A nuclear power maintenance work package generation method based on a domain large model

Abstract

The application discloses a nuclear power maintenance work package generation method based on a domain large model, including the following steps: obtaining a nuclear power maintenance standard data set; performing nuclear power maintenance knowledge extraction, constructing a nuclear power maintenance domain data set and a nuclear power maintenance constraint graph; performing domain adaptation training combined with full-parameter fine-tuning and LoRA low-rank adaptation, and designing a constraint-aware loss function and a reasoning guide mechanism to obtain a nuclear power maintenance domain large model; receiving a corrective work order preparation request to form a work package generation task context; using the nuclear power maintenance domain large model to perform retrieval enhancement generation to form a work package generation evidence set; performing maintenance process disassembly and resource element matching to generate a candidate nuclear power maintenance work package; performing consistency checking and correcting according to the checking result to obtain the nuclear power maintenance work package, thereby improving the automation level, accuracy and consistency of the maintenance work package generation.

Description

Technical Field

[0001] This invention relates to the field of nuclear power equipment maintenance management, and in particular to a method for generating nuclear power maintenance work packages based on a large domain model. Background Technology

[0002] Nuclear power equipment is characterized by its complex system structure, diverse equipment types, long operating cycles, and stringent safety requirements. During the operation of nuclear power equipment, it is necessary to develop nuclear power maintenance work packages (NPMS) for equipment failures, abnormal conditions, and unplanned maintenance tasks. These work packages must clearly define the target nuclear power equipment, fault information, maintenance task type, maintenance procedures, resource elements, safety measures, and quality acceptance requirements. Particularly in corrective maintenance scenarios, personnel need to prepare corrective work orders, consult equipment technical documents, maintenance procedures, historical fault records, maintenance records, safety specifications, quality acceptance standards, and equipment drawings, and combine these with the current operating environment to complete the development of the nuclear power maintenance work package. The nuclear power maintenance work package directly affects the execution sequence of maintenance tasks, resource allocation, and operational safety; its completeness, accuracy, and traceability have a significant impact on nuclear power equipment maintenance management.

[0003] Current nuclear power plant maintenance work packages typically rely on manual preparation. Upon receiving a request to prepare a corrective work order, staff must search multiple business systems, document repositories, and historical databases for relevant information on the target nuclear power equipment and faults. They then filter, organize, and combine the retrieved data based on their personal experience. Because nuclear power plant maintenance data comes from diverse sources and uses inconsistent formats—including structured data such as equipment ledgers, fault records, and maintenance records, as well as unstructured data such as maintenance procedures, safety regulations, quality acceptance documents, and equipment drawings—existing processing methods struggle to provide unified preprocessing and standardized representation of multi-source nuclear power plant maintenance data. This results in inconsistencies in equipment names, fault names, maintenance procedure names, and resource element names across different data sets. Staff must repeatedly verify equipment identification and fault descriptions from different sources, increasing work package preparation time and increasing the risk of information omissions or misinterpretations due to differences in data understanding. Summary of the Invention

[0004] One objective of this invention is to propose a method for generating nuclear power plant maintenance work packages based on a domain-wide model. This invention fully utilizes multi-source data preprocessing for nuclear power plant maintenance, nuclear power plant maintenance knowledge extraction, nuclear power plant maintenance constraint graph construction, full parameter fine-tuning of the domain-wide model, and LoRA low-rank adaptation fusion training technology to achieve an automated generation process from corrective work order preparation request to nuclear power plant maintenance work package.

[0005] A method for generating nuclear power plant maintenance work packages based on a domain-wide large model according to an embodiment of the present invention includes the following steps: Obtain multi-source datasets for nuclear power plant maintenance and perform preprocessing to obtain standard datasets for nuclear power plant maintenance; Nuclear power maintenance knowledge is extracted from the nuclear power maintenance standard dataset to construct a nuclear power maintenance domain dataset and a nuclear power maintenance constraint graph. Based on a general large model, we use a dataset from the nuclear power maintenance domain to perform domain adaptation training that combines full parameter fine-tuning with LoRA low-rank adaptation, inject specialized nuclear power maintenance knowledge, and design a constraint-aware loss function and inference guidance mechanism based on the nuclear power maintenance constraint graph. We then perform inference reinforcement training on the general large model after domain adaptation training to obtain a large model for the nuclear power maintenance domain. Receive a corrective work order preparation request, extract the target nuclear power equipment, fault information, maintenance task type and work environment information from the corrective work order preparation request, and form a work package generation task context; The task context for generating work packages is input into a large model in the field of nuclear power maintenance for retrieval enhancement and generation, forming a work package generation evidence set. The large model in the field of nuclear power maintenance is encapsulated into a nuclear power maintenance work package to generate an intelligent agent. The evidence set generated by the work package is combined to decompose the maintenance process and match resource elements to generate candidate nuclear power maintenance work packages. The candidate nuclear power plant maintenance work packages are subjected to consistency verification, and corrections are made based on the verification results to obtain the nuclear power plant maintenance work packages.

[0006] Optionally, the nuclear power maintenance multi-source dataset includes basic information of nuclear power equipment, operating status data of nuclear power equipment, historical fault records, historical maintenance records, maintenance procedures, safety specifications, equipment drawings and historical work packages. The preprocessing includes document format conversion, duplicate data removal, missing data completion, abnormal data correction, unified equipment identification, time information alignment and normalization of nuclear power maintenance terminology.

[0007] Optionally, the construction of the nuclear power plant maintenance dataset and nuclear power plant maintenance constraint diagram specifically includes: The nuclear power plant maintenance standard dataset is divided into content segments to obtain a set of nuclear power plant maintenance knowledge units. Entity identification and entity attribute extraction are performed on the nuclear power plant maintenance knowledge unit set to generate a nuclear power plant maintenance entity set. Based on the set of nuclear power plant maintenance entities, the relationships between these entities are identified to form a set of nuclear power plant maintenance relationships. The nuclear power maintenance knowledge unit set, nuclear power maintenance entity set, and nuclear power maintenance relationship set are linked and organized, and the source mapping relationship between each entity in the nuclear power maintenance entity set and the corresponding nuclear power maintenance knowledge unit is established to obtain the nuclear power maintenance domain dataset. The nuclear power plant maintenance constraint graph is constructed by setting each entity in the nuclear power plant maintenance entity set as a graph node, setting each relation in the nuclear power plant maintenance relation set as a graph edge, and setting the entity attributes corresponding to each entity as the node attributes of the corresponding graph node.

[0008] Optionally, the large-scale model for nuclear power plant maintenance is obtained specifically including: Construct a nuclear power maintenance training sample set based on a dataset from the nuclear power maintenance field; Full parameter fine-tuning of the general large model is performed, and LoRA low-rank adaptation is introduced into the attention parameter structure of the general large model. Through joint training of nuclear power maintenance expertise injection and local efficient adaptation, the domain-adapted general large model is obtained. Based on the nuclear power plant maintenance constraint graph, reasoning path mining and topology construction are performed. Complete nuclear power plant maintenance reasoning paths are marked as forward nuclear power plant maintenance reasoning paths, and nuclear power plant maintenance reasoning paths with missing nodes are marked as reverse nuclear power plant maintenance reasoning paths, thus forming a nuclear power plant maintenance reasoning training sample set. The nuclear power plant maintenance reasoning training sample set is input into the general large model after domain adaptation training. The constraint-aware loss function that integrates the positive reasoning fitting loss and the negative reasoning penalty loss is constructed in combination with the nuclear power plant maintenance constraint graph. The parameters of the general large model after domain adaptation training are updated according to the constraint-aware loss function to obtain the general large model after reasoning reinforcement training. During the training process of the general large model after inference reinforcement training, the generated content is mapped to the nuclear power maintenance constraint graph. The generated entities are determined according to the relationship between nuclear power maintenance entities, and real-time intervention and adjustment are carried out to construct a constraint graph-driven inference guidance mechanism to obtain a large model in the field of nuclear power maintenance.

