Industrial fault knowledge graph construction and retrieval generation method and system based on evolution ontology

Abstract

The invention discloses an industrial fault knowledge graph construction and retrieval generation method and system based on an evolutionary ontology. The method comprises the steps that entities, attributes and relation information in the entities and the attributes are automatically recognized and extracted from an unstructured industrial document according to an ontology meta-model; designing an ontology construction agent, and automatically discovering, abstracting and constructing new concepts and relationships from the extracted knowledge instances in an incremental learning mode so as to drive the dynamic evolution of the ontology; performing quality evaluation on the candidate ontology by using the ontology review agent, identifying potential problems and providing improvement suggestions; integrating the examined candidate ontology into the global ontology to realize dynamic evolution of the knowledge system; after ontology construction is completed, in the query stage, under the guidance of the latest ontology, semantic analysis and expanding writing are conducted on user query, corresponding entities and relations are extracted, retrieval and multi-hop reasoning are conducted through a knowledge graph, retrieved structured knowledge serves as context to be provided for the large model, and the large model is guided to generate answers.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for constructing and retrieving industrial fault knowledge graphs based on evolutionary ontology. Background Technology

[0002] Industrial question-answering systems have evolved from traditional to intelligent methods. Early industrial question-answering systems typically employed rule-based methods, information retrieval techniques, or machine learning-based models. While these methods are effective in specific scenarios, they also have inherent limitations. Rule-based methods rely on predefined rules and logic from experts, suitable for answering certain types of questions, but lack flexibility and struggle to handle emerging or unexpected questions. Information retrieval techniques focus on finding relevant information from documents or text based on user queries; while they can locate useful content, they often fail to directly provide answers or understand the context of the question. Machine learning-based models train algorithms on historical data to identify patterns and generate answers. However, they typically lack contextual reasoning capabilities and cannot fully understand the complex relationships between domain knowledge.

[0003] To alleviate the illusion problem of LLMs and enhance their performance in specialized domains, retrieval augmentation has become an important means of improving model factual consistency. This method expands the input query by retrieving documents from a specific corpus, thereby providing knowledge support to the model, significantly improving the performance of LLMs, especially in knowledge-intensive tasks. RAG systems still face key challenges in industrial applications. First, document segmentation disrupts the continuity of knowledge, making it difficult for the model to synthesize information from multiple sources, particularly in fault diagnosis scenarios requiring complex reasoning. Second, embedding-based retrieval methods may lose fine-grained contextual relationships, weakening the system's ability to track and integrate knowledge across different documents. Finally, the inconsistent fault description terminology and complex relationships between equipment in industrial fields further exacerbate the problems of poor relevance and insufficient accuracy of retrieval results in traditional RAG systems.

[0004] To overcome the fragmentation and logical gaps caused by unstructured text retrieval, a shift has been made towards a structured knowledge representation method—knowledge graphs—that explicitly encodes complex relationships between entities. As a structured semantic representation connecting entities and relationships through nodes and edges, knowledge graphs have become a key tool supporting large-scale information management and reasoning. By organizing and storing fault knowledge in the form of triples, more accurate and logical contextual information is provided for downstream intelligent diagnosis and question-answering tasks. Knowledge retrieval enhancement based on knowledge graphs integrates the understanding and generation capabilities of large language models with the structured knowledge advantages of knowledge graphs. It not only utilizes vector retrieval but also guides the retrieval process through structured information in the knowledge graph, thereby providing richer contextual semantics and multi-hop reasoning capabilities.

[0005] The value of knowledge graphs highly depends on the quality of their construction. However, manually constructing knowledge graphs is not only inefficient but also difficult to implement when faced with the vast and complex professional corpora in industrial fields. Therefore, automated knowledge graph construction methods have become a core research focus. Early research mainly used deep learning models for knowledge extraction, but these methods typically rely on large amounts of labeled data and have limited generalization capabilities. In recent years, the development of large language models has brought about a paradigm shift in automated knowledge graph construction. LLMs, through self-supervised pre-training on massive amounts of text, have gained powerful semantic understanding and context learning capabilities, enabling them to efficiently identify entities and extract complex structured relationships from unstructured text under zero-shot or few-shot conditions, demonstrating enormous application potential. Although LLMs are powerful in knowledge extraction, ensuring the consistency and factual accuracy of extraction results remains a significant challenge when applied to highly specialized industrial fields. To address this issue, ontology, as a core framework for describing and structuring domain knowledge, is becoming increasingly important. Ontologies provide a unified conceptual constraint and standardized relational "blueprint" for the construction of knowledge graphs. They guide and regulate the knowledge extraction process, ensuring that knowledge from different sources is integrated within a unified semantic framework and supporting subsequent complex analysis and reasoning. Traditional ontology construction and updating processes typically rely on manual intervention or semi-automated tools, which are slow and costly, making it difficult to meet the dynamic requirements of real-time and continuous updates in industrial knowledge applications. Especially in the industrial field, new failure modes, equipment types, and their relationships are constantly emerging, making it difficult to form a closed-loop optimization mechanism from knowledge extraction to ontology evolution.

[0006] To address the above issues, this invention proposes a method for constructing and retrieving industrial knowledge graphs based on evolutionary ontology. Summary of the Invention

[0007] To address the problems of lagging knowledge updates, lack of deep logical reasoning, and poor dynamic adaptability in traditional knowledge question answering systems, this invention provides a method and system for constructing and retrieving industrial fault knowledge graphs based on evolutionary ontology.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for constructing and retrieving an industrial fault knowledge graph based on evolutionary ontology, the method comprising: Automatically identify and extract entities, attributes, and relational information from unstructured industrial documents based on ontology meta-model; The design of the ontology constructs an intelligent agent, which automatically discovers, abstracts and constructs new concepts and relationships from continuously extracted knowledge instances in an incremental learning manner, thereby driving the dynamic evolution of the ontology; An ontology review agent is used to assess the quality of candidate ontologies, identify potential problems, and propose improvement suggestions. The reviewed candidate ontologies are integrated into the global ontology, enabling the dynamic evolution of the knowledge system; Once the ontology is constructed, during the query phase, the user query is semantically parsed and expanded under the guidance of the latest ontology. The corresponding entities and relationships are extracted from the expanded query, and retrieval and reasoning are performed through the knowledge graph. The retrieved structured knowledge is provided as context to the large model, guiding the large model to make full use of the retrieved factual knowledge to generate accurate, professional, and interpretable answers.

