Knowledge graph and large language model combined bearing intelligent maintenance decision-making method

Abstract

The invention discloses a knowledge graph and large language model combined bearing intelligent maintenance decision-making method, and the method comprises the steps: constructing a bearing wear knowledge ontology for a bearing maintenance decision-making scene, and carrying out the knowledge extraction based on a bearing wear analysis technology text; constructing an aero-engine bearing wear analysis structured knowledge graph covering inducement-mechanism-degree-maintenance / prevention; designing a knowledge sub-graph retrieval and refining strategy according to a structured knowledge graph description entity and intention entity set, performing explicit modeling on a combination relationship among multi-attribute entities by adopting set operation, and constructing comprehensive constraints to select knowledge sub-graphs strongly related to bearing wear description; reasoning is carried out in cooperation with a large language model and knowledge sub-graph multi-attribute wear, and the bearing intelligent maintenance decision-making method is constructed. According to the method, the knowledge graph is used for extracting strong correlation information as a constraint, the decision precision and interpretability of the reasoning process are improved, and the bearing maintenance decision accuracy is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of mechanical equipment, specifically involving a bearing intelligent maintenance decision-making method that combines knowledge graphs and large language models. Background Technology

[0002] Rolling bearings are core load-bearing and transmission components of aero-engines, and their single-point failure characteristics affect the overall service performance and reliability of the engine. Aero-engine bearings operate under high-speed, heavy-load, and high-temperature conditions for extended periods, and are subjected to extreme conditions such as instantaneous speed changes, load changes, and oil cut-offs. They are prone to early wear failures through various mechanisms, including fatigue spalling, pitting, and scratching, which have become a key factor restricting the service life and reliability of aero-engines. Statistics show that approximately 60% of premature engine replacements and 37.5% of in-flight shutdowns of a certain type of aero-engine are due to aero-engine bearing wear. Therefore, aero-engine bearing maintenance decision-making, through analyzing wear causes and developing matching maintenance and preventative measures, has become a core technology for ensuring the safe operation of aero-engines. This decision-making process is characterized by its knowledge intensity, requiring not only comprehensive inference based on multi-attribute wear descriptions but also a high degree of reliance on expert experience or textual knowledge such as fault diagnosis manuals. Therefore, domain knowledge representation and knowledge-driven reasoning for aero-engine bearing maintenance decision-making are crucial supports for achieving intelligent maintenance decision-making.

[0003] Existing research has developed diverse decision-making technology systems for friction and wear maintenance, based on knowledge-driven methods such as expert systems and knowledge graphs. Expert systems explicitly organize expert knowledge and experiential rules into knowledge bases or rule bases, and use inference engines to symbolically simulate human reasoning and decision-making processes, outputting diagnostic conclusions and maintenance recommendations based on bearing wear descriptions. For example, Li et al. used IF-THEN rules to express angular contact bearing fault diagnosis knowledge and selected a tree-structured reasoning model as the inference engine to identify wear fault types and recommend appropriate treatment measures. However, expert system methods face challenges such as high knowledge acquisition and maintenance costs, limited knowledge visualization, and difficulty in supporting multi-level knowledge representation and multi-hop reasoning. Knowledge graphs, employing a graph model with "entity-relationship-attribute" as its core, explicitly model the physical concepts and their relationships within the knowledge domain related to maintenance decisions, offering significant advantages in interpretability and traceability. Chen Fangjun, based on semantic rules and natural language processing technology, extracted knowledge from expert literature and web page data to construct a rotor fault decision graph, and combined it with the Neo4 graph database to achieve graph-assisted rotor system fault diagnosis and auxiliary decision-making. Shao Tao employed a seven-step top-down approach to construct a knowledge ontology to clarify the relevant knowledge nodes and relationship types for bearing fault inspection. He then combined a bidirectional long short-term memory neural network and Jaccard correlation coefficients to construct a rolling bearing fault knowledge graph. A Naive Bayes classifier was then used to query entities that match the question, completing knowledge retrieval and decision reasoning. Wang et al. used a BiLSTM-CRF model to identify entities from unstructured text and employed an attention-enhanced BiLSTM model to detect relationships between entities, constructing a CNC machine tool fault diagnosis knowledge graph. While these methods rely on knowledge graphs to achieve explicit representation of multi-attribute knowledge, they are limited by semantic understanding and complex logical reasoning capabilities, making it difficult to directly address the implementation of intelligent maintenance decisions for aerospace bearings.

[0004] With the development of natural language processing (NLP) technology, Large Language Models (LLMs) have demonstrated application potential in auxiliary decision-making tasks such as fault diagnosis and maintenance decision-making question answering, thanks to their excellent context learning capabilities and emergent reasoning characteristics. In particular, LLMs and knowledge graphs have complementary advantages in semantic understanding, domain knowledge, and traceable reasoning. Researchers are combining knowledge graphs and LLMs to address the dual demands of dense knowledge representation and complex reasoning in tasks such as intelligent wear and tear diagnosis. The integration methods can be mainly categorized into two types: model fine-tuning and retrieval-enhanced generation. Model fine-tuning paradigms typically utilize knowledge graphs to generate question-answer pairs or multi-turn dialogues, thereby fine-tuning the general-purpose LLM and injecting domain knowledge into the model in a parameterized form. While this approach can improve the ability to respond to domain-specific questions, its fine-tuning process usually requires significant computational resources, and the reasoning evidence and reasoning paths are difficult to expose, resulting in limited traceability of reasoning and interpretability of results. The retrieval-enhanced generation paradigm retrieves knowledge graphs during the reasoning stage, extracts knowledge subgraphs related to the query question as evidence, and injects them into prompts in a structured or semi-structured form to enhance the generation quality and traceability of the large language model. For example, Ma et al.

