target repair data associated with the equipment model. Next, it constructs a round of consultation prompts based on the target repair data, inputs these prompts into a large language model, obtains the initial repair guidance results output by the large language model, and finally optimizes the content of the initial repair guidance results to generate corresponding repair guidance information. This achieves a fully automated generation chain from equipment model to executable repair guidance, significantly reducing the manual analysis costs for repair personnel. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] Figure 1 This is a schematic diagram of a fault repair guidance method according to an embodiment of the present invention; Figure 2 for Figure 1 A schematic diagram of the troubleshooting guidance method for step S300; Figure 3 for Figure 1 A schematic diagram of the troubleshooting guidance method for step S400; Figure 4 This is a schematic diagram of a fault repair guidance method according to another embodiment of the present invention; Figure 5 for Figure 1 A schematic diagram of the troubleshooting guidance method for step S200; Figure 6 This is a schematic diagram of a fault repair guidance method according to another embodiment of the present invention; Figure 7 for Figure 6 A schematic diagram of the troubleshooting guidance method for step S600; Figure 8 This is a schematic diagram of a fault repair guidance method according to another embodiment of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Well-known modules, units, and their connections, links, communications, or operations are not shown or described in detail. Furthermore, the described features, architectures, or functions can be combined in any way in one or more embodiments. Those skilled in the art should understand that the various embodiments described below are only for illustrative purposes and not for limiting the scope of protection of the present invention. It is also readily understood that the modules, units, or processing methods in the various embodiments described herein and shown in the accompanying drawings can be combined and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In blockchain server maintenance scenarios, even with the use of general-purpose fault diagnosis equipment, maintenance personnel still need to rely on their own experience to manually search, compare, and integrate maintenance information from a vast amount of unstructured maintenance manuals, forum posts, and technical announcements. When faced with specific fault phenomena such as decreased computing power of the computing board, abnormal output voltage of the power module, or fan speed deviating from the rated value, maintenance personnel need to read maintenance articles from different sources one by one to piece together the cause of the fault, maintenance steps, and precautions. This process is not only extremely inefficient and prone to missing key details, but also the information from different sources may have version contradictions or be outdated, leading to incorrect maintenance judgments.

[0021] The main solution of this application embodiment is as follows: based on the equipment model information, perform hierarchical matching retrieval in the local professional knowledge base to obtain maintenance article fragments and maintenance reference images associated with the equipment model, construct a first round of consultation prompt words containing maintenance article analysis and processing instruction segments and multi-dimensional output format constraint segments, input the large language model to obtain the initial maintenance guidance results, and then optimize the content through semantic segmentation recognition and image-text position correspondence insertion to generate structured maintenance guidance information.

[0022] In this embodiment, for ease of description, the following description will focus on the control device as the executing entity.

[0023] This application provides a solution that organically combines three stages: equipment model-driven knowledge retrieval, multi-dimensional constraint-driven prompt word construction, and post-processing optimization of large language model output. This achieves a fully automated processing chain from equipment model input to structured maintenance guidance output, solving the technical problems of low efficiency in obtaining maintenance information, uncontrollable output format, and inaccurate text-image association in existing technologies. It enables maintenance personnel to directly obtain text-image corresponding maintenance guidance information containing fault phenomena, cause analysis, maintenance steps, and precautions, without having to manually browse and integrate scattered maintenance materials.

[0024] Therefore, the present invention proposes a fault repair guidance method; it is understood that the fault diagnosis equipment is equipped with a control device for storing and executing the following method. The control device can be implemented by a main controller, such as MCU (Micro Controller Unit), DSP (Digital Signal Processor), FPGA (Field Programmable Gate Array), SOC (System On Chip), etc.

[0025] Because existing general-purpose fault diagnosis equipment can only provide detection results of abnormal hardware parameters (such as voltage deviation values ​​read through the IPMI protocol, temperature over-limit alarms collected through the I2C bus, and hash rate decline data of the computing board obtained through the SPI interface), it cannot automatically generate maintenance guidance information containing fault analysis logic, maintenance operation steps, and reference pictures based on these abnormal data. There is still an information gap between "discovering abnormal parameters" and "performing specific maintenance operations" that maintenance personnel need to manually overcome by relying on human experience and scattered information. For example, when the diagnostic equipment detects that the power supply voltage of a certain S19 model computing board of a certain brand A blockchain server is 15% lower than the rated value, the maintenance personnel need to determine whether the abnormality is caused by one of the many possible causes such as DC-DC converter failure, MOSFET breakdown, or PCB circuit corrosion. They also need to search and piece together the corresponding detection steps and replacement solutions from maintenance information scattered from different sources. The whole process is highly dependent on personal experience accumulation, and novice maintenance personnel often cannot form an effective maintenance approach when faced with the same abnormal parameters.

[0026] Therefore, referring to Figure 1 In one embodiment of the present invention, the fault repair guidance method is applied to a fault diagnosis device, and the fault repair guidance method includes steps S100-S400, wherein: S100: Obtain a fault repair consultation request and determine the target equipment model information and fault description information; S200. Based on the target equipment model information, perform a matching search in the local professional knowledge base to obtain target maintenance data associated with the equipment model. The target maintenance data includes maintenance article excerpts and maintenance reference pictures. S300. Construct a round of consultation prompts based on the target maintenance data, input the round of consultation prompts into the large language model, and obtain the initial maintenance guidance results output by the large language model; S400: Optimize the content of the initial maintenance guidance results to generate corresponding maintenance guidance information.

[0027] The first-round consultation prompt refers to a structured prompt text constructed based on the target maintenance data and aimed at driving the large language model to generate formatted output that meets the requirements of the maintenance guidance scenario. The first-round consultation prompt includes: a maintenance article analysis and processing instruction segment, used to instruct the large language model to analyze and process the target maintenance data; and an output format constraint segment, used to constrain the output format of the large language model. The fault maintenance consultation request refers to a consultation request initiated by maintenance personnel through the human-computer interaction interface (e.g., touch screen, voice input module, or keyboard) of the fault diagnosis equipment, containing the model of the equipment to be maintained and a description of the fault phenomenon; the target equipment model information refers to the brand, series, and model identification information of the equipment to be maintained extracted from the fault maintenance consultation request, such as "Brand A blockchain server S19jPro" or "Brand B blockchain server M30S++"; the fault description information refers to the natural language description of the equipment fault phenomenon extracted from the fault maintenance consultation request, such as "the computing power of the computing board is unstable and frequently disconnects" or "the power supply fan does not turn, and the equipment overheats and shuts down".

[0028] The local professional knowledge base refers to a structured maintenance knowledge database indexed by device model, deployed on the local storage media of the fault diagnosis equipment (e.g., onboard eMMC, SSD, or SD card). Each index corresponds to at least one maintenance article fragment and at least one maintenance reference image. Maintenance article fragments are structured and annotated text paragraphs extracted from maintenance manuals, technical bulletins, and selected posts from maintenance forums. Each fragment includes four dimensions of annotation tags: "fault phenomenon," "cause analysis," "maintenance steps," and "precautions." Maintenance reference images are accompanying images such as physical circuit board diagrams, component location diagrams, and wiring diagrams, in formats including JPEG, PNG, or BMP. Target maintenance data refers to the collection of all maintenance article fragments and maintenance reference images associated with the target device model, obtained through model matching retrieval.

[0029] Among them, the large language model refers to a pre-trained language model that is deployed in a remote management backend cloud server connected to the fault diagnosis equipment and has been fine-tuned. It supports text generation based on prompt words, and its parameter scale is 7B to 70B. It runs on an onboard GPU or NPU acceleration chip. The initial maintenance guidance result refers to the original text content output by the large language model after one round of input consultation prompt words. The maintenance guidance information refers to the structured guidance text that has been optimized and is provided for maintenance personnel to read and refer to directly during maintenance operations. The maintenance guidance information may include a title field, a fault phenomenon description paragraph, a fault analysis paragraph, a solution paragraph, and maintenance reference images embedded in the corresponding text positions. The output format constraints specifically include: a word count constraint, which limits the number of words in the repair solution generated by the large language model to a preset range, such as 1000-1500 words; an image address constraint, which limits the large language model to place the image address of the repair reference image in the text content position corresponding to the image semantics in the returned result, and prohibits placing all image addresses at the end of the text; a content exclusion constraint, which limits the returned result of the large language model to prohibit the inclusion of third-party recommended content, authorized service contact information, and advertising content; a data source constraint, which limits the large language model to generate content only based on the input target repair data and not to introduce external data source content other than the local professional knowledge base; and a concealment constraint, which limits the output format constraints themselves to not be echoed in the output content of the large language model and to remain effective in subsequent consultation rounds.