[0009] Preferably, methods for designing ontology-based intelligent agents that automatically discover, abstract, and construct new concepts and relationships from continuously extracted knowledge instances through incremental learning, thereby driving the dynamic evolution of the ontology, include: New entity sets added to the knowledge graph With relation set When the scale reaches the threshold for dynamic calculation, the system automatically activates the ontology update and fusion process. This threshold is determined by the following formula: ; in, As a minimum threshold, this ensures that enough instances are accumulated in the early stages of map construction to guarantee the effectiveness of cluster analysis. It is an adaptive scaling factor, which defines the proportion of new knowledge required to trigger an update relative to the total amount of existing knowledge. and These represent the total number of entities and relations in the current knowledge graph; For newly added entities, hierarchical clustering is first performed. Then, for each cluster, a centrality analysis method based on semantic similarity is used to select representative entities. Specifically, the names and descriptions of all entities within the cluster are first vectorized using TF-IDF. Then, the cosine similarity between each entity and all other entities within the cluster is calculated, and the entity with the highest average similarity is selected. The entity serves as the most representative sample in the cluster. Subsequently, the name, type, and semantic description of the representative entity are input into the large language model. The model is guided by prompt templates to summarize and abstract the common features of the entities, thereby generating new candidate concepts. By combining the abstracted set of new concepts and the newly added relationships, the large model is guided to deduce the relationships between the concepts based on the design prompts.

[0010] Preferably, methods for using an ontology review agent to perform quality assessment on candidate ontology, identify potential problems, and propose improvement suggestions include: concept deletion, concept merging, concept renaming, hierarchy adjustment, description optimization, and subclass reconstruction.

[0011] Preferably, the methods for integrating the reviewed candidate ontologies into the global ontology to achieve the dynamic evolution of the knowledge system include: For each new concept in the candidate ontology, a recursive search is first performed; Subsequently, the parent node of the concept is analyzed. If the parent node already exists in the ontology, the new concept is inserted as its child node; if the parent node is missing, the system uses LLM reasoning to determine whether the parent class should be the top-level concept, or to find a parent node that meets the preset requirements.

[0012] Preferably, after the ontology is constructed, during the query phase, guided by the latest ontology, the methods for semantic parsing and expansion of user queries to extract corresponding entities and relations include: Based on the synonym relationships, hyponyms and related concepts in the ontology, the large model is guided to generate extended queries with richer semantics and higher coverage. Based on the expanded query, the key entity set is parsed out. and relation set .

[0013] Preferably, methods for retrieving and reasoning using knowledge graphs, providing the retrieved structured knowledge as context to the large model, and outputting the answer include: Based on vectorized semantic retrieval and multi-hop reasoning technology, the extracted entity set is embedded, similar entity sets are found in the knowledge graph, and multi-hop reasoning is performed in combination with their upstream and downstream relationships. This extends to neighboring nodes and constructs a local subgraph, using node information, attribute information and related descriptions as contextual semantic blocks. The extracted relation list is semantically matched with the candidate relation set in the knowledge graph to obtain relation instances that are the same as or similar to the query relation. The associated triples are obtained from the graph through multi-hop reasoning to form a complete causal evidence chain and expand the relevant entity pairs and contextual information. The retrieved knowledge subgraph contains relevant entities, relationships, and descriptive information, which are used as a structured context and input into the generative model along with the user's original query to generate an answer.

[0014] Preferably, based on vectorized semantic retrieval and multi-hop reasoning techniques, the extracted entity set is embedded, similar entity sets are found in the knowledge graph, and multi-hop reasoning is performed based on their upstream and downstream relationships to expand their topological neighbors and construct a local subgraph. Node information, attribute information, and their related descriptions are used as contextual semantic blocks. The methods for constructing contextual semantic blocks include: ; in, This represents the set of entities extracted in the query. Indicates and Similar entities Representing upstream and downstream neighbors of similar entities, This represents a local subgraph centered around these nodes, containing nodes, edges, and corresponding descriptive information.

[0015] Preferably, the method of semantically matching the extracted relation list with the candidate relation set in the knowledge graph to obtain relation instances that are the same as or similar to the query relation, and obtaining associated triples from the graph through multi-hop reasoning to form a complete causal evidence chain, and expanding the relevant entity pairs and contextual information includes: ; in, This represents the set of entities extracted in the query. Indicates and Similar relationships, This represents triples associated with these similarity relationships. A subgraph representing similarity relationships and associated triples.

[0016] This invention also provides an industrial fault knowledge graph construction and retrieval generation system based on evolutionary ontology. The system is used to implement the aforementioned method and includes: an extraction module, a first evolution module, an evaluation module, a second evolution module, and a retrieval module. The extraction module is used to automatically identify and extract entities, attributes, and their relational information from unstructured industrial documents based on the ontology meta-model. The first evolution module is used to design an ontology-building agent that automatically discovers, abstracts, and constructs new concepts and relationships from continuously extracted knowledge instances through incremental learning, thereby driving the dynamic evolution of the ontology. The evaluation module is used to perform quality evaluation on candidate ontologies using an ontology review agent, identify potential problems and propose improvement suggestions. The second evolution module is used to integrate the reviewed candidate ontology into the global ontology, thereby realizing the dynamic evolution of the knowledge system; The retrieval module is used to perform semantic parsing and expansion of user queries during the query phase after the ontology is constructed, under the guidance of the latest ontology. It extracts the corresponding entities and relationships from the expanded query, performs retrieval and reasoning through the knowledge graph, and provides the retrieved structured knowledge as context to the large model, guiding the large model to make full use of the retrieved factual knowledge to generate accurate, professional and interpretable answers.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) The system achieves automation and dynamic evolution of the knowledge system: Through ontology-based intelligent agent construction and incremental learning mechanisms, the system can automatically identify, abstract, and construct new concepts and relationships from continuous unstructured industrial fault documents without relying on frequent manual intervention. This not only significantly reduces the cost of knowledge graph construction and maintenance but also enables the ontology to adapt in real time to the emergence of new fault modes, equipment types, and relationships in the industrial field, overcoming the shortcomings of traditional methods such as lagging updates and poor adaptability.

[0018] 2) Improved quality and consistency of knowledge graphs: An ontology review agent is introduced to perform multi-dimensional quality assessments of candidate ontologies (including concept deletion, merging, and renaming), effectively identifying and correcting potential semantic conflicts, redundancy, or unreasonable hierarchical issues, ensuring the rigor and logical consistency of the ontology structure. This closed-loop optimization mechanism guarantees the reliability and accuracy of knowledge graphs in industrial fault diagnosis.

[0019] 3) Enhanced semantic understanding capabilities for query parsing and retrieval: During the query phase, the system utilizes the latest ontology to perform semantic parsing and intelligent expansion of user queries. By expanding the query scope through synonym relationships and hyponyms, it effectively solves the fragmentation problem caused by inconsistent terminology in the industrial field. Combined with vectorized semantic retrieval and multi-hop reasoning, the system can accurately extract relevant entities, relationships, and context from the knowledge graph, providing rich and structured knowledge subgraphs as context, significantly improving the coverage and relevance of the answers.