[128] constructed a large model of fault diagnosis reasoning knowledge graphs, extracted entity and intent information from the query question by designing prompts, and retrieved knowledge subgraphs using strategies such as neighborhood expansion and fuzzy matching to achieve high-quality responses to fault diagnosis questions. However, the wear analysis knowledge required for aircraft bearing maintenance decisions involves multiple attributes such as wear mechanism, wear degree, wear causes and maintenance measures, and the relationships between the elements are complex, which makes the structured representation of wear analysis knowledge still challenging. In addition, the results of aircraft bearing inspections often contain multiple attributes such as multi-source wear mechanism and wear degree, while existing knowledge retrieval strategies are mainly based on single entity matching, which easily introduces a large number of irrelevant knowledge subgraphs and dilutes effective evidence, affecting the reasoning reliability and generation quality of aircraft bearing maintenance decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a bearing intelligent maintenance decision-making method that combines knowledge graphs and large language models, which solves the problems of low accuracy and insufficient interpretability in existing bearing intelligent maintenance decision-making technologies.

[0006] The technical solution adopted in this invention is a bearing intelligent maintenance decision-making method combining knowledge graph and large language model. The specific operation steps are as follows:

[0007] Step 1: Construct a bearing wear knowledge ontology for bearing maintenance decision-making scenarios, and extract knowledge based on bearing wear analysis technical texts (including maintenance standard manuals and test analysis reports) to build a structured knowledge graph of aero-engine bearing wear analysis covering causes, mechanisms, degrees, and maintenance / prevention. Step 2: Based on the structured knowledge graph describing the entity and intent entity set, design a knowledge subgraph retrieval and refinement strategy, use set operations to explicitly model the combination relationship between multi-attribute entities, and construct comprehensive constraints to select knowledge subgraphs that are strongly related to the bearing wear description. Step 3: Using a collaborative large language model and knowledge subgraph, reason about maintenance decision-making problems involving multi-attribute wear descriptions to construct an intelligent bearing maintenance decision-making method; Step 4: Using the two-dimensional image and three-dimensional morphology of the bearing wear surface as input, construct an application and evaluation method for the intelligent maintenance decision-making method. Evaluate and optimize the method through accuracy, BLEU, BERT scores and generation quality ranking to achieve intelligent maintenance decision-making for the bearing wear surface.

[0008] The invention is further characterized in that, Step 1: Constructing a knowledge graph for aero-engine bearing wear analysis. The specific process is as follows: Step 1.1: Divide the bearing wear analysis technical text into multiple text blocks according to semantic units, and use a large language model to identify entities and relationships between entities that match the knowledge ontology, forming multiple knowledge sub-chains in the form of node-relationship-node; Step 1.2: To address the semantic ambiguity caused by multiple expressions of the same semantic entity, the text embedding model BGE-M3 is used to map the semantics to a high-dimensional vector space. Synonymous entities are aligned and integrated based on vector similarity measurement, and knowledge sub-chains are aggregated to construct a knowledge graph. The knowledge graph is presented as a network topology structure composed of multi-attribute nodes and directed edges. The nodes represent multi-attribute entities, and the directed edges describe the relationships between multi-attribute entities. Step 1.3: Import the constructed knowledge graph into the Neo4j graph database to achieve knowledge management and visualization.

[0009] The knowledge graph described in step 2 uses entity-relationship-attribute as the basic unit; the entity includes eight categories: bearing type, detection area, wear cause, wear mechanism, morphological characteristics, wear degree, maintenance measures, and prevention measures; the relationship represents the association between the various entities.

[0010] Step 2 is as follows: Step 2.1: Construct the query question; The query consists of two parts: wear description information and query intent. The wear description information includes bearing type, detection area, wear mechanism identification results, wear degree, and morphological characteristics. These elements together constitute a multi-attribute set of wear description entities. ; The query intent indicates the type of decision-making information expected, including maintenance measures, preventive measures, and wear-related causes. The corresponding categories of entities together constitute the intent entity set. ; Step 2.2: Based on the structured wear description entity and intent entity set, select the knowledge subgraph related to the query question from the global knowledge graph; Step 2.3: Construct a comprehensive constraint for knowledge chains by jointly querying entities with multiple attributes, and eliminate irrelevant knowledge chains accordingly; thus obtaining an optimized knowledge subgraph.

[0011] Step 2.2 is as follows: A knowledge retrieval and refinement mechanism is constructed to obtain a knowledge subgraph strongly related to the query question; specifically as follows: The wear description entity is constructed using the BGE-M3 embedding model. Similar to other map entities Mapping to a continuous vector space and using vector distance to measure the similarity between the two, the candidate graph entities with the highest similarity are selected as the alignment result for subsequent retrieval and inference; The similarity measure uses cosine similarity as the standard, as shown in formula (1): (1) In the formula — The degree of similarity between entities; — Vectorized encoding based on the text embedding model BGE-M3; The wear description information after entity alignment, together with the query intent, constitutes the final knowledge subgraph related to the query question.

