Method and device for interpreting multi-layer flow model of nuclear power system, electronic equipment and medium

Abstract

The present disclosure provides a method and device for interpreting a multi-layer flow model of a nuclear power system, electronic equipment and medium. The method comprises: constructing a model object library of a multi-layer flow model and an interpretation knowledge base corresponding to technical data of a nuclear power system; obtaining modeling intent information and establishing an intent binding relationship between the model object and the modeling intent information, and an evidence binding relationship between the model object and the interpretation knowledge segment; in response to a model interpretation request, determining a model object to be interpreted, retrieving relevant model structure information, interpretation knowledge segments and modeling intent information; based on the retrieval result and a preset interpretation condition, constructing a model interpretation context and inputting a large language model to generate a semantic interpretation result. The present disclosure can interpret the multi-layer flow model in combination with the model structure, technical basis and modeling intent, improve the reliability and traceability of the interpretation, and reduce the model reading threshold.

Description

Technical Field

[0001] This disclosure relates to the field of model interpretation technology, and in particular to a method, apparatus, electronic device and medium for interpreting a multilayer flow model of a nuclear power system. Background Technology

[0002] In current engineering practice, the construction of multilevel flow models (MFMs) heavily relies on manual operations by domain experts. Experts must manually draw the model topology based on relevant information about the nuclear power system and their own experience. This process not only generates a graphical structure but also embodies the expert's deep understanding of the "mapping logic from equipment to function," the "selection criteria for relational edge types," and the "target decomposition strategy" (i.e., the "modeling kernel"). However, existing modeling tools can only store graphical structure data. The implicit judgment logic of experts during the modeling process is difficult to preserve with the model and is easily lost due to personnel changes. For third-party personnel who did not participate in the modeling (such as operators, reviewers, trainers, and subsequent modelers), it is often difficult to determine from the graphical structure of the MFM itself which device or process a certain functional node corresponds to, why a certain edge is valid, why a certain target is decomposed in this way, and along which functional paths a certain anomaly will affect the upper-level target. This poses a significant challenge to the reusability and interpretability of the MFM.

[0003] In recent years, large language models have demonstrated outstanding performance in natural language understanding, knowledge extraction, and text generation. The nuclear power field has also accumulated massive amounts of unstructured data, including design documents, operating procedures, and expert experience records. However, existing technologies cannot effectively solve the interpretability problem of MFM models, mainly due to limitations in the following three categories of solutions: Traditional MFM modeling software or graphical tools: These focus on graphical editing, model management, and functional reasoning, lacking a mechanism for semantically binding model elements with design documents, procedural clauses, and modeling intent, and thus cannot automatically generate natural language interpretations that conform to engineering specifications.

[0004] Existing knowledge graph-based retrieval enhancement generation methods, such as GraphRAG-like methods, can organize, retrieve, and generate technical data using entities, relationships, and graph structures. However, they are typically geared towards entity relationship extraction and question answering for general document knowledge. When directly applied to the interpretation of multi-layer flow models in nuclear power systems, without a dedicated representation and processing mechanism for MFM model objects, functional types, target levels, and directed relationships, it is difficult to ensure that the retrieved technical data accurately corresponds to specific MFM objects, and it is also difficult to ensure that the node types, relationship directions, and target support paths in the generated results are consistent with the original MFM model. Furthermore, existing methods typically do not simultaneously establish traceable binding relationships between model objects and nuclear power technology evidence, or between model objects and modeling intentions, thus potentially leading to problems such as mismatched bases, incorrect relationship directions, or interpretation content exceeding the current model boundaries.

[0005] General large model auxiliary tools lack in-depth adaptation to nuclear power field knowledge and understanding of MFM-specific semantics. Without specific constraints, they directly generate interpretations, which can easily lead to "illusion" problems such as deviating from the actual model, reversing the direction of relationships, confusing node types, or lacking a source of evidence. Summary of the Invention

[0006] The technical problem to be solved by this disclosure is to overcome the shortcomings of the prior art in that it is impossible to interpret a multi-layer flow model based on a completed multi-layer flow model, and to provide a method, apparatus, electronic device and medium for interpreting a multi-layer flow model of a nuclear power system.

[0007] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0008] The first aspect of this disclosure provides a method for interpreting a multilayer flow model of a nuclear power system, the method comprising:

[0009] Construct a model object library for multi-layer flow models of nuclear power systems and an explanatory knowledge base for technical data of nuclear power systems;

[0010] Obtain modeling intent information for the multi-layer flow model of the nuclear power system;

[0011] Establish the intent binding relationship between the modeling intent information and the model objects in the model object library, and the evidence binding relationship between the model objects and the explanatory knowledge fragments in the explanatory knowledge base;

[0012] In response to receiving a user's model interpretation request, determine the model object to be interpreted corresponding to the model interpretation request;

[0013] Retrieve association information related to the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship, and the evidence binding relationship;

[0014] Based on the associated information and preset interpretation conditions, a model interpretation context corresponding to the model object to be interpreted is constructed;

[0015] The model interpretation context is input into the large language model to generate the semantic interpretation result of the multi-layer flow model of the nuclear power system.

[0016] Optionally, the interpretation method further includes:

[0017] Based on the model object library, the intent binding relationship, and the evidence binding relationship, the semantic interpretation results of the multi-layer flow model of the nuclear power system are verified for consistency, and the verified semantic interpretation results are obtained.

[0018] Optionally, the interpretation method further includes:

[0019] The interpretation record is recorded in the process of interpreting the multilayer flow model of the nuclear power system. The interpretation record includes at least one of the following: user's model interpretation request, model object to be interpreted, model interpretation context, technical data called, generated semantic interpretation result, consistency verification result, and manual review conclusion.

[0020] In response to the semantic interpretation result being manually reviewed, the semantic interpretation result and the interpretation record are added to the interpretation knowledge base.

[0021] Optionally, the interpretation method further includes:

[0022] In response to detecting a version update of the multi-layer flow model, obtain the difference information between the multi-layer flow model before and after the version update;

[0023] In response to detecting changes in the difference information involving node type, relationship direction, target level, path structure, or key binding data version, the semantic interpretation result associated with the change is marked as expired or pending review; or, in response to detecting changes in the difference information involving name, coordinates, or non-key attributes, the corresponding semantic interpretation result is retained and the version mark of the semantic interpretation result is updated.

[0024] Optionally, the model object library for constructing the multi-layer flow model of the nuclear power system includes:

[0025] Acquire multi-layer flow model data of nuclear power systems, wherein the multi-layer flow model data is derived from modeling software, graph databases, model exchange files, structured tables, or image recognition results after manual correction;

[0026] Attribute information is extracted from the multi-layer flow model data. The attribute information includes at least one of the following: node number, node name, node type, flow type, system to which it belongs, associated device identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler's comments.

[0027] A model object library for constructing a multi-layer flow model of a nuclear power system based on the attribute information is provided. The model object library includes at least one of node objects, edge objects, path objects, target objects, subgraph objects, and version objects.

[0028] Optionally, the explanatory knowledge base corresponding to the technical data for constructing the nuclear power system includes:

[0029] Obtain technical data of the nuclear power system, including at least one of the following: nuclear power system design specifications, operation manual, operating procedures, alarm response procedures, equipment specifications, measurement point list, control logic specifications, training materials, expert confirmation opinions, and version modification records;

[0030] The technical data is preprocessed, including text segmentation, data cleaning, terminology and unit standardization, entity recognition, measurement point information extraction, equipment name normalization, procedure clause annotation, and source number generation.

[0031] A knowledge base for interpreting the technical data of nuclear power systems is constructed based on the preprocessed technical data.

[0032] A second aspect of this disclosure provides an apparatus for interpreting a multilayer flow model of a nuclear power system, the apparatus comprising:

[0033] The first construction module is used to build a model object library for the multi-layer flow model of the nuclear power system and an explanatory knowledge base corresponding to the technical data of the nuclear power system.

[0034] The first acquisition module is used to acquire modeling intent information of the multi-layer flow model of the nuclear power system;

[0035] A module is established to establish the intent binding relationship between the modeling intent information and the model objects in the model object library, as well as the evidence binding relationship between the model objects and the explanatory knowledge fragments in the explanatory knowledge base;

[0036] The determination module is used to determine the model object to be explained corresponding to the model explanation request in response to receiving a model explanation request from a user.

[0037] The retrieval module is used to retrieve related information about the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship, and the evidence binding relationship;

[0038] The second construction module is used to construct a model interpretation context corresponding to the model object to be interpreted based on the association information and preset interpretation conditions;

[0039] The generation module is used to input the model interpretation context into the large language model and generate the semantic interpretation result of the multi-layer flow model of the nuclear power system.