[0020] 4) Improved efficiency of retrieval enhancement generation: By combining structured knowledge graphs with large language models, the system not only relies on unstructured text fragments when generating answers, but also incorporates complex relationships and attribute information between entities. This approach effectively suppresses the "illusion" phenomenon of large models, improves the factual consistency and interpretability of fault diagnosis answers, and is particularly suitable for industrial scenarios that require multi-source information fusion and complex logical reasoning.

[0021] 5) Improved system scalability and efficiency: Adaptive threshold-triggered ontology updates ensure system stability in the early stages of knowledge accumulation while responding efficiently as knowledge volume increases. Furthermore, entity clustering and representative sample selection optimize the concept abstraction process, reducing computational overhead and enabling the system to efficiently process large-scale industrial fault knowledge data.

[0022] In summary, this invention constructs a dynamic, accurate, and efficient industrial fault knowledge graph system through automated ontology evolution, intelligent quality review, and semantically enhanced retrieval, providing reliable technical support for industrial intelligent question answering and fault diagnosis, and has significant practical value and prospects for promotion. Attached Figure Description

[0023] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a method for constructing and retrieving an industrial fault knowledge graph based on an evolutionary ontology, according to an embodiment of the present invention. Figure 2 This is a schematic diagram comparing the experimental results of the ontological evolution ablation in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the results of the ablation experiment of the intelligent agent in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the results of the relationship retrieval ablation experiment in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the results of entity retrieval ablation experiments in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1 like Figure 1 As shown, this invention provides a method for constructing and retrieving industrial fault knowledge graphs based on evolutionary ontology. This method is applicable to complex fault diagnosis and intelligent decision-making scenarios in dynamic knowledge ecosystems such as industrial equipment operation and maintenance and intelligent manufacturing. The method includes the following steps: Knowledge Extraction: Under the semantic constraints provided by the domain ontology, the knowledge extraction module is responsible for automatically identifying and extracting high-quality entities, attributes, and relational information from unstructured industrial documents, providing data support for subsequent knowledge base construction and ontology evolution. First, the document is segmented. The system then uses carefully designed prompts to invoke a large language model to extract entities and relations defined in the ontology from the text.

[0028] Ontology Construction: In a dynamically updated industrial knowledge environment, domain ontologies must possess the ability to continuously learn and update to adapt to the constant emergence of new equipment and new failure modes. Traditional static ontology construction methods clearly cannot meet this requirement. Therefore, this framework designs an ontology construction agent that uses incremental learning to automatically discover, abstract, and construct new concepts and relationships from continuously extracted knowledge instances, thereby driving the dynamic evolution of the ontology. This process is not triggered in real time, but rather automatically activates the ontology construction process when the number of processed documents reaches a preset threshold.

[0029] Multi-agent ontology review: The ontology review agent acts as a "quality evaluator," responsible for conducting a comprehensive quality assessment of candidate ontologies. It reviews candidate ontologies from multiple dimensions, including semantic consistency, structural integrity, and logical rationality, identifying potential problems and proposing improvement suggestions. Through collaboration between the ontology construction agent and the ontology review agent, a closed-loop iterative mechanism of "recommendation-review-correction" is formed. The ontology construction agent iteratively optimizes based on the review comments until the candidate ontology passes the quality check or reaches the preset maximum number of iterations.

[0030] Ontology Update: The reviewed candidate ontologies are seamlessly integrated into the global ontology, enabling the dynamic evolution of the knowledge system. This updated ontology immediately serves as the latest semantic constraint, guiding and optimizing subsequent knowledge extraction tasks. This forms a closed-loop mechanism for continuous learning and self-improvement from data, ensuring the continuous iteration and accuracy of the system's knowledge.

[0031] The enhanced generation strategy for graph-based reasoning: After the ontology is constructed, during the query phase, guided by the latest ontology, the user query is semantically parsed and expanded to extract the corresponding entities and relationships. Retrieval and reasoning are then performed using the knowledge graph, and the retrieved structured knowledge is provided as context to the large model, outputting accurate and reliable answers. Specific implementation process: Knowledge Extraction Intelligent Agent This invention addresses the field of faulty equipment maintenance. Therefore, in the initial stage of knowledge graph construction, it adopts a top-down approach, prioritizing the definition of core concepts and relationships within this domain to construct an ontology meta-model. This ontology aims to provide unified and standardized schema-level constraints for subsequent knowledge extraction, graph construction, and even dynamic evolution. Initial Ontology Concepts and relational structure ,in and The terms "concept" and "relation" represent the concepts and relations in the ontology model, respectively. The initial ontology metamodels are shown in Tables 1 and 2. Considering the challenge of accurately defining ontology in complex domains, providing an ontology metamodel is not a mandatory requirement of this framework. Although it can provide semantic constraints for knowledge extraction, the core advantage of this framework lies in its dynamic evolution and adaptive learning mechanism, enabling it to learn and construct knowledge systems from unstructured text in a bottom-up manner.

[0032] Table 1. List of concepts in the ontology metamodel

[0033] Table 2. List of Relationships in the Ontology Metamodel

[0034] In terms of entity information extraction, the Big Oracle model needs to identify the name and type of the entity and automatically generate a clear and accurate semantic description based on the context. In terms of relation information extraction, in addition to identifying the source node, target node, and type of the relation, it also needs to extract the contextual meaning of the relation and generate a detailed description that reflects logical dependencies to support subsequent relation retrieval and reasoning. The specific process includes: Constraint rule definition: Defines a set of explicit rules used to guide and constrain the behavior of large language models. These include: a) Entity naming standardization rules: The entity names extracted by the model should be generalized concepts and should not contain specific parameters or modifiers.

[0035] b) Ontology mapping rules: The model is required to map the identified entity and relation types to a predefined set of ontologies. and In this context, for new concepts that appear in the text but do not exist in the ontology, the model is required to assign them a special candidate label (such as 'CandidateConcept') rather than creating a new type itself.

[0036] LLM Processing and Joint Generation: The original text to be processed is combined with the constructed dynamic prompt word templates to form the final input, which is then submitted to the large language model for processing. In this stage, the LLM is guided to perform a joint generation task, simultaneously completing the following operations: a) Entity Recognition and Semantic Description Generation: LLM identifies text fragments that conform to entity definitions. For each identified entity, the model does more than just extract its name; it processes it according to entity naming normalization rules and combines this with its context in the text to generate a natural language description. This description aims to summarize the entity's key features, function, or state in the current context.