[0009] Optionally, the formation of the task context for generating the work package specifically includes: Receive corrective work order preparation requests and perform structured parsing on the corrective work order preparation requests to obtain structured representation data of the work order; Based on the structured representation of work orders, the target nuclear power equipment identifier, fault information, maintenance task type information, and operating environment information are extracted. The target nuclear power equipment identifier is matched with the equipment entities in the nuclear power maintenance standard dataset to generate standard equipment entity identifiers; The fault information is standardized based on the standard equipment entity identifier to obtain a standard fault information representation; The standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information are structurally combined to generate a work package generation task context.

[0010] Optionally, the formation of the evidence set generated by the working package specifically includes: Input the task context of the work package generation task into the large model of the nuclear power maintenance domain, generate a semantic representation of the work package generation task, and determine the set of nuclear power maintenance entities based on the semantic representation of the work package generation task. Map the set of nuclear power plant maintenance entities to the corresponding graph nodes in the nuclear power plant maintenance constraint graph, and perform path expansion to obtain the set of target retrieval paths; Based on the target retrieval path set, the corresponding nuclear power maintenance knowledge unit set is extracted from the nuclear power maintenance domain dataset, and the candidate evidence unit set is formed by association extraction based on the source mapping relationship between the nuclear power maintenance knowledge unit set and the graph nodes. The candidate evidence unit set is aligned sequentially and redundantly removed to obtain a structured evidence unit set. The structured evidence unit set is integrated and organized with the work package generation task context to generate the work package generation evidence set.

[0011] Optionally, the generation of the candidate nuclear power plant maintenance work package specifically includes: Functionally encapsulate a large model in the field of nuclear power maintenance, and construct a nuclear power maintenance work package to generate an intelligent agent; The intelligent agent that generates the nuclear power plant maintenance work package calls the work package to generate an evidence set and generates a structured semantic representation of the evidence; Based on the semantic representation of structured evidence, and combined with the relationship between nuclear power plant maintenance entities and the information on maintenance task types, a maintenance procedure sequence is generated. Based on the maintenance process sequence and resource element information in the nuclear power maintenance dataset, a mapping relationship between maintenance processes and resource elements is established, and resource element allocation results are generated. The maintenance procedure sequence and resource allocation results are processed into task arrangement to generate candidate nuclear power plant maintenance work packages.

[0012] Optionally, obtaining the nuclear power plant maintenance work package specifically includes: The structural analysis of the maintenance procedure sequence, resource element allocation results and work package generation task context in the candidate nuclear power maintenance work package is performed to generate a set of consistency verification objects. The consistency verification object set is compared step by step to determine the consistency status between the maintenance process sequence and the resource element allocation result; Based on the set of consistency verification objects, determine the conformity status between the maintenance procedure sequence and the work package generation task context; Based on the consistency and compliance status, the maintenance procedure sequence and resource element allocation results in the candidate nuclear power maintenance work package are adjusted for consistency, and the corrected candidate nuclear power maintenance work package is generated, thus forming the nuclear power maintenance work package.

[0013] The beneficial effects of this invention are: This invention constructs a unified preprocessing mechanism for multi-source nuclear power plant maintenance data. It standardizes data from various sources, such as equipment ledgers, historical fault records, maintenance procedures, safety regulations, and equipment drawings. Furthermore, it extracts nuclear power plant maintenance knowledge to form a dataset and constraint graph for the nuclear power plant maintenance field. This achieves a unified organization of nuclear power plant maintenance data from unstructured and weakly structured to structured representation, effectively solving the problems of complex data sources, inconsistent formats, and difficulty in collaborative utilization in existing technologies.

[0014] This invention introduces a training method combining full-parameter fine-tuning of a domain-wide model with low-rank LoRA adaptation, and incorporates a reasoning guidance mechanism built using nuclear power plant maintenance constraint graphs. This allows the model to retain general language capabilities while injecting specialized nuclear power plant maintenance knowledge, thereby enhancing its understanding and reasoning abilities regarding nuclear power equipment failure mechanisms, maintenance procedure logic, and resource allocation relationships. This effectively overcomes the problems of insufficient adaptability and low reliability of generated content in existing general-purpose models within nuclear power plant scenarios. Furthermore, this invention constructs a work package generation task context through structured parsing of corrective work order preparation requests and employs a retrieval-enhanced generation mechanism to form a work package generation evidence set. This allows the model generation process to be constrained and supported by real knowledge units in the nuclear power plant maintenance domain dataset, reducing unfounded generation and information omissions. In addition, the work package generation agent orchestrates and executes tasks related to maintenance procedure decomposition and resource element matching, establishing a structured correspondence between maintenance procedures, resource elements, and the working environment, thus improving the completeness and executability of the work package content. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a method for generating nuclear power plant maintenance work packages based on a large domain model, as proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction of a trained nuclear power maintenance domain-wide model, which is part of the nuclear power maintenance work package generation method based on a domain-wide model proposed in this invention. Figure 3 This diagram illustrates the construction of the task context for a nuclear power plant maintenance work package generation method based on a domain-wide large model, as proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A method for generating nuclear power plant maintenance work packages based on a domain-wide large model includes the following steps: Obtain multi-source datasets for nuclear power plant maintenance and perform preprocessing to obtain standard datasets for nuclear power plant maintenance; Nuclear power maintenance knowledge is extracted from the nuclear power maintenance standard dataset to construct a nuclear power maintenance domain dataset and a nuclear power maintenance constraint graph. Based on a general large model, we use a dataset from the nuclear power maintenance domain to perform domain adaptation training that combines full parameter fine-tuning with LoRA low-rank adaptation, inject specialized nuclear power maintenance knowledge, and design a constraint-aware loss function and inference guidance mechanism based on the nuclear power maintenance constraint graph. We then perform inference reinforcement training on the general large model after domain adaptation training to obtain a large model for the nuclear power maintenance domain. Receive a corrective work order preparation request, extract the target nuclear power equipment, fault information, maintenance task type and work environment information from the corrective work order preparation request, and form a work package generation task context; The task context for generating work packages is input into a large model in the field of nuclear power maintenance for retrieval enhancement and generation, forming a work package generation evidence set. The large model in the field of nuclear power maintenance is encapsulated into a nuclear power maintenance work package to generate an intelligent agent. The evidence set generated by the work package is combined to decompose the maintenance process and match resource elements to generate candidate nuclear power maintenance work packages. The candidate nuclear power plant maintenance work packages are subjected to consistency verification, and corrections are made based on the verification results to obtain the nuclear power plant maintenance work packages.

[0018] In this embodiment, the multi-source dataset for nuclear power plant maintenance includes basic information of nuclear power equipment, operating status data of nuclear power equipment, historical fault records, historical maintenance records, maintenance procedures, safety specifications, equipment drawings and historical work packages. Preprocessing includes document format conversion, duplicate data removal, missing data completion, abnormal data correction, unified equipment identification, time information alignment and normalization of nuclear power plant maintenance terminology.