[0040] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement the method for interpreting the multilayer flow model of a nuclear power system as described in the first aspect.

[0041] The fourth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for interpreting the multilayer flow model of a nuclear power system as described in the first aspect.

[0042] The fifth aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the method for interpreting a multilayer flow model of a nuclear power system as described in the first aspect.

[0043] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0044] The positive and progressive effects of this disclosure are as follows: Based on the multi-layer flow model of nuclear power system, this disclosure retrieves the associated information related to the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship, and the evidence binding relationship, and inputs the model explanation context constructed by combining it with the preset explanation conditions into the large language model. This enables the generation of semantic explanation results of the multi-layer flow model in response to the user's model explanation request, thereby improving the reliability of the explanation and reducing the reading threshold of the multi-layer flow model. Attached Figure Description

[0045] Figure 1 A flowchart illustrating the method for interpreting a multilayer flow model of a nuclear power system provided in Embodiment 1 of this disclosure.

[0046] Figure 2 A schematic diagram of the module for interpreting the multilayer flow model of a nuclear power system provided in Embodiment 2 of this disclosure.

[0047] Figure 3 This is a schematic diagram of the electronic device used to implement the interpretation method of the multilayer flow model of a nuclear power system according to Embodiment 3 of this disclosure. Detailed Implementation

[0048] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0049] In this embodiment of the disclosure, prefixes such as "first" and "second" are used only to distinguish different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects.

[0050] Example 1

[0051] Nuclear power plant systems are characterized by a wide variety of equipment, high coupling in process flows, complex operating conditions, stringent safety standards, and strong constraints from operating procedures. Deep physical and logical coupling exists between pumps, valves, heat exchangers, pressure vessels, piping, measuring points, control loops, protection logic, and operating procedures within a nuclear power system. To effectively characterize the functional mechanisms of such complex systems, multi-level flow (MFM) models are commonly used in engineering. This model, from an "objective-function-flow" perspective, uses functional elements such as source, sink, transport, storage, balance, and barrier, combined with means-end relationships, flow connections, and causal relationships, to describe how the system achieves higher-level objectives through specific functions, as well as the transmission paths and anomaly propagation mechanisms of mass / energy flow. Thanks to its clear functional semantics and causal expression capabilities, MFM has been widely applied in nuclear power plant fault diagnosis, consequence analysis, operator support, and accident simulation.

[0052] Figure 1 A flowchart illustrating an interpretation method for a multilayer flow model of a nuclear power system provided in Embodiment 1 of this disclosure is shown below. Figure 1 As shown, the interpretation method includes:

[0053] S1. Construct a model object library for multi-layer flow models of nuclear power systems and an explanatory knowledge base for technical data of nuclear power systems;

[0054] In this embodiment, the completed MFM model is subjected to object-oriented analysis, the nuclear power system data (i.e., the technical data of the nuclear power system) is preprocessed, and a model object library and an interpretation knowledge base are formed respectively.

[0055] Specifically, two types of basic knowledge bases are established: a model object base that carries the structural knowledge of the MFM model itself, and an interpretation knowledge base that carries information from the nuclear power field.

[0056] It should be noted that the knowledge base for explanation can use relational databases, document databases, graph databases, full-text search systems, vector search systems, or a combination of multiple storage methods.

[0057] S2. Obtain the modeling intent information of the multi-layer flow model of the nuclear power system;

[0058] In this embodiment, structured modeling intent information is recorded, transforming implicit knowledge that originally relied on verbal explanations into explicit knowledge that can be retrieved, referenced, and verified. The modeling intent information is represented as: I_q = (iid, intention_type, statement, author, review_status, version), where I_q represents the q-th structured modeling intent information, q represents the sequence number of the modeling intent information record in the modeling intent record set, iid represents the unique identifier of the modeling intent information record, used for retrieval, reference, binding, and version management, intention_type represents the type of modeling intent information, including node abstraction, relationship connection, target decomposition, reasons for not modeling, or version modification, statement represents the modeler's specific explanation of the corresponding modeling intent information, used to record why nodes are abstracted in this way, relationships are connected in this way, or targets are decomposed in this way, review_status represents the review status of the modeling intent information record, including pending review, confirmed, rejected, or other preset statuses, author represents the modeler, domain expert, or other responsible entity who proposed, recorded, or confirmed the modeling intent information, and version represents the model version, record version, or version identifier corresponding to the modeling intent information record.

[0059] S3. Establish the intent binding relationship between modeling intent information and model objects in the model object library, as well as the evidence binding relationship between model objects and explanatory knowledge fragments in the explanatory knowledge base;

[0060] In this implementation, each binding relationship can be independently represented as: B_r=(bid, object_id, bound_id, bound_type, bind_type, score, status), where bid represents the unique identifier of the binding relationship, object_id represents the identifier of the bound model object, bound_id represents the identifier of the explanatory knowledge fragment (kid) or the identifier of the modeling intent information (iid) that establishes the binding with the model object, bound_type represents evidence binding or intent binding, bind_type represents rule binding, entity binding, semantic binding, manual confirmation binding, or graph neighborhood extension binding, score represents the binding confidence of the binding relationship, and status represents the status of the binding relationship as pending confirmation, confirmed, requiring review, or abolished. It should be noted that evidence binding sets and intent binding sets for a certain model object can be formed according to bound_type, and each binding relationship has an independent binding method, confidence level, and status. Binding can be achieved through one or more of the following methods: rule binding (precise matching based on device number, measurement point number, system code, procedure clause number, etc.), entity binding (matching data entity attributes after identification), comprehensive binding (vector similarity recall), manual confirmation binding, and graph neighborhood expansion binding (borrowing data from upstream and downstream, the same path, or the same target when data is insufficient, but labeled as indirect basis).

[0061] The model interpretation system assigns a binding confidence score to each binding relationship and differentiates the strength of the basis based on the binding confidence score: the highest value is used for manual confirmation, a relatively high value for exact rule matching, a medium value for entity matching, the similarity value for semantic recall, and the value for graph neighborhood expansion is the source object confidence score multiplied by a decay coefficient. Data with a binding confidence score not lower than the threshold and a confirmed status is used as direct basis; data with a lower threshold, a pending status, or data from graph neighborhood expansion can only be used as indirect basis and are marked as "insufficient basis" or "only indirect basis" in the interpretation. This binding confidence score is also used for ranking multiple candidate objects and verifying deterministic expressions. The purpose of distinguishing between direct and indirect basis is to enable the large language model, which functions as a virtual modeling expert in a multi-layered flow model, to make deterministic statements where the basis is sufficient and to automatically downgrade the expression where the basis is insufficient, thereby achieving a balance between the sufficiency and prudence of the expression.

[0062] For example, the "water supply pump transport" transport node can be bound to the water supply pump equipment description, water supply system design description, relevant measurement point description, and the modeler's comments on "abstracting the water supply pump mainly as a transport function"; the conditional relationship edge "the opening degree of the regulating valve affects the water supply flow" can be bound to the control logic description, valve position measurement point description, and the clauses on water level regulation in the operation procedures.

[0063] It should be noted that the binding of model objects with modeling intent information and the binding of model objects with explanatory knowledge fragments can be achieved through manual annotation, rule matching, entity recognition, semantic similarity retrieval, knowledge graph relation extraction, graph neighborhood expansion, or a combination of the above methods.

[0064] S4. In response to receiving a user's model interpretation request, determine the model object to be interpreted corresponding to the model interpretation request;

[0065] In this embodiment, the user's reading request (i.e., the user's model interpretation request) is transformed into one or more explicit model objects to be interpreted, and all the materials required for subsequent fusion are gathered for them.

[0066] Specifically, users can submit reading requests via a graphical interface (clicking nodes or relationship edges, selecting subgraphs, choosing paths or target levels, or selecting anomaly propagation paths) or natural language. When selecting from the interface, the model interpretation system directly obtains the object number; when asking questions in natural language, the system performs intent recognition, entity recognition, object recall, and interpretation type determination. Interpretation types cover node meaning, edge meaning, path, target decomposition, modeling basis, procedural basis, anomaly impact, role-playing, alternative modeling, and version changes. Object location rules are as follows: when the question contains node names, device names, measurement point numbers, or target names, the corresponding object is matched first; if an object is already selected on the interface, it is used as the primary object; when the question involves relationships such as "upstream," "downstream," "why is it connected this way," or "which target is affected," related edges and paths are expanded within the neighborhood of the primary object; when multiple candidates are matched, they are sorted by view range, name similarity, binding confidence, and historical context.

