An intelligent listing decision method and system for maintenance tasks

Abstract

The present application relates to a kind of intelligent item decision method and system for overhaul task, belong to overhaul management and artificial intelligence application technical field, solve the problem of high artificial dependence of existing technology in overhaul project, defect response lag.Specific steps include: obtaining the overhaul planning information to be audited by user, including overhaul background knowledge text, artificial planning overhaul project text and pre-repair equipment defect description text;Based on the overhaul background knowledge, search in pre-constructed knowledge base, obtain the first overhaul project list, combined with the second overhaul project list obtained by the pre-repair equipment defect description text driven matching, complete overhaul item is obtained;The complete overhaul item is compared with artificial planning overhaul project, obtains the missing overhaul project in the overhaul project text to be audited by user, and generates decision item for the missing overhaul project, realizes the quick, accurate extraction of standard overhaul project and defect-driven project.

Description

Technical Field

[0001] This invention relates to the fields of maintenance management and artificial intelligence application technology, and in particular to an intelligent itemized decision-making method and system for maintenance tasks. Background Technology

[0002] In the life-cycle management of thermal power plant equipment, maintenance project initiation is a crucial step in ensuring the safe operation of the unit and controlling maintenance costs. It requires the comprehensive development of maintenance projects based on equipment condition, historical defects, and standard guidelines. Current technologies primarily rely on manual review or simple keyword-based searches. However, with the increasing refinement of maintenance standards, traditional methods are no longer sufficient to meet engineering requirements, specifically in the following ways:

[0003] First, keyword retrieval cannot understand professional semantics, resulting in a large discrepancy between search results and actual scenarios.

[0004] Second, the differences between professional knowledge systems are significant, and a single model or unified knowledge base is prone to semantic confusion, making it difficult to guarantee the accuracy of cross-professional tasks.

[0005] Third, the lack of a transparent and explainable mechanism for the sources of maintenance projects makes it impossible to distinguish between standard items and defect-based governance items, which is not conducive to auditing and tracing.

[0006] Fourth, traditional natural language processing technology struggles to handle extremely long and hierarchical maintenance standard documents, and long-distance dependencies are easily lost, resulting in incomplete project extraction. Summary of the Invention

[0007] Based on the above analysis, the embodiments of the present invention aim to provide an intelligent itemization decision-making method and system for maintenance tasks, in order to solve the problems of high reliance on manual maintenance item initiation, delayed defect response, and low review efficiency in the prior art.

[0008] The objective of this invention is mainly achieved through the following technical solutions: On one hand, embodiments of the present invention provide an intelligent itemization decision-making method for maintenance tasks, comprising the following steps: Obtain user-submitted maintenance planning information, including maintenance background information text, manually planned maintenance project text, and pre-maintenance equipment defect description text; Based on the maintenance background knowledge, a first maintenance item list is obtained by searching in a pre-built knowledge base. Combined with the second maintenance item list obtained by the pre-maintenance equipment defect description text-driven retrieval, a complete maintenance list is obtained. The complete maintenance list is compared with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user, and a decision list is generated for the missing maintenance items.

[0009] Furthermore, the first maintenance item list is obtained, including: The input maintenance background knowledge text is standardized to obtain structured identifiers for maintenance levels and professional directions. The professional category position of the maintenance level in the pre-set hierarchical maintenance branch structure is determined based on the structured identifier of the maintenance level and the structured identifier of the professional direction, and all equipment types corresponding to the professional category position are taken as the candidate object set; wherein, the hierarchical maintenance branch structure contains three levels of labels in sequence: maintenance level, professional category and equipment type; Multiple query vectors are constructed from the candidate object set, and semantic similarity retrieval is performed in the knowledge base to obtain a preliminary candidate document set. The preliminary candidate documents are matched with the corresponding query vectors in terms of semantics, professionalism, and maintenance level. The top K documents are extracted according to the relevance score to obtain the candidate context set. The candidate context set is input into the large language model, and information extraction and structured transformation are performed under the constraints of the preset prompt word template to obtain the first maintenance item list.