[0019] In this embodiment, the construction of the nuclear power plant maintenance dataset and the nuclear power plant maintenance constraint diagram specifically includes: The nuclear power plant maintenance standard dataset is divided into content segments to obtain a set of nuclear power plant maintenance knowledge units. The process of obtaining the nuclear power maintenance knowledge unit set is as follows: The data is parsed according to the document structure information in the nuclear power maintenance standard dataset. The basic information of nuclear power equipment, historical fault records, maintenance records, maintenance procedures, safety specifications, equipment drawings and historical work packages are segmented according to title level, chapter division, paragraph boundary and field area. Each independent semantic expression unit or structured field content is divided into an independent knowledge unit, and the knowledge units are uniformly numbered and structured, so that each knowledge unit corresponds to only a single maintenance semantic, thus forming the nuclear power maintenance knowledge unit set. Entity identification and entity attribute extraction are performed on the nuclear power plant maintenance knowledge unit set to generate a nuclear power plant maintenance entity set. The specific process for generating the nuclear power plant maintenance entity set is as follows: Lexical segmentation and semantic boundary analysis are performed sequentially on the text content of the nuclear power plant maintenance knowledge unit set to locate key nouns or phrases representing equipment, faults, maintenance procedures, resource elements, safety measures, and quality acceptance. These key nouns or phrases are then used as entity recognition results. For each entity recognition result, relevant descriptive information, parameter information, and limiting information are extracted from its corresponding knowledge unit context, including equipment model, fault type, procedure sequence information, resource category, safety level, and acceptance index information. The entity recognition results are then bound one-to-one with their corresponding attribute information to form the nuclear power plant maintenance entity set. Based on the set of nuclear power plant maintenance entities, the relationships between these entities are identified to form a set of nuclear power plant maintenance relationships. The process of forming a nuclear power plant maintenance relationship set is as follows: Based on the attribute information and semantic descriptions between equipment entities, fault entities, maintenance procedure entities, resource element entities, safety measure entities, and quality acceptance entities in the nuclear power plant maintenance entity set, a matching analysis is performed on the correspondence between different entities. Fault associations are determined through the correspondence information between equipment entities and fault entities; procedure matching relationships are determined through the correspondence information between fault entities and maintenance procedure entities; procedure sequence relationships are determined through the time sequence and execution order between maintenance procedure entities; resource matching relationships are determined through the configuration requirements between maintenance procedure entities and resource element entities; safety association relationships are determined through the safety requirements between maintenance procedure entities and safety measure entities; and acceptance association relationships are determined through the correspondence of acceptance indicators between maintenance procedure entities and quality acceptance entities. All of the above association relationships are then uniformly structured and organized to form a nuclear power plant maintenance relationship set. The nuclear power maintenance knowledge unit set, nuclear power maintenance entity set, and nuclear power maintenance relationship set are linked and organized, and the source mapping relationship between each entity in the nuclear power maintenance entity set and the corresponding nuclear power maintenance knowledge unit is established to obtain the nuclear power maintenance domain dataset. The nuclear power plant maintenance dataset is obtained as follows: Based on the knowledge unit numbers in the nuclear power plant maintenance knowledge unit set, each entity in the nuclear power plant maintenance entity set is associated with the corresponding nuclear power plant maintenance knowledge unit, and each relationship in the nuclear power plant maintenance relationship set is associated with the corresponding entity, forming a structured data record composed of nuclear power plant maintenance knowledge units, entities, entity attributes, and entity association relationships. Based on the document source, chapter position, paragraph position, table field position, or drawing annotation position of each entity in the nuclear power plant maintenance knowledge unit, corresponding source location information is generated, and the source location information is bound to the corresponding entity, establishing a source mapping relationship between each entity and the corresponding nuclear power plant maintenance knowledge unit. The structured data records are classified and stored according to equipment entities, fault entities, maintenance procedure entities, resource element entities, safety measure entities, and quality acceptance entities, retaining the nuclear power plant maintenance entity association relationships and source mapping relationships between each entity, thus obtaining the nuclear power plant maintenance dataset. The nuclear power plant maintenance constraint graph is constructed by setting each entity in the nuclear power plant maintenance entity set as a graph node, setting each relation in the nuclear power plant maintenance relation set as a graph edge, and setting the entity attributes corresponding to each entity as the node attributes of the corresponding graph node.