[0012] Step 2.3 is as follows: Comprehensive constraints are constructed from multi-attribute query entities through multi-level set operations. Specifically, as shown in formulas (2) and (3), the set relationship between entities describing wear of similar and dissimilar attributes is explicitly characterized. Finally, it is determined whether each candidate knowledge chain in the retrieved knowledge subgraph satisfies the above comprehensive constraints. Thus, from the initial knowledge subgraph The knowledge chains that are completely related to the query question are filtered out and an optimized knowledge subgraph is formed, as shown in formula (4); (2) (3) (4) In the formula, —Constraints constituted by the wear mechanism and wear degree query entities; • A collection of entities querying other attributes besides wear mechanism and wear degree; — The i-th wear mechanism entity present on the surface to be diagnosed; —Other attribute queries for entities, such as non-wear mechanisms and wear levels; — Wear mechanism entity Corresponding degree of wear; — Overall wear and tear physical condition; — Query the wear mechanism entity in the question gather; P i Yes—Knowledge Graph The Middle i A chain of reasoning.

[0013] In step 2.2, each query entity... Starting from the first step, we traverse and search along the relation edges in the knowledge graph with a finite step size, enumerate the paths that can connect the wear description entity to the intention entity, and use them as candidate knowledge chains. The resulting knowledge chains cover the reasoning paths from different wear description entities to the target decision conclusion, and their union constitutes a candidate knowledge subgraph for the current query problem.

[0014] In step 3, a prompt template is constructed by integrating query questions, knowledge graphs, and decision rules based on the CRISPE framework. The prompt template includes six elements: background, role, instruction, topic, preset, and exception. The background focuses the issue on bearing maintenance decisions; The role model should be set up to play the role of a bearing wear maintenance decision-making expert with professional knowledge and rich experience; Instruction integration and maintenance issues and knowledge subgraph; The theme is used to limit the reasoning of large language models to only revolve around wear analysis and maintenance decision-making tasks; The pre-defined rules for maintaining decision-making and the requirements for outputting decisions are clearly stated. Exceptions are used to specify processing strategies when wear descriptions are incomplete or knowledge graph rules are missing.

[0015] The beneficial effects of this invention are: This invention constructs a domain knowledge ontology for bearing maintenance decision-making scenarios and extracts knowledge from unstructured texts such as maintenance standard manuals to build a wear analysis knowledge graph. Secondly, based on the characteristics of bearing wear description, a hierarchical knowledge retrieval and refinement strategy is designed. Domain knowledge strongly correlated with bearing wear description is extracted from the knowledge graph as external traceable evidence to enhance the reasoning ability of the large language model, enabling intelligent maintenance decisions for aero-engine bearings. This invention uses the extraction of strongly relevant information from the knowledge graph as a constraint to improve the decision-making accuracy and interpretability of the reasoning process. This method can accurately retrieve domain knowledge strongly correlated with multi-attribute bearing wear descriptions and eliminate redundant knowledge, significantly improving the accuracy of bearing maintenance decisions and the quality of generated responses. Attached Figure Description

[0016] Figure 1 This is the knowledge ontology of bearing wear analysis in this invention; Figure 2 This is the visualization result of the wear analysis knowledge graph of this invention; Figure 3 It is an architecture for intelligent maintenance decision-making method of aero-engine bearings that combines knowledge graphs and large language models; Figure 4 This is a typical example of a worn surface sample. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This invention relates to a bearing intelligent maintenance decision-making method that combines knowledge graphs and large language models, such as... Figure 3 As shown, Includes the following steps: S1. Based on the definition of knowledge ontology and the extraction of domain knowledge, a knowledge graph for bearing wear analysis covering "cause-mechanism-degree-maintenance / prevention" is established, which transforms the text-based wear analysis knowledge into structured information that can be understood and queried by machines. S2. Based on the structured knowledge graph describing the entity and intent entity set, design a knowledge subgraph retrieval and refinement strategy, use set operations to explicitly model the combination relationship between multi-attribute entities, and construct comprehensive constraints to select knowledge subgraphs that are strongly related to the bearing wear description. S3. A method for constructing intelligent maintenance decision-making queries by combining wear description information and query intent, and a bearing intelligent maintenance decision-making method by collaborating with a large language model and knowledge subgraph, and improving decision-making accuracy through multi-attribute wear reasoning chains in knowledge graphs. S4. Using two-dimensional images and three-dimensional morphology of the bearing wear surface as input, construct an application and evaluation method for intelligent maintenance decision-making. Evaluate and optimize the method through accuracy, BLEU, BERT scores, and generation quality ranking to achieve intelligent maintenance decision-making for the bearing wear surface.

[0019] Example 2 Based on Example 1, step S1 is as follows: Step S1: Construct a bearing wear knowledge ontology for bearing maintenance decision-making scenarios, and extract knowledge based on bearing wear analysis technical texts to build a structured knowledge graph of aero-engine bearing wear analysis covering causes, mechanisms, degrees, and maintenance / prevention. Wear analysis knowledge is a collection of rules and criteria for aircraft bearing maintenance decisions, characterized by multiple knowledge attributes and highly complex interrelationships. Knowledge graphs, using "entity-relationship-attribute" as basic units, can structurally represent multi-attribute knowledge and its complex relationships, supporting knowledge querying and interpretable decision-making. This representation paradigm is highly compatible with the multi-attribute information and complex relationships of wear analysis knowledge: the entity-attribute structure can accommodate multi-attribute wear analysis information; while the entity-relationship-entity structure can explicitly depict the relationship path from wear description to maintenance measures, forming an interpretable and traceable multi-hop reasoning chain. Therefore, this section transforms textual wear analysis knowledge into a machine-understandable and queryable structured knowledge graph through knowledge ontology definition and domain knowledge extraction, providing reliable knowledge support for bearing maintenance decisions.

