Maintenance work order generation method and apparatus, electronic device, and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and provides a maintenance work order generation method and device, electronic equipment and a storage medium, the method comprising the following steps: according to the semantic similarity between a fault description text of a target device and historical fault description texts of each historical maintenance work order, and the hierarchical structure distance between a function position code and historical position codes of each historical maintenance work order, a reference work order set is obtained from each historical maintenance work order; the fault description text, the function position code and the reference work order set are input into a work order generation model, and a maintenance work order of the target device is generated. The application utilizes a two-dimensional accurate screening strategy, realizes efficient reuse of historical maintenance experience, and implements logical blocking and path re-planning in the generation link, so that the work order meets the requirements of complex fault logic and safety interlocking.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for generating maintenance work orders. Background Technology

[0002] Maintenance work orders are the core carrier of nuclear power plant equipment maintenance. Their generation quality and efficiency directly affect the safety of maintenance work and are crucial to ensuring the safe and stable operation of nuclear power plants.

[0003] Currently, maintenance work orders are mostly generated by calling a General Large Language Model (LLM). However, this method mainly relies on semantic similarity of text for matching, which can easily confuse work orders with similar descriptions but belonging to different system levels or different safety levels, resulting in inapplicable reference information. In addition, when generating work order content, the LLM cannot guarantee that the output strictly conforms to the complex fault logic and safety interlocking conditions in the nuclear power field. It is prone to generating seemingly reasonable but actually erroneous instructions that violate the equipment operating logic, and lacks an effective real-time interception mechanism, posing potential safety risks to nuclear power maintenance. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for generating maintenance work orders, in order to overcome the deficiencies existing in the prior art.

[0005] This invention provides a method for generating maintenance work orders, comprising the following steps:

[0006] Obtain the fault description text of the target device, as well as the functional location code used to characterize the position of the target device in the device hierarchy;

[0007] Based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order, a set of reference work orders is obtained from each historical maintenance work order.

[0008] The fault description text, the functional location code, and the reference work order set are input into the work order generation model. The draft maintenance work order generated in real time by the work order generation model is logically verified by a preset fault logic graph. If the logical verification fails, the corresponding draft maintenance work order is corrected until the corrected draft maintenance work order passes the logical verification. The draft maintenance work order that passes the logical verification is then used as the maintenance work order for the target equipment.

[0009] According to a maintenance work order generation method provided by the present invention, the step of determining the hierarchical distance includes:

[0010] The functional location code and the historical location code are respectively mapped to nodes on a tree topology structure, and the tree topology structure is used to characterize the association relationship between each functional location in the device hierarchy structure;

[0011] Calculate the shortest path length between the node corresponding to the functional location code and the node corresponding to the historical location code in the tree topology, and use the shortest path length as the hierarchical distance.

[0012] According to a maintenance work order generation method provided by the present invention, the steps for constructing the tree-like topology include:

[0013] Obtain the full set of codes containing the functional location codes of all known devices in the device hierarchy;

[0014] Based on the preset hierarchical parsing rules, the functional position codes in the full encoding set are decomposed into multiple character segments with parent-child hierarchical relationships;

[0015] The tree-like topology is constructed based on the multiple character fragments, where each node in the tree-like topology corresponds to one of the character fragments, and each node is used to represent a hierarchical unit in the device hierarchy.

[0016] According to a maintenance work order generation method provided by the present invention, the step of filtering a reference work order set from historical maintenance work orders based on the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order, includes:

[0017] The semantic similarity and hierarchical distance of each historical maintenance work order are weighted and summed to obtain a comprehensive matching score that represents the overall matching degree between the target equipment and the corresponding historical maintenance work order.

[0018] The historical maintenance work orders are sorted in descending order of their comprehensive matching scores, and a preset number of the top-ranked historical maintenance work orders are selected to form the reference work order set.

[0019] According to a maintenance work order generation method provided by the present invention, the preset fault logic diagram includes a fault tree model, wherein the fault tree model includes equipment fault events as top events, and maintenance operation actions and equipment operating status as bottom events;

[0020] The steps of the logic verification include:

[0021] The draft maintenance work order is parsed, the currently generated maintenance procedures are extracted, and the target operation actions involved in the maintenance procedures are identified.

[0022] In the fault tree model, a reverse traversal is performed starting from the top event to detect whether the target operation is on the logical path that led to the occurrence of the top event.

[0023] The logical verification of the draft maintenance work order is determined based on the test results.

[0024] According to a maintenance work order generation method provided by the present invention, the step of revising the corresponding maintenance work order draft when the logic verification fails includes:

[0025] If the logic check fails, a mask blocking signal is sent to the work order generation model to block the output content corresponding to the target operation action.

[0026] The work order generation model is driven to regenerate alternative maintenance procedures that do not include the target operation action;

[0027] By replacing the corresponding maintenance procedures in the draft maintenance work order with the alternative maintenance procedures, a revised draft maintenance work order is obtained.

[0028] According to a maintenance work order generation method provided by the present invention, the step of determining the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order includes:

[0029] The fault description text is segmented to obtain keywords;

[0030] Based on the index mapping relationship between keywords and historical maintenance work orders, candidate historical maintenance work orders containing the keywords are retrieved from each historical maintenance work order;

[0031] The fault description text and the historical fault description text of the candidate historical maintenance work order are converted into semantic vectors, and the cosine similarity between the semantic vectors is calculated as the semantic similarity.

[0032] According to a maintenance work order generation method provided by the present invention, after obtaining the fault description text of the target equipment, the method further includes:

[0033] The fault description text is semantically parsed based on a preset high-risk sensitive word library, and the risk entropy value of the fault description text is calculated.

[0034] If the risk entropy value exceeds the emergency threshold, the currently executing generation task with a corresponding risk entropy value lower than the emergency threshold is suspended, and the computing resources released by the suspended generation task are used to execute a maintenance work order generation task for the target device.

[0035] According to a maintenance work order generation method provided by the present invention, the step of inputting the fault description text, the functional location code, and the reference work order set into the work order generation model includes:

[0036] Acquire domain standard specifications and security constraint rules;

[0037] Based on a pre-set structured template, the domain standard specification data, the safety constraint rules, the reference work order set, and the fault description text are integrated to construct prompt words that contain specification constraints and contextual information;

[0038] Input the prompt word into the work order generation model.

[0039] According to a maintenance work order generation method provided by the present invention, after using the draft maintenance work order that has passed the logical verification as the maintenance work order for the target equipment, the method further includes:

[0040] Obtain modification operation data for the maintenance work order, and determine the modification type by comparing the modified content corresponding to the modification operation data with the original content of the maintenance work order draft;

[0041] When the modification type indication only involves changes in device parameter values, the entity attributes in the knowledge graph are updated using the modified device parameter values.

[0042] When the modification type indicates a change in the order of maintenance procedures, the structured template used to guide the generation of the work order generation model is updated using the modified maintenance procedures.

[0043] The present invention also provides a maintenance work order generation device, comprising the following modules:

[0044] The acquisition module is used to acquire the fault description text of the target device and the functional location code used to characterize the position of the target device in the device hierarchy structure.