[0020] In this embodiment, the large-scale model for nuclear power plant maintenance is obtained specifically through: Construct a nuclear power maintenance training sample set based on a dataset from the nuclear power maintenance field; The construction process of the nuclear power plant maintenance training sample set is as follows: Nuclear power plant maintenance knowledge units, entities, entity attributes, entity relationships, and source mapping relationships are read from the nuclear power plant maintenance domain dataset. The entities, entity attributes, and entity relationships are then backfilled into the corresponding nuclear power plant maintenance knowledge units according to the source mapping relationships, forming nuclear power plant maintenance sample content that maintains the correspondence between the original semantics and structured knowledge. Based on the equipment information, fault information, maintenance procedures, resource elements, safety measures, and quality acceptance content contained in the nuclear power plant maintenance knowledge units, tasks for nuclear power plant maintenance knowledge completion, entity attribute generation, entity relationship identification, and maintenance content generation are constructed respectively. The task instructions are combined with the corresponding nuclear power plant maintenance knowledge units as sample inputs, and the corresponding nuclear power plant maintenance entities, entity attributes, entity relationships, or maintenance content are used as sample outputs. The sample inputs and outputs are then formatted, length-segmented, and duplicate samples are removed to obtain the nuclear power plant maintenance training sample set. Full parameter fine-tuning of the general large model is performed, and LoRA low-rank adaptation is introduced into the attention parameter structure of the general large model. Through joint training of nuclear power maintenance expertise injection and local efficient adaptation, the domain-adapted general large model is obtained. The process of obtaining the general-purpose large model after domain adaptation training is as follows: The general-purpose large model is a generative language model trained on a large-scale general text corpus, consisting of a word embedding layer, multiple self-attention layers, a feedforward network layer, and an output layer, and possessing text understanding and text generation capabilities; the sample input from the nuclear power plant maintenance training sample set is converted into a word sequence and input into the general-purpose large model. Input features are generated through the word embedding layer, semantic relationships between words are calculated through multiple self-attention layers, feature transformation is performed through the feedforward network layer, and a predicted output is generated through the output layer. The predicted output is compared word-by-word with the corresponding sample output to calculate the content generation error. Based on the content generation error, the parameters of the word embedding layer, self-attention layer, feedforward network layer, and output layer are updated in reverse, completing the injection of specialized nuclear power plant maintenance knowledge; After completing the injection of specialized nuclear power plant maintenance knowledge, the updated parameters of the general large model are kept unchanged. A LoRA low-rank adaptation structure is introduced into the query projection parameters, key projection parameters, value projection parameters, and output projection parameters of the self-attention layer. The LoRA low-rank adaptation structure includes a dimension reduction parameter matrix and a dimension increase parameter matrix. The dimension reduction parameter matrix compresses the high-dimensional input features in the self-attention layer into a low-rank feature space with a dimension smaller than the high-dimensional input features, and extracts local features related to the nuclear power plant maintenance task. The dimension increase parameter matrix restores the local features in the low-rank feature space to a high-dimensional feature space with the same dimension as the output of the self-attention layer, and then superimposes the restored high-dimensional features with the original output of the corresponding projection parameters. The dimension reduction parameter matrix and the dimension increase parameter matrix are updated using the nuclear power plant maintenance training sample set to obtain the general large model after domain adaptation training. Based on the nuclear power plant maintenance constraint graph, reasoning path mining and topology construction are performed. Complete nuclear power plant maintenance reasoning paths are marked as forward nuclear power plant maintenance reasoning paths, and nuclear power plant maintenance reasoning paths with missing nodes are marked as reverse nuclear power plant maintenance reasoning paths, thus forming a nuclear power plant maintenance reasoning training sample set. The process of forming the nuclear power plant maintenance reasoning training sample set is as follows: Based on the graph connection relationships between equipment entities, fault entities, maintenance procedure entities, resource element entities, safety measure entities, and quality acceptance entities in the nuclear power plant maintenance constraint graph, the paths in the graph are traversed and searched. Complete paths formed by sequentially connecting multiple entity nodes are extracted as candidate reasoning paths. Paths that satisfy continuous connection relationships between equipment entities and fault entities, fault entities and maintenance procedure entities, maintenance procedure entities and resource element entities, maintenance procedure entities and safety measure entities, and maintenance procedure entities and quality acceptance entities are marked as forward reasoning paths for nuclear power plant maintenance. Paths with missing entity nodes, broken connections, or missing associated edges during the path traversal are marked as reverse reasoning paths for nuclear power plant maintenance. The forward reasoning paths and reverse reasoning paths are then organized together to form the nuclear power plant maintenance reasoning training sample set. The nuclear power plant maintenance reasoning training sample set is input into the general large model after domain adaptation training. The constraint-aware loss function that integrates the positive reasoning fitting loss and the negative reasoning penalty loss is constructed in combination with the nuclear power plant maintenance constraint graph. The parameters of the general large model after domain adaptation training are updated according to the constraint-aware loss function to obtain the general large model after reasoning reinforcement training. The process of obtaining the general large model after inference reinforcement training is as follows: Input the nuclear power plant maintenance inference training sample set into the domain-adapted general large model; calculate the hidden state corresponding to the nuclear power plant maintenance inference path content before the current entity node; perform an inner product operation between the hidden state and the output parameters corresponding to each candidate entity node in the nuclear power plant maintenance constraint graph to obtain the generation score of each candidate entity node; perform exponential normalization on the generation score to obtain the generation probability of each candidate entity node as the next generated entity; extract the generation probability of each next entity node according to the connection order of entity nodes in the forward inference path of nuclear power plant maintenance; calculate the average value after taking the negative logarithm of each generation probability to obtain the forward inference fitting loss; for the reverse inference path of nuclear power plant maintenance, determine the missing entities. The current entity node before the node, the missing entity node, and the subsequent entity nodes after the missing entity node are considered. The backward inference penalty loss is calculated based on the difference between the generation probability of the current entity node generating the missing entity node and the generation probability of the current entity node directly generating the subsequent entity node. The weighted forward inference fitting loss is added to the weighted backward inference penalty loss to obtain the constraint-aware loss function. The loss gradients of the dimensionality reduction parameter matrix and the dimensionality increase parameter matrix are calculated based on the constraint-aware loss function. The dimensionality reduction parameter matrix and the dimensionality increase parameter matrix are updated in the direction of decreasing constraint-aware loss function. When the constraint-aware loss function does not decrease for 10 consecutive training rounds, it is determined that the constraint-aware loss function has reached the convergence state, and the general large model after inference reinforcement training is obtained. During the training process of the general large model after inference reinforcement training, the generated content is mapped to the nuclear power maintenance constraint graph. The generated entities are determined according to the relationship between nuclear power maintenance entities, and real-time intervention and adjustment are carried out to build a constraint graph-driven inference guidance mechanism to obtain a large model in the field of nuclear power maintenance. The process of obtaining the large-scale model in the nuclear power plant maintenance field is as follows: During the training of the general large-scale model after inference reinforcement training, entity recognition is performed on the currently generated content. The identified nuclear power plant maintenance entities are mapped to corresponding entity nodes in the nuclear power plant maintenance constraint graph, and these corresponding entity nodes are identified as the current entity nodes. Adjacency search is performed along the nuclear power plant maintenance entity association relationships corresponding to the current entity node to extract entity nodes directly connected to the current entity node, forming a graph candidate entity set. The candidate generated entities and their corresponding generation probabilities output by the general large-scale model after inference reinforcement training are read. Each candidate generated entity is matched with the graph candidate entity set. The generation probability of candidate generated entities belonging to the graph candidate entity set is increased, while the generation probability of candidate generated entities not belonging to the graph candidate entity set is decreased. The adjusted generation... The probability is normalized to determine the generated entity. When there is no nuclear power maintenance entity association between the generated entity and the current entity node, the output of the corresponding generated content is stopped, and the process is backtracked to the most recent fully connected current entity node. The missing entity node is then filled in according to the nuclear power maintenance constraint graph, and the generated entity is re-determined. The path guidance for equipment entities, fault entities, maintenance procedure entities, resource element entities, safety measure entities, and quality acceptance entities is completed in the above manner in sequence, so that the generated content gradually unfolds along the fully connected nuclear power maintenance reasoning path. The entity mapping, adjacency search, generation probability adjustment, node missing identification, and path backtracking and filling process constitute a constraint graph-driven reasoning guidance mechanism. The reasoning guidance mechanism is configured in the training process of the general large model after reasoning reinforcement training to obtain the large model of the nuclear power maintenance domain.