[0030] Secondly, the fault diagnosis equipment receives fault repair consultation requests input by maintenance personnel through its human-machine interface, extracts the target equipment model information and fault description information from the request, and the extraction methods include model pattern matching based on regular expressions (such as matching model coding patterns such as "A\d+""S\d+""M\d+") and fault phenomenon phrase extraction based on named entity recognition models. Subsequently, the fault diagnosis equipment uses the target equipment model information as the search key value to perform a full-text search based on an inverted index in the local professional knowledge base, matching all maintenance article fragments and maintenance reference images associated with the equipment model, and generating target maintenance data. Finally, the fault diagnosis equipment assembles the target maintenance data into a maintenance article analysis and processing instruction segment containing a first round of consultation prompts. It also adds an output format constraint segment containing word count constraints, image address constraints, content exclusion constraints, data source constraints, and hidden constraints, forming a complete first round of consultation prompts. This is input into a large language model for reasoning to obtain initial maintenance guidance results. The initial maintenance guidance results are then subjected to content optimization processing, including semantic segmentation recognition, structured recombination, and image positional insertion, generating the final maintenance guidance information, which is then presented to maintenance personnel on the fault diagnosis equipment's display interface.

[0031] Through the above technical solution, this embodiment establishes a fully automated process from equipment model matching and retrieval to multi-dimensional constraint prompt word construction and post-processing of the large language model output. This compresses the manual information integration process, which originally required maintenance personnel to manually sift through multiple sources and take tens of minutes, into an automatic generation process that takes only seconds. Specifically, the automatic retrieval of target maintenance data replaces the manual search of maintenance manuals and forum posts by maintenance personnel. The five-dimensional constraint system ensures the format standardization, accuracy of image and text association, content purity, and data security of the output content. Semantic segmentation and image insertion further transform the raw output of the large language model into structured maintenance guidance information that can be directly used on-site.

[0032] The core advantage of this embodiment lies in linking device model-driven knowledge retrieval, multi-dimensional constraint-driven prompt word construction, and structured post-processing into a closed-loop automated pipeline, upgrading the fault diagnosis equipment from a simple "parameter detection tool" to a "maintenance guidance generation platform." Existing technologies, when encountering the problem of low efficiency in obtaining maintenance information, typically use large language models directly as question-and-answer systems—that is, maintenance personnel directly ask questions to the large language model in natural language, and the large language model directly answers based on the general knowledge memorized in its pre-trained parameters. However, this approach has fundamental flaws: the pre-trained knowledge of large language models is usually general domain knowledge, lacking in-depth maintenance details for specific brands and models (e.g., the difference in fault characteristics caused by changes in capacitor models in specific batches of a certain model of computing board), and the output content cannot guarantee traceability, uniform format, and correspondence between text and images. The reason why those skilled in the art cannot conceive of linking model-level matching retrieval, multi-dimensional output constraints, and content post-processing into an automated pipeline is because there is a technical bias in the field of fault diagnosis technology that equates "large language model = general question-answering tool"—that is, the value of a large language model lies in the broad coverage of its pre-trained parameters, and the role of accessing the knowledge base is merely to provide reference context for the large language model, while ignoring that in the highly specialized scenario of maintenance guidance, the structured constraints of prompt words can themselves serve as a core engineering means to control output quality and format, rather than merely providing content context.

[0033] This embodiment differs fundamentally from conventional "maintenance question-and-answer systems based on large language models." Conventional systems typically use large language models as the direct dialogue interface for maintenance personnel. Personnel ask questions in natural language, and the large language model generates answers directly based on general maintenance knowledge in pre-trained parameters and the context provided by the user. This approach relies on the large language model's own knowledge generalization ability without performing local knowledge retrieval and constraint system construction specific to a particular device model. In contrast, this embodiment combines precise retrieval from a local professional knowledge base driven by the device model with a five-dimensional output format constraint system. This ensures a strict match between the output content and the target device model while separating constraint instructions from the output content through implicit constraint terms—constraint instructions only apply to the generation process of the large language model and do not contaminate the output text. Existing technologies for solving the problem of "automatic generation of maintenance guidance information" typically employ a retrieval-enhanced generation (RAG) architecture. This approach involves concatenating relevant document fragments into a context and inputting it into a large language model, which then generates a response based on the concatenated context. However, the RAG approach does not address the multidimensional constraints of the output format, does not pre-annotate the retrieved documents with structure (such as paragraph annotations in four dimensions: "fault phenomenon," "cause analysis," "maintenance steps," and "precautions"), and does not handle the positional correspondence of images within the output content. As a result, the generated maintenance guidance text lacks structural hierarchy and the correlation between text and images. The reason why existing technologies do not adopt the "multi-dimensional constraint-driven prompt word construction" approach is because there is a technical bias in the field of natural language processing that "prompt words are only used for content guidance rather than format control." That is, it is generally believed that the role of prompt words is limited to specifying the answer topic and content direction to the large language model, while format control should be achieved through post-processing rules or template filling. This embodiment breaks this bias by embedding format control instructions in five dimensions—character count constraints, image address constraints, content exclusion constraints, data source constraints, and concealment constraints—within the prompt words. This makes format control an intrinsic capability of the prompt word engineering, rather than an external post-processing feature. This "constraint-intrinsic prompt word construction" mechanism is non-obvious in the field of fault diagnosis technology.

[0034] This embodiment describes a fault repair guidance method applied to fault diagnosis equipment. The method obtains a fault repair consultation request, determines the target equipment model information and fault description information, then performs a matching search in a local professional knowledge base based on the target equipment model information to obtain target repair data associated with the equipment model. Next, it constructs a round of consultation prompts based on the target repair data, inputs these prompts into a large language model, obtains the initial repair guidance results output by the large language model, and finally optimizes the content of the initial repair guidance results to generate corresponding repair guidance information. In this way, a fully automated generation chain from equipment model to executable repair guidance is realized, significantly reducing the manual analysis costs for repair personnel.

[0035] Although Figure 1 The illustrated embodiment can drive a large language model to generate initial maintenance guidance results by constructing a one-round consultation prompt word that includes a maintenance article analysis and processing instruction segment and an output format constraint segment. This achieves automatic generation of maintenance guidance from equipment model. However, there is still considerable room for optimization in the internal construction method of the one-round consultation prompt word. The maintenance article fragment itself is unstructured or semi-structured text. Maintenance articles from different sources (such as standardized paragraphs in maintenance manuals and free descriptions in forum posts) have significant differences in paragraph organization. If the original maintenance article text is simply spliced ​​directly into the instruction segment without recognizing and reorganizing its internal structure when constructing the prompt word, the large language model may not be able to accurately distinguish the functional roles of each paragraph when faced with messy text that mixes fault phenomenon descriptions, technical parameter listings, and operation step descriptions. This can lead to semantic overlap and logical confusion between the fault cause analysis and the operation guidance of maintenance steps in the generated initial maintenance guidance result. For example, content belonging to the "cause analysis" paragraph may be mistakenly used as part of the "maintenance steps" output, causing maintenance personnel to troubleshoot the fault according to incorrect causal logic when performing operations.

[0036] Therefore, referring to Figure 2 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The embodiment shown constructs a round of consultation prompts based on the target maintenance data, specifically in steps S310-S340, wherein: S310. Determine the preset paragraph titles contained in the maintenance article fragment; S320. Based on the type and arrangement order of the preset paragraph titles, determine the corresponding multiple analysis instruction text segments, and concatenate the multiple analysis instruction text segments in the recognition order to generate a maintenance article analysis and processing instruction segment. S330. Obtain preset constraint requirement information, and determine the character limit, maintenance reference image processing requirements, content exclusion rules, and data source limitation clauses contained in the constraint requirement information. S340. The word count limit, the maintenance reference image processing requirements, the content exclusion rules, and the data source limitation terms are arranged in a preset constraint order to generate the first round of consultation prompts.