[0045] The filtering module is used to filter a set of reference work orders from each historical maintenance work order based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order.

[0046] The generation module is used to input the fault description text, the functional location code, and the reference work order set into the work order generation model, and to perform logical verification on the maintenance work order draft generated in real time by the work order generation model through a preset fault logic graph. If the logical verification fails, the corresponding maintenance work order draft is corrected until the corrected maintenance work order draft passes the logical verification. The maintenance work order draft that passes the logical verification is then used as the maintenance work order for the target equipment.

[0047] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the maintenance work order generation method as described above.

[0048] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the maintenance work order generation method as described above.

[0049] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the maintenance work order generation method as described above.

[0050] The maintenance work order generation method, apparatus, electronic device, and storage medium provided by this invention acquire the fault description text and functional location code of the target device, and comprehensively utilize the semantic similarity of the fault description text and the hierarchical distance of the functional location code to accurately select a set of reference work orders from historical maintenance work orders in a two-dimensional manner. This achieves efficient reuse and accurate matching of historical maintenance experience, effectively avoiding erroneous references across systems and security levels caused by relying solely on semantic matching. Furthermore, a work order generation model is used to generate maintenance work orders by combining the input fault information, location information, and the set of reference work orders. During the generation process, a logic verification mechanism based on the fault logic graph is introduced in real time to conduct strict compliance review and dynamic correction of the generated draft maintenance work orders. Because logic blocking and path replanning are implemented in the generation stage, it is ensured that the final generated maintenance work orders not only fully draw on historical best practices but also strictly comply with the complex fault logic and safety interlocking requirements of the domain. This effectively avoids the illusion phenomenon and illegal operation instructions that may be generated by generative models, significantly improving the accuracy, compliance, and security of maintenance work order generation. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the maintenance work order generation method provided by the present invention.

[0053] Figure 2 This is a flowchart illustrating another maintenance work order generation method provided by the present invention.

[0054] Figure 3 This is a schematic diagram of the maintenance work order generation device provided by the present invention.

[0055] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0057] Maintenance work orders, as the core carrier of nuclear power plant equipment maintenance, record key information such as equipment defects, maintenance procedures, and risk control measures. Their generation quality and efficiency directly affect the safety and timeliness of maintenance work. Currently, the generation of nuclear power plant maintenance work orders faces numerous technical challenges, such as difficulties in integrating professional knowledge. Traditional systems struggle to effectively integrate equipment functional location codes, maintenance process standards, and risk control requirements, leading to high error rates in manual entry and low efficiency in human-machine collaboration. Engineers must repeatedly switch between multiple systems, which is time-consuming and relies heavily on personal experience. Furthermore, when faced with batch tasks, traditional sequential processing lacks dynamic allocation capabilities, resulting in slow system response.

[0058] To address the aforementioned challenges, traditional maintenance work order generation solutions mainly include rule engine solutions, template matching solutions, traditional machine learning solutions, general large model solutions, and workflow engine solutions. However, rule engine solutions rely on preset rules, cannot handle unstructured text, and have poor generalization capabilities; template matching solutions lack semantic understanding and require extensive manual modification; traditional machine learning solutions are limited to classification and cannot generate complete work orders; while general large model solutions have strong generation capabilities, they lack deep integration of nuclear power expertise, have a high error rate in understanding professional content such as functional location coding, and pose data security risks; and workflow engine solutions are difficult to adapt to the heterogeneity of work orders.

[0059] Furthermore, the aforementioned solutions often rely solely on semantic matching to retrieve historical work orders, which can easily confuse work orders with similar descriptions but belonging to different system levels or safety grades, rendering the reference information inapplicable. Simultaneously, the generated content lacks strict constraints on the complex fault logic and safety interlocking conditions in the nuclear power field, potentially generating seemingly reasonable but actually illogical instructions, and lacking real-time interception and correction mechanisms, thus posing potential safety risks to nuclear power plant maintenance.

[0060] To address this, the present invention provides a method for generating maintenance work orders. This method aims to achieve precise two-dimensional filtering of historical work orders by calculating the semantic similarity of fault description text and the hierarchical distance of functional location codes in the equipment hierarchy. The method then uses a work order generation model combined with a retrieved set of reference work orders to generate maintenance work orders. During the generation process, a fault logic graph is introduced in real time for logical verification and dynamic blocking, driving the model to correct the generation path. This achieves intelligent and compliant generation of maintenance work orders that takes into account both historical experience references and strict logical constraints.

[0061] It should be noted that the method provided by this invention can be applied to the generation of maintenance work orders for nuclear power plants, equipment overhaul work orders for chemical plants, and maintenance work orders for production equipment in large manufacturing enterprises. For ease of explanation, the following embodiments are all illustrated using the generation of maintenance work orders for nuclear power plants as an example.

[0062] in, Figure 1 This is a flowchart illustrating the maintenance work order generation method provided by the present invention, as shown below. Figure 1 As shown, the method includes steps 110, 120 and 130.

[0063] Step 110: Obtain the fault description text of the target device, and the functional location code used to characterize the position of the target device in the device hierarchy.

[0064] In this embodiment, the target equipment refers to the nuclear power plant equipment for which a maintenance work order is to be generated. For example, the main feedwater pump of Unit 1, the safety injection pump of Unit 2, or a specific type of valve. The fault description text can be understood as a natural language description of the fault phenomenon, location, and severity of the fault in the target equipment. For example, "The temperature of bearing A of the main feedwater pump of Unit 1 rises abnormally, accompanied by abnormal noise," or "The auxiliary feedwater system valve cannot be completely closed, and there is internal leakage." The functional location code is used to uniquely identify the position of the target equipment in the equipment hierarchy. This code usually follows specific industry standards or enterprise specifications. The equipment hierarchy refers to a hierarchical system composed of various types of equipment and their connections, such as the hierarchy from plant area, unit, system, subsystem, equipment to component. The functional location code can accurately reflect the system to which the target equipment belongs, its functional location, and its logical relationship with other equipment.

[0065] As an optional implementation, fault description text and functional location codes can be obtained by receiving a work order generation request input by the user. For example, when nuclear power plant operators or inspectors discover an equipment fault, they can input the fault description text through the terminal interface of the maintenance management system and select or scan the functional location code of the target equipment. Alternatively, the fault description text can be obtained through an automated monitoring system. For instance, when a sensor detects abnormal equipment parameters, it automatically triggers a work order generation request and converts the abnormal information into fault description text, along with the equipment's functional location code.

[0066] Step 120: Based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the function location code and the historical location code of each historical maintenance work order, a set of reference work orders is obtained from each historical maintenance work order.