[0020] The knowledge ontology is used to define the types of nodes involved in wear analysis knowledge and the relationships between these nodes, forming the structural framework of the bearing wear analysis knowledge graph. This section uses a seven-step method to construct the bearing wear analysis knowledge ontology: First, based on the needs of intelligent bearing maintenance decision-making, and combining the experience of domain experts and relevant wear analysis standards and specifications, the knowledge scope is defined and core physical concepts are extracted; second, by sorting out the inherent relationships between various physical concepts, a structured knowledge framework is constructed, forming the initial knowledge ontology; finally, the ontology is iteratively optimized and improved based on expert review results and engineering application feedback.

[0021] Figure 1 The constructed knowledge ontology for bearing wear analysis is presented. This ontology, centered on wear mechanisms, encompasses eight entity categories: bearing type, detection area, wear causes, wear mechanisms, morphological characteristics, wear degree, maintenance measures, and preventative measures. Directed edges describe the relationships between these multi-attribute entities, forming a knowledge representation framework covering "cause-mechanism-degree-maintenance / prevention."

[0022] Knowledge extraction in the wear analysis domain uses the aforementioned knowledge ontology as a prior constraint. It extracts relevant domain knowledge from unstructured texts such as maintenance standard manuals and testing analysis reports to construct a bearing wear analysis knowledge graph. The specific process is as follows: First, long technical documents are segmented into multiple text blocks according to semantic units. A large language model is used to identify entities and relationships that match the knowledge ontology, forming multiple knowledge sub-chains in the form of "node-relationship-node". Second, to address semantic ambiguity caused by multiple expressions for the same semantic entity, an embedding model is used to map the semantics to a high-dimensional vector space. Synonymous entities are aligned and integrated based on vector similarity metrics, aggregating knowledge sub-chains to construct the knowledge graph. Finally, the constructed bearing wear analysis knowledge graph is imported into the Neo4j graph database for knowledge management and visualization.

[0023] Figure 2 The visualization results of the constructed bearing wear analysis knowledge graph are presented. The graph is presented as a network topology composed of multi-attribute nodes and directed edges, intuitively revealing the organization and association of multi-attribute knowledge in wear analysis. Taking fatigue spalling on the inner ring surface of a separable bearing as an example, the corresponding knowledge sub-chains in the graph start from detection information, progressively linking to wear mechanisms and wear levels, and mapping to multi-objective maintenance conclusions such as repair measures and wear causes, forming a complete and traceable reasoning path. Therefore, this wear analysis knowledge graph can provide structured and interpretable knowledge support for bearing maintenance decisions.

[0024] Example 3 Based on Example 2, step S2 is as follows: Step S2: Based on the structured knowledge graph describing the entity and intent entity set, design a knowledge subgraph retrieval and refinement strategy, use set operations to explicitly model the relationship between entities, and construct comprehensive constraints to select knowledge subgraphs that are strongly related to the bearing wear description. 1) Knowledge Subgraph Retrieval Knowledge subgraph retrieval is based on a set of structured wear description entities and intent entities, selecting knowledge subgraphs relevant to the query question from the global knowledge graph. This paper uses the k-hop path search method to extract knowledge subgraphs: based on each query entity... Starting with the first step, the search proceeds by traversing the knowledge graph along relation edges with a finite step size, enumerating paths that connect wear description entities to the intent entity, and treating these as candidate knowledge chains. To address the issue of redundant chains arising from this path enumeration, the retrieval process performs post-processing on the obtained knowledge chain set through inclusion screening. The resulting knowledge chains encompass reasoning paths from different wear description entities to the target decision conclusion, and their union constitutes a candidate knowledge subgraph for the current query question.

[0025] 2) Knowledge Subgraph Refinement The initial knowledge subgraph obtained based on path search often contains redundant knowledge chains, i.e., knowledge chains inconsistent with the multi-attribute bearing wear description. For example, in the maintenance measure decision for the inner raceway surface of a separable bearing, path search may recall knowledge chains that involve the separable bearing but whose detection area is not the inner raceway surface. This is because the path search method cannot explicitly model the coupling constraints of multi-attribute wear description entities, resulting in knowledge chains inconsistent with the overall wear description being mixed into the knowledge subgraph retrieval results, thus affecting the accuracy of maintenance decisions. To address this, this method designs a knowledge subgraph refinement strategy based on the initial knowledge subgraph retrieval results, jointly constructs comprehensive constraints for knowledge chains with multi-attribute query entities, and eliminates irrelevant knowledge chains accordingly, providing concise and highly relevant domain knowledge support for maintenance decisions.

[0026] The wear description in the query involves multi-attribute entities with explicit set relationships. Especially in maintenance decision-making scenarios involving surfaces with coexisting multi-source wear mechanisms, candidate knowledge chains should cover the wear mechanism and its corresponding wear degree, or involve conclusions about the overall wear degree. Simultaneously, the knowledge chain should match entities such as bearing type and inspection type in the query. Therefore, knowledge subgraph refinement should comprehensively utilize multi-attribute wear description information to further filter candidate knowledge chains, thereby obtaining a knowledge subgraph highly relevant to the query.

