Manufacturing process knowledge multi-path retrieval question and answer method based on knowledge graph enhancement

Abstract

The invention discloses a manufacturing process knowledge multi-path retrieval question and answer method based on knowledge graph enhancement. The method comprises the following steps: firstly, performing knowledge structured extraction on a manufacturing process document; constructing a multi-modal index and a knowledge graph; intelligent analysis of the query intention is carried out; then, multi-path mixed retrieval of knowledge graph enhancement is carried out; then depth reordering and context enhancement are carried out; and finally completing answer generation based on the enhanced context. According to the method, the defect that a traditional analysis mode is insufficient in understanding of a complex document structure is overcome, a solid structured foundation is laid for cross-document complex technology reasoning, and the reasoning ability and the answering accuracy of the system are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph technology, specifically relating to a multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement. Background Technology

[0002] With the deepening of industrial digitalization and intelligent manufacturing, enterprises have accumulated a massive amount of manufacturing process documents, including unstructured technical documents such as technical standards and specifications, process flow documents, operating procedures, and equipment maintenance manuals. These documents are the core knowledge source for process design, production decisions, and troubleshooting. However, due to the highly specialized, logically rigorous, and complex parameter relationships inherent in manufacturing process knowledge, traditional information retrieval methods are insufficient to meet practical needs. When faced with complex engineering problems involving multiple documents, technical personnel often need to sort out multi-step causal chains, trace process parameter dependencies, and conduct comparative analyses of multiple constraints. However, due to fragmented knowledge storage, information retrieval efficiency is low, easily leading to retrieval blind spots or logical gaps, posing potential risks to production safety and technical decision-making.

[0003] Existing intelligent question-answering technologies primarily rely on single vector similarity retrieval or keyword matching. While these can locate text fragments, they struggle to capture the deep topological structures and logical connections between knowledge points. Keyword retrieval cannot understand semantic variations, and simple vector retrieval often returns only fragmented paragraphs, lacking an effective structured expression and storage mechanism for the "entity-attribute-relationship" triple structure (such as "process parameter-numerical range-constraints") prevalent in manufacturing process knowledge. When faced with complex queries requiring multi-step logical reasoning (such as "what defects will occur when a certain material is processed using a certain process at a certain temperature"), existing methods cannot effectively integrate related knowledge scattered across different document chapters, and the relevance ranking of search results is often inaccurate. This results in generative large models lacking rigorous and complete contextual support when answering questions, easily leading to "machine illusions" or generating unfounded incorrect answers. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement. First, it extracts structured knowledge from manufacturing process documents; next, it constructs a multimodal index and knowledge graph; then, it intelligently parses the query intent; next, it performs multi-path hybrid retrieval enhanced by the knowledge graph; then, it performs deep re-ranking and context enhancement; finally, it generates answers based on enhanced context. This invention overcomes the shortcomings of traditional parsing methods in understanding complex document structures, lays a solid structured foundation for complex technical reasoning across documents, and significantly improves the system's reasoning ability and answer accuracy.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Knowledge structure extraction from manufacturing process documents; Transform unstructured process documents into structured knowledge entities; Step 2: Multimodal indexing and knowledge graph construction; Establish a composite index structure that supports multiple retrieval methods, and construct a knowledge graph to represent the logical connections of process knowledge; Step 3: Intelligent parsing of query intent; Deconstruct the user's natural language query into structured search instructions; Step 4: Knowledge graph-enhanced multi-path hybrid retrieval; Retrieve relevant process knowledge through multiple parallel retrieval paths; Step 5: Deep reordering and context enhancement; The candidate knowledge is sorted and a structured context is constructed. Step 6: Answer generation based on enhanced context; Use a language generation model to generate answers based on enhanced context.

[0006] Preferably, step 1 specifically comprises: Step 1.1: Parse and segment the process document, dividing it into independent text units according to the document's hierarchical structure; Step 1.2: Identify the knowledge type of the text unit and determine the category of process knowledge it contains; Step 1.3: Based on the identified knowledge type, extract structured knowledge elements using the corresponding extraction strategy; Step 1.4: Verify and standardize the extracted structured data to form standardized process knowledge entries.