[0067] In this embodiment, historical maintenance work orders refer to work order records that have completed maintenance processes and been archived, stored in a historical work order database. Each historical maintenance work order typically includes information such as historical fault description text, historical location code, maintenance measures, tools and spare parts used, and risk level. Semantic similarity is used to measure the semantic closeness between the fault description text of the target equipment and the fault description text in the historical maintenance work order. The higher the semantic similarity, the more similar the fault phenomena described by the two are. Hierarchical distance is used to measure the spatial or logical distance between the functional location code of the target equipment and the historical location code in the historical maintenance work order in the equipment hierarchy. The smaller the hierarchical distance, the closer the two are in the equipment hierarchy, such as belonging to the same system, the same unit, or having similar functional attributes. The reference work order set refers to several historical maintenance work orders selected from a massive number of historical maintenance work orders that are most relevant to the target equipment and the current fault.

[0068] Specifically, semantic similarity can be calculated using natural language processing techniques, such as converting text into semantic vectors and calculating the cosine similarity between vectors. Hierarchical distance can be calculated based on the topological relationships of the device hierarchy, by determining the shortest path length between two nodes in a tree structure.

[0069] Filtering historical work orders solely based on semantic similarity might retrieve work orders from equipment at different system levels or safety grades, even if the fault descriptions are similar. This could lead to misapplying maintenance experience from conventional island equipment to nuclear island equipment, resulting in inapplicable or even misleading information. Filtering solely based on hierarchical distance might retrieve work orders located close to each other but with completely different fault types, such as different fault modes of the same pump, which would also lack reference value.

[0070] Therefore, in order to accurately locate the most valuable historical maintenance work orders, this embodiment combines semantic similarity and hierarchical distance for filtering. By comprehensively considering the similarity of fault phenomena and the proximity of equipment locations, erroneous referencing across systems and security levels can be effectively avoided, ensuring that the selected reference work orders are similar in both fault type and equipment attributes.

[0071] As an optional implementation, a preliminary screening based on semantic similarity can be performed to obtain a set of candidate work orders, followed by a secondary screening based on hierarchical distance within the candidate work order set. Alternatively, a comprehensive scoring model can be constructed to weight and fuse semantic similarity and hierarchical distance, calculate a comprehensive score for each historical maintenance work order, and then sort and filter them according to their scores.

[0072] Step 130: Input the fault description text, function location code, and reference work order set into the work order generation model. Logically verify the maintenance work order draft generated in real time by the work order generation model through the preset fault logic diagram. If the logical verification fails, revise the corresponding maintenance work order draft until the revised maintenance work order draft passes the logical verification. Then, use the maintenance work order draft that passes the logical verification as the maintenance work order for the target equipment.

[0073] In this embodiment, the work order generation model refers to a trained generative artificial intelligence model capable of generating maintenance work order content based on input information, such as a large language model based on the Transformer architecture. This model can understand the input fault description text, functional location codes, and reference work order sets, and generate a draft maintenance work order containing maintenance procedures, safety measures, required spare parts, etc., based on this contextual information.

[0074] A fault logic graph is a pre-built, structured knowledge base used to describe the logical relationships between equipment failures, maintenance operations, and system states. For example, a fault logic graph can include a fault tree model, where the top event is equipment failure and the bottom events are specific maintenance operations or equipment states. Logical verification refers to using the fault logic graph to check the compliance and security of the generated draft maintenance work orders.

[0075] Specifically, the maintenance procedures output by the work order generation model can be parsed in real time and mapped to the fault logic graph to determine whether the currently generated maintenance operation violates safety rules, will lead to more serious faults, or meets the current system's operating state constraints, such as interlocking logic.

[0076] During the process of generating maintenance work orders, the work order generation model does not wait for the entire work order to be generated before verification; instead, it employs a real-time verification strategy. If a logical verification fails, such as when the model generates an operation step that violates safety interlocking logic, a blocking mechanism is immediately triggered to prevent the output of the erroneous content and drive the work order generation model to replan its generation path. This replanning of the generation path can be understood as prompting the model to abandon the current erroneous generation direction and attempt to find other generation strategies that conform to logical constraints. Revising the draft maintenance work order refers to the model generating new, logically valid content to replace or continue the previous draft after replanning the path. This process may repeat until the generated draft maintenance work order completely passes logical verification. The final draft maintenance work order that passes verification is confirmed as the official maintenance work order.

[0077] As an optional implementation, logical verification can be implemented using a logical review agent. This agent runs in parallel with the work order generation process, acquires the generated word segmentation stream, and performs reverse derivation in the fault logic graph in real time. If it is found that the generated action is on the logical path that leads to the top event and the triggering condition is met, the verification is determined to fail, a mask signal is sent to block the current generation, and error information is fed back to the generation model to guide its self-correction.

[0078] The maintenance work order generation method provided in this embodiment obtains the fault description text and functional location code of the target equipment, and comprehensively utilizes the semantic similarity of the fault description text and the hierarchical distance of the functional location code to accurately select a set of reference work orders from historical maintenance work orders from two dimensions. This achieves efficient reuse and accurate matching of historical maintenance experience, effectively avoiding erroneous references across systems and safety levels caused by relying solely on semantic matching. Furthermore, the work order generation model is used to generate maintenance work orders by combining the input fault information, location information, and the set of reference work orders. During the generation process, a logic verification mechanism based on the fault logic graph is introduced in real time to conduct strict compliance review and dynamic correction of the generated draft maintenance work order. Because logic blocking and path replanning are implemented in the generation stage, it is ensured that the final generated maintenance work order not only fully draws on historical best practices, but also strictly complies with the complex fault logic and safety interlocking requirements of the nuclear power field. This effectively avoids the illusion phenomenon and illegal operation instructions that may be generated by the generative model, significantly improving the accuracy, compliance, and safety of nuclear power maintenance work order generation.

[0079] Considering that functional location coding is essentially a type of structured data with hierarchical attributes, it is difficult to accurately quantify the physical or logical relationship between two devices through simple string matching alone. To accurately measure the degree of association between the target device and devices in historical maintenance work orders within a hierarchical structure, this embodiment employs a distance calculation method based on a tree-like topology.

[0080] Specifically, the steps for determining the hierarchical distance in this embodiment include:

[0081] Functional location codes and historical location codes are mapped to nodes on a tree topology, which is used to characterize the relationships between functional locations in the device hierarchy.

[0082] Calculate the shortest path length between the node corresponding to the functional location code and the node corresponding to the historical location code in the tree topology, and use the shortest path length as the hierarchical distance.

[0083] First, considering the need to compare the positions of different devices within a unified structural framework, this embodiment maps the functional location code and historical location code to nodes on a tree topology.

[0084] A tree topology can be understood as a data structure that graphically represents the parent-child relationships or logical connections between various hierarchical units (such as units, systems, equipment, and components) in a device hierarchy. In this structure, each node represents a specific functional location, and the lines connecting nodes represent hierarchical relationships. The functional location code is a unique identifier for each node. By parsing the encoding rules, each linear encoded string can be located at a specific node position in the tree topology. The mapping process is the process of finding the corresponding node in the tree based on the encoded content. Associations refer to the mutual relationships between devices in terms of physical connection, system affiliation, or functional logic.

[0085] As an alternative implementation, the hierarchical separators or fixed-length character segments in the functional location code can be parsed, and a search and match can be performed level by level in the tree topology until the leaf node or the specified level node is located. For example, for the code "10PAA10AA001", the "10" node is first found under the root node, then the "PAA" node is found at the next level, and so on, until the code is finally mapped to a specific node in the tree, thus realizing the transformation of the abstract character code into a concrete topological location.