[0027] Considering the multi-attribute description of bearing wear, this section employs set operations to explicitly model the relationships between entities such as wear mechanisms and wear degrees, and constructs comprehensive constraints to select knowledge chains strongly related to bearing wear description. First, for maintenance decisions on surfaces with coexisting multi-source wear mechanisms, the set attribute annotation rules for wear description entities are defined as follows: each wear mechanism and its corresponding wear degree form an intersection relationship; different wear mechanisms form a union relationship; and other attribute entities form an intersection relationship. Second, based on the above annotation rules, comprehensive constraints are constructed from multi-attribute query entities through multi-level set operations. Specifically, as shown in formulas (2) and (3), the set relationship between entities describing wear of similar and dissimilar attributes is explicitly characterized. Finally, it is determined whether each candidate knowledge chain in the retrieved knowledge subgraph satisfies the above comprehensive constraints. Thus, from the initial knowledge subgraph The knowledge chains that are completely related to the query question are filtered out and formed into an optimized knowledge subgraph, as shown in formula (4).

[0028] (2) (3) (4) In the formula — Constraints consisting of wear mechanism and wear degree query entities; — A collection of query entities for other attributes (non-wear mechanism and wear degree); — Wear mechanism entity Corresponding degree of wear; — Overall wear and tear physical condition; — Querying the mechanistic entities in the problem gather.

[0029] Example 4 Based on Example 3, step S3 is as follows: Step S3: Based on the structured knowledge graph describing the entity and intent entity set, design a knowledge subgraph retrieval and refinement strategy, use set operations to explicitly model the relationship between entities, and construct comprehensive constraints to select knowledge subgraphs that are strongly related to the bearing wear description. 1) Query question construction Regarding bearing maintenance decision-making, the query consists of two parts: wear description information and query intent. Wear description information includes data entered by the testing system, such as bearing type and testing area, as well as wear mechanism identification results and wear degree assessment conclusions derived from the evaluation model. These elements together constitute a multi-attribute set of wear description entities. The query intent indicates the type of decision-making information desired, mainly involving categories such as maintenance measures, preventive measures, and wear-related causes. Entities in these categories collectively constitute the intent entity set. .

[0030] For wear description entities With standardized entities in knowledge graphs There may be inconsistencies in the descriptions between different entities. This section addresses entity alignment based on distance metrics. Specifically, this entity alignment process uses an embedding model to align wear-description entities. Similar to other map entities The similarity between the two is measured by mapping to a continuous vector space and using vector distance to select the candidate graph entity with the highest similarity as the alignment result for subsequent retrieval and inference. The similarity measure uses cosine similarity as the standard, as shown in formula (1).

[0031] (1) In the formula — The degree of similarity between entities; — Vectorized coding based on the BGE-M3 model.

[0032] The wear description information after entity alignment and the query intent together constitute the final query question, which is further encoded into two forms: one is a structured set of entities, used to support the retrieval and refinement of knowledge subgraphs; the other is a semi-structured text paragraph, which provides contextual information for reasoning and decision generation based on a large language model.

[0033] 2) Decision generation The decision generation process uses a large language model as its reasoning engine, deriving maintenance decision suggestions based on the query question and an optimized knowledge subgraph. This process employs semi-structured hints as model input, integrating the optimized knowledge subgraph with the query question, and supplementing it with maintenance decision rules and typical examples to constrain the reasoning process. For the reasoning of maintenance measures, since wear descriptions may involve multi-source wear mechanisms and multi-level wear degree assessments, different knowledge chains may provide inconsistent maintenance measure suggestions. To address this decision conflict, the hints explicitly stipulate that maintenance measure reasoning should adopt a conservative principle, prioritizing more stringent maintenance measures to ensure the overall reliability of the machine. Furthermore, the hints require that the supporting knowledge chain be provided along with the output conclusion, thereby improving the traceability and interpretability of the reasoning process.

[0034] To standardize the construction of bearing maintenance decision prompts, this paper is based on the CRISPE framework. By integrating the query question, knowledge graph, and decision rules, a prompt template as shown in Table 1 was constructed. This framework decomposes the prompt into six key elements: background, role, instruction, topic, preset, and exception. The background focuses the problem on the bearing maintenance decision-making domain; the role sets the large language model to act as a bearing wear maintenance decision-making expert with professional knowledge and rich experience; the instruction integrates the maintenance question and knowledge subgraph; the topic limits the large language model's reasoning to focus solely on wear analysis and maintenance decision-making tasks; the preset clarifies the maintenance decision rules and output decision requirements; and the exception specifies the processing strategy under conditions of incomplete wear description or missing knowledge graph rules.

[0035] Table 1. Prompt Templates for Bearing Maintenance Decision Generation

[0036] Example 5 Based on Example 4, step S4 is as follows: Step S4: Using the two-dimensional image and three-dimensional morphology of the bearing wear surface as input, construct an application and evaluation method for the intelligent maintenance decision-making method. The method is evaluated and optimized through accuracy, BLEU score, BERT score, and generation quality ranking to achieve intelligent maintenance decision-making for the bearing wear surface; details are as follows: 1) Model Configuration To address the confidentiality requirements of bearing wear data, the proposed maintenance decision-making method relies on the Ollam framework for localized model deployment and invocation, primarily involving the BGE-M3 embedding model and the Qwen3-14B large language model. The BGE-M3 embedding model possesses multilingual text representation capabilities, encoding textual descriptions of wear entities into vector representations for entity alignment during knowledge subgraph retrieval. The Qwen3-14B large language model possesses long-contextual understanding capabilities to comprehend textual domain knowledge, wear descriptions, and inference rules, thereby deriving reliable maintenance decision conclusions and corresponding knowledge chains. Furthermore, this method employs a knowledge retrieval enhancement paradigm to inject domain knowledge into inference hints in textual form, and combined with the zero-shot inference capability of the Qwen3-14B large language model, maintenance decision inference can be achieved without parameter fine-tuning, reducing the complexity of model deployment.