[0021] In this embodiment, the formation of the work package generation task context specifically includes: Receive corrective work order preparation requests and perform structured parsing on the corrective work order preparation requests to obtain structured representation data of the work order; The process of obtaining structured representation data for work orders is as follows: Unstructured text in the corrective work order preparation request is cleaned to remove redundant symbols and invalid tags. The text is then segmented into sentences and words, dividing continuous text into text fragments with independent semantic meaning. Semantic units are located for each text fragment. Fragments describing the same event or object are aggregated, and their corresponding time expressions, location expressions, state change descriptions, and operation descriptions are identified. Semantic units are reorganized and numbered according to field positional relationships to generate hierarchical data records. These hierarchical data records are then encoded and stored according to a unified data format, forming structured representation data for work orders with hierarchical structure and field identifiers. Based on the structured representation of work orders, the target nuclear power equipment identifier, fault information, maintenance task type information, and operating environment information are extracted. The specific process for extracting target nuclear power equipment identification, fault information, maintenance task type information, and operating environment information is as follows: Field parsing and semantic positioning are performed on each level of the work order structured representation data; the identification fields related to the equipment in the structured representation data are matched and compared with the unique code of the nuclear power equipment to determine the target nuclear power equipment identification; semantic aggregation is performed on the descriptive fields related to abnormal states or phenomena in the structured representation data to synthesize scattered descriptive information into a unified fault information representation; categorization is performed on the fields related to maintenance behavior or operation category in the structured representation data to summarize maintenance behaviors with different expressions into standardized maintenance task type information; and information extraction and classification are performed on the fields related to environmental conditions in the structured representation data to integrate descriptions related to temperature, pressure, space limitations, and operating status into operating environment information. The target nuclear power equipment identifier is matched with the equipment entities in the nuclear power maintenance standard dataset to generate standard equipment entity identifiers; The specific process for generating standard equipment entity identifiers is as follows: The target nuclear power equipment identifier undergoes character normalization processing to unify the identifier encoding format and extract equipment name and structural features; after retrieving a set of candidate equipment entities from the nuclear power maintenance standard dataset, the target nuclear power equipment identifier and each candidate equipment entity are vectorized, mapping equipment name features, equipment type features, and structural features into multi-dimensional feature vectors; the similarity matching degree between the target nuclear power equipment identifier vector and each candidate equipment entity vector is calculated, obtained by calculating the cosine similarity between two vectors, dividing the dot product of the two vectors by the product of their respective vector magnitudes; based on the similarity matching degree, the equipment entity corresponding to the maximum matching value is selected from the candidate equipment entity set as the matching result, and the standard encoding of the equipment entity is bound and replaced with the target nuclear power equipment identifier to generate the standard equipment entity identifier; The fault information is standardized based on the standard equipment entity identifier to obtain a standard fault information representation; The process of obtaining standard fault information representation is as follows: Based on the standard equipment entity identifier, the corresponding equipment entity is located in the nuclear power maintenance standard dataset, and historical fault records associated with the equipment entity are extracted to form a candidate fault information set. The fault information is then segmented into sentences, terms, and synonyms are merged, and units of measurement are standardized. Fault phenomena, fault locations, state changes, alarm codes, and occurrence times are extracted. Each candidate fault information in the candidate fault information set is represented as a term weight vector. The cosine similarity between term weight vectors is calculated to obtain the fault semantic similarity. The fault attribute consistency is obtained by calculating the proportion of consistent terms in fault location, state changes, and alarm codes to the total number of corresponding fields. The fault semantic similarity and fault attribute consistency are weighted and summed to obtain the fault matching degree for each candidate fault information. The candidate fault information with the highest fault matching degree is determined as the target fault information. The standard fault name, standard fault code, fault location, and state change from the target fault information are used to replace the corresponding descriptions in the fault information. The replaced fault information is then associated with the standard equipment entity identifier to obtain the standard fault information representation. The standard equipment entity identifier, standard fault information representation, maintenance task type information, and work environment information are structurally combined to generate a work package generation task context; The process of generating a task context for a work package is as follows: Standard equipment entity identifiers, standard fault information representations, maintenance task type information, and operating environment information are encoded into fields, and different information units are uniformly mapped to structured field vector representations. Semantic alignment is performed on each field vector based on the field type, aggregating field information representing the same nuclear power plant maintenance object into the same data record unit. The aggregated data record unit's field order is rearranged, with the equipment identifier field serving as the primary key field, the fault information field establishing an association mapping with the equipment identifier field, the maintenance task type field serving as the process control field, and the operating environment information serving as the constraint description field, all participating in the structured organization. The rearranged structured data records are then encapsulated and stored according to a unified data format, forming a task context for generating a work package with a primary key association structure and semantically aligned fields.

[0022] In this embodiment, the formation of the evidence set for the work package specifically includes: Input the task context of the work package generation task into the large model of the nuclear power maintenance domain, generate a semantic representation of the work package generation task, and determine the set of nuclear power maintenance entities based on the semantic representation of the work package generation task. The specific process for determining the nuclear power plant maintenance entity set is as follows: The context of the work package generation task is divided into a sequence of lexical units. Lexical features corresponding to each lexical unit are generated through the word embedding layer of the nuclear power plant maintenance domain model. Semantic relationships between lexical units are calculated through a self-attention layer, allowing the lexical features corresponding to standard equipment entity identifiers, standard fault information representations, maintenance task type information, and operating environment information to be fused together to obtain the semantic representation of the work package generation task. Utilizing the nuclear power plant maintenance entities and their relationships learned during the domain adaptation training process of the nuclear power plant maintenance domain model, the starting position, ending position, and entity category of entities in the semantic representation of the work package generation task are identified. Equipment entities, fault entities, maintenance procedure entities, resource element entities, safety measure entities, and quality acceptance entities directly corresponding to the work package generation task context are extracted. The identified entities are aligned with the nuclear power plant maintenance entities in the nuclear power plant maintenance domain dataset in terms of name, entity attributes, and nuclear power plant maintenance entity relationships. The entity identifiers of the corresponding entities in the nuclear power plant maintenance domain dataset are used to replace the identified entity descriptions, and duplicate entities are merged to form the nuclear power plant maintenance entity set. Map the set of nuclear power plant maintenance entities to the corresponding graph nodes in the nuclear power plant maintenance constraint graph, and perform path expansion to obtain the set of target retrieval paths; The specific process for obtaining the target retrieval path set is as follows: Read the entity identifier and entity category of each entity in the nuclear power plant maintenance entity set; compare the entity identifier with the entity identifier of each node in the nuclear power plant maintenance constraint diagram item by item; determine the corresponding node as the node with the same entity identifier and entity category; using the corresponding node as the starting node of the path, sequentially visit adjacent nodes according to the connection direction of the nuclear power plant maintenance entity association relationship, expanding from the node corresponding to the equipment entity to the node corresponding to the fault entity, from the node corresponding to the fault entity to the node corresponding to the maintenance procedure entity, and from the node corresponding to the maintenance procedure entity to the nodes corresponding to the resource element entity, safety measure entity, and quality acceptance entity, respectively; record the nodes traversed, the nuclear power plant maintenance entity association relationship, and the node connection order during the path expansion process; terminate the cyclic expansion containing duplicate nodes; merge duplicate paths where the node sequence and the nuclear power plant maintenance entity association sequence are the same; and retain continuous paths that connect the corresponding nodes in the nuclear power plant maintenance entity set and contain the node corresponding to the maintenance procedure entity, thus forming the target retrieval path set. Based on the target retrieval path set, the corresponding nuclear power maintenance knowledge unit set is extracted from the nuclear power maintenance domain dataset, and the candidate evidence unit set is formed by association extraction based on the source mapping relationship between the nuclear power maintenance knowledge unit set and the graph nodes. The process of forming the candidate evidence unit set is as follows: First, sequentially read the graph nodes, graph node connection order, and nuclear power plant maintenance entity associations contained in each target retrieval path set. Then, based on the entity identifiers of the entities corresponding to each graph node, locate the nuclear power plant maintenance knowledge units associated with the entity identifiers in the nuclear power plant maintenance domain dataset and aggregate them to form a nuclear power plant maintenance knowledge unit set. Based on the source mapping relationship between each entity and its corresponding nuclear power plant maintenance knowledge unit, read the document source, chapter position, paragraph position, table field position, or drawing annotation position corresponding to each nuclear power plant maintenance knowledge unit. Then, associate and bind the content, corresponding graph node, target retrieval path, graph node connection order, and source position of the nuclear power plant maintenance knowledge unit to form candidate evidence units. Finally, aggregate all candidate evidence units to form a candidate evidence unit set. The candidate evidence unit set is aligned sequentially and redundantly removed to obtain a structured evidence unit set. The process of obtaining the structured evidence unit set is as follows: Read the target retrieval path, graph node connection order, nuclear power plant maintenance entity association, and source location corresponding to each candidate evidence unit. Arrange the candidate evidence units according to the connection order of graph nodes corresponding to equipment entities, fault entities, maintenance process entities, resource element entities, safety measure entities, and quality acceptance entities in the target retrieval path. Then, arrange the candidate evidence units corresponding to the same graph node in order of document source, chapter position, paragraph position, table field position, or drawing annotation position. Compare each of the arranged candidate evidence units. When the entity identifier, nuclear power plant maintenance entity association, and text content of two candidate evidence units are identical, retain the candidate evidence unit with the complete source location. When the text content of two candidate evidence units has an inclusion relationship, retain the candidate evidence unit with more information. Associate and store the candidate evidence units that have completed the sequential arrangement and duplicate content removal with their corresponding graph nodes, target retrieval paths, graph node connection order, and source location to obtain the structured evidence unit set. The structured evidence unit set is integrated and organized with the work package generation task context to generate a work package generation evidence set; The process of generating the work package evidence set is as follows: Read the standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information from the work package generation task context. Assign the standard equipment entity identifier and standard fault information representation to the entity identifiers in the structured evidence unit set. Associate the maintenance task type information with the structured evidence units corresponding to the maintenance procedure entities. Associate the operating environment information with the structured evidence units corresponding to the resource element entities, safety measure entities, and quality acceptance entities. Arrange the structured evidence units associated with the standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information sequentially according to the target retrieval path and graph node connection order corresponding to the structured evidence units. Bind the text content, entity identifier, nuclear power maintenance entity association relationship, and source location of each structured evidence unit to the corresponding work package generation task context field. Remove structured evidence units that are not associated with the work package generation task context. Encapsulate the associated and arranged structured evidence units together with the work package generation task context to generate the work package generation evidence.