[0010] Furthermore, the omitted maintenance items are obtained, including: The complete maintenance list is semantically matched with the manually planned maintenance items to identify several first missing items; combined with the directed acyclic graph based on equipment type, the first missing items are prioritized and sorted to obtain the omitted maintenance items.

[0011] Furthermore, the knowledge base is pre-built based on the following process: Collected multi-source maintenance technical data, resulting in multiple standard documents; The aforementioned standard documents are categorized and labeled with corresponding maintenance levels, professional categories, and equipment types; The categorized documents are segmented into semantically complete text units according to the labels, and then vectorized using a pre-trained language model to obtain multiple vectorized texts. Using maintenance level as the first-level label, professional category as the second-level label, and equipment type as the third-level label, a hierarchical maintenance branch structure is constructed and the vectorized text is reorganized to obtain the vectorized knowledge base.

[0012] Furthermore, the decision items include the content of the missing maintenance items, the basis for the omission, and the decision action; based on the decision action taken by the user each time, the prompt word template during subsequent searches is adjusted to optimize the process of obtaining the complete maintenance items.

[0013] Furthermore, the decision-making actions include adoption, modification, or deletion. The process involves statistically analyzing the modification or deletion decisions made by users, identifying the differences between the omitted maintenance items corresponding to the decision and the preset prompt word template, and iteratively updating the task role description, output specifications, or keyword matching rules of the prompt word template to optimize the process of obtaining the complete maintenance items.

[0014] Furthermore, the second maintenance item list is obtained, including: Semantic recognition is used to parse the pre-repair equipment defect description text, identify equipment objects and abnormal features, and obtain structured context objects; A query vector is constructed using the structured context object. Semantic similarity retrieval is performed in the knowledge base. The retrieval results within the candidate object set are extracted and structurally transformed to obtain the second maintenance item list.

[0015] Furthermore, the first missing items are prioritized, including: The acyclic graph is constructed by taking each equipment type in the hierarchical maintenance branch structure as a node and the operational dependencies between each equipment type as edges. For missing project nodes, the missing top impact value of the node is obtained by normalization based on the basic weight of the node and the cumulative weight of all its downstream dependent nodes, and then prioritized according to the order of the missing top impact values.

[0016] Furthermore, the maintenance background knowledge text includes a description of the maintenance plan name, a description of the maintenance level, and a description of the professional direction.

[0017] On the other hand, embodiments of the present invention provide an intelligent itemized decision-making system for maintenance tasks, comprising: The human-machine collaborative interaction module is used to input the maintenance planning information to be verified by the user, including maintenance background knowledge text, manually planned maintenance project text, and pre-maintenance equipment defect description text; it also includes a visual review interface to generate decision items for omitted maintenance projects; The structured parsing module is used to parse the maintenance planning information to obtain a structured context object; The intelligent retrieval module is used to retrieve a standard maintenance item list from a pre-built knowledge base based on the structured context object. This list is then combined with a second maintenance item list obtained by the pre-maintenance equipment defect description text-driven retrieval to obtain a complete maintenance list. The item comparison module is used to compare the complete maintenance items with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention proposes to automatically identify and aggregate necessary maintenance items by matching professional branches and retrieving vectorized knowledge bases, combining equipment defect descriptions with professional scope, accurately mapping defect information into executable maintenance tasks, avoiding omissions in the maintenance of occasional faults, improving the targeting and completeness of maintenance, realizing the rapid and accurate extraction of standard maintenance items and defect-driven items, and significantly reducing the workload of manual sorting.

[0019] 2. By comparing standard maintenance items with actual items, the system automatically identifies uncovered tasks and generates a structured missing list. It also uses DAG to identify the importance score of missing items and prioritizes them. Combined with the provided evidence of the missing items, the system enables maintenance items to be traceable and verifiable, thereby improving the level of management refinement.