[0037] 2) Model Application Examples This bearing maintenance decision-making method can be applied to various tasks, including recommending maintenance measures, suggesting preventative measures, and analyzing wear causes. Furthermore, by combining structured domain knowledge representation in the form of a knowledge graph with the general reasoning capabilities of a large language model, this method can enable the querying of wear analysis knowledge. Table 2 provides examples of responses to bearing maintenance decisions and knowledge queries. The results show that the responses generated by the method highly match the query questions and are consistent with wear analysis knowledge, preliminarily validating its effectiveness in different tasks. Moreover, the generated responses contain the knowledge chain upon which the reasoning decision relies, ensuring the interpretability and traceability of the decision results.

[0038] Table 2 Examples of Maintenance Decision-Making and Knowledge Retrieval Applications

[0039] 3) Model performance evaluation To evaluate the performance of the intelligent maintenance decision-making method, this paper uses bearings subjected to accelerated wear testing as the object, acquiring two-dimensional images and three-dimensional morphology of the wear surface for method verification. Specifically, a multi-light source imaging system is used to acquire images of the bearing wear surface, and the wear surface is characterized in two dimensions using full-light source images; simultaneously, relying on photometric stereo vision technology, the three-dimensional morphology of the wear surface is reconstructed through photometric image sequences, achieving a holographic representation of the typical morphology of the wear surface. Using this data acquisition process, a total of 150 sets of wear surface samples were collected. Some of the collected wear surface samples are shown below. Figure 4 As shown; #1 wear type is spalling, #2 and #3 wear types are spalling and scratching, #4 wear type is scratching and pitting, #5 wear type is scratching, and #6 wear type is pitting and scratching. It can be observed that different wear mechanisms lead to different morphological characteristics of the worn surface, and a single wear area often involves the combined effect of multiple wear mechanisms.

[0040] Using the obtained wear surface morphology data as test samples, a semi-structured bearing wear description is generated based on the wear mechanism and wear degree annotation results using a prompt template. A quantitative performance evaluation is then conducted by comparing the inference conclusions from maintenance measures with the manually annotated results. The manually annotated results include the maintenance measure conclusions and the corresponding reasoning chain. This validation uses the ratio of correct samples to the total number of samples as the accuracy rate and introduces the following multiple indicators to comprehensively evaluate the quality of the maintenance decision conclusions generated by the method from different dimensions.

[0041] (1) Bilingual Evaluation Understudy (BLEU): BLEU is based on n-grams ( The similarity between the generated text and the reference text is measured by the matching degree of a segment consisting of consecutive words, in order to evaluate the quality of the generated text. Specifically, the generated text and the reference text (manually annotated results) are decomposed into n-grams of different orders, the matching accuracy of each level of n-gram is calculated, and a length penalty factor (Brevity Penalty, BP) is introduced to alleviate the problem of the generated text being too short and thus having a high score. The calculation process of BLEU is shown in formulas (5) and (6), and its value range is as follows: A higher score indicates higher quality generated text.

[0042] (5) (6) In the formula — The vocabulary length of the text generated by the model; — The vocabulary length of the reference text; — The accuracy calculation weights of n-grams; — The ratio of the number of n-grams that match the generated text to the number of n-grams derived from the generated text.

[0043] (2) The BERT score assesses the similarity between the generated text and the reference text by comparing their contextual semantic representations. This metric utilizes the BERT (Bidirectional Encoder Representations from Transformers) model to encode the generated and reference texts into token-level context vectors, and calculates the maximum similarity match based on cosine similarity to solve for the accuracy. and recall rate The F1 score is then synthesized as the final BERT score. The specific calculation is shown in formulas (7) to (9). Compared with the BLEU index, the BERT score can capture semantic similarity and contextual relevance to a certain extent, and is suitable for evaluating the overall quality of complex texts derived by the method.

[0044] (7) (8) (9) In the formula — Generate the set of terms corresponding to the text; — The set of terms corresponding to the reference text; -- gather Lexical elements in; -- gather Lexical elements in; — Lexical units and The cosine similarity.

[0045] (3) Quality ranking: Using the manually annotated results as a reference, the maintenance decision conclusions generated by different models are relatively evaluated and ranked using the large language model (Qwen3). Specifically, this indicator comprehensively evaluates the generated text from the dimensions of correctness of conclusions, relevance to the problem, and rationality of reasoning path by designing prompts that include clear ranking rules and review criteria.