[0007] Preferably, the process knowledge category includes at least one or more of the following: terminology definition, process parameter, process step, constraint, and causal relationship.

[0008] Preferably, step 2 specifically comprises: Step 2.1: Create a document context node in the graph database to record the document's metadata; Step 2.2: Instantiate the structured data extracted in Step 1 into process knowledge nodes and establish the association relationship with the corresponding document context nodes; Step 2.3: Perform entity normalization on the core elements in the process knowledge nodes to create concept nodes; Step 2.4: Based on the inherent logic of process knowledge, establish semantic relationship edges between process knowledge nodes and concept nodes to form a process knowledge graph; Step 2.5: Use semantic embedding technology to convert the text description of process knowledge nodes into vector representations and construct a semantic vector index; Step 2.6: Build an exact matching index for the key fields in the process knowledge node.

[0009] Preferably, the semantic relation edges include at least one or more of the following: definition relation, parameter relation, step relation, causal relation, and constraint relation.

[0010] Preferably, step 3 specifically comprises: Step 3.1: Receive the natural language query input from the user; Step 3.2: Perform intent analysis on the query to identify the query type, query entities, and query constraints; Step 3.3: Extract generalized search terms for fuzzy matching and entity names for exact matching from the query; Step 3.4: Encapsulate the extracted search terms, entities, and query types into structured search instructions.

[0011] Preferably, step 4 specifically comprises: Step 4.1: Perform semantic similarity retrieval based on generalized search terms to recall semantically related process knowledge nodes; Step 4.2: Perform keyword matching retrieval based on precise entities to recall process knowledge nodes containing specific terms; Step 4.3: For complex queries involving multiple entities, perform relational reasoning in the knowledge graph, search for logical paths connecting different entities, and recall intermediate process knowledge nodes on the paths; Step 4.4: For cases where single entity queries or relational reasoning do not sufficiently recall nodes, expand the graph starting from the precisely recalled nodes and filter neighbor nodes based on semantic similarity. Step 4.5: Merge and deduplicate the process knowledge nodes retrieved by multi-way retrieval to generate a candidate knowledge set.

[0012] Preferably, step 5 specifically comprises: Step 5.1: Obtain the complete text description of each process knowledge node in the candidate knowledge set; Step 5.2: Construct text pairs between queries and process knowledge; Step 5.3: Calculate the relevance score between the query and process knowledge using a deep semantic matching model; Step 5.4: Sort the data according to relevance scores and filter out the highly relevant process knowledge; Step 5.5: Extract the selected process knowledge nodes and their relationship edges from the knowledge graph, and construct a reasoning subgraph; Step 5.6: Transform the reasoning subgraph into natural language text containing knowledge content and relational descriptions to form an enhanced context.

[0013] Preferably, step 6 specifically comprises: Step 6.1: Construct a generation prompt template that includes role definitions, task instructions, and constraints; Step 6.2: Fill the enhanced context and original user query generated in Step 5 into the prompt template; Step 6.3: Input the complete prompt information into the language generation model for reasoning; Step 6.4: The language generation model performs logical fusion and reasoning on the process knowledge in the enhanced context to generate an answer that includes the conclusion, reasoning process, and knowledge source.

[0014] The beneficial effects of this invention are as follows: 1. This invention constructs an efficient industrial knowledge tracing and multimodal fusion system. Through intelligent segmentation and multimodal knowledge extraction technology based on a large language model, the system can accurately capture key elements, technical parameters, and process logic in unstructured technical documents, and achieve low-cost incremental iteration of massive amounts of data with a file hash comparison mechanism. Based on this, a multimodal joint index constructed using graph and vector databases achieves deep fusion of knowledge in semantic space and topological structure. This not only intuitively demonstrates the hierarchical belonging between discrete knowledge points, but also reveals the implicit logic and causal chains scattered across different document chapters through the knowledge graph, thus overcoming the shortcomings of traditional parsing methods in understanding complex document structures and laying a solid structured foundation for complex technical reasoning across documents.