[0020] 3. By introducing a semantically aligned maintenance project comparison and feedback closed-loop mechanism, continuous self-optimization of the maintenance planning process was achieved, improving the completeness, accuracy, and engineering applicability of the maintenance project list.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0023] Figure 1 This is a flowchart of an intelligent itemized decision-making method for maintenance tasks according to an embodiment of the present invention. Detailed Implementation

[0024] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0025] Example 1 A specific embodiment of the present invention discloses an intelligent itemization decision-making method for maintenance tasks, such as... Figure 1 As shown, it includes the following steps: Step S1: Obtain the maintenance planning information to be reviewed by the user, including maintenance background knowledge text, manually planned maintenance project text, and pre-maintenance equipment defect description text. Step S2: Based on the maintenance background knowledge, search in the pre-built knowledge base to obtain the first maintenance item. Combine the maintenance items retrieved by the pre-maintenance equipment defect description text to obtain the complete maintenance list. Step S3: Compare the complete maintenance list with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user, and generate decision items for the missing maintenance items.

[0026] By using the above method, the search scope is first narrowed quickly by using professional branches as constraints through the background-defect semantic bijective channel. Then, the nearest neighbor search is performed using a vectorized knowledge base. Historical experience standard maintenance items are extracted through background knowledge retrieval, and occasional fault items are extracted by using defect text semantic matching. Finally, a set of maintenance items without redundancy or omission is generated, which improves the efficiency and accuracy of maintenance item listing.

[0027] Specifically, in step S1, the maintenance plan information to be reviewed by the user is obtained, including maintenance background knowledge (such as maintenance plan name, maintenance level and professional direction), text of manually planned maintenance projects and pre-maintenance equipment defect description information.

[0028] The maintenance plan name describes the full name of the entire maintenance plan, the maintenance level describes the maintenance level corresponding to the plan name, and the professional direction describes the professional field to which the maintenance plan belongs. Both the maintenance project text and the defect description are entered in natural language, allowing for non-standard expressions, overlapping content, and semantic overlap.

[0029] It should be noted that before obtaining complete maintenance items by searching and matching the maintenance background knowledge text and the pre-maintenance equipment defect description text, a local knowledge base needs to be pre-built. The specific process is as follows: Multiple sources of maintenance technical data were collected, including maintenance guidelines, historical maintenance records, equipment defect and project mapping data, and manufacturer technical information, resulting in a number of standard documents. The standard documents are categorized according to professional dimensions, the scope of equipment and maintenance levels are clearly defined, and data tags for maintenance level, professional category and equipment type are added; The standard document is segmented into semantically complete text units according to tags using a semantic segmentation strategy, and then vectorized using a pre-trained language model to obtain multiple vectorized texts. Using maintenance level as the first-level label, professional category as the second-level label, and equipment type as the third-level label, a hierarchical maintenance branch structure is constructed, and the vectorized text is reorganized to pre-build the vectorized knowledge base. The hierarchical maintenance branch structure has a nested first-level-second-level-third-level arrangement. In other words, if the maintenance level is level 3, each level will contain a corresponding professional category; similarly, each professional category will contain different equipment types.

[0030] Specifically, in step S2, based on the user input information and the constructed vectorized knowledge base, the complete maintenance items are extracted through the following process: S21. Based on maintenance background knowledge, a search is performed in a pre-built knowledge base to obtain the first maintenance item list. Specifically, S211. By combining domain roles, industry standard semantic background, and maintenance process consistency to limit prompt words, the input maintenance background knowledge text (i.e., maintenance plan name, maintenance level, and professional direction, etc.) is standardized. Different expressions of professional names and maintenance levels are mapped to standardized identifiers to obtain structured identifiers for professional directions and maintenance levels. This is used to eliminate naming differences for the same professional or maintenance level in different input scenarios.