[0067] S5. Retrieve related information about the model object to be explained from the model object library, explanation knowledge base, intent binding relationship and evidence binding relationship;

[0068] In this embodiment, the associated information includes, but is not limited to, model structure information, explanatory knowledge fragments, and modeling intent information.

[0069] S6. Construct a model interpretation context corresponding to the model object to be interpreted based on the associated information and preset interpretation conditions;

[0070] In this embodiment, after locating the model object to be explained, a model explanation context is constructed. This context includes a model structure layer, a nuclear power field evidence layer, and a modeling intent layer. The model explanation system retrieves and organizes the associated information constituting these three information layers based on intent binding relationships and evidence binding relationships. The model structure layer includes the attribute information, local graph structure, target level, relationship direction, and model version of the model object to be explained. The nuclear power field evidence layer includes design specifications, operating procedures, equipment specifications, measurement point data, and other explanatory knowledge fragments obtained through evidence binding relationships. The modeling intent layer includes reasons for node abstraction, relationship connection, target decomposition, non-modeling, and version modification obtained through intent binding relationships. The local graph structure depends on the type of model object. Nodes are defined by their direct upstream and downstream nodes, their path and target level, and related measurement point equipment. Edges are defined by their start point, end point, relationship type, direction, conditions for establishment, and adjacent relationships. Paths are defined by their start and end points, the sequence of nodes traversed, the sequence of relationship types, and the target support relationships. Targets are defined by their lower-level sub-targets, supporting functional paths, and their association with safety or operational targets. The data in the model interpretation context is divided into direct data (explicitly bound to the current model object and confirmed) and indirect data (from adjacent model objects, the same path, or the same target level). The two are always distinguished in subsequent interpretation and verification.

[0071] S7. Input the model interpretation context into the large language model to generate the semantic interpretation results of the multi-layer flow model of the nuclear power system.

[0072] In this embodiment, based on intent binding relationships and evidence binding relationships, model structure information, explanatory knowledge fragments, and modeling intent information related to the model object to be explained are retrieved. The retrieved related information is organized according to the model structure layer, nuclear power domain evidence layer, and modeling intent layer, and a model explanation context is constructed in conjunction with preset explanation conditions. The model explanation context is input into the large language model to generate semantic explanation results for the multi-flow model of the nuclear power system. Furthermore, through the above-mentioned retrieval enhancement generation process, the large language model generates semantic explanation results based on the MFM model structure, nuclear power domain evidence, and modeling intent in the current model explanation task, thereby functionally acting as a virtual modeling expert for the multi-flow model. It should be noted that this embodiment binds the scattered nuclear power domain data and the implicit judgments of the modeling expert to specific model objects, providing a traceable basis "for the model object" for the large language model that functionally acts as a virtual modeling expert for the multi-flow model, and making the modeling kernel explicitly saved with the model.

[0073] Specifically, a large language model hint constrained by the MFM model graph structure is constructed based on the model interpretation context. The prompt should include at least: a task description (interpreting the specified MFM model object only from the perspective of a virtual modeling expert of this multi-layer flow model, not an open-ended nuclear power Q&A), attribute information in the multi-layer flow model data (e.g., node number, node name, node type, flow type, system to which it belongs, associated equipment identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler's comments), a local graph structure given in a structured form, modeling intent information for constructing the model object (constituting the expert intent layer, e.g., why nodes are so abstract, why relationships are so connected, why targets are so decomposed, why some objects are not modeled, why the version has been modified), knowledge basis with source numbers and reliability markers (constituting the nuclear power field data layer, e.g., design specifications, operating procedures, alarm response procedures, equipment specifications, measurement point lists, etc.), user roles, and prohibitive constraints (prohibiting the introduction of nodes that do not exist in the current multi-layer flow model, prohibiting changes to relationship directions, prohibiting the presentation of indirect basis as direct basis, prohibiting the omission of key source numbers, and prohibiting the presentation of insufficient basis as definitive conclusions). The prompt can be generated from a template or jointly generated by the rule module and the prompt generation model. The large language model acquires the knowledge required to act as a virtual modeling expert, and prohibitive constraints and graph structure limit its generated content to fall within the scope of the model and its basis. That is, based on intent binding relationships and evidence binding relationships, it retrieves model structure information, explanatory knowledge fragments, and modeling intent information related to the model object to be explained; it organizes the retrieved related information according to the model structure layer, nuclear power domain evidence layer, and modeling intent layer, and constructs a model explanation context in combination with preset explanation conditions. The model explanation context is then input into the large language model, enabling the large language model to not only understand the MFM model graph itself, but also to understand the modeling kernel of "why such a graph is constructed under this nuclear power system," thereby generating semantic explanation results as a virtual modeling expert of this multi-layer flow model.

[0074] The large language model generates semantic interpretation results based on this. The semantic interpretation results include at least the meaning of multi-layer flow model objects, nuclear power system correspondence, modeling rationale, upstream and downstream relationships, source of basis, role-based explanation (emphasizing operational concerns and functional impact for operators, emphasizing source of basis and relationship direction for reviewers, emphasizing the order from objectives to functions for trainers, and emphasizing abstract rationale and version changes for modelers), and insufficient basis prompts.

[0075] To ensure the reliability of the aforementioned large language model, which functions as a virtual modeling expert in the multi-level flow model, in nuclear power safety scenarios and facilitates subsequent automatic verification, this disclosure also requires the output of the semantic interpretation results in a pre-defined structured format. The output is then converted into relation quadruples and data citations, and machine-verified line by line within the MFM model graph structure, ensuring that each part can be parsed item by item. All expressions involving relationships between objects or upstream / downstream relationships are uniformly expressed as relation quadruples of "source object, relation type, target object, relation direction." Where data is cited, the data segment number is uniformly marked, indicating whether it is a direct or indirect basis. Content lacking sufficient basis or requiring verification is given with independent deterministic markers, without being mixed into the main text. Thus, each relation expression and each citation in the semantic interpretation results can be extracted and compared one by one with the multi-level flow model objects, edge directions, and bound data during consistency verification, thereby transforming the prohibitive constraints at the prompt end from "soft constraints that only rely on prompts" into "hard constraints that can be machine-verified at the output end." This structural requirement is the key technical means to make the large language model, which functions as a virtual modeling expert in a multi-layer flow model, "acceptable" in the nuclear power scenario.

[0076] Understandably, based on intent-binding and evidence-binding relationships, the system retrieves model structure information, explanatory knowledge fragments, and modeling intent information related to the model object to be explained. The retrieved related information is organized according to the model structure layer, nuclear power domain evidence layer, and modeling intent layer, and a model explanation context is constructed in conjunction with preset explanation conditions. This model explanation context is input into a large language model to generate semantic explanation results for the multi-layer flow model of the nuclear power system. This allows the large language model to not only understand the multi-layer flow model diagram itself but also the modeling kernel of "why such a model was constructed under this nuclear power system." Thus, it can explain the model's meaning, modeling rationale, and scope of application to third parties, much like a modeling expert, rather than simply reiterating the graphical structure. Its knowledge sufficiency comes from the model object library, explanation knowledge base, intent-binding relationships, and evidence-binding relationships, while its controllability comes from the MFM model diagram structure verification. These two aspects are two sides of the same mechanism, ensuring that the explanation is both sufficient and grounded in the current multi-layer flow model and its basis, meeting the requirements of accuracy, traceability, and reviewability in the nuclear power scenario.

[0077] In this embodiment, the large language model can be replaced by a domain-fine-tuned language model, a retrieval-enhanced generation model, a multi-model combination system, a rule template generation system, a knowledge graph question-answering system, or a small local language model.

[0078] In this embodiment, based on the completion of the nuclear power system MFM model, the model interpretation context constructed by retrieving related information about the model object to be interpreted from the model object library, interpretation knowledge base, intent binding relationship, and evidence binding relationship, combined with preset interpretation conditions, is input into the large language model. This allows the large language model to both inherit the judgment of the modeling expert and answer like the virtual modeling expert of the multi-layer flow model, and be constrained by the current multi-layer flow model structure, generating natural language interpretations that have a source of evidence, have passed consistency verification, and are suitable for reading by different user roles.

[0079] This implementation transforms the MFM model graph structure into a natural language interpretation result through a constrained large language model, enabling third-party personnel who were not involved in the modeling to understand the functional logic of the nuclear power system expressed by the multilayer flow model, thus lowering the reading threshold of the MFM model.