[0015] 2. This invention innovatively proposes a topology-semantic collaborative multi-path hybrid retrieval and enhanced generation mechanism, significantly improving the system's reasoning ability and answer accuracy. Through a goal-driven path reasoning algorithm, the system can forcibly recall the optimal logical links connecting discrete entities, solving the problem of broken links in multi-hop reasoning. Simultaneously, combined with a semantic-aware expansion strategy, the graph expansion process is dynamically pruned and denoised, effectively solving the semantic gaps and long-tail entity omissions problems inherent in traditional retrieval. Finally, enhanced context is constructed using local relational subgraph serialization technology, transforming the retrieval results into natural language descriptions with logical derivation capabilities. This effectively suppresses the "illusion" phenomenon of generative models, ensuring the rigor and traceability of answers, and providing a powerful intelligent tool for fault diagnosis, technical decision-making, and standard querying in industrial production. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2Example of knowledge graph visualization at the instance layer in this invention Figure 1 ; Figure 3 Example of knowledge graph visualization at the instance layer in this invention Figure 2 . Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] This invention aims to provide a multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement. By integrating semantic understanding, precise matching and relational reasoning into a multi-path retrieval mechanism, it constructs a high-precision, structured knowledge context to achieve accurate answers and logical tracing of complex technical questions.

[0019] 1. Knowledge structure extraction from manufacturing process documents; 1.1 Construct a domain knowledge dataset containing 184 standard specification documents (covering GB national standards, HB aviation industry standards, GJB national military standards) as well as several process documents, professional books and other technical materials. All original documents have been preprocessed into a unified Markdown (.md) format to preserve the document's title hierarchy structure. 1.2 The Markdown document parser is used to read the contents of the .md file, and non-text characters are cleaned based on Markdown syntax features and structured segmentation is performed to break down long documents into semantic text chunks with independent contexts. 1.3 For extremely long texts where the content of a single chapter exceeds the processing limit, an overlapping segmentation strategy is adopted for secondary segmentation. Specifically, the target block size is set to 4000 characters, which is large enough to accommodate a complete technical description paragraph without exceeding the effective processing window of the subsequent large language model. The overlap interval is set to 1000 characters, meaning there is a 1000-character overlap between adjacent text blocks. 1.4 Input the text block into the large language model for "triage" and prediction. Construct specific prediction prompts to instruct the model to identify the specific knowledge type contained in the text block, including technical requirements (such as material performance indicators, process parameter ranges), terminology definitions (such as explanations of professional terms, abbreviations), process steps (such as operation procedures, processing sequence), or question-and-answer pairs (such as common questions and answers). After processing, the model outputs a standard JSON format list, with each list item containing a knowledge type label and the corresponding text segment position. For example, if the input text contains "The isothermal forging temperature should be controlled within the range of 920-950°C, and the temperature deviation should not exceed ±5°C", the model outputs the following JSON: [{"type": "Technical Requirements", "segment":"The isothermal forging temperature should be controlled within 920-950°C..."}]. 1.5 Based on the knowledge type predicted in step 1.4, the corresponding extraction template is dynamically matched from the pre-designed Prompt template library, and the large language model is called again to accurately extract key information elements. Different knowledge types correspond to different extraction templates and output format requirements.

[0020] For knowledge categorized as "technical requirements," construct a prompt template for mandatory formatted output. The model must output strictly JSON-formatted data, including the subject (e.g., "isothermal forging temperature"), property (e.g., "temperature range"), value (e.g., "920-950"), unit (e.g., "°C"), and constraint (e.g., "deviation not exceeding ±5°C"). The template should explicitly state: "Should output in JSON format, must not contain any explanatory text," and provide 3-5 examples to guide the model's output specification.

[0021] For knowledge in the "Term Definition" category, the prompt word template requires the output to include a JSON structure containing term (term name), definition (definition content), abbreviation (abbreviation, if any), and synonyms (synonyms list). For example, from "TC4 titanium alloy, also known as Ti-6Al-4V, is an α+β type two-phase titanium alloy", the following can be extracted: {"term": "TC4 titanium alloy", "definition": "α+β type two-phase titanium alloy", "abbreviation": "Ti-6Al-4V", "synonyms": ["TC4", "Ti-6Al-4V"]}.

[0022] For knowledge related to "process steps", the prompt word template requires the output to include a JSON structure containing process_name, steps, preconditions, and postconditions. Steps is an array, and each step contains fields such as step_order, step_description, and key_parameters. 1.6 Parse the JSON data returned by the model, perform field integrity verification, and store the verified structured knowledge point data into the MySQL database.