[0037] b) Relationship Identification and Semantic Description Generation: LLM identifies the interaction relationships between entity pairs. For each identified relationship, the model first determines its type according to ontology mapping rules, and then, combining the relationship's context in the text and following generative instructions, generates a natural language description. The core objective of this description is to reveal the logical dependencies that cause the relationship, for example, explaining the specific reasons and processes that "lead" to the relationship, or describing the specific components of the "contains" relationship.

[0038] Add the extracted entities and relationships to the current knowledge graph. Among them These represent the entity set and relation set of the current knowledge graph, respectively. The system first performs a uniqueness check on the existing graph to determine if an entity already exists. If the entity already exists, the system will not create a new node but will instead add the extracted context description information to the existing entity's description. If the entity is a completely new node, the system will create that entity in the graph. Similarly, the system will also perform an existence check on the extracted relations to avoid redundancy and ensure the graph's conciseness and consistency.

[0039] After adding or updating entities and relationships, the system invokes a text embedding model to vectorize entity names and descriptions, as well as relationship names and descriptions, to support subsequent semantic retrieval and knowledge reasoning. The specific process is as follows: 1) Data preparation For entities, the system concatenates each entity's "name" with its "semantic description" text to form a complete entity semantic block. For example, for a fault phenomenon entity, the concatenated text might be: "Bearing overheating: During equipment operation, the temperature of the bearing part rises abnormally, exceeding the normal operating range." For relations, the system concatenates the "type (name)" of each relation with its generated "semantic description" text. Furthermore, to enhance the contextual semantics of the relation, the system also includes the "source entity name" and "target entity name" of the relation in the concatenation. For example, for a "cause" relation, the concatenated text might be: "Cause: Due to prolonged lack of lubrication, ... [detailed description].... Entities involved: [Cause of failure] Insufficient lubrication -> [Fault phenomenon] Bearing overheating."

[0040] 2) Generate vectors The system calls a text embedding model (such as qwen3-embeddding, a model specifically optimized for semantic similarity tasks). The concatenated entity text and relation text are submitted to the embedding model as independent inputs in batches, and the output is a fixed-dimensional high-dimensional floating-point vector.

[0041] Ontology-based intelligent agents During each update, only the set of entities added since the last update is considered. and relation set .

[0042] The ontology update process is not triggered in real time, but rather employs an adaptive strategy driven by knowledge increment. This occurs when new entity sets are added to the knowledge graph. With relation set When the scale reaches the threshold for dynamic calculation, the system automatically activates the ontology update and fusion process. This threshold is determined by the following formula: ; in, As a minimum starting threshold, this ensures that enough instances are accumulated in the early stages of map construction to guarantee the effectiveness of cluster analysis. It is an adaptive scaling factor that defines the proportion of new knowledge required to trigger an update relative to the total amount of existing knowledge. and These represent the total number of entities and relations in the current knowledge graph.

[0043] This dynamic threshold strategy lies in its intelligent adaptability. In the early stages of knowledge graph construction, the system iterates rapidly with a low threshold to optimize the ontology structure and quickly adapt to domain knowledge. As the graph matures and becomes large in scale, the threshold will be increased accordingly, avoiding high-cost computation caused by a small amount of new knowledge, thereby ensuring the efficiency and scalability of the system in long-term operation and large-scale data processing.

[0044] Concept generation process: For each cluster ,Include An entity, denoted as Each entity corresponds to a vector. ,in Let be the dimension of the vector. The system first calculates the cluster centers. Then select the representative cluster closest to the cluster center. Each node is a set The sample, along with its contextual description, serves as the most representative sample of the cluster. The formula is as follows: ; in Representing entities With cluster center The Euclidean distance between them ,in This represents the Euclidean norm.

[0045] These representative examples are input into a large language model, which guides the model to summarize and refine the common features of the examples, thereby generating a new candidate concept and semantic description. For example, the concept of "blade fatigue failure" is abstracted from representative examples such as "local blade wear," "blade crack propagation," and "blade deformation."

[0046] Relational reasoning: Combining the newly abstracted set of concepts with the newly added relationships. Based on carefully designed prompts, the large language model infers the possible relationships between concepts.

[0047] Multi-agent collaborative review framework A multi-agent collaborative review framework is a key mechanism for ensuring ontology quality and consistency. This framework enables multi-dimensional quality assessment and iterative optimization of candidate ontologies through collaborative work between ontology-building agents and ontology-reviewing agents.

[0048] The ontology-building agent, acting as a "concept producer," is responsible for generating candidate ontology concepts and relations from the raw data. This agent analyzes the massive amounts of entities extracted from the raw data, identifies their semantic commonalities, and abstracts them into higher-level concepts. The ontology review agent, acting as a "quality evaluator," is responsible for conducting a comprehensive quality assessment of the candidate ontology. It reviews the candidate ontology from multiple dimensions, including semantic consistency, structural integrity, and logical rationality, identifies potential problems, and proposes improvement suggestions. The suggestions from the ontology review agent include: Concept removal: Remove overly broad concepts that lack specific industrial significance. For example, remove abstract concepts such as "parameter," "state," and "electrical parameter" because they do not provide specific semantic information in industrial fault diagnosis scenarios.

[0049] Concept merging: Integrating semantically redundant or highly similar concepts. The merging process is not a simple deletion, but rather an intelligent fusion of the subclass structures and attribute sets of the two concepts. For example, when merging "pump equipment" with the existing concept of "pump," the system will retain all subclasses of both (such as centrifugal pumps and screw pumps) and integrate them after deduplicating the attribute lists.

[0050] Concept renaming: Standardize concept names to conform to industry terminology standards. For example, unify "centrifugal pump" as "centrifugal pump" to eliminate naming inconsistencies.

[0051] Hierarchical adjustment: Correcting logical errors in "is-a" relationships. This is one of the most common and important types of correction. For example, when the system incorrectly sets the parent class of "steam turbine startup process" to "equipment", the correction mechanism will recognize this semantic error (process is not equipment) and adjust its parent class to the correct "operation process".

[0052] Description optimization: Supplementing domain knowledge enhances the clarity and distinctiveness of the description. For example, the simple description "a steam turbine is a type of equipment" can be optimized to "a steam turbine is a rotary thermal power machine that uses steam as its working fluid to convert the thermal energy of steam into mechanical energy, and is widely used in thermal power generation and nuclear power, among other fields."

[0053] Subclass refactoring: Re-dividing or adjusting the subclass structure. When it is found that the granularity of subclass division is inappropriate or there is overlap, the system can reorganize the subclass hierarchy while keeping the parent concept unchanged.