[0080] It should be noted that the output format of the semantic interpretation results of the multi-layer flow model can be text, graph annotation, path animation, hierarchical cards, review reports, or model handover documents.

[0081] In an optional implementation, the interpretation method further includes:

[0082] Consistency verification of the semantic interpretation results of the multi-layer flow model of nuclear power system is performed based on the model object library, intent binding relationship and evidence binding relationship, and the verified semantic interpretation results are obtained.

[0083] In this embodiment, the semantic interpretation results are subjected to consistency verification. If the consistency verification fails, the semantic interpretation results are regenerated, partially corrected, or downgraded, and the consistency verification is performed again on the processed semantic interpretation results, or the semantic interpretation results are transferred to manual review and a manual review conclusion is obtained.

[0084] Furthermore, if the semantic interpretation result passes the consistency check or is confirmed by manual review to be output, the semantic interpretation result is output, and the semantic interpretation result, consistency check result, manual review conclusion, applicable model version, and source version or validity period of the cited knowledge fragment are recorded.

[0085] In this implementation, the free expression of the large language model, which functions as a virtual modeling expert for the multi-layered flow model, is constrained within the MFM model graph structure and binding criteria: the structured semantic interpretation results are automatically verified, and only those that pass are output or submitted for manual review. Consistency verification includes, but is not limited to, model object consistency verification, relation direction consistency verification, and node type consistency verification. Specifically, for model object consistency verification: the nodes, edges, targets, paths, and subgraph numbers of the multi-layered flow model mentioned in the semantic interpretation results are extracted and searched one by one in the model object library; if any of these cannot be found, it is determined that a non-existent object has been introduced, and the model object consistency verification fails; if all of these are found in the model object library, the model object consistency verification passes.

[0086] Relationship direction consistency check: Extract the "source object, relationship type, target object, relationship direction" quadruple from the semantic interpretation results of the multi-layer flow model, find the corresponding edge E_j connecting the same two objects in the edge set E, and compare the interpretation direction with the direction attribute of the corresponding edge E_j; if any of the descriptions such as upstream, downstream, causality, condition, means-end, and target level are inconsistent with the edge direction of the multi-layer flow model, it is marked as a relationship direction conflict, and the relationship direction consistency check fails; this relationship direction consistency check is also used to intercept the application of reverse relationships that only exist on specific edges such as feedback control to objects that do not have such reverse edges.

[0087] Node type consistency check: Map the description of node function in the semantic interpretation result to source, sink, transport, storage, balance, barrier enumeration and compare it with node type; if any is inconsistent (such as interpreting a transport node as having storage function), it is marked as a node type conflict and the node type consistency check fails.

[0088] In addition, the consistency verification of the semantic interpretation results of the multi-layer flow model also includes reference verification (for example, the reference number must belong to the evidence binding relationship between the model object and the model object in the established modeling intent information and the model object in the model object library, and between the model object and the explanatory knowledge fragment in the explanatory knowledge base, and the "direct / indirect" label must be consistent with the binding confidence), terminology consistency verification, deterministic expression verification (for example, content without direct evidence or below the threshold must use non-deterministic expressions such as "maybe", "needs to be reviewed", "insufficient evidence"), and role suitability verification.

[0089] If the consistency check of the semantic interpretation results fails, the system will perform actions such as regeneration, partial correction, deletion of unsubstantiated content, downgrade to a "insufficient basis" warning, or submission for manual review, depending on the error type. For interpretations involving safety-critical functions, conflicting procedures, conflicting relationship directions, or multiple consecutive failures, manual review will be initiated by default. Once the consistency check of the semantic interpretation results passes, the semantic interpretation results of the multi-layer flow model will be output in the form of text descriptions, graphical highlights, path expansions, basis fragment displays, and report paragraphs. Users can continue to submit model interpretation requests based on the current semantic interpretation results (such as "explain the upstream node", "change to the trainer's perspective", "what is the procedural basis for this edge", "why not use a storage node"). The model interpretation system inherits the current model interpretation context and session state, redetermines the model object to be interpreted or expands the local subgraph related to the current object according to the new model interpretation request, and re-retrieves related information based on the existing model object library, interpretation knowledge base, intent binding relationship, and evidence binding relationship. The model interpretation system updates the model interpretation context according to the re-retrieved related information, inputs the updated model interpretation context into the large language model to generate new semantic interpretation results, and performs consistency checks on the new semantic interpretation results. When the model version, technical documents, or binding relationships change, the model interpretation system incrementally updates the corresponding model object library, interpretation knowledge base, or binding relationship.

[0090] The semantic interpretation results generated in this implementation are not a free description of the graph, but are completed under the joint constraints of model objects, local graph structure, bound data, modeling intent, and output format. This disclosure further adopts a two-layer mechanism of "soft constraints at the prompt end + hard verification at the output end." At the prompt end, generation is guided by graph structure, binding, and prohibitive constraints. At the output end, a structured interpretation is required, which is then restored to relation quadruples and data references, returning to the MFM model to compare edge directions, node types, and bound data line by line. Therefore, errors such as fabricated nodes, altered relation directions, confused function types, or references to irrelevant data no longer rely solely on prompt constraints, but can be automatically detected, intercepted, and triggered for regeneration or manual review after output. This transforms illusion control from qualitative guidance to measurable and traceable interception, improving interpretation accuracy and controllability, and reducing the risk of illusions in large language models. Furthermore, each interpretation can be linked to design specifications, operating procedures, equipment specifications, measurement point lists, expert annotations, or modeler instructions, facilitating review and verification and enhancing the traceability of the basis.

[0091] It should be noted that consistency verification can employ rule-based verification, graph traversal verification, structured output verification, data number verification, large language model review, expert manual review, or a combination of these methods.

[0092] In addition, the input format of MFM models can be graph database, XML file, JSON file, tabular node and edge list, file exported from modeling software, manually organized structured model, or model manually corrected after image recognition.

[0093] In an optional implementation, the interpretation method further includes:

[0094] Record the interpretation process of the multi-layer flow model of the nuclear power system. The interpretation record includes at least one of the following: user's model interpretation request, model object to be interpreted, model interpretation context, technical documents called, generated semantic interpretation results, consistency verification results, and manual review conclusions.

[0095] In response to the semantic interpretation results being manually reviewed, the semantic interpretation results and interpretation records are added to the interpretation knowledge base.

[0096] In this implementation, each interpretation and its verification and review process are precipitated as a reusable knowledge asset, and the validity of existing interpretations is maintained during the evolution of the multi-layer flow model.

[0097] Record user questions (i.e., user model interpretation requests), the model object to be interpreted, the model interpretation context, the technical materials called, the generated semantic interpretation results, the consistency verification results, the manual review conclusions, and the output version to form an interpretation record; interpretations that have been manually confirmed or passed multiple reviews can be incorporated into the subsequent interpretation knowledge, further enriching the knowledge sources of virtual modeling experts.

[0098] In an optional implementation, the interpretation method further includes:

[0099] In response to the detection of a version update of the multi-layer flow model, obtain the difference information between the multi-layer flow model before and after the version update;

[0100] In response to the detection of changes in node type, relationship direction, target level, path structure, or key binding data version in the discrepancy information, the semantic interpretation results associated with the change are marked as expired or pending review; or, in response to the detection of changes in name, coordinates, or non-key attributes in the discrepancy information, the corresponding semantic interpretation results are retained and the version markers of the semantic interpretation results are updated.

[0101] In this implementation, when the MFM model is updated, the existing semantic interpretation results are determined based on the object number, object attributes, edge relationships, path sequence, and binding data version to determine whether they are still applicable. If the node type, relationship direction, target level, path structure, or key binding data changes, the semantic interpretation results associated with the change are marked as expired or pending review. If only the name, coordinates, or non-key attributes change, the corresponding semantic interpretation results are retained and the version marking is updated.

[0102] In an optional implementation, the model object library for constructing the multilayer flow model of the nuclear power system in S1 includes:

[0103] Acquire multi-layer flow model data for nuclear power systems. The multi-layer flow model data comes from modeling software, graph databases, model exchange files, structured tables, or image recognition results that have been manually corrected.

[0104] Extract attribute information from the multi-layer flow model data. The attribute information includes at least one of the following: node number, node name, node type, flow type, system to which it belongs, associated device identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler's comments.

[0105] A model object library for constructing a multi-layer flow model of a nuclear power system based on attribute information. The model object library includes at least one of node objects, edge objects, path objects, target objects, subgraph objects, and version objects.