[0037] Among them, the preset paragraph titles refer to the structured tags pre-labeled on each repair article segment in the local professional knowledge base. The tag system includes four standard categories: "Fault Phenomenon," "Cause Analysis," "Repair Steps," and "Precautions." Each tag corresponds to a logical paragraph in the repair article. The analysis instruction text segment refers to the natural language instruction segment dynamically generated based on the identified preset paragraph title type. Each paragraph title type corresponds to a preset analysis instruction template. For example, the title "Fault Phenomenon" corresponds to the analysis instruction "Please read the following fault phenomenon description paragraph and extract the key fault manifestations." The title "Cause Analysis" corresponds to the analysis instruction "Please read the following cause analysis paragraph and extract the root cause and related factors of the fault." The title "Repair Steps" corresponds to the analysis instruction "Please list the operation sequence and specific methods in the following repair step paragraphs in detail." The title "Precautions" corresponds to the analysis instruction "Please extract the safety warnings and operation taboos in the following precautions paragraphs." The constraint requirements information refers to the set of constraint parameters pre-stored in the fault diagnosis equipment configuration module. These include word limits (e.g., limiting the word count of the repair solution text to between 500 and 2000 words), requirements for processing repair reference images (e.g., "the image address of each repair reference image shall be placed after the text paragraph corresponding to the semantics of the image, and it is forbidden to put all image addresses together at the end of the output content"), content exclusion rules (e.g., "it is forbidden to include recommendations for third-party repair tools, contact information of authorized service points, and any form of commercial advertising in the output content"), and data source restriction clauses (e.g., "the answer shall be strictly based only on the content of the repair article excerpts provided above, and no external knowledge not mentioned in the above materials shall be cited"). The preset constraint order refers to the fixed arrangement of word limit → repair reference image processing requirements → content exclusion rules → data source restriction clauses. This arrangement order is designed according to the progressive relationship of the constraint hierarchy from the text form layer → image layer → content security layer → data source layer.

[0038] Secondly, the fault diagnosis equipment identifies paragraph titles for each maintenance article segment in the acquired target maintenance data. This identification can be based on regular expressions matching the paragraph start markers in the maintenance article segment (e.g., title lines marked with square brackets such as "[Fault Phenomenon]", "[Cause Analysis]", "[Maintenance Steps]", "[Precautions]"). If a maintenance article segment lacks a certain type of preset paragraph title, the generation of the corresponding analysis instruction is skipped. Subsequently, based on the types of preset paragraph titles actually identified, the fault diagnosis equipment retrieves the corresponding analysis instruction text segments from the instruction template library. Following the original order of the paragraph titles in the maintenance article segment, each analysis instruction text segment is sequentially concatenated with the corresponding maintenance article text content in the format "[Analysis Instruction Text Segment] + [Corresponding Paragraph Original Text]" to generate a complete maintenance article analysis and processing instruction segment. Finally, the fault diagnosis equipment reads the preset constraint requirements information from the configuration module, extracts four types of constraint parameters: word limit, maintenance reference image processing requirements, content exclusion rules, and data source limitation clauses, and concatenates them into an output format constraint segment according to the preset constraint order (word limit → maintenance reference image processing requirements → content exclusion rules → data source limitation clauses). The maintenance article analysis and processing instruction segment is placed first, followed by the output format constraint segment, and connected by delimiters (such as ", " or "===") to generate a complete round of consultation prompts.

[0039] Through the above technical solution, this embodiment identifies preset paragraph titles within maintenance article fragments, generates targeted analysis instruction text segments for paragraphs of different functional roles, and concatenates these analysis instruction text segments according to the original paragraph order to form a maintenance article analysis and processing instruction segment. This allows the large language model to understand the content structure of the maintenance article segment by segment according to the logical order of the original text when receiving prompt words, avoiding the mixed output of paragraph content from different functions. Furthermore, the four types of constraint parameters are organized into output format constraint segments according to a hierarchical progression, creating a hierarchical relationship from formal constraints to data source constraints within the prompt words, improving the completeness of constraint coverage and execution consistency.

[0040] The core advantage of this embodiment lies in integrating the paragraph structure recognition of maintenance text fragments, the dynamic generation of analysis instructions, and the hierarchical organization of constraint parameters into a unified prompt word construction pipeline. Existing technologies, when encountering unstructured maintenance text processing problems, typically employ text preprocessing—that is, first segmenting the maintenance text fragments into independent paragraphs using preset rules or classification models, and then inputting them separately into a large language model for segmentation and merging. This approach treats content splitting and result merging as two independent processing stages. However, this increases the complexity of the system pipeline, and segmentation processing severs the semantic connections between different paragraphs (for example, safety warnings in "Precautions" often have a one-to-one correspondence with specific operating steps in "Maintenance Procedures," and independent segmentation processing leads to the loss of this correspondence). The reason why those skilled in the art cannot conceive of processing semi-structured maintenance text through "dynamic analysis of instruction generation + paragraph order preservation" within the prompt words is because there is a technical bias in the field of prompt word engineering that "prompt words = fixed templates"—that is, it is believed that prompt words should be a pre-set static text, and only the variable placeholders in it are replaced according to different input content, while ignoring that prompt words themselves can also be dynamically reorganized and spliced ​​according to the structural characteristics of the input content. This mechanism of "structure-aware prompt word dynamic construction" is not obvious in the field of fault diagnosis technology.

[0041] This embodiment differs fundamentally from conventional fixed-template prompt construction. Conventional fixed-template prompt construction typically uses one or more predefined static instruction texts as the prompt skeleton, filling only the original text of the maintenance article fragments in variable placeholders. This approach treats the maintenance article fragments as indiscriminate plain text blocks and fills them entirely into the template. In contrast, this embodiment first identifies the internal paragraph structure of the maintenance article fragment, then generates targeted analysis instructions for each paragraph type, and finally assembles them according to the original text order. This construction method ensures a precise one-to-one correspondence between the distribution of analysis instructions for the prompts and the content structure of the maintenance article. Existing technologies, when addressing the problem of "prompt-based conversion of unstructured knowledge documents," typically employ preprocessing schemes such as summary generation or key information extraction. These schemes first use an extractive summarization model to extract key sentences from the maintenance article fragments, then use these key sentences as input context for a large language model. However, this approach loses the original paragraph logical structure and the citation relationships between paragraphs. The reason why existing technologies do not adopt the "structure-aware dynamic instruction construction" approach is because there is a technical bias in prompt word engineering that "prompt words should be determined before the task begins." That is, prompt words are considered to be a static, holistic, pre-set concept, and dynamic selection and order adjustment of instruction fragments based on the internal structural features of the input content are not allowed during the construction process. This embodiment overcomes the above-mentioned technical bias by using the paragraph title recognition result as the trigger condition for instruction template selection, so that the internal instruction distribution of the prompt words adapts to the structural features of the input content.

[0042] There is still significant room for improvement in the post-processing of the initial maintenance guidance results output by the large language model. When the large language model receives prompt words and generates initial maintenance guidance results, although it is constrained by the output format constraints, the original text output may still exhibit certain format deviations at the level of paragraph organization. For example, the separators between the title and the body text may be inconsistent, the fault phenomenon description paragraph may contain content that should belong to the cause analysis paragraph, and the step descriptions in the solution paragraph may not be arranged strictly in the order of operation. If the initial maintenance guidance results are directly output to maintenance personnel for reading, the inconsistency in format and the cross-mixing of content areas may reduce the efficiency of maintenance personnel in extracting key information. Especially in emergency maintenance scenarios, maintenance personnel need to quickly locate the three core information areas of "what fault → what cause → how to repair" within a few seconds. The messy output will significantly increase the cognitive load of information extraction.

[0043] Therefore, referring to Figure 3 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The illustrated embodiment optimizes the initial maintenance guidance results to generate corresponding maintenance guidance information, including steps S410-S460, wherein: S410. Perform semantic segmentation recognition on the initial maintenance guidance results, and identify the title area, fault phenomenon description area, fault analysis area and solution area based on the preset paragraph semantic tag set; S420. Extract the title text from the identified title region and generate the corresponding title field; S430. Extract the fault phenomenon description text from the fault phenomenon description area and generate the corresponding fault phenomenon description segment. S440. Extract the fault analysis text from the fault analysis area and generate the corresponding fault analysis segment; S450. Extract the solution text from the solution area and generate the corresponding solution segment; S460. Reorganize the title field, the fault phenomenon description section, the fault analysis section, and the solution section according to a preset order to generate corresponding maintenance guidance information.