[0054] The collaboration between these two agents forms a closed-loop iterative mechanism of "recommendation-review-correction." The ontology construction agent iteratively optimizes based on review comments until the candidate ontology passes the quality check or reaches the preset maximum number of iterations. Specifically: To ensure the quality of dynamically constructed ontology, this framework designs a closed-loop "review-correction" mechanism involving collaboration between an ontology review agent and an ontology construction agent. This mechanism efficiently resolves disagreements in concept evolution through a hierarchical conflict resolution strategy and an iterative correction process based on structured feedback. The core of this mechanism is ensuring the logical consistency and semantic rationality of decisions. Specifically: 1) Evaluation and conflict diagnosis of the review agent: preliminary diagnosis to quickly identify obvious problems such as non-standard naming and overly broad names; in-depth review to examine the naming and description of each candidate concept, its actual industrial significance, the rationality of its attributes, and the correctness of its hierarchical structure; conflict identification and classification: during the review process, the agent will actively compare the problems found with the existing ontology to perform conflict diagnosis.

[0055] Semantic overlap conflicts are quantified by calculating the cosine similarity of the vector embeddings of candidate concepts and existing concepts. For example, if the similarity is higher than a threshold, it is determined to be highly semantically overlapping. Hierarchical logical conflict diagnosis involves an agent that checks whether newly added hierarchical relationships will create circular inheritance within the existing ontology.

[0056] 2) Priority-based decision-making strategies and feedback generation: For different types of conflicts, the system will provide different handling strategies with clear priorities.

[0057] Merging Strategy: For "semantic overlap and conflict", the system will recommend performing a merging operation. For example, "lack of lubricating oil" can be used as a synonym or alias for "insufficient lubricating oil", and its related knowledge instance information (such as contextual description) can be added to the node attributes of "lack of lubricating oil" to enrich the existing concepts and avoid redundancy.

[0058] Hierarchical Adjustment Strategy: For "hierarchical logic conflicts," the system recommends performing hierarchical adjustments. The system will determine that "bearing overheating fault" is a more specific subclass of "rotating part overheating fault." In this case, the "inspection agent" will generate explicit correction instructions, requiring the "construction agent" to set the parent node of "bearing overheating fault" to "rotating part overheating fault," thereby establishing the correct inheritance relationship.

[0059] Restructuring Request: When a conflict cannot be resolved through the simple adjustments described above (e.g., when a new concept partially overlaps with multiple existing concepts), the review agent will initiate a restructuring request, requiring the building agent to trace back to the original set of entities that generated the concept, and, in conjunction with the conflict context information, to re-perform cluster analysis or concept abstraction.

[0060] 3) Iterative Negotiation and Manual Adjudication: After receiving structured review comments, including explicit instructions on conflict type, conflict objects, and recommendation strategies, the ontology-building agent takes different corrective actions based on the instruction type. Description correction directly modifies superficial issues such as naming and description. Backtracking and re-abstracting, for hierarchical conflicts or reconstruction requests, the agent backtracks to the original entity set, adjusts the abstract logic or clustering strategy, and uses the review comments as strong constraints to generate a new, more reasonable candidate concept.

[0061] If a conflict cannot be resolved within the preset maximum number of iterations, it will be marked as "pending" by the system and submitted to a human review interface for final arbitration by domain experts. This human-machine collaborative adjudication mechanism ensures that the quality of the ontology is effectively guaranteed even in the face of complexity and ambiguity. Through this closed-loop negotiation mechanism, this framework can efficiently and accurately resolve disagreements during the concept evolution process, thereby guaranteeing the overall quality of the dynamically updated ontology.

[0062] Hierarchical Awareness of Dynamic Ontology Update For the newly generated concepts and relationships, the system employs a hierarchical-aware adaptive integration strategy for ontology updates. The core of this strategy is a top-down, layered processing approach, prioritizing the correctness of the high-level conceptual structure before gradually integrating specific leaf node concepts.

[0063] Specifically, for each new concept in the candidate ontology, the system first performs a recursive search to prevent concept redundancy and ensure its uniqueness. The formula is as follows: ,in This indicates a newly added concept. As the main body, For the threshold, The similarity function is used. Subsequently, the system analyzes the parent node of the concept. If the parent node already exists in the ontology, the new concept is inserted as its child node; if the parent node is missing, the system uses LLM reasoning to determine whether the parent class should be the top-level concept, or to find a suitable parent node for it.

[0064] The system establishes a closed-loop iterative process of "ontology guidance - knowledge extraction - graph construction - instance clustering - ontology update". In each iteration cycle, the updated ontology serves as a new semantic constraint to guide the knowledge extraction process, ensuring that the extraction results remain consistent with the latest ontology structure. Simultaneously, newly extracted knowledge instances provide richer data support for ontology updates, driving continuous ontology updates. This closed-loop mechanism enables the system to continuously learn and improve while processing new documents, adapting to the dynamic changes in domain knowledge.

[0065] Image retrieval User query understanding User query understanding is the first step in RAG question answering, and its core goal is to transform users' natural language questions into structured query representations. This process mainly includes two key steps: ontology-based query expansion and entity relation extraction.

[0066] Ontology-based query expansion: When the system receives a user's natural language query, it first uses the domain ontology to semantically expand and understand the query. This process leverages the semantic understanding capabilities of the large model, combined with the conceptual hierarchy and semantic relationships within the ontology, to deeply understand and expand the original query. Specifically, based on synonym relationships, hyponyms, and related concept associations within the ontology, the system guides the large model to generate semantically richer and more comprehensive expanded queries. This expansion mechanism effectively handles ambiguous expressions, incomplete descriptions, and diverse expressions of technical terms in user queries, ensuring accurate understanding of the query intent. The formula is as follows: ,in These are prompt words that guide the semantic expansion of large models.

[0067] Based on the expanded query, the system parses out the key entity set. and relation set The extracted entity and relationship information will serve as the core basis for subsequent knowledge graph retrieval, ensuring a high degree of match between the retrieval results and the user's query intent.

[0068] Retrieval and generation After extracting entities and relationships from the query, the system enters the knowledge retrieval stage to fully utilize relevant information from the knowledge graph to support question-answer generation. Considering the diverse and flexible terminology in the industrial field, traditional string-matching-based retrieval methods often fail to capture the user's true intent. Therefore, the system utilizes vectorized semantic retrieval technology and multi-hop reasoning, combining upstream and downstream relationships for multi-hop reasoning, extending to neighboring nodes and constructing a local subgraph, using node information, attribute information, and related descriptions as contextual semantic blocks. The specific formula is shown below: ; in This represents the set of entities extracted in the query. Indicates and Similar entities Representing upstream and downstream neighbors of similar entities, This represents a local subgraph centered around these nodes, containing nodes, edges, and corresponding descriptive information.