[0023] In this embodiment, the generation of candidate nuclear power plant maintenance work packages specifically includes: Functionally encapsulate a large model in the field of nuclear power maintenance, and construct a nuclear power maintenance work package to generate an intelligent agent; The specific process of constructing the intelligent agent for generating nuclear power maintenance work packages is as follows: The input interfaces, output interfaces, and internal reasoning call interfaces of the large-scale model in the nuclear power maintenance domain are modularly encapsulated. During the encapsulation process, the parameter weights, nuclear power maintenance entity recognition capabilities, and nuclear power maintenance entity relationship reasoning capabilities of the large-scale model are functionally bound, enabling it to provide work package generation services. Furthermore, a work package generation evidence set call interface and a maintenance procedure generation execution interface are configured for the large-scale model in the nuclear power maintenance domain. Upon receiving a work package generation task, the model automatically calls the work package generation evidence set to read and reason about nuclear power maintenance entity information, outputting maintenance procedure sequences and resource allocation results, thus forming a nuclear power maintenance work package generation intelligent agent with task scheduling and reasoning execution capabilities. The intelligent agent that generates the nuclear power plant maintenance work package calls the work package to generate an evidence set and generates a structured semantic representation of the evidence; The structured evidence semantic representation generation process is as follows: The nuclear power maintenance work package generating agent reads the candidate evidence units in the evidence set generated by the work package, and parses the nuclear power maintenance entity identifier, nuclear power maintenance entity association relationship and source location corresponding to each candidate evidence unit. The candidate evidence units corresponding to the same nuclear power maintenance entity identifier are aggregated, and the candidate evidence units are sequentially organized according to the graph node connection order corresponding to the nuclear power maintenance entity association relationship. The maintenance process related information, resource element related information, safety measure related information and quality acceptance related information in the candidate evidence units are extracted to form a structured evidence semantic representation consistent with the maintenance task structure. Based on the semantic representation of structured evidence, and combined with the relationship between nuclear power plant maintenance entities and the information on maintenance task types, a maintenance procedure sequence is generated. The maintenance procedure sequence generation process is as follows: the equipment entity, fault entity, and maintenance task type information in the structured evidence semantic representation are parsed to determine the corresponding maintenance task category. Based on the nuclear power maintenance entity association relationship, the association link between the equipment entity and the fault entity is determined. The entity nodes in the association link are sequentially expanded according to the connection direction in the nuclear power maintenance constraint diagram. At each entity node, the maintenance procedure information corresponding to the maintenance task type information is extracted. The continuously associated maintenance procedure information is arranged according to the connection order of the entity nodes to form a maintenance procedure sequence from the equipment entity to the fault entity and then to the maintenance execution process. Based on the maintenance process sequence and resource element information in the nuclear power maintenance dataset, a mapping relationship between maintenance processes and resource elements is established, and resource element allocation results are generated. The specific process for generating resource element allocation results is as follows: Read each maintenance process node in the maintenance process sequence and extract the corresponding process type, execution order, and process input / output requirements for each maintenance process node; retrieve resource element information corresponding to each maintenance process node from the nuclear power maintenance dataset, including maintenance tools, spare parts models, personnel qualifications, and operating conditions; match the resource element information with each maintenance process node one by one; bind the matched resource elements sequentially according to the execution order of the maintenance process nodes, ensuring that each maintenance process node corresponds to a unique or multiple sets of resource element combinations; and record the structured relationship between maintenance process nodes and resource elements to generate resource element allocation results. The maintenance procedure sequence and resource allocation results are processed into task arrangement to generate candidate nuclear power maintenance work packages. The specific process for generating candidate nuclear power plant maintenance work packages is as follows: Each maintenance procedure in the maintenance procedure sequence is organized into a process flow according to its execution order; the corresponding resource elements in the resource element allocation results are bound to each maintenance procedure node; the execution time of each maintenance procedure node is sorted and the work units are divided, so that each maintenance procedure node forms a standard work unit containing execution order and resource configuration; based on the standard work unit, the maintenance process is arranged as a whole to form a complete maintenance execution process structure with sequential constraints and resource configuration relationships; and the maintenance execution process structure is encapsulated as a candidate nuclear power plant maintenance work package.