[0031] For example, the pre-set prompt word is "As a thermal power plant maintenance process expert, possessing the following knowledge background and capabilities: proficient in the 'Equipment Maintenance Guidelines for Power Generation Enterprises' and the maintenance compliance requirements of the thermal power industry; familiar with on-site maintenance scenarios in thermal power plants (such as isolation of electrical, steam, water, and flue gas systems, equipment disassembly and assembly, commissioning monitoring, etc.); concise and professional language style, output content conforming to the on-site operation instruction writing specifications, without redundant expressions, etc." When maintenance personnel input different expressions such as "thermal control," "thermal engineering control," or "DCS," under the constraint of this prompt word, they are uniformly mapped to the same standard professional identifier; when inputting "A-level maintenance," "Level 1 maintenance," or "major overhaul," they are uniformly mapped to the same standard maintenance level identifier.

[0032] Through the above process, the deterministic fields in maintenance planning information can be rapidly standardized without relying on real-time semantic reasoning, thereby ensuring the consistency and stability of subsequent processing flows.

[0033] S212. Using the structured identifiers of professional directions and maintenance levels as search scope constraints, a search is conducted in the pre-constructed vectorized maintenance knowledge base to obtain a set of candidate contexts. Specifically, based on the structured identifiers of the maintenance level and the structured identifiers of the professional direction, the professional category position in the pre-set hierarchical maintenance branch structure is determined, and all equipment types corresponding to the professional category position are used as a candidate object set; multiple query vectors are constructed using the candidate object set, and semantic similarity retrieval is performed in the knowledge base to obtain a preliminary candidate document set; the preliminary candidate documents and their corresponding query vectors are matched for relevance from the dimensions of semantics, system, profession, and maintenance level, respectively, and the top K documents are extracted according to the relevance score to obtain the first candidate context set.

[0034] For example, a user inputs a maintenance plan name as "Maintenance of Unit 2," a maintenance level of A, and a professional category of "Steam Turbine." After determining the corresponding search branch based on the user input, equipment categories related to the steam turbine profession are selected as the candidate search object set, including but not limited to steam turbines, feedwater pump turbines, main steam valves, speed regulating valves, high-pressure heaters, condensers, and other equipment systems. Each equipment system is used as a search vector, and semantic similarity is calculated with the vectorized text in the knowledge base to obtain a preliminary candidate document set with a similarity greater than a preset threshold. Subsequently, a re-ranking model is introduced to refine the preliminary candidate document set. The re-ranking model takes the query vector and the original text content of the candidate documents as input, and comprehensively evaluates the relevance of the candidate documents from dimensions such as semantic matching degree, professional consistency, and maintenance level suitability, generating a refined relevance score to correct any ranking bias that may arise from simply calculating based on vector similarity. Based on the re-ranking results, the top K documents with the highest scores are selected as the candidate context set for the current equipment system. The candidate contexts contain only standardized text content from the unified maintenance knowledge base and do not contain any non-standard or speculative information.

[0035] S213. Input the candidate context set into the large language model for analysis, and extract the first maintenance item list that meets the maintenance level. Here, a pre-set prompt word template for the large language model is used to constrain and control the task role, information source, and output format of the large language model, thereby limiting the reasoning boundaries and output specifications when performing the maintenance item extraction task.

[0036] For example, a matching prompt word template is loaded based on the current professional direction. The prompt word template predefines the large language model's role as an information extraction character, specifying that it can only perform maintenance project extraction tasks based on the retrieval context and the unified maintenance knowledge base content, and must not introduce any external knowledge or subjective inference information. Simultaneously, the applicable professional scope and maintenance level conditions of the model are limited to avoid erroneous extraction across professional or level boundaries. The output format is structurally constrained, explicitly requiring the model to output only the names of standard maintenance projects that meet the maintenance level requirements and their corresponding categories, and to output them numbered according to a predefined data structure to ensure consistency and parsability of the results. Under the constraints of the above prompt word template, the large language model performs semantic analysis and filtering on the candidate context set, automatically extracting standard maintenance projects that meet the current maintenance planning conditions, and returning them in a structured result format. Subsequently, the model output results are deduplicated, sorted, and validated for completeness, ultimately forming a standardized first maintenance project list.