[0069] For retrieving relational information, the system performs semantic matching between the extracted relation list and the candidate relation set in the knowledge graph to obtain relation instances that are the same as or similar to the queried relation. Based on the matched relations, it obtains associated triples from the graph through multi-hop reasoning, forming a complete causal evidence chain, expanding the relevant entity pairs and contextual information, thereby providing factual evidence on a larger scale. The specific formula is shown below: ; in This represents the set of entities extracted in the query. Indicates and Similar relationships, This represents triples associated with these similarity relationships. A subgraph representing similarity relationships and associated triples.

[0070] Answer generation: The system uses the relevant entities, relationships, and descriptions from the retrieved knowledge subgraph as structured context, and inputs them along with the user's original query into the large language model to generate the answer. The formula is: ,in Represents the probability distribution of a large language model. This indicates the user's query. This represents the answer with the highest probability. The system guides the large model to fully utilize retrieved factual knowledge to generate accurate, professional, and interpretable answers. During generation, the system ensures that the answer content is consistent with the retrieved knowledge graph information, avoiding the generation of content that contradicts the facts, while maintaining the naturalness and readability of the answer.

[0071] Example 2 In this embodiment, 496 question-answer pairs were extracted and manually reviewed from industrial equipment fault documents provided by a manufacturing company to form an evaluation dataset. These documents mainly focus on equipment fault diagnosis and handling, covering various types of power generation equipment. Each document contains information from multiple dimensions, including fault phenomenon description, cause analysis, solutions, and preventive measures, and includes rich professional terminology, technical parameters, and industry standards.

[0072] The knowledge extraction module in this invention has the following internal processing flow: 1. Initialization and Input Reception: The module first receives two core inputs: one is the predefined concepts and relational schemas (i.e., "ontology semantic constraints") provided by the domain ontology; the other is the unstructured industrial text to be processed.

[0073] 2. Entity Recognition and Description Generation: The module performs deep semantic analysis on the input industrial text. Based on the ontology pattern received in step 1, it automatically identifies all entity instances in the text that conform to predefined types. For each identified entity, the module automatically generates a summary description based on its context in the text. This description summarizes the entity's key features, functions, or specific state.

[0074] 3. Relationship Identification and Description Generation: After entity identification, the module further analyzes the potential relationships between the identified entities. Based on the relationship types defined in the ontology schema, it determines and extracts specific relationship instances connecting the source and target entities. Simultaneously, the module generates a brief text description of this relationship based on the statements describing the relationship in the text.

[0075] 4. Structured Output: Finally, the module formats and integrates all extracted information, outputting a structured data object. This object contains two lists: an "entity list," where each element contains the entity's name, type, and description; and a "relationship list," where each element contains the source entity name, target entity name, relationship type, and relationship description. This structured data provides standardized input for subsequent knowledge graph construction.

[0076] Once the number of processed files reaches a threshold, the internal entity aggregation phase will be initiated. This phase uses a clustering algorithm to generate common upper-level concepts and candidate relationships for the cluster, thus starting the ontology review process.

[0077] The ontology review module in this invention has the following internal processing flow: 1. Receiving Review Objects: The module receives a set of "candidate concepts" and "candidate relationships" generated by the ontology construction module as input.

[0078] 2. Perform multi-dimensional quality assessment: The module automatically evaluates the input candidate ontology elements based on a set of pre-defined, standardized review rules. This evaluation process includes, but is not limited to, the following aspects: Concept quality review: Automated checks on the standardization of concept naming (whether it is concise and conforms to industry terminology), semantic clarity (whether the description is accurate and unambiguous), and the practical application value of the concept. Hierarchical structure review: Verifies whether the parent class setting of the candidate concept is the most logically direct superior concept, and checks for structural errors such as circular inheritance, hierarchical redundancy, or missing hierarchies. Relationship definition review: Evaluates whether the semantics of the candidate relationship are accurate, whether the direction is logically consistent, and determines the necessity of the relationship for enhancing the expressive power of the knowledge graph.

[0079] 3. Generate and execute processing strategies: Based on the evaluation results, the module automatically makes decisions and generates processing instructions according to preset priorities (e.g., modification and optimization > hierarchical adjustment > concept merging > deletion). This mechanism ensures that the iterative optimization process of candidate ontologies is efficient and closed-loop until all candidate elements pass quality review.

[0080] The final two-level ontology structure generated by this invention based on the fault documents provided by the enterprise is as follows: The ontology defines six top-level conceptual classes: Device; Fault Phenomenon; Fault Cause; Solution. Equipment category: Rotating machinery, stationary equipment, electrical and power equipment, control and instrumentation systems, transmission and conveying equipment, hydraulic and pneumatic systems, load-bearing and support structures, heat exchange and environmental control equipment, etc. Fault phenomena include: deviations from performance indicators, abnormal physical conditions, breaches of structural integrity, abnormal electrical characteristics, abnormal control signals and logic, and abnormal media conditions. Failure causes include: mechanical wear and fatigue, material corrosion and aging, lubrication and cooling failure, design or manufacturing defects, installation or assembly errors, improper operation or exceeding limits, improper maintenance, and external environmental influences. Solutions: Inspection and diagnostics, adjustment and calibration, repair and reinforcement, replacement and upgrades, cleaning and maintenance, operation and process changes, etc. The user query understanding module in this invention has the following internal query rewriting process: 1. Input Reception: The module receives the user's raw natural language query and loads the current industry domain ontology knowledge base. 2. Concept Identification and Ontology Matching: The module first parses the user query to identify the core entities or concepts it contains. Then, it retrieves these identified concepts from the loaded ontology knowledge base.

[0081] 3. Query Expansion Decision: If a match is successful: If the core concept in the query exists in the ontology, the module will utilize the ontology's hierarchical structure (such as subclass relationships) and associations to automatically find other related concepts and supplement them into the original query, thereby generating a new query with richer semantics and more comprehensive coverage. This process strictly maintains the user's core intent while expanding the query. If a match fails: If the concept in the query does not exist in the ontology, the module will not make any modifications and will directly return to the original query.

[0082] The user query understanding module in this invention uses the following keyword extraction process when parsing a query: 1. Input reception: The module receives a user query text.

[0083] 2. Multi-dimensional Extraction: The module analyzes the query text and extracts two types of keywords with different search values: Entity keywords: words in the query that explicitly point to specific devices, fault phenomena, components, etc. Relationship keywords: words in the query used to describe the interaction and influence between entities; these words correspond to "edges" in the knowledge graph.

[0084] 3. Structured Output: The module will organize and output the extracted two types of keywords in a structured manner, so that the subsequent retrieval system can more accurately match nodes and edges in the knowledge graph based on entity and relationship information.