[0086] After successfully mapping the functional location code and the historical location code to nodes in a tree topology, this embodiment takes into account that the relative position of the node in the tree can objectively reflect the closeness or distance between devices, and then calculates the shortest path length between the node corresponding to the functional location code and the node corresponding to the historical location code in the tree topology, and uses the shortest path length as the hierarchical distance.

[0087] The shortest path length refers to the minimum number of edges required to connect two nodes in a tree topology. In a tree structure, there is usually only one unique simple path between two points, and the length of this path is the shortest path length. This metric effectively quantifies the distance between two devices at the system level. For example, two devices belonging to the same system have a lower level nearest common parent node in the tree, resulting in a shorter path; devices belonging to different units have the root node as their nearest common parent node, resulting in a longer path. Using the shortest path length as the hierarchical distance can intuitively reflect the similarity between devices: the shorter the distance, the closer the devices are physically or logically, and the more reliable their maintenance experience is usually.

[0088] Considering the need for a data structure foundation that can completely and accurately reflect the overall picture of the physical system devices before calculating the hierarchical distance, this embodiment requires constructing a tree-like topology structure covering all known devices. To achieve this goal, the steps for constructing the tree-like topology structure in this embodiment include:

[0089] Obtain the full set of codes containing the functional location codes of all known devices in the device hierarchy;

[0090] Based on the preset hierarchical parsing rules, the encoding of each functional position in the full encoding set is decomposed into multiple character segments with parent-child hierarchical relationships;

[0091] A tree-like topology is constructed based on multiple character fragments. Each node in the tree-like topology corresponds to a character fragment, and each node is used to represent a hierarchical unit in the device hierarchy.

[0092] First, considering that the integrity of the tree-like topology depends on the coverage of the source data, in order to ensure that the constructed structure can support the query and location of any device, this embodiment obtains a full set of codes that includes the functional location codes of all known devices in the device hierarchy.

[0093] The full set of codes can be understood as a database or list containing the functional location codes of all registered and filed equipment within the target physical system (such as the entire nuclear power plant). This ensures that the tree-like topology will not miss any equipment node that may be involved in maintenance, thereby guaranteeing the comprehensiveness of subsequent searches.

[0094] As an optional implementation, a complete set of codes can be formed by exporting a list of all currently active functional location codes through the database interface of the device management system.

[0095] After obtaining the full set of codes, considering that the functional position code is usually a continuous string, it is difficult to reflect its internal hierarchical logic directly. Therefore, this embodiment then decomposes each functional position code in the full set of codes into multiple character segments with parent-child hierarchical relationships based on the preset hierarchical parsing rules.

[0096] Hierarchical parsing rules refer to the logical rules established according to industry standards or enterprise specifications for decomposing the functional position coding structure. For example, the coding rule may specify that the first digit is the unit number, the second to fourth digits are the system code, and the fifth and sixth digits are the subsystem code, etc. Character fragments refer to substrings extracted from the coded string according to these rules, representing specific hierarchical meanings. Parent-child hierarchical relationships refer to the logical inclusion and containment relationships of these character fragments; for example, "system code" is a child of "unit number," and "equipment serial number" is a child of "system code." Through this decomposition step, the flat string is transformed into hierarchical elements with structured semantics.

[0097] As an optional implementation, for encodings conforming to the "A-BBB-CC" format, the hierarchical parsing rule can be set to segment using hyphens "-" as delimiters. Assuming the functional location code is "10-PAA-01", based on this rule, it can be decomposed into three fields: "10" (first level, representing the unit), "PAA" (second level, representing the system), and "01" (third level, representing the equipment). Here, "10" is the parent of "PAA", and "PAA" is the parent of "01".

[0098] After decomposing the encoding into ordered character segments, considering the need to organize these discrete segments into a unified whole to facilitate path calculation, this embodiment constructs a tree-like topology based on multiple character segments. Each node in the tree-like topology corresponds to a character segment, and each node is used to represent a hierarchical unit in the device hierarchy.

[0099] A tree-like topology is an efficient hierarchical indexing structure. In this structure, the root node typically represents the highest level of the physical system, such as the entire plant. Each path originating from the root node represents a complete or partial functional location code. Each node stores a character fragment, and the depth of the node in the tree corresponds to the hierarchical position of the character fragment within the code. A hierarchical unit refers to an entity with specific physical or logical meaning in the equipment hierarchy, such as a specific system, plant, or piece of equipment. Through this construction method, codes with the same prefix will share the same upper-level node; for example, all equipment belonging to unit "10" will be attached to the node "10," thus greatly saving storage space and intuitively demonstrating the hierarchical relationships between devices.

[0100] As an alternative implementation, a tree-like topology can be constructed using a line-by-line insertion method. For the decomposed character segment sequence S = {s1, s2, ..., s...} n Starting from the root node of the tree, check if a child node corresponding to s1 exists; if it does, move to that child node; if it doesn't exist, create a new child node to store s1. Continue in this manner until all character fragments in the sequence have been processed. Repeat this process for each code in the full encoding set, ultimately forming a complete tree-like topology. For example, when inserting "10-PAA" and "10-PAB", they share the same parent node "10", and under the "10" node, two child nodes, "PAA" and "PAB", branch out, thus accurately representing their hierarchical relationship as belonging to the same unit 10 but belonging to different systems.

[0101] Considering that semantic similarity and hierarchical distance represent the similarity of fault phenomena and the proximity of equipment locations, respectively, they have different emphases and importance when evaluating the value of reference work orders. In order to comprehensively and objectively evaluate the reference value of each historical maintenance work order for the current generated task, this embodiment adopts a weighted fusion strategy for comprehensive scoring.

[0102] Specifically, in this embodiment, a reference work order set is obtained by filtering from each historical maintenance work order based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order. This set includes:

[0103] The semantic similarity and hierarchical distance of each historical maintenance work order are weighted and summed to obtain a comprehensive matching score that represents the overall matching degree between the target equipment and the corresponding historical maintenance work order.

[0104] The historical maintenance work orders are sorted in descending order of their comprehensive matching scores, and a set of reference work orders is formed by selecting a predetermined number of the top-ranked historical maintenance work orders.

[0105] First, considering that a single indicator is difficult to fully reflect the applicability of historical work orders, such as some work orders having very similar descriptions but the equipment locations being far apart, and some work orders having equipment locations being very close but the fault descriptions being slightly different, this embodiment performs a weighted summation of the semantic similarity and hierarchical distance corresponding to each historical maintenance work order to obtain a comprehensive matching score that characterizes the overall matching degree between the target equipment and the corresponding historical maintenance work order.

[0106] Semantic similarity is typically a value between 0 and 1, representing the closeness of text meaning; hierarchical distance is an integer representing path length, with smaller values ​​indicating closer proximity. To unify both into a single evaluation system, hierarchical distance is usually normalized, such as converting it to a value between 0 and 1, with values ​​closer to 1 indicating shorter distances. The comprehensive matching score is a comprehensive quantitative indicator used to intuitively reflect the fit between historical repair work orders and current needs. Weighted summation refers to merging the two dimensions of indicators into a single scalar based on preset weighting coefficients.