[0037] By combining professional branch constraints, structured retrieval context, and rule-based loading of prompt word templates, a large language model extraction mechanism was developed to achieve controlled automatic extraction of maintenance projects. This reduced the uncertainty caused by free model generation and improved the accuracy, standardization, and consistency of the generated standard maintenance project results.

[0038] S22. Based on the text-driven retrieval of pre-repair equipment defect descriptions, obtain the second maintenance item list. Specifically, S221. A text parsing method based on semantic recognition performs semantic analysis on the unstructured natural language content of the pre-repair equipment defect description text, automatically identifies the equipment objects, abnormal phenomena and abnormal parts involved, and obtains a structured context object.

[0039] For example, when maintenance personnel input "#1 feedwater pump vibrates excessively, accompanied by abnormal noise during operation" or "feedwater pump operates unsteadily with obvious shaking" in the pre-repair defect description, semantic recognition can uniformly identify different expressions as the same equipment object "#1 feedwater pump" and classify them into the standard defect type of "abnormal vibration". This method effectively eliminates the impact of differences in natural language descriptions on the understanding of maintenance projects.

[0040] S222. Construct a query vector using a structured context object, perform semantic similarity retrieval in the knowledge base, and extract the retrieval results within the candidate object set to obtain the second maintenance item list.

[0041] Specifically, based on the structured identifier of the professional direction, the equipment objects from which defects are extracted are constrained by professional branches, limiting them to the equipment category range under that profession. After determining the equipment affiliation, the equipment object and abnormal characteristics (abnormal phenomena and abnormal locations) are used as joint search conditions to perform vector semantic search in the knowledge base, recalling maintenance guidelines and historical maintenance case texts that are highly related to the current defect semantics, and using them as the defect reference context for subsequent reasoning and analysis from defect to maintenance project.

[0042] The defect reference context is input into the large language model. Following the same analysis method as step S213, a prompt word template matching the current professional direction is loaded to strictly limit the model's reasoning scope. Under the constraint of the prompt word template, the large language model performs joint reasoning on the defect semantics and reference context to determine the name of the maintenance project required to eliminate or handle the defect, and directly extracts the corresponding maintenance project from the knowledge base. If a single defect involves multiple maintenance projects, the model only selects the two to three maintenance projects with the highest correlation to the defect for output; if multiple defects semantically point to the same maintenance intention, the model merges duplicate or highly similar maintenance projects through semantic consistency judgment.

[0043] Finally, following the predefined output specifications in the prompts, structured results are generated solely in the form of maintenance item names, and a defect-driven maintenance item list, i.e., the second maintenance item list, is output in standard JSON data format. This method achieves automatic semantic mapping between defect descriptions and maintenance items, avoiding manual reliance on experience to supplement maintenance content and improving the relevance of maintenance plans to actual equipment operation problems.

[0044] S23. Merge the first and second maintenance item lists to obtain the complete maintenance items. To avoid generating duplicate maintenance items, semantic matching is performed between the first and second maintenance item lists. Using a pre-set coverage threshold, when the similarity exceeds this threshold, items that are semantically identical or express different meanings but point to the same maintenance intent are identified and merged, providing a basis for subsequent comparison with manual input.

[0045] Specifically, in step S3, based on the comparative analysis results of the complete maintenance items and the manually planned maintenance items, the missing maintenance items in the maintenance item text to be reviewed by the user are obtained, and a decision list is generated for the missing maintenance items, including the content of the missing maintenance items, the basis for the omission, and the decision action; the user can make a decision action to adopt, modify or delete the missing maintenance items based on the basis for the omission.

[0046] The process of comparing and analyzing the complete maintenance list with the manually planned maintenance items includes: performing semantic matching between the complete maintenance list and the manually planned maintenance items to identify several first missing items; and combining the directed acyclic graph based on equipment type to prioritize each first missing item, thereby obtaining the omitted maintenance items.