[0046] Based on the aforementioned evaluation metrics, this section quantitatively compares the proposed maintenance decision-making model with several comparative methods. These methods include the DeepSeek-R1 model, the GPT-5 model, and the LightRAG inference framework. Both the DeepSeek-R1 and GPT-5 models employ text knowledge injection to enhance inference decision-making, directly inputting wear analysis-related documents into the model. The LightRAG inference framework encompasses knowledge graph construction and inference decision-making, automatically extracting the knowledge graph from wear analysis knowledge documents. During the inference decision-making stage, it employs multiple retrieval strategies to extract query-related knowledge subgraphs: local retrieval focuses on specific entities involved in the query question; global retrieval emphasizes extracting knowledge content semantically related to the overall query question; and hybrid retrieval combines local and global retrieval to obtain more complete domain knowledge. To ensure objectivity, the LightRAG framework uses the same Qwen3-14B large language model and the same decision generation prompt template as the method presented in this paper.

[0047] Table 3 lists the quantitative evaluation results of the maintenance decision-making methods and control methods in deriving maintenance recommendations. It can be observed that DeepSeek-R1 and GPT-5, two large language models based on unstructured text injection, only achieved maintenance recommendation accuracy of 47.33% and 44.67%, respectively, with relatively low BLEU and BERT scores. This indicates that relying solely on the model's own contextual understanding ability to process unstructured domain knowledge is insufficient to fully capture the complex relationships between multi-attribute wear analysis knowledge, leading to insufficient decision accuracy and decreased quality of generated text. In contrast, the LightRAG framework, which incorporates structured representations from knowledge graphs, improved accuracy, with the global retrieval strategy achieving 70.67%, higher than local retrieval (52.00%) and hybrid retrieval (66.67%). This suggests that structured knowledge representation and a holistic semantic-based retrieval mechanism help to more effectively match multi-attribute wear descriptions, thereby improving the reliability of decisions. The maintenance decision-making method proposed in this paper achieves a decision accuracy of 98.67%, significantly higher than the control model (47.33%-70.67%). Furthermore, its BLEU and BERT scores, as well as its ranking of generated quality, are superior to the control model. These results demonstrate that the proposed method can achieve high-quality and highly reliable maintenance decisions. In addition, the supporting reasoning chain output by this method is highly consistent with the manually labeled results, exhibiting good traceability and interpretability.

[0048] Table 3. Quantitative evaluation results of maintenance measures under different maintenance decision-making methods

[0049] The proposed maintenance decision-making method, after initial retrieval of the knowledge subgraph, further introduces knowledge subgraph refinement based on multi-attribute entity set relationships to filter knowledge chains highly relevant to the overall wear description, thereby improving the accuracy of inference. To verify the effectiveness of this hierarchical retrieval strategy, this section compares the model's performance in terms of the accuracy of maintenance measure conclusions and the quality of responses under two settings: with and without knowledge subgraph refinement. The results are shown in Table 4. Comparative analysis shows that after introducing knowledge subgraph refinement, the number of inference chains retrieved by the model is significantly reduced from 131.00 to 7.56, a decrease of 94.23%, effectively filtering redundant information. At the same time, the accuracy of maintenance decisions significantly increases from 46.67% to 98.67%. In addition, the quality of the generated text also improves simultaneously, with the BLEU score increasing from 0.12 to 0.26 and the BERT score increasing from 0.73 to 0.79. The results show that the knowledge subgraph refinement strategy can accurately identify knowledge chains that perfectly match the multi-attribute bearing wear description, reduce the interference of irrelevant knowledge on the reasoning process, and improve the reliability and output quality of maintenance decisions.

[0050] Table 4. Quality Comparison of Maintenance Decision Conclusions Derived from the Model in This Paper and the Control Model

[0051] Example 6 Based on Example 1, step 2 is as follows: Step 2.1: Construct the query question; The query consists of two parts: wear description information and query intent. The wear description information includes bearing type, detection area, wear mechanism identification results, wear degree, and morphological characteristics. These elements together constitute a multi-attribute set of wear description entities. ; The query intent indicates the type of decision-making information expected, including maintenance measures, preventive measures, and wear-related causes. The corresponding categories of entities together constitute the intent entity set. ; Step 2.2: Based on the structured wear description entity and intent entity set, select the knowledge subgraph related to the query question from the global knowledge graph; Step 2.3: Construct a comprehensive constraint for knowledge chains by jointly querying entities with multiple attributes, and eliminate irrelevant knowledge chains accordingly; thus obtaining an optimized knowledge subgraph.

[0052] Comprehensive constraints are constructed from multi-attribute query entities through multi-level set operations. This approach explicitly characterizes the set relationships between entities with similar and dissimilar attribute wear descriptions. Finally, it determines whether each candidate knowledge chain in the retrieved knowledge subgraph satisfies the aforementioned comprehensive constraints. Thus, from the initial knowledge subgraph The knowledge chains that are completely relevant to the query question are filtered out, and an optimized knowledge subgraph is formed.

[0053]

[0054]

[0055] In the formula, —Constraints constituted by the wear mechanism and wear degree query entities; — A collection of query entities for attributes other than wear mechanism and wear degree; — The first one present on the surface to be diagnosed i A wear mechanism entity; — Wear mechanism entity Corresponding degree of wear; —Other attribute query entities besides wear mechanism and wear degree; — Overall wear and tear physical condition; — Query the wear mechanism entity in the question gather; P i —— Knowledge Graph The Middle i A chain of reasoning.

[0056] It should be noted that the present invention is not limited to the specific embodiments described above. For those skilled in the art, several improvements or substitutions can be made based on the concept of the present invention, and all such improvements or substitutions should be considered to fall within the scope of the present invention.