[0107] The weighting coefficients can be adjusted according to actual business needs. For example, in scenarios with a wide variety of equipment and fault modes that are strongly location-dependent, such as where pump maintenance strategies differ significantly at different locations, the weighting coefficient for hierarchical distance can be appropriately increased. In scenarios where fault symptom descriptions are highly specific and location has little impact, the weighting coefficient for semantic similarity can be appropriately increased.

[0108] After calculating the comprehensive matching score of all candidate historical maintenance work orders, considering the user's reading ability and the input length limit of the model, it is impossible to use all historical records as references. Therefore, in this embodiment, the historical maintenance work orders are then sorted in descending order of comprehensive matching score, and a preset number of historical maintenance work orders at the top of the sorting are selected to form a reference work order set.

[0109] Sorting refers to rearranging historical maintenance work orders based on their comprehensive matching scores, ensuring that work orders with higher scores are listed first. The preset quantity refers to the upper limit of work order entries set by the system as a reference input. The reference work order set is the final selection of a group of high-quality historical work orders. Because the comprehensive matching score integrates semantic and location-based evaluation information, it can accurately distinguish the quality of historical work orders, thereby eliminating noisy data with low relevance and retaining the most valuable core cases.

[0110] As an optional implementation, a preset quantity K=5 can be set. After calculating the comprehensive matching score of all historical maintenance work orders, the five historical maintenance work orders with the highest scores are selected and input into the work order generation model to help the model learn from excellent historical maintenance experience, thereby generating more accurate and practical new work orders.

[0111] While generative AI models possess powerful text generation capabilities, they lack a deep understanding of the causal logic of the physical world, making them prone to producing seemingly coherent but illogical content. This is particularly relevant in nuclear power plant maintenance, where a single erroneous maintenance action can trigger a severe chain reaction. To ensure that the generated work order content is logically safe and compliant, this embodiment introduces a logic verification mechanism based on a fault tree model.

[0112] Specifically, the preset fault logic diagram in this embodiment includes a fault tree model, which includes equipment fault events as top events, and maintenance operation actions and equipment operating status as bottom events.

[0113] The steps of logical verification include:

[0114] Analyze the draft maintenance work order, extract the currently generated maintenance procedures, and identify the target operation actions involved in the maintenance procedures;

[0115] In the fault tree model, a reverse traversal is performed starting from the top event to check whether the target operation is on the logical path that led to the top event.

[0116] The logical verification of the draft repair work order is determined based on the test results.

[0117] Specifically, the fault tree model is a top-down deductive logic analysis model. In this model, the top event refers to an undesirable, serious system failure or safety incident. The bottom events refer to the fundamental causes that lead to the top event; in this embodiment, the bottom events are specified as maintenance operations and equipment operating states. Logic gates connect the bottom events and the top events, forming a tight logical network.

[0118] As an alternative implementation, based on the nuclear power plant's design documents and safety analysis reports, knowledge graph construction tools can be used to transform various failure modes and their triggering conditions into nodes and edges in the graph.

[0119] With the fault tree model in place, and considering that work order generation is a dynamic process, in order to monitor the security of the generated content in real time, this embodiment parses the draft maintenance work order, extracts the currently generated maintenance procedures, and identifies the target operation actions involved in the maintenance procedures.

[0120] A maintenance work order draft refers to the text content that is being generated in the model and has not yet been finalized. A maintenance procedure refers to the specific steps described in the draft. A target operation action refers to the operation instruction entity with clear semantics extracted from the procedure described in natural language. For example, if the maintenance procedure is "Step 3: Place channel A as a bypass", the corresponding target operation action is "place as a bypass".

[0121] As an optional implementation, named entity recognition or dependency parsing can be used to scan the segmented word stream generated by the model in real time. Once a verb-object phrase, such as "open the valve" or "cut off the power," is detected, it is extracted as the target action, and the object of the operation is determined in combination with the context.

[0122] After identifying the target operation, in order to determine whether the operation is safe, this embodiment then performs a reverse traversal in the fault tree model, starting from the top event, to detect whether the target operation is on the logical path that led to the top event.

[0123] Reverse traversal refers to the process of starting from the top event of the fault tree and searching downwards along the input direction of the logic gates until the bottom event is found. A logical path refers to a series of logical connections from the top event to the bottom event. If an operation is on a logical path, it means that the operation is one of the potential causes of the top event.

[0124] As an optional implementation, the identified target action can be input into a logic review agent. This agent searches the fault tree to determine if a corresponding bottom event node exists. If it does, it traces the path from that node to the top event.

[0125] Finally, considering that simply confirming that the action is on the path is not enough to determine a violation, this embodiment also determines whether the logical verification of the draft maintenance work order passes based on the detection results.

[0126] The detection results should include not only whether the action is on the path, but also a comprehensive judgment based on the current system state. The logical verification result is a Boolean value. If the verification passes, the action is safe under the current operating conditions; if the verification fails, the action has triggered a safety interlock condition and must be intercepted.

[0127] Considering that when logic verification detects a safety risk in the currently generated maintenance action, a mere warning is insufficient; mandatory technical measures are needed to prevent the output of erroneous content and guide the model back to the correct path. To achieve this real-time, closed-loop correction capability, this embodiment provides a correction strategy based on dynamic blocking and regeneration.

[0128] Specifically, in this embodiment, if the logic check fails, the corresponding draft maintenance work order is revised, including:

[0129] If the logic check fails, a mask blocking signal is sent to the work order generation model to block the output content corresponding to the target operation action.

[0130] The work order generation model is driven to regenerate alternative maintenance procedures that do not include the target operation action;

[0131] By replacing the corresponding maintenance procedures in the draft maintenance work order with alternative maintenance procedures, a revised draft maintenance work order is obtained.

[0132] Considering that a failed validation means the currently generated word segmentation sequence has deviated from the safe path, its output needs to be immediately cut off to prevent erroneous information from entering the final work order text, this embodiment sends a mask blocking signal to the work order generation model in the event of a logical validation failure, blocking the output content corresponding to the target operation action.

[0133] A masking signal can be understood as a control command applied to the model's inference layer. If the work order generation model is a generative model (such as Transformer), the generation of the next word segment is based on a probability distribution. A mask can force the generation probability of a specific word segment to negative infinity or zero, thus physically blocking the output of that content. The output content corresponding to the target operation action refers to the word or phrase that the model originally intended to generate but was judged to be in violation.

[0134] As an optional implementation, when the logic review agent determines that the "bypass" action is in violation, it immediately modifies the vector of the model's output layer, setting the value of the word segmentation index position corresponding to "bypass" and its synonyms to negative infinity. This way, after normalization, the selection probability of these words becomes 0. The model will then reselect the output based on the probability distribution of the remaining words, thereby avoiding the generation of inappropriate content.