[0044] The preset paragraph semantic tag set refers to four predefined sets of semantic tags, including title tags (corresponding to semantic tags such as "[Fault Title]" "[Repair Topic]"), fault phenomenon tags (corresponding to semantic tags such as "[Fault Phenomenon]" "[Problem Description]" "[Symptom Manifestation]"), fault analysis tags (corresponding to semantic tags such as "[Cause Analysis]" "[Fault Cause]" "[Problem Diagnosis]"), and solution tags (corresponding to semantic tags such as "[Repair Steps]" "[Solution]" "[Handling Measures]" "[Operation Guidelines]"). Each tag type contains multiple semantically equivalent tag aliases to cover possible wording differences in the output of the large language model. Semantic segmentation recognition refers to the process of detecting paragraph boundaries and classifying semantic types of the initial maintenance guidance result text based on a preset paragraph semantic tag set. The recognition method includes a cascade of two methods: tag line matching based on regular expressions and paragraph semantic classification based on a pre-trained text classification model (such as the BERT-base fine-tuning model). First, the candidate paragraph boundaries are determined by matching explicit tag lines with regular expressions, and then the semantic types of the text paragraphs between the boundaries are confirmed by the text classification model to eliminate the ambiguity of regular expression matching. The title field refers to the plain text string extracted from the title area as the title of the maintenance guidance information, such as "Maintenance guidance for unstable computing power fault of A brand blockchain server S19j Pro computing power board". The preset arrangement order refers to the fixed output order of title field → fault phenomenon description paragraph → fault analysis paragraph → solution paragraph.

[0045] Secondly, after the fault diagnosis equipment obtains the initial maintenance guidance results output by the large language model, it uses a cascaded recognition strategy for semantic segmentation: first, it scans the entire text using regular expressions to match all lines that conform to the label category patterns in the preset paragraph semantic label set (e.g., lines starting with "

" and ending with "

[0046] Through the above technical solution, this embodiment uses a cascaded semantic segmentation and recognition strategy (regular expression initial screening + classification model confirmation) to automatically segment and classify the unstructured initial maintenance guidance results output by the large language model according to four semantic dimensions, and then reorganize them into structured maintenance guidance information according to a preset arrangement order. This ensures that when maintenance personnel open the guidance information, the first thing they see is the equipment model title, followed by a progressive presentation of fault phenomena, cause analysis, and maintenance steps. This keeps the cognitive path of information extraction consistent with the maintenance personnel's natural troubleshooting logic (first confirm the fault → then analyze the cause → finally perform maintenance).

[0047] The core advantage of this embodiment lies in replacing simple text segmentation based on delimiters with semantic segmentation recognition, enabling the post-processing pipeline to understand the semantic structure of the text output by the large language model. Existing technologies, when faced with the need for standardized output formats from large language models, typically employ rule-based segmentation schemes based on fixed delimiters—that is, forcibly inserting specific delimiters before and after each paragraph type (e.g., "#####fault phenomenon#####") in the output format constraints of the large language model, and then performing segmentation through delimiter matching in the post-processing stage. This approach completely entrusts paragraph boundary marking to the large language model. However, this approach is unreliable: the large language model may omit delimiters, insert incorrect delimiters, or insert inconsistent delimiter formats in different answers during the generation process, leading to rule-based segmentation failure in the post-processing stage. The reason why existing technologies do not adopt the cascaded recognition strategy of "regular expression screening + classification model confirmation" for semantic segmentation is because there is a technical bias in the field of prompt word engineering that "post-processing should rely entirely on the formatted output generated by prompt word constraints"—that is, it is believed that as long as the prompt word constraints are written strictly enough, the large language model will definitely output text that fully conforms to the format specifications, and post-processing can be simplified to simple string matching, while underestimating the degradation problem of format consistency when the large language model generates long text.

[0048] The semantic segmentation recognition post-processing in this embodiment differs fundamentally from conventional fixed-delimiter segmentation post-processing. Conventional fixed-delimiter segmentation post-processing relies entirely on delimiter markers inserted by the large language model during generation for segmentation. This approach ties the correctness of post-processing to the accuracy of the large language model's output format; if the large language model's output deviates from the expected delimiter format, post-processing fails. In contrast, this embodiment introduces independent semantic recognition capabilities in the post-processing stage (regular expression screening is only used for quickly locating candidate boundaries; the classification model is the core of semantic confirmation). This allows post-processing to move beyond solely relying on format markers in the large language model's output, enabling correct segmentation based on semantic features of the text content even when format markers are missing or incorrect. Existing technologies for solving the problem of "structured reorganization of unstructured text" typically employ supervised sequence labeling models (such as BiLSTM-CRF) to perform token-level semantic role labeling on the text before aggregating it by role. This approach requires a large amount of manually labeled maintenance text training data, with labeling granularity down to the token level, resulting in extremely high labeling costs. The reason why existing technologies do not adopt the "cascaded semantic segmentation recognition" scheme is because there is a technical bias in the field of natural language processing that "end-to-end models are superior to cascaded schemes." This embodiment reduces the input volume and false alarm rate of the classification model through regularization screening, and reduces the computational overhead of the classification model while ensuring accuracy with the cascaded architecture. This cascaded lightweight semantic segmentation strategy is non-obvious in the specific scenario of post-processing of maintenance guidance information.

[0049] There is still significant room for improvement in how maintenance reference images are inserted into maintenance guidance information. In the context of maintenance guidance for blockchain servers, the value of maintenance reference images is highly dependent on their position within the guidance information. For example, a physical image of a circuit board labeled "Test point of the 3rd DC-DC converter on the computing board" can only be directly located by maintenance personnel and the measurement operation performed if it appears after the text paragraph describing "measuring the output voltage of the 3rd DC-DC converter with a multimeter." If the image is placed at the end of the guidance information (a common practice in traditional maintenance manuals and most AI-generated maintenance guidance), maintenance personnel must look away, flip to the end of the text to find the image, and then return to the corresponding text location when reading "measuring the 3rd DC-DC converter." This viewpoint jump not only increases the cognitive switching cost, but also interrupts the maintenance process when maintenance personnel are holding measuring tools in both hands, as the page-turning or scrolling operation itself will disrupt the maintenance workflow. More importantly, the semantic association between different repair reference images and the repair text in this section is uneven. Some images (such as the overall layout diagram of the circuit board) may be associated with multiple descriptions in the fault analysis section, while other images (such as the pin label diagram of a specific component) are only associated with a specific measurement action in the solution section. Simply inserting images according to their storage order in the knowledge base without considering the semantic correspondence between the images and the text cannot achieve the best matching between images and text in the repair guidance information.

[0050] Therefore, referring to Figure 4 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The embodiment shown optimizes the initial maintenance guidance results to generate corresponding maintenance guidance information, including steps S470-S490, wherein: S470. Extract the title field, fault phenomenon description section, fault analysis section, and solution section from the initial maintenance guidance result; S480. Determine the original text location information of each repair reference image in the target repair data within the repair article fragment; S490. Insert the repair reference image into the corresponding text content position in the initial repair guidance result according to the original text position information to generate the repair guidance information.

[0051] The original text location information refers to the embedded location identifier of the maintenance reference image in the original text of the source maintenance article, which is recorded when the maintenance reference image is included in the local professional knowledge base. This includes the paragraph title type (such as "maintenance steps"), the relative offset of the image within the paragraph (with the paragraph starting at 0, counted according to the number of Unicode characters), and the text context interval of 256 characters before and after the image (used for text alignment in the post-processing stage). The corresponding text content location refers to the location in the initial maintenance guidance result with the highest similarity to the text context interval of the original text location information. The similarity is calculated as follows: extract TF-IDF vectors (feature dimension of 5000, using character-level 2-gram and 3-gram features) from the text context intervals before and after the original text location information and the text before and after each candidate location in the initial maintenance guidance result, calculate the cosine similarity, and take the candidate location with the cosine similarity exceeding the threshold (such as 0.75) and the largest similarity value as the corresponding text content location.