[0047] In this process, similar to the aforementioned coverage threshold, the similarity assessment of each item in the overall maintenance list and each item in the manually planned maintenance projects is performed to identify any missing maintenance items in the user's pending maintenance project text. Furthermore, by combining the directed acyclic graph (DAG) structure of equipment-system-component, the upstream and downstream dependencies of each missing (absent) item at the equipment level are analyzed, and its importance score to the overall availability is automatically calculated. This importance score is used as the priority ranking criterion for missing items, providing a quantitative reference for maintenance personnel in manual decision-making, ensuring the organic integration of the generated results and manually input maintenance plans, and improving the completeness and executability of the maintenance plan.

[0048] Specifically, a directed acyclic graph (DAG) of equipment, systems, and components is constructed based on the equipment types in the knowledge base. In this DAG, each node represents a specific maintenance object, including the entire machine or equipment system (e.g., steam turbine system, feedwater pump system), subsystems or key devices (e.g., main steam valve, speed regulating valve), and specific components (e.g., valve core, bearing, blades). Directed edges between nodes represent upstream and downstream dependencies; that is, the completion of maintenance on a node directly affects the function of downstream nodes or the overall machine availability, thus forming a machine-level DAG, ensuring that dependencies can be topologically ordered and are acyclic. Based on the constructed DAG, a leak importance analysis is performed on each missing item to quantify its potential impact on the overall machine availability. Specifically, firstly, based on the topological relationships of the DAG, the downstream dependency order of the missing item node is determined, the cumulative weight of all its downstream dependent nodes is calculated, and combined with the basic weight of the missing item node itself, the preliminary leak impact value of that node is obtained, expressed as: , in, It is the base weight of the current node. This represents all direct or indirect downstream nodes in the DAG. For the basic weights, core equipment or systems have higher weights, while auxiliary or non-critical components have lower weights. The initial leak impact values ​​are normalized to the range [0,1] to generate the final leak importance score. This score reflects the criticality of the missing item in the equipment-system-component structure; a higher score indicates a greater impact of the missing item on the overall availability.

[0049] After completing the project comparison and leak calculation, the comparison results are presented to maintenance personnel in a visual format. In the visualization interface, missing items are sorted from highest to lowest importance score, with high-impact items highlighted to remind maintenance personnel to prioritize them. The basis for each missing item is also provided, indicating the corresponding clause or defect causing the missing item, thus providing a complete chain of evidence for traceability. Maintenance personnel can make decisions on missing items with a single click, manually adjust, or delete them. By combining the leak impact score, maintenance personnel can scientifically select priority maintenance items while considering equipment availability risks, achieving a synergy between manual experience and quantitative analysis, thereby improving the rationality and effectiveness of the maintenance plan.

[0050] Furthermore, each user's decision-making action can be used as feedback data for closed-loop optimization to optimize the process of obtaining a complete maintenance project. This involves adjusting the prompt word template configuration and the inference weights of the large language model based on the feedback data, making the model more accurate in generating subsequent maintenance projects.

[0051] Specifically, for the prompt word template, the system analyzes the types of decisions that users frequently modify or delete, identifies the differences between the omitted maintenance items corresponding to those decisions and the pre-set prompt word template, i.e., identifies parts of the template where restrictions are not precise enough or key instructions are not fully reflected. It then iteratively updates the task role description, output specifications, or keyword matching rules of the prompt word template to optimize the process of obtaining the complete maintenance item. For example, if the model frequently generates maintenance items for non-standard components in the steam turbine field, the system can add a clear instruction to the template stating "limited to steam turbine, feedwater pump, valve, and other equipment systems" to reduce out-of-bounds output.

[0052] For the inference weights of the large language model, the system calculates the acceptance rate of each project attribute (such as project category, professional direction, and defect type) in human feedback and converts this data into weight adjustment factors. For example, if a certain type of project has a low human acceptance rate, the inference weight of the model when generating similar projects will be reduced; conversely, the weight of categories with high acceptance rates will be appropriately increased to encourage the model to output validated, highly relevant projects. Through iterative training, the model will favor project types with high acceptance rates in subsequent generation tasks, reducing erroneous or low-relevance projects.