[0135] After successfully blocking the output of erroneous content, considering that the model cannot fall into a stagnant state, it is also necessary to guide it to generate a compliant operation to replace the erroneous operation. Therefore, in this embodiment, the work order generation model is driven to regenerate an alternative maintenance procedure that does not contain the target operation action.

[0136] Driven regeneration refers to sending a fallback or retry signal to the model, possibly accompanied by a prompt message informing the model that the previous attempt failed and the reason for the failure. An alternative repair procedure refers to a logically safe alternative operation that can achieve the same repair objective. The fact that an alternative repair procedure does not include the target operation means that the new generation path must avoid the previously blocked action.

[0137] As an optional implementation, the generated context can be rolled back to the position before the target operation was generated, and a system prompt can be inserted, such as "A bypass operation has been detected that violates the safety interlock logic. Please try other operation methods that comply with the Technical Specifications." After receiving this feedback, the model, based on the remaining probability distribution, will re-search for the compliant path with the highest probability, thereby generating new process content.

[0138] Finally, after the model successfully generated the new process, considering the need to ensure the integrity and consistency of the draft work order, this embodiment uses alternative maintenance processes to replace the maintenance processes in the draft maintenance work order, resulting in a revised draft maintenance work order.

[0139] The replacement operation refers to updating the original incomplete or incorrect sentence in the draft text with a newly generated compliant sentence. The revised draft maintenance work order is the version that has undergone logical cleansing and path replanning, which eliminates known security vulnerabilities while maintaining semantic coherence.

[0140] As an optional implementation, a state tree for the draft work order can be maintained in memory. Once the alternative procedure is generated and passes a new round of verification, the text block corresponding to that procedure is written to the corresponding node in the state tree, overwriting the previous erroneous branch, thereby completing the physical correction of the draft. This process may be transparent to the user, such as directly seeing the correct result, or it may be visible, such as seeing AI thinking and correcting, ultimately outputting a robust and safe work order.

[0141] Considering that performing full semantic vector calculation directly on a massive amount of historical maintenance work orders would consume huge computing resources and result in high response latency, this embodiment adopts a two-stage retrieval strategy of first coarse screening and then fine calculation in order to achieve efficient and accurate similarity matching.

[0142] Specifically, in this embodiment, the steps for determining the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order include:

[0143] Perform word segmentation on the fault description text to obtain keywords;

[0144] Based on the index mapping relationship between keywords and historical maintenance work orders, candidate historical maintenance work orders containing keywords are retrieved from each historical maintenance work order.

[0145] The fault description text and the historical fault description text of the candidate historical maintenance work order are converted into semantic vectors, and the cosine similarity between the semantic vectors is calculated as the semantic similarity.

[0146] Specifically, word segmentation refers to dividing continuous natural language text into independent lexical units according to linguistic rules. Keywords are content words with clear referential meaning after removing stop words from the segmented vocabulary. By extracting keywords, complex text can be simplified into a set of core features, facilitating rapid indexing.

[0147] As an alternative implementation, open-source word segmentation tools can be used to process the fault description text, and combined with professional dictionaries in the nuclear power field for optimization, to ensure that professional terms are correctly identified as a complete keyword, rather than being broken down.

[0148] After obtaining the keywords, in order to achieve efficient retrieval, this embodiment retrieves candidate historical maintenance work orders containing the keywords from each historical maintenance work order based on the index mapping relationship between the keywords and historical maintenance work orders.

[0149] The index mapping relationship is a pre-built data structure that records which historical maintenance work orders each keyword appears in. Candidate historical maintenance work orders refer to a subset of historical work orders containing at least one keyword. This step allows for the filtering of hundreds or thousands of relevant candidate work orders from millions of historical records within milliseconds, significantly narrowing the scope of subsequent calculations.

[0150] After determining a small set of candidate historical repair work orders, considering that keyword matching cannot capture deep semantic information, such as "loud noise" and "abnormal sound" being semantically similar despite their different literal meanings, this embodiment converts the fault description text and the historical fault description text of the candidate historical repair work orders into semantic vectors, and calculates the cosine similarity between the semantic vectors as the semantic similarity.

[0151] Semantic vectors are numerical vectors that map text to a high-dimensional space, and the distance between vectors represents the semantic similarity. Cosine similarity is a metric that measures the difference in direction between two vectors; the closer the value is to 1, the more semantically similar they are.

[0152] As an optional implementation, a pre-trained language model can be used to encode the fault description text and the historical fault description text of each candidate work order into fixed-dimensional vectors. and Then, the cosine similarity is calculated using the following formula. :

[0153] .

[0154] Given the significant priority differences among nuclear power plant maintenance tasks, some tasks involving nuclear safety or major equipment failures are highly time-sensitive and must be prioritized. However, the traditional first-come, first-served scheduling model may cause emergency tasks to be blocked by backlogged routine tasks, creating safety hazards. To ensure that high-risk tasks can receive immediate responses, this embodiment introduces a preemptive scheduling mechanism based on risk entropy values.

[0155] Specifically, after obtaining the fault description text of the target device, this embodiment further includes:

[0156] Semantic parsing of fault description text is performed based on a pre-defined high-risk sensitive word library, and the risk entropy value of fault description text is calculated.

[0157] If the risk entropy value exceeds the emergency threshold, suspend the currently executing generation task whose corresponding risk entropy value is below the emergency threshold, and use the computing resources released by the suspended generation task to execute the maintenance work order generation task for the target device.

[0158] Specifically, the high-risk sensitive word database refers to a database containing high-risk keywords in the nuclear power field and their risk weights. Semantic parsing refers to identifying whether a text contains these sensitive words and analyzing its context. Risk entropy is a numerical indicator reflecting the degree of potential risk of a task; the higher the entropy value, the greater the risk involved in the task and the higher the urgency of its handling.

[0159] As an optional implementation, the fault description text can be scanned and matched against a sensitive word database. Assume the set of matched sensitive words is... ={ , ,..., The risk weight for each word is} Risk entropy value It can be calculated using the following formula:

[0160] ;

[0161] in, Sensitive words The amount of information or frequency factor.

[0162] After calculating the risk entropy value, considering that computing resources are limited, in order to make room for high-risk tasks, this embodiment suspends the currently executing generation task whose corresponding risk entropy value is lower than the emergency threshold when the risk entropy value exceeds the emergency threshold, and uses the computing resources released by the suspended generation task to execute the maintenance work order generation task for the target device.

[0163] The emergency threshold is a set critical value used to distinguish between high-risk emergency tasks and regular tasks. A suspension operation refers to temporarily interrupting the inference process of an ongoing low-priority task, saving its context state, and releasing the GPU memory and computing units it occupies. The released computing resources are then immediately allocated to the current high-risk task, allowing it to begin generation without queuing.

[0164] As an optional implementation, when a new task is detected... When the emergency threshold is reached and the current video memory is full, the system iterates through the running task queue, suspends tasks with risk entropy values ​​less than the emergency threshold, and transfers their video memory data to RAM. Then, the freed-up video memory is used to load the model parameters and context of the high-risk task and initiate its generation. Once the high-risk task is completed, the suspended task is swapped back into video memory for continued execution, ensuring zero-wait response for critical tasks and effectively improving the system's emergency response capabilities.