[0106] In this embodiment, the multi-layer flow model data of the nuclear power system that has been drawn up is acquired. The multi-layer flow model data may come from modeling software, graph databases, model exchange files, structured node edge lists, or structured image recognition results after manual correction of images;

[0107] The attribute information extracted from the multi-layer flow model data includes at least the node number, node name, node type, flow type, system to which it belongs, associated device identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler comments.

[0108] The MFM model is parsed into an object-oriented structure, forming a model object library consisting of node objects, edge objects, path objects, target objects, subgraph objects, and version objects. The multi-layer flow model M can be represented as: M = (N, E, G, H, A, P, V); where N represents the set of functional nodes, E represents the set of relational edges, G represents the set of target objects, H represents the set of hierarchical relationships between target objects and functions, A represents the set of attributes of model objects, P represents the set of path objects and subgraph objects composed of functional nodes and relational edges, and V represents the set of model versions, used to identify the state of the model and model objects in different versions.

[0109] Node object N_i and edge object E_j are represented as follows: N_i = (id, name, type, flow, equipment, system, state, tags); E_j = (id, source, target, relation_type, direction, condition, confidence); where i represents the index of the functional node object in the functional node set N, id represents the unique identifier of the i-th functional node object and the unique identifier of the j-th relation edge object, name represents the name of the functional node, used to describe the system function expressed by the node, type represents the MFM model function type including source, sink, transport, storage, balance, barrier, etc., flow represents the flow type corresponding to the functional node including mass flow, energy flow, information flow, control flow, or other types; equipment represents the name, equipment number, or equipment identifier of the nuclear power equipment corresponding to or associated with the functional node, system represents the nuclear power system, subsystem, or functional partition to which the functional node belongs or is involved, state represents the functional state attribute of the functional node under the corresponding model version or operating condition, and tags represent the tags used for linking nodes. This set of auxiliary labels constructs the context for point classification, retrieval, semantic matching, and model interpretation. `source` represents the identifier of the starting object of the relation edge; the starting object can be a node, target, or other connectable model object. `target` represents the identifier of the ending object of the relation edge; the ending object can also be a node, target, or other connectable model object. `relation_type` represents the relation type of the relation edge, including means-end, condition, causality, influence, flow connection, or hierarchical relationship. `direction` indicates that the relation points from the starting object to the ending object and serves as the basis for verifying the consistency of the relation direction. `condition` represents the operating conditions, equipment status, or other constraints under which the relation edge is established, activated, or applicable. `confidence` represents the modeling confidence level of the relation edge and its relation type, which can be represented by numerical values, levels, or manual confirmation markers. Path objects are generated by traversing the graph, recording the starting and ending objects, the sequence of objects traversed, the sequence of relation types, and the applicable operating conditions. Through object-oriented parsing, the type, position, upstream and downstream relationships, target level, and version status of each model object to be interpreted in the model can be clearly retrieved; this is the structural basis for subsequent verification.

[0110] In an optional implementation, the explanatory knowledge base corresponding to the technical data for constructing the nuclear power system in S1 includes:

[0111] Obtain technical data for the nuclear power system, including at least one of the following: nuclear power system design specifications, operation manual, operating procedures, alarm response procedures, equipment specifications, measurement point list, control logic specifications, training materials, expert confirmation opinions, and version modification records.

[0112] Technical data is preprocessed, including text segmentation, data cleaning, terminology and unit standardization, entity recognition, measurement point information extraction, equipment name normalization, specification clause annotation, and source number generation.

[0113] A knowledge base for interpreting the technical data of nuclear power systems is constructed based on the preprocessed technical data.

[0114] In this embodiment, the technical data is preprocessed (e.g., text segmentation, data cleaning, terminology and unit standardization, entity recognition, measurement point information extraction, equipment name normalization, specification clause annotation, source numbering, and index construction) to form an explanatory knowledge base. Each knowledge segment is represented as:

[0115] K_m=(kid, source_type, source_name, paragraph_id, content, entities, system, equipment, measurement, condition, reliability, valid_time);

[0116] Wherein, K_m represents the m-th knowledge fragment in the explanation knowledge base, kid represents the identifier of the knowledge fragment used for retrieval, referencing, binding, and verification based on references, source_type represents the source type of the knowledge fragment (e.g., design specifications, operating procedures, alarm response procedures, equipment specifications, measurement point lists, control logic specifications, training materials, or expert confirmation opinions), source_name represents the name of the specific document, procedure, manual, list, or record from which the knowledge fragment originates, paragraph_id represents the paragraph number, chapter number, clause number, or other location identifier of the knowledge fragment in the source material, content represents the main text content of the knowledge fragment, i.e., the text content used for explanation generation, display based on, and consistency verification, and entities represent the sources from... This knowledge fragment is a collection of identified equipment, systems, measurement points, actions, targets, and other entities in the nuclear power field. `system` represents the nuclear power system, subsystem, or functional area to which the knowledge fragment relates or applies. `equipment` represents the equipment name, equipment number, or equipment identifier involved in the knowledge fragment. `measurement` represents the measurement point name, measurement point number, measurement parameter, or monitoring quantity involved in the knowledge fragment. `condition` represents the applicable operating conditions, equipment status, control mode, or other boundary conditions of the knowledge fragment. `reliability` represents the reliability level or manual verification mark of the knowledge fragment and its source, used to determine whether it can be used as a basis for deterministic interpretation. `valid_time` represents the valid time, applicable version, or validity period of the corresponding data for the knowledge fragment. The interpretation knowledge base can be implemented using relational databases, document databases, graph databases, full-text indexes, vector indexes, or combinations thereof. For safety-critical content, manual verification status, source level, and usage restrictions can be set.

[0117] This implementation method, without altering the multi-layer flow model of the nuclear power system, retrieves relevant information about the model object to be explained from the model object library, interpretation knowledge base, intent binding relationship, and evidence binding relationship. This information, combined with a model interpretation context constructed using preset interpretation conditions, is input into a large language model to generate constrained prompts and semantic interpretation results. After graph structure verification, a semantic interpretation result with supporting evidence and consistency verification is output. This method, through the combined effect of knowledge fusion and graph structure constraints, allows the large language model to effectively act as a virtual modeling expert for the multi-layer flow model. The multi-layer flow model is parsed into nodes, edges, paths, targets, subgraphs, and version objects. Each object is linked to knowledge bases such as design specifications, operating procedures, equipment specifications, measurement point lists, and expert annotations, as well as structured modeling intents (node ​​abstraction reasons, relationship connection basis). The system establishes traceable bindings for (reasons for target decomposition, reasons for not modeling, and version modification descriptions), and distinguishes between direct and indirect evidence based on binding confidence. This ensures that the multi-layer flow model retains not only graphical results but also the modeling kernel, becoming a knowledge source for virtual modeling experts and accumulating the thinking of modelers. Furthermore, it generates semantic interpretation results with different granularities and focuses based on the different needs of operators, reviewers, trainers, and modeling maintenance personnel. Users can gradually understand the node, edge, path, and target levels through point selection, box selection, path selection, or natural language questioning, realizing the transformation from static graph reading to dynamic question-and-answer reading. In addition, for explanations with insufficient evidence, conflicting relationships, or safety-critical aspects, it can automatically mark and submit them for manual review, preventing unconfirmed explanations from directly entering the usage stage, which is beneficial to the safety review of nuclear power systems.

[0118] This implementation method associates the MFM model object and its local graph structure, nuclear power system technical data, and modeling intent information. It dynamically retrieves associated information, constructs model interpretation context, and generates semantic interpretation results for specific model interpretation requests. Then, it uses the MFM model structure, intent binding relationship, and evidence binding relationship to verify the generated semantic interpretation results. Thus, without changing the original MFM model reasoning logic, it improves the readability, traceability, inheritability, and engineering reuse value of the nuclear power system MFM model.

[0119] In a specific example, such as a nuclear power plant feedwater system MFM model, we first import the already completed feedwater system MFM model. This feedwater system MFM model includes the top-level objective "maintaining the steam generator water level," as well as functional nodes such as feedwater supply, feedwater pump transportation, feedwater pressurization, feedwater flow regulation, feedwater heating, deaeration, and steam generator water level balancing. The feedwater system MFM model also includes the correspondence between equipment such as feedwater pumps, regulating valves, deaerators, heaters, pipelines, and steam generators and functional nodes, as well as mass flow, control flow, and conditional relationships.