[0053] For example, suppose that in the generation of defect-driven maintenance items, the system finds that items related to "high-pressure heater pipeline leakage" are frequently manually deleted, while items related to "main steam valve regulating mechanism maintenance" have a high adoption rate. The system will adjust the prompt word template through feedback, emphasizing that maintenance items are extracted only within the scope of "critical steam system components," while reducing the weight of items like "high-pressure heater pipeline leakage" in model inference. As a result, in the next generation process, the maintenance items output by the model are closer to the standards accepted by humans, reducing duplicate or irrelevant items and improving overall accuracy and practicality.

[0054] Furthermore, the system can dynamically update the weight parameters of the equipment-system-component DAG and the leak calculation logic to reflect actual operating experience and engineering adjustments. Through the above closed-loop optimization, the system can continuously improve the generation of defect-driven maintenance items, the prioritization of leaks, and the manual comparison process during actual use, thereby enhancing the completeness, accuracy, and engineering applicability of standard maintenance item lists, while reducing reliance on manual experience and improving the scientific nature and feasibility of maintenance plans.

[0055] Compared with existing technologies, this embodiment provides an intelligent itemization decision-making method for maintenance tasks. Through a dual-pathway semantic analysis of specialties and defects, it first rapidly narrows the search scope by using specialty branches as constraints, then performs nearest neighbor search using a vectorized knowledge base. Simultaneously, it automatically clusters semantic defect content to generate a set of maintenance items without redundancy or omissions. On one hand, by introducing semantically aligned maintenance item comparison, it automatically identifies uncovered tasks and generates a structured missing list. Combining this with a Directed Acyclic Graph (DAG) to identify the importance score of missing items, it prioritizes them, providing a scientific quantitative reference for manual decision-making. On the other hand, it utilizes a feedback loop mechanism to continuously self-optimize the maintenance planning process, ultimately improving the completeness, accuracy, and engineering practicality of the maintenance item list.

[0056] Example 2 Another specific embodiment of the present invention discloses an intelligent itemized decision-making system for maintenance tasks, comprising: The human-machine collaborative interaction module is used to input the maintenance planning information to be verified by the user, including maintenance background knowledge text, manually planned maintenance project text, and pre-maintenance equipment defect description text; it also includes a visual review interface to generate decision items for omitted maintenance projects; The structured parsing module is used to parse the maintenance planning information to obtain a structured context object; The intelligent retrieval module is used to retrieve standard maintenance items from a pre-built knowledge base based on the structured context object. The maintenance items are then combined with the maintenance items matched by the pre-maintenance equipment defect description text to obtain a complete maintenance list. The item comparison module is used to compare the complete maintenance items with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user.

[0057] The system can perform intelligent itemization decisions for maintenance tasks according to any of the methods described in Embodiment 1. Related aspects can be referenced from each other, but are not repeated in this embodiment.

[0058] Compared with the prior art, the intelligent item listing and decision-making system for maintenance tasks provided in this embodiment, through the linkage and cooperation of various modules, lists the maintenance items and decision actions (which can be accepted or rejected) that users are concerned about, so as to facilitate review and improve the completeness, accuracy and engineering applicability of maintenance item listing.

[0059] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent itemization decision-making method for maintenance tasks, characterized in that, Includes the following steps: Obtain user-submitted maintenance planning information, including maintenance background information text, manually planned maintenance project text, and pre-maintenance equipment defect description text; Based on the maintenance background knowledge, a first maintenance item list is obtained by searching in a pre-built knowledge base. Combined with the second maintenance item list obtained by the pre-maintenance equipment defect description text-driven retrieval, a complete maintenance list is obtained. The complete maintenance list is compared with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user, and a decision list is generated for the missing maintenance items.