[0085] To discuss the performance level of the methods, this invention selects three methods for comparison: Native-RAG, GraphRAG, and Self-RAG. To comprehensively evaluate the effectiveness of graph retrieval enhancement generation based on dynamic ontology in industrial equipment fault question answering, this invention mainly focuses on three core dimensions: relevance, fidelity, and completeness of the answers, allowing the large model to assign scores between 1 and 10 for each dimension.

[0086] Relevance: Measures the degree to which the generated answer matches the user's query question, and assesses whether the system can accurately understand the user's intent and provide the corresponding answer content.

[0087] Fidelity: Evaluate whether the system's answers are based on reliable knowledge sources to avoid generating "illusionary" content that does not conform to the facts.

[0088] Completeness: Based on the relevance of the answer, assess whether the generated answer adequately answers the user's question, and measure the comprehensiveness and information coverage of the answer content.

[0089] Table 3 shows the evaluation results of all methods on the dataset.

[0090] Table 3 Comparison results with baseline experiment

[0091] This invention's in-depth analysis of baseline methods yielded the following results: Native RAG performed the worst in relevance and completeness. This is mainly because it relies solely on raw text retrieval, lacking a deep semantic understanding of industry-specific terminology and guidance from structured knowledge, resulting in insufficient ability to handle complex professional problems. Self RAG, by introducing self-reflection and enhancement mechanisms, improved the quality of answers to some extent, showing a slight performance improvement compared to Native RAG. However, its general design still has limited adaptability in highly specialized industrial fields. GraphRAG utilizes graph structures to enhance information retrieval capabilities, performing well in relevance and completeness, but lacks semantic-level constraint mechanisms.

[0092] The significant advantages of the method of this invention are mainly reflected in three aspects: First, through ontology-driven semantic constraints, the system can accurately understand professional concepts and terms in the industrial field, significantly improving the relevance of the answer; second, the structured representation of the knowledge graph effectively supports the fusion of multi-source information and relational reasoning, improving the completeness of the answer; finally, the generation mechanism based on ontology semantic verification and knowledge graph constraints ensures the fidelity and traceability of the answer.

[0093] Example 3 To verify the effectiveness of the dynamic ontology, we designed an ablation experiment to compare the performance of the complete DO-GRAG method with that of a static version, SO-GRAG, which only uses the ontology meta-model without subsequent updates. The experimental results are as follows: Figure 2 As shown, DO-GRAG significantly outperforms the static version across all key evaluation metrics. Specifically, after removing the dynamic update mechanism, the system's relevance, completeness, and fidelity decreased by 53.33%, 56.32%, and 36.23%, respectively. This gap demonstrates that continuous ontology updates play a crucial role in improving the overall system performance. The fundamental reason lies in DO-GRAG's ability to continuously enrich and optimize the ontology structure through instance clustering and higher-level concept expansion mechanisms. This, in turn, drives the dynamic expansion of the knowledge graph, significantly improving its coverage and semantic consistency, whereas a static ontology can only extract limited information. A higher-quality knowledge graph supports subsequent question-answering generation, ultimately enhancing the system's usability in real-world applications.

[0094] Example 4 To verify the role of the review agent in ensuring ontology quality, this invention designed an ablation experiment, removing the iterative correction step between the ontology construction agent and the review agent. Under this setting, candidate ontology generated by the construction agent will be directly integrated into the global ontology without verification.

[0095] DO-GRAG-w / o-rwa indicates the removal of the censorship agent in the test, and the experimental results are as follows: Figure 3 As shown, after removing the review agent, the system's relevance, completeness, and fidelity decreased by 10.54%, 9.29%, and 11.33%, respectively. This result indicates that the review agent is crucial for ensuring the quality of knowledge graphs. Its core value lies in effectively filtering and correcting errors and "illusions" that may be introduced during the knowledge extraction and ontology construction stages through logical consistency and semantic rationality checks. Without this step, the knowledge graph will contain illogical triples or triples that contradict domain common sense, thereby misleading the retrieval and generation process during the question-answering stage, leading to a significant reduction in the accuracy and credibility of the final answer.

[0096] Example 5 Figure 4 The results of ablation experiments on the relation retrieval module are presented, where DO-GRAG-w / o-re indicates the test with the relation retrieval module removed. The results show that removing the relation retrieval module reduces relevance by approximately 16.83%, completeness by approximately 4.87%, and fidelity by approximately 16.54%. This demonstrates that relation retrieval plays an indispensable role in constructing semantic networks between entities and ensuring the integrity of logical links in answers. Its absence weakens the model's ability to capture upstream and downstream semantic relationships, thus affecting the rigor and coherence of information organization.

[0097] Example 6 Figure 5 The ablation study results for the entity retrieval module are presented, with DO-GRAG-w / o-en indicating the test results when only the entity retrieval module is removed. The comparison shows a decrease of approximately 38.44% in relevance, 36.72% in completeness, and 33.06% in fidelity. This indicates that entity information plays a crucial role in accurately identifying user query intent and generating high-quality answers; relying solely on relational information is insufficient to fully cover the key information required for the question. The rich descriptive information and upstream / downstream links provided by entity retrieval are fundamental to constructing accurate answers; their absence will hinder the model's understanding of the core semantics of the question.

[0098] Example 7 This invention also provides an industrial fault knowledge graph construction and retrieval generation system based on evolutionary ontology. The system is used to implement the method described in Embodiment 1. The system includes: an extraction module, a first evolution module, an evaluation module, a second evolution module, and a retrieval module. The extraction module is used to automatically identify and extract entities, attributes, and their relational information from unstructured industrial documents based on the ontology meta-model; The first evolution module is used to design an ontology-building agent that automatically discovers, abstracts, and constructs new concepts and relationships from continuously extracted knowledge instances through incremental learning, thereby driving the dynamic evolution of the ontology. The evaluation module is used to perform quality assessments on candidate ontologies using an ontology review agent, identify potential problems, and propose improvement suggestions. The second evolution module is used to integrate the reviewed candidate ontology into the global ontology, thereby realizing the dynamic evolution of the knowledge system. The retrieval module is used during the query phase after the ontology is built. Guided by the latest ontology, it performs semantic parsing and expansion of user queries, extracts corresponding entities and relationships from the expanded queries, performs retrieval and reasoning through knowledge graphs, and provides the retrieved structured knowledge as context to the large model. This guides the large model to make full use of the retrieved factual knowledge to generate accurate, professional, and interpretable answers.