[0165] While general-purpose models possess broad knowledge, they often lack awareness of specific nuclear power industry standards, easily generating work orders with incorrect formats or omissions of safety points. To inject industry standards and safety red lines into the model's generation logic, this embodiment inputs fault description text, functional location codes, and a set of reference work orders into the work order generation model, including:

[0166] Acquire domain standard specifications and security constraint rules;

[0167] Based on a pre-built structured template, domain standard specification data, safety constraint rules, reference work order set and fault description text are integrated to construct prompt words that contain specification constraints and contextual information;

[0168] Input the prompt words into the work order generation model.

[0169] Specifically, domain standard specifications refer to the technical documents that nuclear power plant maintenance operations must follow, such as the "Maintenance Management Outline" and "Standard Operating Procedures." These specifications define requirements for work order formats, terminology, and process granularity. Safety constraints refer to mandatory red lines related to industrial safety and radiation protection, such as "high-voltage equipment must be de-energized before operation" and "protective clothing must be worn when entering controlled areas."

[0170] As an optional implementation, standard clauses related to the current fault type can be retrieved from the enterprise knowledge base. For example, for electrical equipment faults, relevant chapters from the "Electrical Maintenance Safety Regulations" can be automatically retrieved as regulatory data.

[0171] After obtaining the domain standard specification data and safety constraint rules, considering that directly inputting scattered information into the model may lead to distraction, it is necessary to organize it into a form that the model can easily understand. Therefore, this embodiment integrates the domain standard specification data, safety constraint rules, reference work order set and fault description text based on a pre-set structured template to construct prompt words that contain specification constraints and contextual information.

[0172] The structured template is a predefined text framework with different slots for filling specific information. The cue words are the complete instruction text ultimately input into the model. The integration process involves filling various types of information into their corresponding slots, forming a logically clear task specification.

[0173] After receiving prompts containing rich constraints and context, the model reasones according to the established roles and rules, thereby generating a repair work order that conforms to professional standards and is relevant to the actual fault.

[0174] Considering that even after rigorous logical verification, the generated work order may still have slight discrepancies in details compared to the experience of human experts, and in order to continuously learn and optimize from each human-computer interaction, this embodiment, after using the logically verified draft maintenance work order as the maintenance work order for the target equipment, also includes:

[0175] Obtain modification operation data for maintenance work orders, and determine the modification type by comparing the modified content corresponding to the modification operation data with the original content of the maintenance work order draft;

[0176] When the modification type indicator only involves changes in device parameter values, the entity attributes in the knowledge graph are updated using the modified device parameter values.

[0177] When the modification type indication involves changes in the order of maintenance procedures, the structured template used to guide the generation of the work order generation model is updated using the modified maintenance procedures.

[0178] Specifically, modification operation data refers to the records of editing actions such as adding, deleting, and modifying the content of AI-generated work orders performed by engineers during the final review stage. Modification type is a classification of the intent behind these actions; different modification intents correspond to the optimization needs of different components in the system.

[0179] As an optional implementation, a text difference comparison algorithm can be used to calculate the differences between the original content and the modified content. If the difference is only a change in numbers, the modification type is determined to be "numerical correction"; if the difference is a displacement of paragraphs or sentences, the modification type is determined to be "logical order correction".

[0180] After determining the type of modification, in order to achieve refined updates and avoid catastrophic model amnesia, this embodiment adopts different update strategies for different types.

[0181] Considering that the numerical correction type implies that the static data in the knowledge graph is inaccurate, this embodiment updates the entity attributes in the knowledge graph using the modified device parameter values ​​when the modification type indication only involves changes in device parameter values.

[0182] For example, the modified value "60 N.m" can be automatically extracted, and a knowledge graph service can be called via API to locate the entity "a certain type of bolt" and update its "standard tightening torque" attribute value to "60 N.m". The next time a work order involving this bolt is generated, the retrieved knowledge will be the latest accurate value.

[0183] Considering that the logical sequence correction reflects the optimization of the maintenance process, this embodiment uses the modified maintenance procedure to update the structured template used to guide the generation of the work order generation model when the modification type indication involves changes in the order of maintenance steps.

[0184] For example, the corrected process sequence can be extracted and added as a small sample case to a slot in a structured template, or used to update the process orchestration instructions in the Prompt. This way, when the model generates a similar task next time, it will tend to mimic this better process, which has been corrected by human experts.

[0185] Figure 2 This is a flowchart illustrating another maintenance work order generation method provided by the present invention, as shown below. Figure 2 As shown, firstly, the user inputs a work order description, which includes a fault description text for the target device and a functional location code representing the target device's position in the device hierarchy. At the field granularity level, structured fields such as functional location codes are processed using a lightweight index query engine; unstructured fields such as fault description text are processed through a large language model semantic retrieval pipeline, and similar semantic requests are merged using dynamic batch processing technology.

[0186] Then, based on the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order, a reference work order set is obtained from each historical maintenance work order. Specifically, by combining semantic similarity and hierarchical distance, relevant historical work orders are retrieved, and valid historical work orders are obtained through field information integration, deduplication, and sorting. Key information from these work orders is then extracted as the reference work order set.

[0187] Next, the fault description text, functional location code, and reference work order set are input into the work order generation model, which then generates a maintenance work order for the target equipment. During this process, the nuclear power professional knowledge base is invoked to obtain domain standard and specification data and safety constraint rules. Based on a pre-set structured template, the domain standard and specification data, safety constraint rules, reference work order set, and fault description text are integrated to construct a prompt word containing specification constraints and contextual information. The work order generation model generates the maintenance work order based on the prompt word. During the work order generation process, the generated maintenance work order draft is mapped to a pre-set fault logic graph in real time for logical verification. If the logical verification fails, the work order generation model is driven to replan the generation path, correct the maintenance work order draft, and return for logical verification until the corrected maintenance work order draft passes the logical verification. The logically verified maintenance work order draft is then used as the maintenance work order, and manual verification is performed again to output the final compliant maintenance work order.

[0188] In addition, after outputting compliant maintenance work orders, the system captures the modification data of engineers on the maintenance work orders and fine-tunes the work order generation model based on the modification data to achieve continuous optimization of the model.

[0189] The maintenance work order generation device provided by the present invention is described below. The maintenance work order generation device described below and the maintenance work order generation method described above can be referred to in correspondence.

[0190] Based on any of the above embodiments Figure 3 This is a schematic diagram of the maintenance work order generation device provided by the present invention, as shown below. Figure 3 As shown, the device includes:

[0191] The acquisition module 310 is used to acquire the fault description text of the target device and the functional location code used to characterize the position of the target device in the device hierarchy structure.

[0192] The filtering module 320 is used to filter a set of reference work orders from each historical maintenance work order based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, as well as the hierarchical distance between the function location code and the historical location code of each historical maintenance work order.