2. The method according to claim 1, characterized in that, The first maintenance item list is obtained, including: The input maintenance background knowledge text is standardized to obtain structured identifiers for maintenance levels and professional directions. The professional category position of the maintenance level in the pre-set hierarchical maintenance branch structure is determined based on the structured identifier of the maintenance level and the structured identifier of the professional direction, and all equipment types corresponding to the professional category position are taken as the candidate object set; wherein, the hierarchical maintenance branch structure contains three levels of labels in sequence: maintenance level, professional category and equipment type; Multiple query vectors are constructed from the candidate object set, and semantic similarity retrieval is performed in the knowledge base to obtain a preliminary candidate document set. The preliminary candidate documents are matched with the corresponding query vectors in terms of semantics, professionalism, and maintenance level. The top K documents are extracted according to the relevance score to obtain the candidate context set. The candidate context set is input into the large language model, and information extraction and structured transformation are performed under the constraints of the preset prompt word template to obtain the first maintenance item list.

3. The method according to claim 2, characterized in that, The omitted maintenance items are as follows: The complete maintenance list is semantically matched with the manually planned maintenance items to identify several first missing items; combined with the directed acyclic graph based on equipment type, the first missing items are prioritized and sorted to obtain the omitted maintenance items.

4. The method according to claim 3, characterized in that, The knowledge base is pre-built based on the following process: Collected multi-source maintenance technical data, resulting in multiple standard documents; The aforementioned standard documents are categorized and labeled with corresponding maintenance levels, professional categories, and equipment types; The categorized documents are segmented into semantically complete text units according to the labels, and then vectorized using a pre-trained language model to obtain multiple vectorized texts. Using maintenance level as the first-level label, professional category as the second-level label, and equipment type as the third-level label, a hierarchical maintenance branch structure is constructed and the vectorized text is reorganized to obtain the vectorized knowledge base.

5. The method according to any one of claims 1-4, characterized in that, The decision items include the content of the missing maintenance items, the basis for the omission, and the decision action; based on the decision action taken by the user each time, the prompt word template during subsequent searches is adjusted to optimize the process of obtaining the complete maintenance items.

6. The method according to claim 5, characterized in that, The decision-making actions include adoption, modification, or deletion. The process involves statistically analyzing the modification or deletion decisions made by users, identifying the differences between the missing maintenance items corresponding to the decision and the preset prompt word template, and iteratively updating the task role description, output specifications, or keyword matching rules of the prompt word template to optimize the process of obtaining the complete maintenance items.

7. The method according to claim 2, characterized in that, The second maintenance item list is obtained, including: Semantic recognition is used to parse the pre-repair equipment defect description text, identify equipment objects and abnormal features, and obtain structured context objects; A query vector is constructed using the structured context object. Semantic similarity retrieval is performed in the knowledge base. The retrieval results within the candidate object set are extracted and structurally transformed to obtain the second maintenance item list.

8. The method according to claim 3, characterized in that, Prioritize each of the first missing items, including: The acyclic graph is constructed by taking each equipment type in the hierarchical maintenance branch structure as a node and the operational dependencies between each equipment type as edges. For missing project nodes, the missing top impact value of the node is obtained by normalization based on the basic weight of the node and the cumulative weight of all its downstream dependent nodes, and then prioritized according to the order of the missing top impact values.

9. The method according to claim 1, characterized in that, The maintenance background information text includes a description of the maintenance plan name, a description of the maintenance level, and a description of the professional direction.

10. An intelligent itemized decision-making system for maintenance tasks, characterized in that, include: The human-machine collaborative interaction module is used to input the maintenance planning information to be verified by the user, including maintenance background knowledge text, manually planned maintenance project text, and pre-maintenance equipment defect description text; it also includes a visual review interface to generate decision items for omitted maintenance projects; The structured parsing module is used to parse the maintenance planning information to obtain a structured context object; The intelligent retrieval module is used to retrieve a standard maintenance item list from a pre-built knowledge base based on the structured context object. This list is then combined with a second maintenance item list obtained by the pre-maintenance equipment defect description text-driven retrieval to obtain a complete maintenance list. The item comparison module is used to compare the complete maintenance items with the manually planned maintenance items to obtain the missing maintenance items in the maintenance item text to be reviewed by the user.

Images