[0024] In this embodiment, obtaining the nuclear power plant maintenance work package specifically includes: The structural analysis of the maintenance procedure sequence, resource element allocation results and work package generation task context in the candidate nuclear power maintenance work package is performed to generate a set of consistency verification objects. The process of generating the consistency verification object set is as follows: Read the maintenance procedure sequence, resource element allocation results, and work package generation task context from the candidate nuclear power plant maintenance work package; sequentially split each maintenance procedure node in the maintenance procedure sequence; decompose each resource element allocation relationship in the resource element allocation results step-by-step; and parse the equipment entity information, fault information, maintenance task type information, and operating environment information in the work package generation task context; aggregate and align the decomposed maintenance procedure nodes, resource element allocation relationships, and task context fields according to the same maintenance procedure node; and generate a unified number and structured identifier for the aggregation results, so that each maintenance procedure node corresponds to its associated resource element information and task context information, forming a consistency verification object set. The consistency verification object set is compared step by step to determine the consistency status between the maintenance process sequence and the resource element allocation result; The consistency status determination process is as follows: The maintenance process information and resource allocation information corresponding to each maintenance process node in the consistency verification object set are read sequentially. The maintenance process information and resource allocation results under the same maintenance process node are compared item by item to determine whether the resource allocation results cover the tool, spare parts, and personnel configuration requirements of the maintenance process. When the resource allocation results completely match the resource requirement information of the corresponding maintenance process node, it is recorded as a consistent state. When there are missing resources or mismatched resource types, it is recorded as an inconsistent state. The comparison results of all maintenance process nodes are summarized to form the consistency status between the maintenance process sequence and the resource allocation results. Based on the set of consistency verification objects, determine the conformity status between the maintenance procedure sequence and the work package generation task context; The compliance status determination process is as follows: The maintenance procedure information corresponding to each maintenance procedure node in the consistency verification object set is read sequentially, along with the standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information from the work package generation task context. The maintenance procedure information is then compared item by item with the equipment maintenance scope corresponding to the standard equipment entity identifier, the fault handling requirements corresponding to the standard fault information representation, the maintenance operation content corresponding to the maintenance task type information, and the operating conditions corresponding to the operating environment information. When the maintenance procedure information simultaneously meets the equipment maintenance scope, fault handling requirements, maintenance operation content, and operating conditions, the corresponding maintenance procedure node is recorded as compliant; if any item fails to meet these requirements, the corresponding maintenance procedure node is recorded as non-compliant. The comparison results of all maintenance procedure nodes in the maintenance procedure sequence are then summarized to determine the compliance status between the maintenance procedure sequence and the work package generation task context. Based on the consistency and compliance status, the maintenance procedure sequence and resource element allocation results in the candidate nuclear power maintenance work package are adjusted for consistency, and the corrected candidate nuclear power maintenance work package is generated, thus forming the nuclear power maintenance work package. The specific process for forming a nuclear power plant maintenance work package is as follows: Based on the consistency status, locate maintenance process nodes recorded as inconsistent in the maintenance process sequence. Then, based on the resource requirement information of the corresponding maintenance process nodes in the consistency verification object set, supplement, delete, or replace resource elements in the resource element allocation results to address resource deficiencies and resource type mismatches. Based on the compliance status, locate maintenance process nodes recorded as non-compliant in the maintenance process sequence. Then, based on the standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information in the task context generated by the work package, add, delete, replace, or adjust the order of the corresponding maintenance process nodes. Update the resource element allocation results corresponding to the adjusted maintenance process nodes synchronously. Re-determine the consistency and compliance status of the adjusted maintenance process sequence and resource element allocation results until all maintenance process nodes in the maintenance process sequence are in a consistent and compliant state. Finally, write the adjusted maintenance process sequence and resource element allocation results into the revised candidate nuclear power plant maintenance work package to form the nuclear power plant maintenance work package.

[0025] Example 1: Taking the corrective maintenance scenario of a large nuclear power plant's operation and maintenance center as the application background, this nuclear power plant includes two units with a total of approximately 3,860 sets of equipment, of which about 820 sets are critical nuclear island equipment. Common corrective work order types include pump failures, valve jamming, heat exchanger leaks, and electrical control anomalies. This example selects 12,480 corrective work order preparation request data generated within 6 consecutive months as the experimental data source to verify the practical effect of the present invention in the generation of nuclear power plant maintenance work packages.

[0026] In this scenario, approximately 3.2TB of raw data was first collected from multiple sources, including the equipment ledger system, historical fault database, maintenance procedure document library, safety specification library, and equipment drawing library. The data was then cleaned using a unified preprocessing method, including equipment name standardization, fault code standardization, maintenance procedure terminology standardization, and redundant record removal. This resulted in a nuclear power plant maintenance standard dataset, containing approximately 1.8 million records of valid structured data and approximately 950,000 segments of unstructured text.

[0027] In the knowledge extraction stage, entity recognition and relationship extraction were performed on the standard dataset, resulting in approximately 42,600 nuclear power plant maintenance entities, including 8,300 equipment entities, 6,200 fault entities, 18,500 maintenance procedure entities, 6,100 resource element entities, and 3,500 safety and quality related entities. At the same time, a nuclear power plant maintenance constraint graph was constructed, which contains approximately 1.25 million entity association edges to express the logical relationships between equipment, faults, procedures, and resources.

[0028] During the model training phase, a general large model was used as the base. Through full parameter fine-tuning and LoRA low-rank adaptation joint training, approximately 3.2 million nuclear power plant maintenance training samples were learned. A constraint-aware reasoning mechanism was introduced, which improved the model's accuracy in sequencing maintenance procedures from 71.3% of the baseline model to 92.8%.

[0029] During the corrective work order preparation and processing phase, a typical work order is selected from the actual production system: "No. 2 Unit Main Feedwater Pump Vibration Abnormality Alarm". The system automatically extracts the target equipment number as HN-2-FW-013 Main Feedwater Pump, the fault type as bearing vibration exceeding the limit, the maintenance task type as emergency shutdown maintenance, and the working environment as the high radiation restriction area of ​​the nuclear island.

[0030] During the work package generation phase, the retrieval enhancement generation module retrieves approximately 48 relevant evidence units from the nuclear power maintenance dataset, including historical maintenance records, standard dismantling and inspection procedures, safety isolation requirements, and spare parts replacement specifications, forming a work package generation evidence set. Subsequently, the nuclear power maintenance work package generation agent completes the process breakdown and resource matching, generating candidate nuclear power maintenance work packages containing 17 standard maintenance procedures and 23 resource element configurations.

[0031] During the consistency verification and correction phase, the candidate work packages were verified for process consistency and task compliance. It was found that there were two mismatches in resource configuration of processes and one process sequence deviation. Adjustments were made through the automatic correction mechanism to finally form a complete nuclear power maintenance work package.

[0032] To verify the effectiveness of the present invention, the results of comparing the traditional manual compilation method with the method of the present invention are shown in Table 1 below.

[0033] Table 1 Comparison of Electrical Maintenance Work Package Generation Results

[0034] As shown in Table 1, the present invention significantly outperforms traditional manual methods in both the efficiency and quality of nuclear power plant maintenance work package generation. In terms of generation time, it is reduced from 6.8 hours to 1.2 hours, indicating that the method based on a large domain model and retrieval enhancement mechanism can significantly reduce the time spent on manual data retrieval and organization. Regarding process completeness, it is improved from 87.5% to 98.9%, demonstrating that the present invention can effectively complete the maintenance process chain and reduce omissions.

[0035] In terms of resource matching accuracy, this invention achieves 96.7%, an improvement of approximately 20.6% compared to traditional methods, indicating that the matching mechanism based on nuclear power plant maintenance constraint diagrams and entity relationships can effectively improve the rationality of resource allocation. Regarding safety compliance rate, it is improved to 99.1%, demonstrating that this invention has stronger capabilities in safety constraint information fusion, which helps reduce safety risks during nuclear power plant maintenance.

[0036] Furthermore, the work order error rate decreased from 9.3% to 1.8%, indicating that the consistency verification and correction mechanism can effectively reduce inconsistencies between maintenance procedures and resource allocation, significantly improving the reliability of work packages. This embodiment demonstrates that the present invention can achieve high-efficiency, high-accuracy, and high-safety automatic generation of work packages in complex nuclear power plant maintenance scenarios, possessing significant engineering application value and promotional significance.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating nuclear power plant maintenance work packages based on a domain-wide large model, characterized in that, Includes the following steps: Obtain multi-source datasets for nuclear power plant maintenance and perform preprocessing to obtain standard datasets for nuclear power plant maintenance; Nuclear power maintenance knowledge is extracted from the nuclear power maintenance standard dataset to construct a nuclear power maintenance domain dataset and a nuclear power maintenance constraint graph. Based on a general large model, we use a dataset from the nuclear power maintenance domain to perform domain adaptation training that combines full parameter fine-tuning with LoRA low-rank adaptation, inject specialized nuclear power maintenance knowledge, and design a constraint-aware loss function and inference guidance mechanism based on the nuclear power maintenance constraint graph. We then perform inference reinforcement training on the general large model after domain adaptation training to obtain a large model for the nuclear power maintenance domain. Receive a corrective work order preparation request, extract the target nuclear power equipment, fault information, maintenance task type and work environment information from the corrective work order preparation request, and form a work package generation task context; The task context for generating work packages is input into a large model in the field of nuclear power maintenance for retrieval enhancement and generation, forming a work package generation evidence set. The large model in the field of nuclear power maintenance is encapsulated into a nuclear power maintenance work package to generate an intelligent agent. The evidence set generated by the work package is combined to decompose the maintenance process and match resource elements to generate candidate nuclear power maintenance work packages. The candidate nuclear power plant maintenance work packages are subjected to consistency verification, and corrections are made based on the verification results to obtain the nuclear power plant maintenance work packages.