[0023] 2. Multimodal indexing and knowledge graph construction; 2.1 Create the `documents` and `chapters` tables in the MySQL database to store document-level and chapter-level metadata, respectively. Simultaneously, initialize a graph data structure (a directed graph built using the NetworkX library) in memory, creating "context nodes" corresponding to the database records. These context nodes serve as the root nodes for knowledge tracing, providing original source information for subsequent answer tracing. 2.2 Traverse the `knowledge` table in the MySQL database and create a corresponding "knowledge point node" in the memory graph for each structured knowledge record. Simultaneously, establish a `BELONGS_TO` relationship edge between the knowledge point node and its source chapter node. The edge's attribute includes `extraction_confidence` (extraction confidence, determined by the probability value output by the large model). 2.3 Entity normalization is performed on core elements of knowledge points (such as terminology names, process components, material names, equipment names, etc.). The system maintains an entity mapping dictionary to unify the same concept in different expressions into standard concept nodes; 2.4 Establish logical relationship edges between knowledge point nodes and concept nodes based on semantic rules, and preset the weights of the edges according to the strength of the relationship. Establish directed edges such as DEFINES (definition), HAS_PROPERTY (having attributes), IS_STEP_OF (belonging to steps), CAUSES (causing), CONSTRAINS (constraints) to form a complete knowledge graph network with causal and logical expression capabilities; 2.5 Using the BAAI / bge-large-zh-v1.5 deep semantic embedding model, the text description of each knowledge point is converted into a 1024-dimensional high-dimensional dense vector and stored in the ChromaDB vector database to support semantic retrieval; 2.6 In the knowledge table of the MySQL database, a functional inverted index is built for the JSON field in the database and stored in the MySQL database to support precise keyword matching based on SQL.

[0024] 3. Intelligent interpretation of query intent; 3.1 Receive the user's natural language query input, perform basic text cleaning to remove extra spaces, special characters, etc., and then pass the standardized query to the intent analysis module; 3.2 The query is input into the LLM-based intent analysis module, which uses Few-Shot Prompting (FSP) technology to deconstruct the deep semantic structure of the query. The prompt words are designed as follows: "You are a query intent analysis expert. Please analyze user queries and extract key information. Output in JSON format, including: query_type (query type: definition query / parameter query / process query / causal query), entities (list of specific entity names involved), and keywords (list of generalized keywords for semantic retrieval).

[0025] Example: Input 'What is the heat treatment temperature of TC4 material?' Output: {"query_type": "parameter query","entities": ["TC4", "heat treatment"], "keywords": ["TC4 material", "heat treatment temperature", "temperature parameter"]}; Now let's analyze: What problems might arise if the forging temperature is too high during the isothermal forging of TC4 titanium alloy blades? The large language model, based on example learning, outputs the following JSON result: {"query_type": "causal query", "entities": ["TC4 titanium alloy", "isothermal forging", "forging temperature"], "keywords": ["TC4", "titanium alloy blade", "isothermal forging", "overheating", "problem", "defect"]} 3.3 Extract generalized search terms for vector fuzzy matching from the query, and identify and extract two or more specific entity names involved in the query to support subsequent path reasoning. In this example, three entities are identified: "TC4 titanium alloy", "isothermal forging", and "forging temperature". The system standardizes these entities and compares them with the entity mapping dictionary in step 2.3, standardizing "TC4 titanium alloy" as the "TC4 titanium alloy" concept node in the graph. Simultaneously, the keywords field is extracted, which contains generalized search terms for vector fuzzy matching. These keywords are broader than entity names, including "TC4", "titanium alloy blade", "isothermal forging", "excessive temperature", "problem", and "defect", etc., to broaden the recall scope during semantic retrieval.