[0052] Secondly, after obtaining the initial maintenance guidance results, the fault diagnosis equipment first extracts four types of text regions: title field, fault phenomenon description section, fault analysis section, and solution section. Then, the fault diagnosis equipment reads the original text location information of each maintenance reference image obtained by the control device from its local professional knowledge base. For each maintenance reference image, it traverses the preceding and following text context intervals recorded in the original text location information. In the initial maintenance guidance results, it calculates the cosine similarity between the TF-IDF vector of each window position and the TF-IDF vector of the original text context interval using a sliding window (window size is the same as the length of the context interval, with a step size of 32 Unicode characters). It then records the maximum similarity value and corresponding position of all windows. Finally, the maintenance reference images are sorted from highest to lowest based on the maximum similarity value. For each image, if the maximum similarity value exceeds the threshold of 0.75, the image address marker (e.g., "[Image: Computing Board DC-DC Test Point.jpg]") is inserted at the position of the maximum similarity in the initial maintenance guidance result to generate the final maintenance guidance information. If the maximum similarity value of an image does not exceed the threshold, it is inserted at the end of a text area with the same paragraph title type as its original position information.

[0053] Through the above technical solution, this embodiment establishes a TF-IDF vector similarity alignment mechanism between the original text location information of the maintenance reference image and the initial maintenance guidance result text. This enables the semantic positional insertion of the maintenance reference image in the maintenance guidance information, ensuring that each maintenance reference image appears after the text paragraph most relevant to its semantics. This allows maintenance personnel to achieve a smooth reading experience of "reading a paragraph of text and looking at a corresponding image" when reading the maintenance guidance information, eliminating the problems of image-text fragmentation and frequent viewpoint jumps caused by traditional concentrated image placement at the end of text.

[0054] The core advantage of this embodiment lies in introducing a text-image semantic alignment mechanism based on the original text's location information into the post-processing workflow of maintenance guidance information. This upgrades image insertion from "batch attachment at fixed positions" to "precise embedding one by one based on semantic correspondence." Existing technologies, when encountering issues with text-image association in maintenance guidance, typically use a method of listing all maintenance reference images as attachments at the end of the document—this involves listing all maintenance reference images as attachments at the end of the maintenance guidance information, using phrases like "as shown in the image." Figure 1 "As shown" "See appendix" Figure 2 Textual references such as "see Figure X" refer to preceding text. This practice stems from the layout inertia of traditional paper repair manuals—due to the technical limitations of printing and typesetting, paper manuals had to insert "see Figure X" references into the main text and uniformly concentrate images at the end of chapters or in appendices. However, in the scenario of displaying digital repair guidance information, the screen is a dynamic medium that can scroll freely, without the physical limitations of paper layout, and images can be directly embedded into their semantically corresponding text positions. The reason why existing technology cannot think of semantic alignment insertion based on the original text position information is because there is a technical bias in the field of technical document generation that "separation of text and images is the standard format for repair manuals"—that is, mistakenly equating the layout constraints of paper repair manuals with the format specifications of repair guidance information, ignoring the need for real-time text-image association in digital display scenarios.

[0055] The semantically aligned image insertion in this embodiment differs fundamentally from conventional end-of-text image insertion. Conventional end-of-text image insertion leaves the entire association between images and text to the maintenance personnel. They must interrupt reading based on image references (e.g., "see Figure X"), scroll to the end of the text to find the corresponding image, and then scroll back to continue reading. This multi-step viewpoint switching process is extremely impractical when maintenance personnel are using tools. In contrast, this embodiment uses TF-IDF semantic alignment driven by the original text's location information, automatically placing the image after its semantically most relevant text position. Maintenance personnel can see the corresponding image the instant they see the text description without any additional effort. Existing technologies for solving the "image-text association display" problem typically employ object detection or visual localization models to extract visual features from images and perform cross-modal alignment with the text. This approach requires deploying computationally expensive visual-language pre-trained models (such as CLIP or BLIP-2), making it impractical for embedded devices. The reason why existing technologies do not adopt the "text-side alignment based on original text location information" scheme is because there is a technical bias in the field of multimedia information retrieval that "image-text alignment must rely on visual features." However, this embodiment uses the original text location information recorded during the construction of the knowledge base as a bridge to complete the alignment calculation entirely on the text side, avoiding the computational bottleneck of image feature extraction on embedded devices. This mechanism of "using original text location information to replace visual features for image-text alignment" can achieve the desired purpose in a convenient way in the specific scenario of maintenance guidance, because the maintenance reference image is naturally bound to its source text when it is collected.

[0056] There is still significant room for optimization in the acquisition of target maintenance data—that is, how to quickly and accurately match maintenance data associated with the target device model from the local professional knowledge base. In the maintenance scenario of blockchain servers, the naming system of device models exhibits a highly hierarchical characteristic: for example, in "A Brand Blockchain Server S19j Pro", "A Brand Blockchain Server" is the brand name, "S19" is the series name, and "j Pro" is the model number variant. Different variants within the same series (such as S19, S19j, S19j Pro, S19 XP) share some common series-level maintenance knowledge (such as the common computing board architecture and heat dissipation design of the series), as well as some differentiated maintenance points unique to the model variants (such as the specific power module layout of S19jPro and the specific water cooling interface design of S19 XP). If the retrieval system simply uses the complete model number string as the search key for exact matching, the search will fail because there is no completely identical record for the target model in the knowledge base. If only brand words are used for fuzzy matching, the search results will contain a large amount of maintenance data for a series of equipment that is unrelated to the target model, resulting in too much noise information being input into the subsequent prompt word construction, which affects the relevance and accuracy of the large language model output.

[0057] Therefore, referring to Figure 5 In another embodiment of the present invention, a fault repair guidance method is provided, based on the above. Figure 1 The illustrated embodiment involves performing a matching search in a local professional knowledge base based on the target equipment model information to obtain target maintenance data associated with the equipment model, including steps S210-S250, wherein: S210. The target device model information is divided into three levels of keyword units: brand words, series words, and model serial number. S220. Using the brand term as the first matching level, retrieve all device model sets belonging to the brand from the local professional knowledge base; S230. Using the series words as the second matching level, filter out the first candidate model subset belonging to the corresponding series from the set of all equipment models; S240. Using the model number as the third matching level, match a unique target model in the first candidate model subset. S250. Extract all repair article fragment indexes and all repair reference image indexes corresponding to the unique target model in the local professional knowledge base, concatenate the original text content corresponding to all the repair article fragment indexes, and associate the image files corresponding to all the repair reference image indexes to generate the target repair data.

[0058] Among them, the brand term refers to the part of the equipment model name that identifies the manufacturer's brand, such as splitting "A Brand Blockchain Server S19j Pro" into "A Brand Blockchain Server" or "Antminer". The brand term is identified through a pre-set brand term dictionary (containing standard names and aliases of known blockchain server brands such as "A Brand Blockchain Server / Antminer", "B Brand Blockchain Server / Whatsminer", "C Brand Blockchain Server / Avalon", and "D Brand Blockchain Server / Leopard"). The series term refers to the part of the equipment model name that identifies the product series, usually consisting of letters followed by numbers, such as "S19", "M30", and "A12". The series term is extracted through pattern matching based on the regular expression "[A-Za-z]+\d+". The model number refers to the suffix in the equipment model name that identifies the specific variant under the series, such as "j". "Pro", "S++", "XP", "Hydro"; the model number is extracted from the remaining string after sequentially removing identified brand and series words from the complete model string; edit distance refers to the minimum number of single-character editing operations required to transform one string into another, including insertion, deletion, and replacement.

[0059] Secondly, after obtaining the target device model information, the fault diagnosis equipment first performs prefix matching on the model string through a preset brand word dictionary to identify and extract the brand word. If the brand word has multiple aliases in the dictionary, they are uniformly mapped to the standard brand name. After removing the brand word from the original string, the remaining part is matched using the regular expression "[A-Za-z]+\d+" to extract the series words. The remaining string part is then used as the model number. Subsequently, using brand terms as the first matching level, the brand tag field is retrieved from the inverted index of the local professional knowledge base to obtain the set of all included device models under that brand. Using series terms as the second matching level, prefix matching is performed on all device models (matching whether the series terms appear as substrings in the device models) to filter out the first candidate model subset belonging to the corresponding series. Using model number as the third matching level, exact string matching is first performed on the first candidate model subset. If an exact match is found, the result is directly used as the unique target model. If no exact match is found, the edit distance between the target model number and each candidate model number in the first candidate model subset is calculated, and the candidate model corresponding to the model number with the smallest edit distance (not exceeding 2) is selected as the final matching result. Finally, based on the unique device ID of the target model obtained by matching, the local professional knowledge base is used to query all maintenance article fragment indexes and maintenance reference image indexes associated with the Device ID. The full text content of each maintenance article fragment is read and concatenated using the indexes, and the image file paths corresponding to each maintenance reference image index are added to the image file list of the target maintenance data to generate complete target maintenance data.