2. The method for generating nuclear power plant maintenance work packages based on a domain-wide model according to claim 1, characterized in that, The nuclear power maintenance multi-source dataset includes basic information of nuclear power equipment, operating status data of nuclear power equipment, historical fault records, historical maintenance records, maintenance procedures, safety specifications, equipment drawings and historical work packages. The preprocessing includes document format conversion, duplicate data removal, missing data completion, abnormal data correction, unified equipment identification, time information alignment and normalization of nuclear power maintenance professional terminology.

3. The method for generating nuclear power plant maintenance work packages based on a domain-wide model according to claim 1, characterized in that, The construction of the nuclear power maintenance dataset and nuclear power maintenance constraint diagram specifically includes: The nuclear power plant maintenance standard dataset is divided into content segments to obtain a set of nuclear power plant maintenance knowledge units. Entity identification and entity attribute extraction are performed on the nuclear power plant maintenance knowledge unit set to generate a nuclear power plant maintenance entity set. Based on the set of nuclear power plant maintenance entities, the relationships between these entities are identified to form a set of nuclear power plant maintenance relationships. The nuclear power maintenance knowledge unit set, nuclear power maintenance entity set, and nuclear power maintenance relationship set are linked and organized, and the source mapping relationship between each entity in the nuclear power maintenance entity set and the corresponding nuclear power maintenance knowledge unit is established to obtain the nuclear power maintenance domain dataset. The nuclear power plant maintenance constraint graph is constructed by setting each entity in the nuclear power plant maintenance entity set as a graph node, setting each relation in the nuclear power plant maintenance relation set as a graph edge, and setting the entity attributes corresponding to each entity as the node attributes of the corresponding graph node.

4. The method for generating nuclear power plant maintenance work packages based on a large domain model according to claim 1, characterized in that, The specific details of obtaining the large-scale model in the field of nuclear power plant maintenance include: Construct a nuclear power maintenance training sample set based on a dataset from the nuclear power maintenance field; Full parameter fine-tuning of the general large model is performed, and LoRA low-rank adaptation is introduced into the attention parameter structure of the general large model. Through joint training of nuclear power maintenance expertise injection and local efficient adaptation, the domain-adapted general large model is obtained. Based on the nuclear power plant maintenance constraint graph, reasoning path mining and topology construction are performed. Complete nuclear power plant maintenance reasoning paths are marked as forward nuclear power plant maintenance reasoning paths, and nuclear power plant maintenance reasoning paths with missing nodes are marked as reverse nuclear power plant maintenance reasoning paths, thus forming a nuclear power plant maintenance reasoning training sample set. The nuclear power plant maintenance reasoning training sample set is input into the general large model after domain adaptation training. The constraint-aware loss function that integrates the positive reasoning fitting loss and the negative reasoning penalty loss is constructed in combination with the nuclear power plant maintenance constraint graph. The parameters of the general large model after domain adaptation training are updated according to the constraint-aware loss function to obtain the general large model after reasoning reinforcement training. During the training process of the general large model after inference reinforcement training, the generated content is mapped to the nuclear power maintenance constraint graph. The generated entities are determined according to the relationship between nuclear power maintenance entities, and real-time intervention and adjustment are carried out to construct a constraint graph-driven inference guidance mechanism to obtain a large model in the field of nuclear power maintenance.

5. The method for generating nuclear power plant maintenance work packages based on a large domain model according to claim 1, characterized in that, The formation of the task context for generating the work package specifically includes: Receive corrective work order preparation requests and perform structured parsing on the corrective work order preparation requests to obtain structured representation data of the work order; Based on the structured representation of work orders, the target nuclear power equipment identifier, fault information, maintenance task type information, and operating environment information are extracted. The target nuclear power equipment identifier is matched with the equipment entities in the nuclear power maintenance standard dataset to generate standard equipment entity identifiers; The fault information is standardized based on the standard equipment entity identifier to obtain a standard fault information representation; The standard equipment entity identifier, standard fault information representation, maintenance task type information, and operating environment information are structurally combined to generate a work package generation task context.

6. The method for generating nuclear power plant maintenance work packages based on a domain-wide model according to claim 1, characterized in that, The formation of the evidence set generated by the work package specifically includes: Input the task context of the work package generation task into the large model of nuclear power maintenance domain, generate a semantic representation of the work package generation task, and determine the set of nuclear power maintenance entities based on the semantic representation of the work package generation task; Map the set of nuclear power plant maintenance entities to the corresponding graph nodes in the nuclear power plant maintenance constraint graph, and perform path expansion to obtain the target retrieval path set; Based on the target retrieval path set, the corresponding nuclear power maintenance knowledge unit set is extracted from the nuclear power maintenance domain dataset, and the candidate evidence unit set is formed by association extraction based on the source mapping relationship between the nuclear power maintenance knowledge unit set and the graph nodes. The candidate evidence unit set is aligned sequentially and redundantly removed to obtain a structured evidence unit set. The structured evidence unit set is integrated and organized with the work package generation task context to generate the work package generation evidence set.

7. The method for generating nuclear power plant maintenance work packages based on a domain-wide model according to claim 1, characterized in that, The generation of the candidate nuclear power plant maintenance work package specifically includes: Functionally encapsulate a large model in the field of nuclear power maintenance, and construct a nuclear power maintenance work package to generate an intelligent agent; The intelligent agent that generates the nuclear power plant maintenance work package calls the work package to generate an evidence set and generates a structured semantic representation of the evidence; Based on the semantic representation of structured evidence, and combined with the relationship between nuclear power plant maintenance entities and the information on maintenance task types, a maintenance procedure sequence is generated. Based on the maintenance process sequence and resource element information in the nuclear power maintenance dataset, a mapping relationship between maintenance processes and resource elements is established, and resource element allocation results are generated. The maintenance procedure sequence and resource allocation results are processed into task arrangement to generate candidate nuclear power plant maintenance work packages.

8. The method for generating nuclear power plant maintenance work packages based on a domain-wide model according to claim 1, characterized in that, The acquisition of the nuclear power plant maintenance work package specifically includes: The structural analysis of the maintenance procedure sequence, resource element allocation results and work package generation task context in the candidate nuclear power maintenance work package is performed to generate a set of consistency verification objects. The consistency verification object set is compared step by step to determine the consistency status between the maintenance process sequence and the resource element allocation result; Based on the set of consistency verification objects, determine the conformity status between the maintenance procedure sequence and the work package generation task context; Based on the consistency and compliance status, the maintenance procedure sequence and resource element allocation results in the candidate nuclear power maintenance work package are adjusted for consistency, and the corrected candidate nuclear power maintenance work package is generated, thus forming the nuclear power maintenance work package.

Images