[0193] The generation module 330 is used to input the fault description text, function location code and reference work order set into the work order generation model. It performs logical verification on the maintenance work order draft generated in real time by the work order generation model through the preset fault logic diagram. If the logical verification fails, the corresponding maintenance work order draft is corrected until the corrected maintenance work order draft passes the logical verification. The maintenance work order draft that passes the logical verification is used as the maintenance work order for the target equipment.

[0194] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a maintenance work order generation method.

[0195] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0196] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the maintenance work order generation method provided by the above methods.

[0197] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the maintenance work order generation method provided by the above methods.

[0198] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0199] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating maintenance work orders, characterized in that, include: Obtain the fault description text of the target device, as well as the functional location code used to characterize the position of the target device in the device hierarchy; Based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order, a set of reference work orders is obtained from each historical maintenance work order. The fault description text, the functional location code, and the reference work order set are input into the work order generation model. The generated word segmentation stream is obtained by the logic review agent running in parallel during the work order generation process. The model performs reverse traversal in real time in the preset fault logic graph. If it is found that the generated action is on the logic path that leads to the top event and the triggering condition is met, the logic verification is determined to fail. A mask signal is sent to block the current generation and an error message is fed back to the work order generation model so that the work order generation model can correct the corresponding maintenance work order draft. The corrected maintenance work order draft passes the logic verification and is then used as the maintenance work order for the target equipment. The fault logic graph refers to a pre-built, structured knowledge base used to describe the logical relationships between equipment faults, maintenance operations, and system states. The fault logic graph includes a fault tree model, which contains equipment fault events as top events, maintenance operation actions as bottom events, and equipment operating states. The reverse traversal refers to the process of searching downwards from the top event of the fault tree along the input direction of the logic gates until the bottom event is found. The logical path refers to a series of logical connections from the top event to the bottom event.

2. The maintenance work order generation method according to claim 1, characterized in that, The steps for determining the hierarchical distance include: The functional location code and the historical location code are respectively mapped to nodes on a tree topology structure, and the tree topology structure is used to characterize the association relationship between each functional location in the device hierarchy structure; Calculate the shortest path length between the node corresponding to the functional location code and the node corresponding to the historical location code in the tree topology, and use the shortest path length as the hierarchical distance.

3. The maintenance work order generation method according to claim 2, characterized in that, The steps for constructing the tree-like topology include: Obtain the full set of codes containing the functional location codes of all known devices in the device hierarchy; Based on the preset hierarchical parsing rules, the functional position codes in the full encoding set are decomposed into multiple character segments with parent-child hierarchical relationships; The tree-like topology is constructed based on the multiple character fragments, where each node in the tree-like topology corresponds to one of the character fragments, and each node is used to represent a hierarchical unit in the device hierarchy.

4. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, The step of filtering a set of reference work orders from historical maintenance work orders based on the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order, includes: The semantic similarity and hierarchical distance of each historical maintenance work order are weighted and summed to obtain a comprehensive matching score that represents the overall matching degree between the target equipment and the corresponding historical maintenance work order. The historical maintenance work orders are sorted in descending order of their comprehensive matching scores, and a preset number of the top-ranked historical maintenance work orders are selected to form the reference work order set.

5. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, The steps of the logic verification include: The draft maintenance work order is parsed, the currently generated maintenance procedures are extracted, and the target operation actions involved in the maintenance procedures are identified. In the fault tree model, a reverse traversal is performed starting from the top event to detect whether the target operation is on the logical path that led to the occurrence of the top event. The logical verification of the draft maintenance work order is determined based on the test results.

6. The maintenance work order generation method according to claim 5, characterized in that, The revised draft maintenance work order includes: Send a mask blocking signal to the work order generation model to block the output content corresponding to the target operation action; The work order generation model is driven to regenerate alternative maintenance procedures that do not include the target operation action; By replacing the corresponding maintenance procedures in the draft maintenance work order with the alternative maintenance procedures, a revised draft maintenance work order is obtained.

7. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, The steps for determining the semantic similarity between the fault description text and the historical fault description texts of each historical maintenance work order include: The fault description text is segmented to obtain keywords; Based on the index mapping relationship between keywords and historical maintenance work orders, candidate historical maintenance work orders containing the keywords are retrieved from each historical maintenance work order; The fault description text and the historical fault description text of the candidate historical maintenance work order are converted into semantic vectors, and the cosine similarity between the semantic vectors is calculated as the semantic similarity.

8. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, After obtaining the fault description text of the target device, the following is also included: The fault description text is semantically parsed based on a preset high-risk sensitive word library, and the risk entropy value of the fault description text is calculated. If the risk entropy value exceeds the emergency threshold, the currently executing generation task with a corresponding risk entropy value lower than the emergency threshold is suspended, and the computing resources released by the suspended generation task are used to execute a maintenance work order generation task for the target device.

9. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, The step of inputting the fault description text, the functional location code, and the reference work order set into the work order generation model includes: Acquire domain standard specifications and security constraint rules; Based on a pre-set structured template, the domain standard specification data, the safety constraint rules, the reference work order set, and the fault description text are integrated to construct prompt words that contain specification constraints and contextual information; Input the prompt word into the work order generation model.

10. The maintenance work order generation method according to any one of claims 1 to 3, characterized in that, After using the draft maintenance work order that has passed the logical verification as the maintenance work order for the target equipment, the following is also included: Obtain modification operation data for the maintenance work order, and determine the modification type by comparing the modified content corresponding to the modification operation data with the original content of the maintenance work order draft; When the modification type indication only involves changes in device parameter values, the entity attributes in the knowledge graph are updated using the modified device parameter values. When the modification type indicates a change in the order of maintenance procedures, the structured template used to guide the generation of the work order generation model is updated using the modified maintenance procedures.

11. A maintenance work order generation device, characterized in that, include: The acquisition module is used to acquire the fault description text of the target device and the functional location code used to characterize the position of the target device in the device hierarchy structure. The filtering module is used to filter a set of reference work orders from each historical maintenance work order based on the semantic similarity between the fault description text and the historical fault description text of each historical maintenance work order, and the hierarchical distance between the functional location code and the historical location code of each historical maintenance work order. The generation module is used to input the fault description text, the functional location code, and the reference work order set into the work order generation model. The generated word segmentation stream is obtained through a logic review agent that runs in parallel during the work order generation process. The module performs reverse traversal in a preset fault logic graph in real time. If it finds that the generated action is on the logic path that leads to the top event and the triggering condition is met, it is determined that the logic verification fails. A mask signal is sent to block the current generation and error information is fed back to the work order generation model so that the work order generation model can revise the corresponding maintenance work order draft. The revised maintenance work order draft passes the logic verification and is then used as the maintenance work order for the target equipment. The fault logic graph refers to a pre-built, structured knowledge base used to describe the logical relationships between equipment faults, maintenance operations, and system states. The fault logic graph includes a fault tree model, which contains equipment fault events as top events, maintenance operation actions as bottom events, and equipment operating states. The reverse traversal refers to the process of searching downwards from the top event of the fault tree along the input direction of the logic gates until the bottom event is found. The logical path refers to a series of logical connections from the top event to the bottom event.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the maintenance work order generation method as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the maintenance work order generation method as described in any one of claims 1 to 10.

Images