[0060] Through the above technical solution, this embodiment breaks down equipment model information into three levels: brand terms, series terms, and model serial numbers, and constructs a three-level matching pipeline that narrows progressively. While ensuring that the retrieval granularity progresses from coarse to fine, it solves the problem of missing matches in the naming of blockchain server equipment models when the models are similar to variant models in the same series but the exact variant models may not be included in the knowledge base through the edit distance fuzzy matching strategy of the last-level model serial number. This enables the system to automatically match the closest known model variant when the target model is not directly included, so as to maximize the reuse of existing maintenance knowledge.

[0061] The core advantage of this embodiment lies in directly mapping the hierarchical naming characteristics of blockchain server device models to the multi-level index structure of the retrieval system, making the retrieval granularity completely aligned with the natural hierarchy of device models. Existing technologies, when encountering device model matching problems, typically employ a single-level string matching scheme—either exact matching of the complete model string or single-dimensional fuzzy matching of brand terms or series terms. However, exact matching has a high failure rate when the knowledge base coverage is incomplete, while single-dimensional fuzzy matching introduces a large amount of noisy data—for example, using only "A brand blockchain server" as the search condition will return repair data for hundreds of models across the entire S9 to S21 series, most of which are of no reference value to the currently awaiting repair S19j Pro and will severely dilute the output accuracy of the large language model. The reason why existing technologies cannot directly map the hierarchical naming system of device models to the three-level progressively narrowing structure of the retrieval system, supplemented by the final edit distance, is because there is a technical bias in the field of information retrieval that "retrieval accuracy and recall need to be balanced by a general word segmenter"—that is, it is believed that the optimization of retrieval results should be achieved by adjusting the word segmentation granularity and stop word list of the general word segmenter, while ignoring the highly structured characteristics of the naming system of blockchain server device models, which makes "brand-series-model number" itself the most accurate word segmentation scheme.

[0062] The three-level progressive narrowing matching in this embodiment differs fundamentally from conventional single-level string matching. Conventional single-level string matching treats the device model as an indivisible atomic string, either matching all or none, ignoring the hierarchical semantic information inherent in the device model. This embodiment, however, employs a strategy of first splitting the levels and then progressively narrowing them. Even if a precise match at the final level fails, the brand and series filtering at the first two levels has already narrowed the search scope from potentially millions of records in the global knowledge base to dozens of records within the same series, significantly reducing the computational scale and risk of mismatches in the final-level fuzzy matching. Existing technologies for solving the "fuzzy retrieval of structured names" problem typically employ inverted indexes based on n-grams or semantic retrieval schemes based on vector embedding. These approaches have high storage and computational resource requirements, making them unsuitable for deployment on embedded fault diagnosis devices. The reason why existing technologies do not adopt the "three-level progressive narrowing + final-level edit distance" scheme is because there is a technical bias in the field of information retrieval that "edit distance calculation is too costly on long strings" - while this embodiment calculates the edit distance only for model serial number strings with a length that usually does not exceed 8 characters at the final level, making its calculation cost negligible.

[0063] For scenarios where a single consultation cannot fully cover the entire troubleshooting process, there is still significant room for optimization. In actual blockchain server maintenance, a single fault may be caused by multiple potential root causes, and these root causes may be linked by a causal chain. For example, when maintenance personnel initially consult about "decreased computing power on the computing board," the maintenance guidance information generated by the control device based on the target maintenance data may initially point to "abnormal output of the DC-DC converter" as a preliminary troubleshooting direction. When maintenance personnel check the DC-DC converter according to the guidance information and find that it is working normally, they need to further inquire about "the DC-DC converter is normal, but the computing power is still decreasing" to investigate the next possible cause (such as poor heat dissipation of the ASIC chip or clock signal jitter). If the control device constructs the exact same prompt word structure in each consultation, it cannot utilize the information confirmed in the first round of consultation (such as "the DC-DC converter is normal") to narrow down the analysis scope of the second round of consultation. This results in the initial maintenance guidance results of the second round of consultation still containing the possibility of faults ruled out in the first round, increasing the redundancy in the maintenance information. In addition, when maintenance personnel need to see a comparison of different possibilities on a single screen, the presentation format of the secondary consultation should be different from the structured output of the first round (such as displaying each key point separately on a separate line) to meet the reading needs of quickly comparing different fault probabilities.

[0064] Optionally, refer to Figure 6 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The illustrated embodiment of the fault repair guidance method further includes steps S500-S800, wherein: S500: Obtain a secondary fault repair consultation request, and determine the first equipment model information and secondary consultation information in the secondary fault repair consultation request; S600. Based on the first device model information, determine the corresponding consultation database, which includes a local professional knowledge base and an external knowledge base; S700: Retrieve maintenance knowledge entries from the consultation database that match the first equipment model information and the secondary consultation information; S800. Construct a second-round consultation prompt corresponding to the maintenance knowledge item. The second-round consultation prompt includes a system role setting section and a secondary output format constraint section.

[0065] The "secondary fault repair consultation request" refers to a supplementary consultation request initiated by maintenance personnel through the human-machine interface of the fault diagnosis equipment after obtaining the initial maintenance guidance information, based on the execution results of the initial maintenance guidance information or newly discovered fault clues. The "first equipment model information" refers to the model of the equipment currently under repair carried in the secondary fault repair consultation request. This model may be the same as the target equipment model in the initial consultation (inquiring about different fault stages of the same equipment) or different (the maintenance personnel may ask about a different model in the interface). The "secondary consultation information" refers to the supplementary question description carried in the secondary fault repair consultation request, such as "The DC-DC converter is normal after checking the above steps, but the computing power is still decreasing. What should I check next?" The "consultation database" refers to the maintenance knowledge data source dynamically selected by the system based on the first equipment model information, including both local professional knowledge bases and external knowledge bases. External knowledge bases refer to those not stored locally on the fault diagnosis equipment but connected via a wireless network (such as Wi-Fi or 4G / 5G cellular). The knowledge obtained from remote maintenance knowledge servers accessed by WoNet or through network searches includes a wider range of equipment models and more frequently updated maintenance knowledge entries. A maintenance knowledge entry refers to one or more maintenance knowledge records retrieved from the consultation database that match the first equipment model information and the second consultation information. Each record contains a maintenance text description and optional maintenance reference images. The system role setting section refers to the instruction paragraph in the second-round consultation prompts used to set the response role of the large language model. Through natural language description, the response identity of the large language model is limited to "a data analysis assistant with professional knowledge of blockchain server maintenance," and the response style is required to be "detailed, professional, and include image references." The line break constraint item refers to a format control instruction in the second-round output format constraint section, requiring the large language model to separate each independent knowledge point or troubleshooting branch with a line break and display them on a separate line when generating the response, to meet the reading needs of maintenance personnel who need to quickly scan and compare different fault probabilities within a single screen. The constraints defined by the secondary output format constraint segment are different from those in the first round of structured prompts. The secondary output format constraint segment includes: line break constraints where each detail point is displayed independently on separate lines, and implicit constraints where the secondary output format constraint segment itself is not echoed in the output content.

[0066] Secondly, the fault diagnosis equipment receives secondary fault repair consultation requests input by maintenance personnel through a human-machine interface, extracting the primary equipment model information and secondary consultation information from the request text (the remaining part of the request text after removing the equipment model information is taken as the secondary consultation information). Then, based on the primary equipment model information, the consultation database is determined: it is determined whether the model corresponding to the primary equipment model information has a corresponding maintenance knowledge entry in the local professional knowledge base. If it exists, the local professional knowledge base is selected; otherwise, an external knowledge base is selected. After determining the consultation database, using the primary equipment model information and secondary consultation information as joint query conditions, semantic retrieval is performed in the selected consultation database (e.g., a hybrid retrieval of keyword matching based on the BM25 algorithm and semantic vector similarity retrieval based on SBERT (Sentence-BERT), with a weight ratio of 0.4:0.6 and a semantic vector dimension of 384; other semantic retrieval methods are also feasible), retrieving the Top-K (e.g., K=5) maintenance knowledge entries that match the query conditions. Finally, construct the second-round consultation prompts: First, concatenate the system role setting segment ("You are a professional blockchain server maintenance data assistant. Please perform fault analysis based on the following maintenance knowledge items. Please answer in detail, including maintenance reference images corresponding to the text semantics."). Then, concatenate the secondary output format constraint segment (including line break constraints and hidden constraints for each detailed point to be displayed independently on separate lines). Finally, concatenate the retrieved maintenance knowledge item text content after the system role setting segment and the secondary output format constraint segment, separating them with delimiters to form the complete second-round consultation prompts. Input this into the large language model to obtain the second-round maintenance guidance results.