[0120] The MFM model of the water supply system is parsed using an object-oriented approach, resulting in a model object library that includes node objects, relation edge objects, path objects, target objects, and version objects. For example, the functional nodes related to the water pump are represented as: N001=(N001, water pump transport, transport, mass flow, water pump, water supply system, effective, {water delivery, water pressurization});

[0121] Define "the regulating valve opening meets the water level regulation requirements" as target object O014, which belongs to the target object set G; define "downstream water supply flow regulation function" as function node N016, which belongs to the function node set N; define the conditional relationship edge between target object O014 and function node N016 as E015, which belongs to the relationship edge set E. Relationship edge E015 is represented as: E015 = (E015, O014, N016, condition, O014 points to N016, automatic or manual water level regulation, confirmed). The incorrect direction quadruple is represented as: N016 (downstream water supply flow regulation node), condition, O014 (regulating valve opening condition), N016 points to O014. The corrected quadruple is represented as: O014 (regulating valve opening condition), condition, N016 (downstream water supply flow regulation node), O014 points to N016.

[0122] The target support path between the objective of "maintaining the steam generator water level" and related functional objects is represented as P003. The path sequentially includes objects such as feedwater supply, feedwater pump transport, feedwater flow regulation, and steam generator water level balancing. The target-to-functional path is represented as: P003 = (Path number: P003, starting point: maintaining the steam generator water level, objects passed through: feedwater supply, feedwater pump transport, feedwater flow regulation, steam generator water level balancing, path type: target support path);

[0123] Simultaneously, technical documents such as water supply system design specifications, operation manuals, alarm response procedures, control logic specifications, measurement point lists, equipment specifications, and modeling personnel's drawings were imported. These documents were then segmented, cleaned, entity-identified, terminology-standardized, source-numbered, and labeled with applicable conditions to form an explanatory knowledge base. For example, the section on water supply pump equipment specifications was labeled as Equipment Specification, Water Supply Pump, Delivery, Outlet Pressure; the water supply flow measurement points in the measurement point list were labeled as Measurement Point, Water Supply Flow, Water Supply System; and the clause regarding steam generator water level regulation in the operation procedures was labeled as Operation Procedures, Water Level Control, Water Supply Regulation.

[0124] Acquire and record the modeling intent information proposed by the modelers for key model objects. For example, for N001, record "The feedwater pump mainly undertakes the function of overcoming flow resistance and maintaining feedwater delivery at the current model granularity, so it is abstracted as a transport node"; for E015, record "The opening degree of the feedwater regulating valve is the condition or control factor for the feedwater flow regulation function"; for the top-level objective, record "Maintaining the steam generator water level is an important objective to support the heat transfer of the secondary loop and the stable operation of the unit".

[0125] Subsequently, the evidence binding relationship between model objects and interpretive knowledge fragments, as well as the intent binding relationship between model objects and modeling intents, are established. For example, the feedwater pump transport node N001 is bound to the feedwater pump equipment description, feedwater system design description, feedwater flow measurement point description, and modeling annotation "The feedwater pump mainly undertakes the function of overcoming flow resistance and maintaining feedwater delivery at this model granularity, therefore it is abstracted as a transport node." The regulating valve condition relationship E015 is bound to the regulating valve control logic description, valve position measurement point description, feedwater flow measurement point description, and the clauses on water level regulation in the operating procedures, and the intents of the above relationships are connected: the goal of "maintaining the steam generator water level" is bound to the water level control requirements in the operating manual, the water level deviation handling instructions in the alarm response procedures, and the modeler's explanation of the goal hierarchy division. At this point, the structural knowledge, domain information, and modeling intent information related to N001 are all in place, constituting the material for the model interpretation system to respond to this model object as a "virtual modeling expert."

[0126] When the operator clicks on N001 and asks "What does this node represent?", the model interpretation system receives the model interpretation request from the user, identifies the user role as the operator, locates the model object to be interpreted as N001, and uses N001 as the retrieval target. The system extracts its upstream water supply node, downstream water flow regulation node, target "maintaining steam generator water level", and local path structure from the model object library. It also retrieves equipment descriptions, design descriptions, measurement point descriptions, and modeling intent information that are directly or indirectly bound to N001 from the interpretation knowledge base, intent binding relationships, and evidence binding relationships.

[0127] The model interpretation system constructs a model interpretation context based on the retrieved N001 object attributes, local graph structure, interpretive knowledge fragments, modeling intent information, operator-posed questions, user roles, and preset interpretation rules (i.e., preset interpretation conditions). These preset interpretation rules include: interpreting only based on the current model object, local graph structure, and bound data; providing source numbers for referenced data; providing relation quadruples for relation descriptions; marking indirect or low-confidence bases with uncertainty; and prohibiting the generation of objects or relations that conflict with the current MFM model structure.

[0128] The model interpretation system inputs the model interpretation context into the large language model, generating a structured semantic interpretation result for N001. The semantic interpretation result explains that N001 in the MFM model represents the transport function undertaken by the feedwater pump, corresponding to the process in the actual system where the feedwater pump overcomes flow resistance and transports upstream feedwater to the downstream feedwater pipeline; N001, together with the downstream feedwater flow regulation function, supports the steam generator's water level maintenance target; the feedwater pump is abstracted as a transport node rather than a storage node because, at the current model object granularity, the main function of the feedwater pump is to drive feedwater flow, not to store feedwater. The semantic interpretation result also lists the corresponding object number, relation quadruple, equipment description number, design description number, measurement point description number, modeling intent number, and deterministic marker.

[0129] Consistency checks are performed on the structured semantic interpretation results. Since the semantic interpretation results are given in a structured form, the model interpretation system first extracts all mentioned object IDs, relation quadruples, data reference IDs, and deterministic tags from the structured semantic interpretation results, and then performs checks on model object existence, relation direction consistency, node type consistency, basis reference consistency, and deterministic expression. If the upstream node, downstream node, and target mentioned in the semantic interpretation results all exist in the model object library, the direction of each relation quadruple is consistent with the direction attribute of the corresponding edge in the multi-layer flow model, the node function description is consistent with the node type field of N001, the referenced data are all directly bound data of N001, and the deterministic tags match the binding confidence and data reliability, then the consistency check passes.

[0130] Taking a failed relation direction check and partial regeneration as an example: E015 records its starting point as O014 (regulating valve opening condition) and ending point as N016 (downstream water supply flow regulation node) in the model object library, with the correct direction being "O014 points to N016". Suppose the large language model incorrectly outputs the relation quadruple "N016 (downstream water supply flow regulation node), condition, O014 (regulating valve opening condition), N016 points to O014". The model interpretation system, during relation direction consistency checks, finds that this quadruple is opposite to the direction field of E015, and therefore marks the interpretation as a relation direction conflict. The model interpretation system adds the correct starting point, ending point, and direction of E015 as strong constraints to the model interpretation context, requiring the large language model to only correct the relevant relation representations. After regeneration, the relation quadruple was corrected to "O014 (regulating valve opening condition), condition, N016 (downstream water supply flow regulation node), O014 points to N016". After the object, direction, node type and reference were checked again and found to be consistent, the verification passed.

[0131] When the reviewer clicks on E015 and asks "What is the basis for this condition relationship?", the model interpretation system determines the interpretation type as relationship basis interpretation, retrieves the control logic description, valve position and flow measurement point description, operating procedure clauses, and relationship connection intent bound to E015, and generates a review-type interpretation. The interpretation result states that E015 indicates that the control valve opening is a condition or control factor for the feedwater flow regulation function, lists the corresponding source number, and explains that the relationship direction is "O014 (control valve opening condition) points to N016 (downstream feedwater flow regulation node)". Unless there is another feedback control relationship edge in the current MFM model, it must not be interpreted in reverse as the feedwater flow regulation determining the control valve opening condition.

[0132] When a trainee selects the objective "Maintain steam generator water level" and requests "Explanation by Functional Link," the model interpretation system extracts the target supporting path P003, along with the knowledge fragments and modeling intents bound to each path object, constructing a model interpretation context for the trainee. The generated semantic interpretation results sequentially explain how feedwater supply provides the water source, feedwater pumps provide flow capacity, regulating valves regulate the feedwater flow rate into the steam generator, and the steam generator water level balance reflects the relationship between feedwater inflow and evaporation consumption. The model interpretation system can also simultaneously highlight path P003 in the MFM diagram and display brief explanations next to each node.