[0026] 3.4 The extracted search terms and entities are encapsulated into a structured JSON search instruction object and passed to the hybrid search engine. The format is as follows: { "query_type": "Causal query", "entities": ["TC4 titanium alloy", "isothermal forging", "forging temperature"], "entity_count": 3, Keyword: ["TC4", "Titanium Alloy Blades", "Isothermal Forging", "Overheating", "Problems", "Defects"] "original_query": "What problems will occur if the forging temperature is too high during isothermal forging of TC4 titanium alloy blades?" "enable_graph_reasoning": true } 4. Knowledge graph-enhanced multi-path hybrid retrieval; 4.1 Path A (Semantic Similarity Retrieval): Input the generalized search terms extracted in step 3 into the BGE embedding model to generate query vectors, perform KNN (K-Nearest Neighbor) search in ChromaDB, and retrieve the top-5 knowledge point IDs with the highest semantic similarity; 4.2 Path B (Precise Keyword Retrieval): Using the entity names extracted in step 3, perform a precise matching query based on the JSON index in the MySQL database to locate the core knowledge point ID containing specific terms; 4.3 Path C-1 (Goal-Driven Path Reasoning): For complex queries involving two or more entities, a goal-driven reasoning mechanism is initiated. In the memory graph, the optimal logical link connecting these scattered entity nodes is calculated using either a shortest path algorithm or a full path search algorithm. The system retains all intermediate nodes on the path, thus resolving the broken link problem in multi-hop reasoning. 4.4 Path C-2 (Semantic Aware Graph Expansion): For queries containing only a single entity or where path reasoning fails, a semantic awareness expansion mechanism is initiated. Using nodes precisely recalled by path B as anchor points, a breadth-first search (BFS, depth=2) is performed. During the expansion process, the cosine similarity between the candidate neighbor node vector and the query vector is calculated in real time. A dynamic threshold is set to retain only high-confidence associated nodes and remove semantically irrelevant noisy neighbors, achieving dynamic pruning of the graph expansion. 4.5 Perform a union operation on all knowledge point IDs recalled by the semantic path (A), the precise path (B), and the two topological collaborative paths (C-1, C-2) to generate a deduplicated set of candidate knowledge points.

[0027] 5. Deep reordering and context enhancement; 5.1 Traverse the candidate knowledge point set and retrieve the complete text description and metadata of each knowledge point from the database; 5.2 Combine the user's original query with the content of each candidate knowledge point into a "query-document" pair; 5.3 Input these pairs into the BAAI / bge-reranker-large cross-encoder model to deeply calculate the semantic relevance score (Score 0~1) between the query and the knowledge point. 5.4 Based on the relevance scores obtained in step 5.3, the candidate knowledge points are sorted in descending order, the top-5 highly relevant knowledge points are selected, and low-scoring noisy data are removed. 5.5 Extract the Top-5 knowledge point nodes and their interconnecting edges from the memory graph to construct a local "reasoning subgraph". Execute the subgraph extraction algorithm: First, add these 5 nodes to the subgraph node set; then, iterate through any two nodes to see if there is a direct connecting edge, and if so, add the edge to the subgraph edge set; simultaneously, check if these 5 nodes are connected to any intermediate nodes (such as common concept nodes), and if so, add the intermediate nodes and related edges to the subgraph as well. 5.6 Employs a subgraph serialization algorithm to convert subgraphs into natural language text. First, the core knowledge points are formatted and output. Then, the topological relationships between knowledge points are explicitly described, generating an enhanced context containing logical chains.

[0028] 6. Answer generation based on enhanced context; 6.1 Construct prompt word templates for generative tasks, define the "domain expert" role of the large model, and embed "thinking chain" instructions and strict "context-based answer" constraints; 6.2 Fill the {context} area of ​​the template with the structured enhanced context generated in step 5, fill the {query} area with the original user query, and attach the source document name for tracing purposes; 6.3 Send the completed prompt words to the locally deployed large language model inference interface; The 6.4 model performs logical fusion and reasoning on scattered knowledge points in the enhanced context, and identifies and eliminates interfering information that is irrelevant to the query intent; The 6.5 model generates natural language responses that include the final conclusion, logical derivation process, and cited sources based on the constraints, and outputs them to the user.