[0067] Through the above technical solution, this embodiment differentiates between the prompt word construction strategies of the first and second rounds of consultation—the first round of consultation employs a five-dimensional constraint system for structured output to ensure the standardization and completeness of the output format, while the second round of consultation uses role setting and line break display constraints to adapt to reading scenarios requiring rapid comparison and supplementary investigation—ensuring that the output style and format of the same large language model in different consultation rounds precisely match the needs of the current consultation stage. Furthermore, by dynamically selecting the consultation database, when the second round of consultation involves device models not covered locally, it automatically switches to an external knowledge base, expanding the system's knowledge coverage boundaries.

[0068] The core advantage of this embodiment lies in its dual-channel adaptive architecture, which establishes a round-aware, differentiated prompt word construction strategy and a dynamic database selection mechanism. Existing technologies, when encountering multi-round consultation scenarios, typically use the same prompt word template across all rounds—reusing the prompt word template from the initial consultation for all subsequent rounds, only replacing the search results in the input content. This approach ignores the fundamental differences in information presentation needs across different consultation rounds: the first round requires a structured, complete fault analysis report, while subsequent rounds require supplementary troubleshooting clues that can be quickly scanned and compared item by item. The reason why those skilled in the art cannot conceive of designing differentiated output format constraints for different consultation rounds is due to the technical bias in human-computer dialogue system design of "using the same prompt word template for the same task type"—that is, assuming that the same maintenance consultation task has the same output format requirements across all rounds, ignoring the fundamentally different information consumption patterns of maintenance personnel in the "initial diagnosis" and "item-by-item troubleshooting" stages.

[0069] The round-aware differentiated prompt construction in this embodiment differs fundamentally from conventional unified template reuse. Conventional unified template reuse outputs text with a completely consistent format across all consultation rounds. This leads to the structured, regionalized format potentially failing to flexibly adapt to the parallel comparison display requirements of "Possible Cause A → Possible Cause B → Possible Cause C" in supplementary investigation rounds. This embodiment, however, ensures structured output (title – fault phenomenon – cause analysis – solution) using a five-dimensional constraint system in the first round of consultation, and switches to a line-by-line key point listing mode in the second round of consultation, achieving precise round-level adaptation between the output format and the information consumption pattern. Existing technologies, when addressing the problem of "multi-round output differentiation in dialogue systems," typically rely on a Dialogue State Tracking (DST) module to maintain the dialogue state between rounds and dynamically assemble prompts based on the state. This approach requires additional state tracking models and complex state management logic. The reason why existing technologies do not adopt the "direct template switching for each round" approach is that there is a problem in the field of dialogue system design where "the granularity of round switching is too coarse, and template switching will cause context breakage." However, this embodiment ensures that the purity of the second-round output is not affected by template switching by retaining the implicit constraint terms that continue to take effect in subsequent rounds. At the same time, the system role setting segment maintains the consistency of the professional assistant identity of the large language model in all rounds, so that template switching only affects the output format and does not affect the continuity of context.

[0070] There is still significant room for optimization in the dynamic database selection strategy for consultation—especially the logic for determining when to choose the local professional knowledge base and when to choose an external knowledge base. In secondary consultations, the consultation intentions of maintenance personnel may fall into two distinct scenarios: Scenario 1: During the initial maintenance guidance process, the maintenance personnel discover a new branch of the fault and need to obtain more maintenance details for the same equipment model—in this case, the most relevant maintenance data remains in the local professional knowledge base, and there is no need to access the external network; Scenario 2: After the initial maintenance, the maintenance personnel switch to repairing or seeking information about another equipment model different from the initial one, and the maintenance data for this equipment model may not be included in the local professional knowledge base—in this case, automatically switching to the external knowledge base can avoid the embarrassment of being unable to respond due to insufficient coverage of the local knowledge base. If the system adopts the same database selection strategy for these two scenarios (e.g., uniformly using the local knowledge base or uniformly using the external knowledge base), either in Scenario 2, an empty result will be returned due to the lack of local data, or in Scenario 1, unnecessary external network access will be triggered, introducing latency and bandwidth consumption.

[0071] Therefore, referring to Figure 7 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The illustrated embodiment determines the corresponding consultation database based on the first device model information, including steps S610-S630, wherein: S610. Determine whether there is at least one maintenance knowledge entry in the local professional knowledge base that matches the model information of the first device; S620. If at least one maintenance knowledge item exists, the consultation database is determined to be a local professional knowledge base; S630. If no maintenance knowledge entry exists, the consultation database is determined to be an external knowledge base.

[0072] Among them, a maintenance knowledge entry refers to a maintenance knowledge record in the local professional knowledge base indexed by the equipment model. Each record contains a text description of a maintenance scenario for that equipment model and optional maintenance reference images; matching with the first equipment model information refers to... Figure 5 The three-level progressive narrowing matching strategy described in the embodiment performs hierarchical matching using brand words, series words, and model serial numbers of the first device model information, and at least one device model record is matched in the local professional knowledge base.

[0073] Secondly, after obtaining the first device model information, the fault diagnosis equipment uses the first device model information as the search key value and then... Figure 5In this embodiment, a three-level progressively narrowing matching strategy (progressive matching of brand term → series term → model number, with the final level using fuzzy matching with an edit distance not exceeding 2 as a fallback) is used to perform a search in the local professional knowledge base to check for the existence of matching device model records. Subsequently, the search results are evaluated: if the search in the local professional knowledge base matches at least one maintenance knowledge item (i.e., the number of returned device model records is ≥1), then the consultation database is identified as the local professional knowledge base, and subsequent operations are performed entirely based on the local professional knowledge base; if the search in the local professional knowledge base does not match any maintenance knowledge items (i.e., the number of returned device model records is 0), then the consultation database is identified as the external knowledge base, and subsequent operations are performed by accessing the external knowledge base via the wireless network interface.

[0074] Through the above technical solution, this embodiment prioritizes checking the matching of the local professional knowledge base when selecting the consultation database. When matching data is available locally, the local knowledge base is used to avoid inaccurate external data or unnecessary network access delays (local SSD read latency is typically on the order of 0.1ms, while the round-trip latency of 4G wireless network access to external servers is typically on the order of 50-200ms). When no matching data is available locally, the system automatically switches to the external knowledge base to ensure service continuity. This achieves a two-level knowledge coverage strategy of prioritizing local and using network as a fallback.

[0075] This embodiment eliminates the synchronization gap between the external configuration list and the actual database content by retrieving the local professional knowledge base during each consultation after the initial consultation, directly basing the database routing decision on a real-time snapshot of the current database state. Existing technologies typically employ a federated query scheme at the database middleware layer to address the problem of "dynamic selection of multiple data sources." This approach sends query requests to both local and external data sources simultaneously, using the first returned result. However, this approach suffers from slowing down the overall response speed due to timeouts in external queries when the external network is unavailable or has high latency, or significant result errors may occur due to inaccuracies in the external data source. The reason why existing technologies do not adopt the serial decision-making scheme of "first searching locally and then deciding whether to access the external network" is that in the field of distributed data system design, most technicians believe that "the delay of serial decision-making is the sum of the delays of the two queries". However, in this embodiment, the delay of local retrieval is extremely low (on the order of 0.1ms), which is much smaller than the delay of external network access (on the order of 50-200ms). When the local retrieval is successful, the external access delay is completely avoided. Only in the rare case of local retrieval failure is the external access delay suffered. The overall average delay is actually lower than the delay model of the parallel query scheme of "always waiting for the slowest data source".

[0076] Optionally, refer to Figure 8 Another embodiment of the present invention provides a fault repair guidance method, based on the above. Figure 1 The illustrated embodiment determines the corresponding consultation database based on the first device model information, including steps S640-S660, wherein: S640. Determine the consistency between the first device model information and the target model information; S650. If the consistency is consistent, determine that the consultation database is a local professional knowledge base; S660. If the consistency is inconsistent, determine that the consultation database is an external knowledge base.