[0133] When the modeling maintenance personnel inquired "why the water pump isn't a storage node," the modeling intent information and node type definition of N001 were retrieved, and a semantic interpretation result was generated in the guise of a virtual modeling expert: Storage nodes are typically used to represent functions with storage capabilities, such as containers or level maintenance; the water pump's primary function at the current model object granularity is to provide transport and pressurization capabilities, not to store water, therefore it is abstracted as a transport node. The model interpretation system also explained alternative modeling possibilities: if pump chamber volume or local fluid stagnation is considered in a finer-grained transient model, relevant storage or state nodes can be added, but this alternative approach is not part of the main modeling assumptions of the current version of the model. The reuse of modeling intent information here reflects the process of the modeling kernel being saved with the model and communicated to third parties by the virtual modeling expert.

[0134] For semantic interpretation results that pass the consistency check, the model interpretation system outputs the interpretation content and its source evidence to the user. For results that fail the consistency check, the model interpretation system performs regeneration, partial correction, downgraded expression, or transfers the result to manual review. The model interpretation system also records the user's question, user role, object to be interpreted, semantic interpretation result, source evidence, consistency check result, manual review conclusion, applicable model version, and the source version or validity period of the cited knowledge fragments to support subsequent review and version applicability management.

[0135] When a multi-layer flow model is updated, for example, when N001 is split from a single transport node into two nodes, "feed pump transport" and "feed pump pressurization," the model interpretation system determines that the original historical interpretation of N001 is no longer fully applicable based on changes in object numbering and relational structure, and marks it as requiring review. After confirmation by the modelers, the model interpretation system inherits or splits the original basis for the two new nodes respectively, and generates new interpretation records.

[0136] This implementation extends the nuclear power system MFM model from a professional graphical model into a semantic model that is interactively readable, interpretable, traceable, verifiable, and maintainable. Instead of simply calling upon a large language model to describe the graphics, this approach constructs and verifies constrained prompts through model objectification, data binding, recording of modeling intent, and knowledge fusion. This allows the large language model to generate semantic interpretation results under the shared constraints of the MFM model's graphical structure and the nuclear power system, acting as a virtual modeling expert. This improves the engineering usability of the MFM model in nuclear power system training, review, diagnostic interpretation, and model handover.

[0137] The application of this disclosure is not limited to feedwater systems, but can also be used in nuclear power systems with MFM modeling requirements, such as waste heat removal systems, safety injection systems, reactor coolant systems, steam generator systems, chemical and volume control systems, main steam systems, and condensate systems.

[0138] In addition to online interactive explanation, this disclosure can also be used for offline batch generation of MFM model manuals, MFM model review reports, training materials, handover documents, and diagnostic result interpretation reports.

[0139] Example 2

[0140] Corresponding to the aforementioned embodiment of the method for interpreting a multilayer flow model of a nuclear power system, this disclosure also provides an embodiment of an apparatus for interpreting a multilayer flow model of a nuclear power system.

[0141] Figure 2 This is a schematic diagram of a module for interpreting a multilayer flow model of a nuclear power system, as provided in Embodiment 2 of this disclosure. Figure 2 As shown, the explanatory apparatus includes:

[0142] The first construction module 21 is used to build a model object library for the multi-layer flow model of the nuclear power system and an explanatory knowledge base corresponding to the technical data of the nuclear power system.

[0143] In this embodiment, the completed MFM model is subjected to object-oriented analysis, the nuclear power system data (i.e., the technical data of the nuclear power system) is preprocessed, and a model object library and an interpretation knowledge base are formed respectively.

[0144] It should be noted that the knowledge base for explanation can use relational databases, document databases, graph databases, full-text search systems, vector search systems, or a combination of multiple storage methods.

[0145] The first acquisition module 22 is used to acquire modeling intent information of the multi-layer flow model of the nuclear power system;

[0146] In this embodiment, structured modeling intent information is recorded, transforming implicit knowledge that originally relied on verbal explanations into explicit knowledge that can be retrieved, referenced, and verified. The representation of modeling intent information is detailed in Embodiment 1 and will not be repeated here.

[0147] Module 23 is established to establish semantic intent binding relationships between modeling intent information and model objects in the model object library, as well as evidence binding relationships between model objects and explanatory knowledge fragments in the explanatory knowledge base.

[0148] In this embodiment, the representation and related description of each binding relationship are detailed in Embodiment 1, and will not be repeated here.

[0149] The determination module 24 is used to determine the model object to be explained corresponding to the model explanation request in response to receiving the user's model explanation request;

[0150] In this embodiment, the user's reading request (i.e., the user's model interpretation request) is transformed into one or more explicit model objects to be interpreted, and all the materials required for subsequent fusion are gathered for them.

[0151] For details on the specific implementation process, please refer to the detailed content in Example 1, which will not be repeated here.

[0152] The retrieval module 25 is used to retrieve related information about the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship and the evidence binding relationship;

[0153] In this embodiment, the associated information includes, but is not limited to, model structure information, explanatory knowledge fragments, and modeling intent information.

[0154] The second construction module 26 is used to construct a model interpretation context corresponding to the model object to be interpreted based on the association information and preset interpretation conditions;

[0155] The generation module 27 is used to input the model interpretation context into the large language model and generate the semantic interpretation results of the multi-layer flow model of the nuclear power system.

[0156] In this embodiment, a model interpretation context is constructed after the model object to be interpreted is located. The relevant details are described in detail in Embodiment 1 and will not be repeated here.

[0157] In this embodiment, the large language model can be replaced by a domain-fine-tuned language model, a retrieval-enhanced generation model, a multi-model combination system, a rule template generation system, a knowledge graph question-answering system, or a small local language model.

[0158] In an optional implementation, the interpretation apparatus further includes:

[0159] The verification module is used to perform consistency verification on the semantic interpretation results of the multi-layer flow model of the nuclear power system based on the model object library, intent binding relationship and evidence binding relationship, and obtain the verified semantic interpretation results.

[0160] In this embodiment, the free expression of the large language model, which functions as a virtual modeling expert in a multi-layered flow model, is constrained within the MFM model graph structure and binding basis: the structured semantic interpretation results are automatically verified, and only those that pass the verification are output or submitted for manual review. Consistency verification includes, but is not limited to, model object consistency verification, relation direction consistency verification, and node type consistency verification. For details on the three types of consistency verification, please refer to the detailed description in Embodiment 1, which will not be repeated here.

[0161] In an optional implementation, the interpretation apparatus further includes:

[0162] The recording module is used to record the interpretation records corresponding to the interpretation process of the multilayer flow model of the nuclear power system. The interpretation records include at least one of the following: user's model interpretation request, model object to be interpreted, model interpretation context, technical documents called, generated semantic interpretation results, consistency verification results, and manual review conclusions.

[0163] The processing module is used to add the semantic interpretation results and interpretation records to the interpretation knowledge base in response to the semantic interpretation results being manually reviewed.

[0164] In this implementation, each interpretation and its verification and review process are precipitated as a reusable knowledge asset, and the validity of existing interpretations is maintained during the evolution of the multi-layer flow model.

[0165] In an optional implementation, the interpretation apparatus further includes:

[0166] The second acquisition module is used to acquire the difference information between the multi-layer flow model before and after the version update in response to the detection of a version update of the multi-layer flow model.

[0167] The detection module is used to mark the semantic interpretation results associated with the detected differences as expired or pending review in response to changes in node type, relationship direction, target level, path structure, or key binding data version; or, in response to changes in name, coordinates, or non-key attributes detected differences, to retain the corresponding semantic interpretation results and update the version mark of the semantic interpretation results.

[0168] In this implementation, when the MFM model is updated, the existing semantic interpretation results are determined based on the object number, object attributes, edge relationships, path sequence, and binding data version to determine whether they are still applicable. If the node type, relationship direction, target level, path structure, or key binding data changes, the semantic interpretation results associated with the change are marked as expired or pending review. If only the name, coordinates, or non-key attributes change, the corresponding semantic interpretation results are retained and the version marking is updated.

[0169] In an optional implementation, the first building module includes:

[0170] The first acquisition unit is used to acquire multi-layer flow model data of the nuclear power system. The multi-layer flow model data comes from modeling software, graph database, model exchange files, structured tables, or image recognition results after manual correction.

[0171] The extraction unit is used to extract attribute information from the multi-layer flow model data. The attribute information includes at least one of the following: node number, node name, node type, flow type, system to which it belongs, associated device identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler's comments.

[0172] The first building unit is a model object library for building a multi-layer flow model of a nuclear power system based on attribute information. The model object library includes at least one of node objects, edge objects, path objects, target objects, subgraph objects, and version objects.

[0173] In this embodiment, the multi-layer flow model data of the nuclear power system that has been drawn up is acquired. The multi-layer flow model data may come from modeling software, graph databases, model exchange files, structured node edge lists, or structured image recognition results after manual correction of images;

[0174] For details on object-oriented parsing of the MFM model, please refer to the description in Example 1, which will not be repeated here.