Claims

1. A bearing intelligent maintenance decision-making method combining knowledge graph and large language model, characterized in that, Step 1: Construct a bearing wear knowledge ontology for bearing maintenance decision-making scenarios, and extract knowledge based on bearing wear analysis technical texts to build a structured knowledge graph of aero-engine bearing wear analysis covering causes, mechanisms, degrees, and maintenance / prevention. Step 2: Based on the structured knowledge graph describing the entity and intent entity set, design a knowledge subgraph retrieval and refinement strategy, use set operations to explicitly model the combination relationship between multi-attribute entities, and construct comprehensive constraints to select knowledge subgraphs that are strongly related to the bearing wear description. Step 3: Use the collaborative large language model and knowledge subgraph to reason about multi-attribute wear and tear, and construct an intelligent bearing maintenance decision-making method.

2. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 1, characterized in that, It also includes step 4, which uses two-dimensional images and three-dimensional morphology of the bearing wear surface as input to construct an application and evaluation method for the intelligent maintenance decision-making method. The method is evaluated and optimized by accuracy, BLEU, BERT score and generation quality ranking to realize intelligent maintenance decision-making for the bearing wear surface.

3. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 1, characterized in that, Step 1: Constructing a knowledge graph for aero-engine bearing wear analysis. The specific process is as follows: Step 1.1: Divide the bearing wear analysis technical text into multiple text blocks according to semantic units, and use a large language model to identify entities and relationships between entities that match the knowledge ontology, forming multiple knowledge sub-chains in the form of node-relationship-node; Step 1.2: To address the semantic ambiguity caused by multiple expressions of the same semantic entity, the embedding model BGE-M3 is used to map the semantics to a high-dimensional vector space. Synonymous entities are aligned and integrated based on vector similarity metrics, and knowledge sub-chains are aggregated to construct a knowledge graph. The knowledge graph is presented as a network topology structure composed of multi-attribute nodes and directed edges. The nodes represent multi-attribute entities, and the directed edges describe the relationships between multi-attribute entities. Step 1.3: Import the constructed knowledge graph into the Neo4j graph database to achieve knowledge management and visualization.

4. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 1, characterized in that, The knowledge graph described in step 2 uses entity-relationship-attribute as the basic unit; the entity includes eight categories: bearing type, detection area, wear cause, wear mechanism, morphological characteristics, wear degree, maintenance measures, and prevention measures; the relationship represents the association between the various entities.

5. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 3, characterized in that, Step 2 is as follows: Step 2.1: Construct the query question; The query consists of two parts: wear description information and query intent. The wear description information includes bearing type, detection area, wear mechanism identification results, wear degree, and morphological characteristics. These elements together constitute a multi-attribute set of wear description entities. ; The query intent indicates the type of decision-making information expected, including maintenance measures, preventive measures, and wear-related causes. The corresponding categories of entities together constitute the intent entity set. ; Step 2.2: Based on the structured wear description entity and intent entity set, select the knowledge subgraph related to the query question from the global knowledge graph; Step 2.3: Construct a comprehensive constraint for knowledge chains by jointly querying entities with multiple attributes, and eliminate irrelevant knowledge chains accordingly; thus obtaining an optimized knowledge subgraph.

6. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 5, characterized in that, Step 2.2 is as follows: A knowledge retrieval and refinement mechanism is constructed to obtain a knowledge subgraph strongly related to the query question; specifically as follows: The wear description entity is constructed using the embedded model BGE-M3. Similar to other map entities Mapping to a continuous vector space and using vector distance to measure the similarity between the two, the candidate graph entities with the highest similarity are selected as the alignment result for subsequent retrieval and inference; The similarity measure uses cosine similarity as the standard, as shown in formula (1): （1） In the formula — The degree of similarity between entities; — Vectorized coding based on the BGE-M3 embedding model; The wear description information after entity alignment, together with the query intent, constitutes the final knowledge subgraph related to the query question.

7. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 5, characterized in that, In step 2.2, starting from each query entity, the knowledge graph is traversed and searched along directed edges with a finite step size. The paths that can connect the wear description entity to the intent entity are enumerated and used as candidate knowledge chains. The resulting knowledge chains cover the reasoning paths from different wear description entities to the target decision conclusion, and their union constitutes a candidate knowledge subgraph for the current query problem.

8. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 5, characterized in that, Step 2.3 is as follows: Comprehensive constraints are constructed from multi-attribute query entities through multi-level set operations. Specifically, as shown in formulas (2) and (3), the set relationship between entities describing wear of similar and dissimilar attributes is explicitly characterized. Finally, it is determined whether each candidate knowledge chain in the retrieved knowledge subgraph satisfies the above comprehensive constraints. Thus, from the initial knowledge subgraph The knowledge chains that are completely related to the query question are filtered out and an optimized knowledge subgraph is formed, as shown in formula (4); （2） （3） （4） In the formula, —Constraints constituted by the wear mechanism and wear degree query entities; — A collection of query entities for attributes other than wear mechanism and wear degree; — The first one present on the surface to be diagnosed i A wear mechanism entity; — Wear mechanism entity Corresponding degree of wear; —Other attribute query entities besides wear mechanism and wear degree; — Overall wear and tear physical condition; — Query the wear mechanism entity in the question gather; P i —— Knowledge Graph The Middle i A chain of reasoning.

9. The bearing intelligent maintenance decision-making method combining knowledge graph and large language model according to claim 7, characterized in that, The technical documents for bearing wear analysis include maintenance standard manuals and testing and analysis reports.

Images