[0099] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing and retrieving industrial fault knowledge graphs based on evolutionary ontology, characterized in that, The method includes: Automatically identify and extract entities, attributes, and relational information from unstructured industrial documents based on ontology meta-model; The design of the ontology constructs an intelligent agent, which automatically discovers, abstracts and constructs new concepts and relationships from continuously extracted knowledge instances in an incremental learning manner, thereby driving the dynamic evolution of the ontology; An ontology review agent is used to assess the quality of candidate ontologies, identify potential problems, and propose improvement suggestions. The reviewed candidate ontologies are integrated into the global ontology, enabling the dynamic evolution of the knowledge system; Once the ontology is constructed, during the query phase, the user query is semantically parsed and expanded under the guidance of the latest ontology. The corresponding entities and relationships are extracted from the expanded query, and retrieval and reasoning are performed through the knowledge graph. The retrieved structured knowledge is provided as context to the large model, guiding the large model to make full use of the retrieved factual knowledge to generate accurate, professional, and interpretable answers.

2. The method according to claim 1, characterized in that, Methods for designing ontology-based intelligent agents that use incremental learning to automatically discover, abstract, and construct new concepts and relationships from continuously extracted knowledge instances, thereby driving the dynamic evolution of the ontology, include: New entity sets added to the knowledge graph With relation set When the scale reaches the threshold for dynamic calculation, the system automatically activates the ontology update and fusion process. This threshold is determined by the following formula: ； in, As a minimum threshold, this ensures that enough instances are accumulated in the early stages of map construction to guarantee the effectiveness of cluster analysis. It is an adaptive scaling factor, which defines the proportion of new knowledge required to trigger an update relative to the total amount of existing knowledge. and These represent the total number of entities and relations in the current knowledge graph; For newly added entities, hierarchical clustering is first performed. Then, for each cluster, a centrality analysis method based on semantic similarity is used to select representative entities. Specifically, the names and descriptions of all entities within the cluster are first vectorized using TF-IDF. Then, the cosine similarity between each entity and all other entities within the cluster is calculated, and the entity with the highest average similarity is selected. The entity serves as the most representative sample in the cluster. Subsequently, the name, type, and semantic description of the representative entity are input into the large language model. The model is guided by prompt templates to summarize and abstract the common features of the entities, thereby generating new candidate concepts. By combining the abstracted set of new concepts and the newly added relationships, the large model is guided to deduce the relationships between the concepts based on the design prompts.

3. The method according to claim 1, characterized in that, Methods for using ontology review agents to assess the quality of candidate ontologies, identify potential problems, and propose improvement suggestions include: concept deletion, concept merging, concept renaming, hierarchy adjustment, description optimization, and subclass reconstruction.

4. The method according to claim 1, characterized in that, The reviewed candidate ontologies are integrated into the global ontology. Methods for achieving dynamic evolution of the knowledge system include: For each new concept in the candidate ontology, a recursive search is first performed; Subsequently, the parent node of the concept is analyzed. If the parent node already exists in the ontology, the new concept is inserted as its child node; if the parent node is missing, the system uses LLM reasoning to determine whether the parent class should be the top-level concept, or to find a parent node that meets the preset requirements.

5. The method according to claim 1, characterized in that, Once the ontology is constructed, during the query phase, guided by the latest ontology, the methods for semantic parsing and expansion of user queries to extract corresponding entities and relationships include: Based on the synonym relationships, hyponyms and related concepts in the ontology, the large model is guided to generate extended queries with richer semantics and higher coverage. Based on the expanded query, the key entity set is parsed out. and relation set .

6. The method according to claim 1, characterized in that, Methods for retrieval and reasoning using knowledge graphs, providing the retrieved structured knowledge as context to the large model, and outputting answers include: Based on vectorized semantic retrieval and multi-hop reasoning technology, the extracted entity set is embedded, similar entity sets are found in the knowledge graph, and multi-hop reasoning is performed in combination with their upstream and downstream relationships. This extends to neighboring nodes and constructs a local subgraph, using node information, attribute information and related descriptions as contextual semantic blocks. The extracted relation list is semantically matched with the candidate relation set in the knowledge graph to obtain relation instances that are the same as or similar to the query relation. The associated triples are obtained from the graph through multi-hop reasoning to form a complete causal evidence chain and expand the relevant entity pairs and contextual information. The retrieved knowledge subgraph contains relevant entities, relationships, and descriptive information, which are used as a structured context and input into the generative model along with the user's original query to generate an answer.

7. The method according to claim 6, characterized in that, Based on vectorized semantic retrieval and multi-hop reasoning techniques, the extracted entity set is embedded, similar entity sets are found in the knowledge graph, and multi-hop reasoning is performed based on their upstream and downstream relationships to expand their topological neighbors and construct a local subgraph. Node information, attribute information, and related descriptions are used as contextual semantic blocks. The methods for constructing contextual semantic blocks include: ； in, This represents the set of entities extracted in the query. Indicates and Similar entities Representing upstream and downstream neighbors of similar entities, This represents a local subgraph centered around these nodes, containing nodes, edges, and corresponding descriptive information.

8. The method according to claim 7, characterized in that, The extracted relation list is semantically matched with the candidate relation set in the knowledge graph to obtain relation instances that are the same as or similar to the query relation. Then, related triples are obtained through multi-hop reasoning from the graph to form a complete causal evidence chain. Methods for expanding related entity pairs and contextual information include: ； in, This represents the set of entities extracted in the query. Indicates and Similar relationships, This represents triples associated with these similarity relationships. A subgraph representing similarity relationships and associated triples.

9. A system for constructing and retrieving industrial fault knowledge graphs based on evolutionary ontology, the system being used to implement the method described in any one of claims 1-8, characterized in that, The system includes: an extraction module, a first evolution module, an evaluation module, a second evolution module, and a retrieval module; The extraction module is used to automatically identify and extract entities, attributes, and their relational information from unstructured industrial documents based on the ontology meta-model. The first evolution module is used to design an ontology-building agent that automatically discovers, abstracts, and constructs new concepts and relationships from continuously extracted knowledge instances through incremental learning, thereby driving the dynamic evolution of the ontology. The evaluation module is used to perform quality evaluation on candidate ontologies using an ontology review agent, identify potential problems and propose improvement suggestions. The second evolution module is used to integrate the reviewed candidate ontology into the global ontology, thereby realizing the dynamic evolution of the knowledge system; The retrieval module is used to perform semantic parsing and expansion of user queries during the query phase after the ontology is constructed, under the guidance of the latest ontology. It extracts the corresponding entities and relationships from the expanded query, performs retrieval and reasoning through the knowledge graph, and provides the retrieved structured knowledge as context to the large model, guiding the large model to make full use of the retrieved factual knowledge to generate accurate, professional and interpretable answers.

Images