[0175] In an optional implementation, the first building module further includes:

[0176] The second acquisition unit is used to acquire technical data of the nuclear power system. The technical data includes at least one of the following: nuclear power system design specifications, operation manual, operating procedures, alarm response procedures, equipment specifications, measurement point list, control logic specifications, training materials, expert confirmation opinions, and version modification records.

[0177] The preprocessing unit is used to preprocess technical data. Preprocessing includes text segmentation, data cleaning, terminology and unit standardization, entity recognition, measurement point information extraction, equipment name normalization, specification clause annotation, and source number generation.

[0178] The second building unit is used to build an explanatory knowledge base for the technical data of the nuclear power system based on the preprocessed technical data.

[0179] In this embodiment, the relevant content on preprocessing technical data is described in detail in Example 1, and will not be repeated here.

[0180] In a specific example, such as the MFM model of a nuclear power plant feedwater system, the detailed implementation process can be found in Example 1, and will not be repeated here.

[0181] The application of this disclosure is not limited to feedwater systems, but can also be used in nuclear power systems with MFM modeling requirements, such as waste heat removal systems, safety injection systems, reactor coolant systems, steam generator systems, chemical and volume control systems, main steam systems, and condensate systems.

[0182] In addition to online interactive explanation, this disclosure can also be used for offline batch generation of MFM model manuals, MFM model review reports, training materials, handover documents, and diagnostic result interpretation reports.

[0183] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components 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 disclosure according to actual needs.

[0184] Example 3

[0185] Figure 3 This is a schematic diagram of the structure of an electronic device according to Embodiment 3 of this disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the interpretation method of the multilayer flow model of the nuclear power system described in any of the above embodiments. Figure 3 The electronic device 90 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0186] like Figure 3 As shown, the electronic device 90 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 90 may include, but are not limited to: at least one processor 91, at least one memory 92, and a bus 93 connecting different system components (including memory 92 and processor 91).

[0187] Bus 93 includes a data bus, an address bus, and a control bus.

[0188] The memory 92 may include volatile memory, such as random access memory (RAM) 921 and / or cache memory 922, and may further include read-only memory (ROM) 923.

[0189] The memory 92 may also include a program tool 925 (or utility) having a set (at least one) program module 924, such program module 924 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0190] The processor 91 executes various functional applications and data processing by running computer programs stored in the memory 92, such as the interpretation method of the multilayer flow model of the nuclear power system provided in any of the above embodiments.

[0191] Electronic device 90 can also communicate with one or more external devices 94 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 95. Furthermore, electronic device 90 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 96. Figure 3 As shown, network adapter 96 communicates with other modules of electronic device 90 via bus 93. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with electronic device 90, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0192] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0193] Example 4

[0194] Embodiment 4 of this disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for interpreting the multilayer flow model of a nuclear power system provided in any of the above embodiments.

[0195] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0196] Example 5

[0197] Embodiment 5 of this disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the interpretation method of the multilayer flow model of the nuclear power system described in any of the preceding claims.

[0198] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0199] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A method for interpreting a multilayer flow model of a nuclear power system, characterized in that, The interpretation method includes: Construct a model object library for multi-layer flow models of nuclear power systems and an explanatory knowledge base for technical data of nuclear power systems; Obtain modeling intent information for the multi-layer flow model of the nuclear power system; Establish the intent binding relationship between the modeling intent information and the model objects in the model object library, and the evidence binding relationship between the model objects and the explanatory knowledge fragments in the explanatory knowledge base; In response to receiving a user's model interpretation request, determine the model object to be interpreted corresponding to the model interpretation request; Retrieve association information related to the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship, and the evidence binding relationship; The associated information is organized into a model structure layer, a nuclear power field evidence layer, and a modeling intent layer, and a model interpretation context corresponding to the model object to be interpreted is constructed in combination with preset interpretation conditions. The model interpretation context is input into the large language model to construct large language model prompt information constrained by the graph structure of the nuclear power system multilayer flow model. Based on the large language model prompt information, the semantic interpretation result of the nuclear power system multilayer flow model is generated.

2. The method for interpreting the multilayer flow model of a nuclear power system as described in claim 1, characterized in that, The interpretation method further includes: Based on the model object library, the intent binding relationship, and the evidence binding relationship, the semantic interpretation results of the multi-layer flow model of the nuclear power system are verified for consistency, and the verified semantic interpretation results are obtained.

3. The method for interpreting the multilayer flow model of a nuclear power system as described in claim 1, characterized in that, The interpretation method further includes: The interpretation record is recorded in the process of interpreting the multilayer flow model of the nuclear power system. The interpretation record includes at least one of the following: user's model interpretation request, model object to be interpreted, model interpretation context, technical data called, generated semantic interpretation result, consistency verification result, and manual review conclusion. In response to the semantic interpretation result being manually reviewed, the semantic interpretation result and the interpretation record are added to the interpretation knowledge base.

4. The method for interpreting the multilayer flow model of a nuclear power system as described in claim 1, characterized in that, The interpretation method also includes: In response to detecting a version update of the multi-layer flow model, obtain the difference information between the multi-layer flow model before and after the version update; In response to detecting changes in the difference information involving node type, relationship direction, target level, path structure, or key binding data version, the semantic interpretation result associated with the change is marked as expired or pending review; or, in response to detecting changes in the difference information involving name, coordinates, or non-key attributes, the corresponding semantic interpretation result is retained and the version mark of the semantic interpretation result is updated.

5. The method for interpreting the multilayer flow model of a nuclear power system as described in claim 1, characterized in that, The model object library for constructing multi-layer flow models of nuclear power systems includes: Acquire multi-layer flow model data of nuclear power systems, wherein the multi-layer flow model data is derived from modeling software, graph databases, model exchange files, structured tables, or image recognition results after manual correction; Attribute information is extracted from the multi-layer flow model data. The attribute information includes at least one of the following: node number, node name, node type, flow type, system to which it belongs, associated device identifier, measurement point number, relationship edge number, relationship type, relationship direction, target level, subgraph affiliation, model version number, and modeler's comments. A model object library for constructing a multi-layer flow model of a nuclear power system based on the attribute information is provided. The model object library includes at least one of node objects, edge objects, path objects, target objects, subgraph objects, and version objects.

6. The method for interpreting the multilayer flow model of a nuclear power system as described in claim 1, characterized in that, The explanatory knowledge base corresponding to the technical data for constructing nuclear power systems includes: Obtain technical data of the nuclear power system, including at least one of the following: nuclear power system design specifications, operation manual, operating procedures, alarm response procedures, equipment specifications, measurement point list, control logic specifications, training materials, expert confirmation opinions, and version modification records; The technical data is preprocessed, including text segmentation, data cleaning, terminology and unit standardization, entity recognition, measurement point information extraction, equipment name normalization, procedure clause annotation, and source number generation. A knowledge base for interpreting the technical data of nuclear power systems is constructed based on the preprocessed technical data.

7. An interpretation apparatus for a multilayer flow model of a nuclear power system, characterized in that, The interpretation device includes: The first construction module is used to build a model object library for the multi-layer flow model of the nuclear power system and an explanatory knowledge base corresponding to the technical data of the nuclear power system. The first acquisition module is used to acquire modeling intent information of the multi-layer flow model of the nuclear power system; A module is established to establish the intent binding relationship between the modeling intent information and the model objects in the model object library, as well as the evidence binding relationship between the model objects and the explanatory knowledge fragments in the explanatory knowledge base; The determination module is used to determine the model object to be explained corresponding to the model explanation request in response to receiving a model explanation request from a user. The retrieval module is used to retrieve related information about the model object to be explained from the model object library, the explanation knowledge base, the intent binding relationship, and the evidence binding relationship; The second construction module is used to organize the associated information according to the model structure layer, the nuclear power field evidence layer and the modeling intent layer, and to construct a model interpretation context corresponding to the model object to be interpreted in combination with preset interpretation conditions; The generation module is used to input the model interpretation context into the large language model, construct large language model prompt information constrained by the graph structure of the nuclear power system multilayer flow model, and generate the semantic interpretation result of the nuclear power system multilayer flow model based on the large language model prompt information.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the method for interpreting the multilayer flow model of the nuclear power system as described in any one of claims 1 to 6.

9. A 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 method for interpreting the multilayer flow model of the nuclear power system as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for interpreting the multilayer flow model of the nuclear power system as described in any one of claims 1 to 6.

Images