[0077] Consistency refers to the judgment result of whether the first equipment model information (equipment model in the second consultation) and the target equipment model information (equipment model in the first consultation, i.e. the extracted model information) are the same model. The consistency judgment criterion is whether the two model strings are the same after brand word standardization (unified alias mapping) and model number fuzzy matching (edit distance not exceeding 2).

[0078] Secondly, after obtaining the first device model information, the fault diagnosis equipment compares it with the target device model information determined in the initial consultation: firstly, it performs brand term standardization on the two model strings (by mapping possible aliases to standard brand names using a preset brand term dictionary, for example, mapping "Brand A", "Server A", and "Antminer" to "Brand A Blockchain Server"), then extracts their respective series words and model serial numbers, and calculates the edit distance between the two model serial numbers. If the edit distance does not exceed 2, the consistency is determined to be consistent, and the consultation database is identified as the local professional knowledge base; if the edit distance exceeds 2, the consistency is determined to be inconsistent, and the consultation database is identified as an external knowledge base.

[0079] Through the above technical solution, this embodiment optimizes the decision-making condition in the database selection for secondary consultation from a retrieval-based judgment of "whether the local knowledge base contains data for this model" to a comparative judgment of "whether the model in the secondary consultation is consistent with the model in the initial consultation." This eliminates the need to repeatedly perform model matching searches when the models are consistent, reducing computational overhead. Using the local knowledge base consistently when the models are consistent ensures the continuity of maintenance knowledge sources for both the initial and secondary consultations, avoiding confusion among maintenance personnel due to inconsistent terminology.

[0080] The core advantage of this embodiment lies in using the device model information confirmed in the initial consultation as prior knowledge for selecting the database in the secondary consultation, compressing the database selection decision from "searching the local knowledge base" to an O(1) level judgment of "comparing two model strings". When encountering the problem of data source selection in multi-round dialogues, those skilled in the art usually execute the data source selection logic independently for each round of dialogue. Their approach is to treat each round of dialogue as an independent request processing without historical state. Even if the device models in the two rounds of dialogue are exactly the same, the complete data source selection process must be executed again. This approach stems from the stateless design inertia in Web service architecture—each HTTP request is processed independently and does not depend on the state of previous requests. However, in the scenario of fault diagnosis equipment deployed on the edge, the maintenance consultation currently being processed by the equipment naturally has a session context (the continuous maintenance process of the same equipment), and the device model state of the initial consultation can be maintained in memory for direct reference in subsequent rounds.

[0081] The present invention also proposes a control device, the control device comprising: a memory, a processor, and a fault repair guidance program stored in the memory and executable on the processor, the fault repair guidance program being configured to implement the fault repair guidance method as described above.

[0082] It is worth noting that, since the control device of the present invention is based on the above-described fault repair guidance method, the embodiments of the control device of the present invention include all the technical solutions of all embodiments of the above-described fault repair guidance method, and the technical effects achieved are exactly the same, so they will not be repeated here.

[0083] The present invention also proposes a fault diagnosis device, which includes the control device as described in the above embodiments.

[0084] It is worth noting that since the fault diagnosis device of the present invention is based on the above-mentioned control device, the embodiments of the fault diagnosis device of the present invention include all the technical solutions of all the embodiments of the above-mentioned control device, and the technical effects achieved are exactly the same, which will not be repeated here.

[0085] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A fault repair guidance method, applied to fault diagnosis equipment, characterized in that, The fault repair guidance method includes: Obtain fault repair consultation requests and determine the target equipment model information and fault description information; Based on the target device model information, a matching search is performed in the local professional knowledge base to obtain target repair data associated with the device model. The target repair data includes repair article excerpts and repair reference pictures. Based on the target maintenance data, a round of consultation prompts is constructed, and the round of consultation prompts is input into the large language model to obtain the initial maintenance guidance results output by the large language model. The initial maintenance guidance results are optimized to generate corresponding maintenance guidance information.

2. The fault repair guidance method as described in claim 1, characterized in that, The process of constructing a round of consultation prompts based on the target maintenance data is as follows: Determine the preset paragraph titles contained in the repair article fragment; Based on the type and arrangement order of the preset paragraph titles, determine the corresponding multiple analysis instruction text segments, and concatenate the multiple analysis instruction text segments in the recognition order to generate a maintenance article analysis and processing instruction segment; Obtain preset constraint requirements information, and determine the word limit, maintenance reference image processing requirements, content exclusion rules, and data source limitation clauses contained in the constraint requirements information; The first round of consultation prompts are generated by following the preset constraint order of the character limit, the requirements for processing the repair reference images, the content exclusion rules, and the data source limitation clauses.

3. The fault repair guidance method as described in claim 1, characterized in that, The step of optimizing the initial repair guidance results to generate corresponding repair guidance information includes: The initial maintenance guidance results are semantically segmented and identified, and the title area, fault phenomenon description area, fault analysis area and solution area are identified based on a preset paragraph semantic tag set. Extract the title text from the identified title region to generate the corresponding title field; Extract the fault phenomenon description text from the fault phenomenon description area to generate the corresponding fault phenomenon description segment; Extract the fault analysis text from the fault analysis area to generate the corresponding fault analysis segment; Extract the solution text from the solution region to generate the corresponding solution segment; The title field, the fault phenomenon description section, the fault analysis section, and the solution section are reorganized according to a preset order to generate corresponding maintenance guidance information.

4. The fault repair guidance method as described in claim 1, characterized in that, The step of optimizing the initial repair guidance results to generate corresponding repair guidance information includes: Extract the title field, fault phenomenon description section, fault analysis section, and solution section from the initial maintenance guidance results; Determine the original text location information of each repair reference image in the target repair data within the repair article fragment; The repair reference image is inserted into the corresponding text content position in the initial repair guidance result according to the original text position information to generate the repair guidance information.

5. The fault repair guidance method as described in claim 1, characterized in that, The step of matching and searching the local professional knowledge base based on the target equipment model information to obtain target maintenance data associated with the equipment model includes: The target device model information is broken down into three levels of keyword units: brand terms, series terms, and model serial number; Using the brand name as the first matching level, retrieve all device models belonging to that brand from the local professional knowledge base; Using the series terms as the second matching level, a first subset of candidate models belonging to the corresponding series is selected from the set of all device models; Using the model number as the third matching level, a unique target model is matched in the first candidate model subset; Extract all repair article fragment indexes and all repair reference image indexes corresponding to the unique target model in the local professional knowledge base, concatenate the original text content corresponding to all the repair article fragment indexes, and associate the image files corresponding to all the repair reference image indexes to generate the target repair data.

6. The fault repair guidance method as described in claim 1, characterized in that, The fault repair guidance method also includes: Obtain a secondary fault repair consultation request, and determine the first equipment model information and secondary consultation information in the secondary fault repair consultation request; Based on the first device model information, a corresponding consultation database is determined, which includes a local professional knowledge base and an external knowledge base; Retrieve maintenance knowledge entries from the consultation database that match the first equipment model information and the secondary consultation information; Construct a second-round consultation prompt phrase corresponding to the maintenance knowledge item. The second-round consultation prompt phrase includes a system role setting section and a secondary output format constraint section.

7. The fault repair guidance method as described in claim 6, characterized in that, The step of determining the corresponding consultation database based on the first device model information includes: Determine whether at least one maintenance knowledge entry exists in the local professional knowledge base that matches the model information of the first device; If at least one maintenance knowledge item exists, the consultation database is determined to be a local professional knowledge base; If no maintenance knowledge entry exists, the consultation database is determined to be an external knowledge base.

8. The fault repair guidance method as described in claim 6, characterized in that, The step of determining the corresponding consultation database based on the first device model information includes: Determine the consistency between the first device model information and the target model information; If the consistency is consistent, the consultation database is determined to be a local professional knowledge base; If the consistency is inconsistent, the consultation database is determined to be an external knowledge base.

9. A control device, characterized in that, The control device includes: a memory, a processor, and a fault repair guidance program stored in the memory and executable on the processor, the fault repair guidance program being configured to implement the fault repair guidance method as described in any one of claims 1 to 8.

10. A fault diagnosis device, characterized in that, Includes the control device as described in claim 9.

Images