Claims

1. A multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement, characterized in that, Includes the following steps: Step 1: Knowledge structure extraction from manufacturing process documents; Transform unstructured process documents into structured knowledge entities; Step 2: Multimodal indexing and knowledge graph construction; Establish a composite index structure that supports multiple retrieval methods, and construct a knowledge graph to represent the logical connections of process knowledge; Step 3: Intelligent parsing of query intent; Deconstruct the user's natural language query into structured search instructions; Step 4: Knowledge graph-enhanced multi-path hybrid retrieval; Retrieve relevant process knowledge through multiple parallel retrieval paths; Step 5: Deep reordering and context enhancement; The candidate knowledge is sorted and a structured context is constructed. Step 6: Answer generation based on enhanced context; Use a language generation model to generate answers based on enhanced context.

2. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Parse and segment the process document, dividing it into independent text units according to the document's hierarchical structure; Step 1.2: Identify the knowledge type of the text unit and determine the category of process knowledge it contains; Step 1.3: Based on the identified knowledge type, extract structured knowledge elements using the corresponding extraction strategy; Step 1.4: Verify and standardize the extracted structured data to form standardized process knowledge entries.

3. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 2, characterized in that, The process knowledge categories include at least one or more of the following: terminology definition, process parameter, process step, constraint, and causal relationship.

4. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 2, characterized in that, Step 2 specifically involves: Step 2.1: Create a document context node in the graph database to record the document's metadata; Step 2.2: Instantiate the structured data extracted in Step 1 into process knowledge nodes and establish the association relationship with the corresponding document context nodes; Step 2.3: Perform entity normalization on the core elements in the process knowledge nodes to create concept nodes; Step 2.4: Based on the inherent logic of process knowledge, establish semantic relationship edges between process knowledge nodes and concept nodes to form a process knowledge graph; Step 2.5: Use semantic embedding technology to convert the text description of process knowledge nodes into vector representations and construct a semantic vector index; Step 2.6: Build an exact matching index for the key fields in the process knowledge node.

5. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 4, characterized in that, The semantic relation edges include at least one or more of the following: definition relation, parameter relation, step relation, causal relation, and constraint relation.

6. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 4, characterized in that, Step 3 specifically involves: Step 3.1: Receive the natural language query input by the user; Step 3.2: Perform intent analysis on the query to identify the query type, query entities, and query constraints; Step 3.3: Extract generalized search terms for fuzzy matching and entity names for exact matching from the query; Step 3.4: Encapsulate the extracted search terms, entities, and query types into structured search instructions.

7. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 6, characterized in that, Step 4 specifically involves: Step 4.1: Perform semantic similarity retrieval based on generalized search terms to recall semantically related process knowledge nodes; Step 4.2: Perform keyword matching retrieval based on precise entities to recall process knowledge nodes containing specific terms; Step 4.3: For complex queries involving multiple entities, perform relational reasoning in the knowledge graph, search for logical paths connecting different entities, and recall intermediate process knowledge nodes on the paths; Step 4.4: For cases where single entity queries or relational reasoning do not sufficiently recall nodes, expand the graph starting from the precisely recalled nodes and filter neighbor nodes based on semantic similarity. Step 4.5: Merge and deduplicate the process knowledge nodes retrieved by multi-way retrieval to generate a candidate knowledge set.

8. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 7, characterized in that, Step 5 specifically involves: Step 5.1: Obtain the complete text description of each process knowledge node in the candidate knowledge set; Step 5.2: Construct text pairs between queries and process knowledge; Step 5.3: Calculate the relevance score between the query and process knowledge using a deep semantic matching model; Step 5.4: Sort the data according to relevance scores and filter out the highly relevant process knowledge; Step 5.5: Extract the selected process knowledge nodes and their relationship edges from the knowledge graph, and construct a reasoning subgraph; Step 5.6: Transform the reasoning subgraph into natural language text containing knowledge content and relation descriptions to form an enhanced context.

9. The multi-path retrieval and question-answering method for manufacturing process knowledge based on knowledge graph enhancement according to claim 8, characterized in that, Step 6 specifically involves: Step 6.1: Construct a generation prompt template that includes role definitions, task instructions, and constraints; Step 6.2: Fill the enhanced context and original user query generated in Step 5 into the prompt template; Step 6.3: Input the complete prompt information into the language generation model for reasoning; Step 6.4: The language generation model performs logical fusion and reasoning on the process knowledge in the enhanced context to generate an answer that includes the conclusion, reasoning process, and knowledge source.

Images