Large language model guided aircraft manufacturing process instruction automatic compilation method

Abstract

The application provides a large language model guided airplane manufacturing process instruction automatic compiling method, constructs a graph structure expert library based on a large language model, analyzes process documents, historical data and field experience through a pre-training model, defines multilevel rules, and designs an artificial feedback mechanism to iteratively optimize the rule library; knowledge storage and dynamic updating are realized through a graph database to form a standardized knowledge base; based on a component object model and parameterized driving non-geometric manufacturing information intelligent extraction, an automatic extraction and multi-modal verification fusion strategy is adopted to accurately extract non-geometric manufacturing information through component object model technology and parameterized rules; the dependence on artificial experience is significantly reduced, and the standardization and consistency of process instructions are improved. The intelligent process design mode overcomes the pain points of low efficiency and insufficient standardization in the traditional process instruction writing process, and provides a reusable technical path for the process digitization and intelligent development in the field of aviation equipment manufacturing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing, in particular to the field of aircraft process data processing technology, and belongs to the cross technology of large language model and aircraft manufacturing, and particularly relates to a large language model guided aircraft manufacturing process instruction automatic compilation method. BACKGROUND

[0002] In the field of aircraft manufacturing, process instructions are an important bridge connecting process personnel and operators, and are the first standard for guiding actual manufacturing work. The traditional form of process instructions is compiled by process personnel manually. Although there are fixed templates, the process personnel need to search for information in a vast knowledge base in combination with product information, so as to fill in the correct process statements into the templates, and make fine adjustments according to the actual situation. This instruction compilation method has a large amount of repetitive work and low efficiency, which greatly affects the development of new models and the batch production speed.

[0003] Based on product information, the construction of a knowledge graph through a large language model in the instruction compilation link can effectively improve the efficiency of instruction compilation, reduce the repetitive workload of process personnel, and enable the intelligence of process personnel to be more fully utilized in creative work. In addition, intelligent interactive design significantly reduces the dependence on human experience and avoids the risk of human error, fundamentally improving the standardization and consistency of process instructions. This "data-driven, knowledge-enabled" intelligent process design mode not only solves the problems of low efficiency and insufficient standardization in the traditional process instruction compilation process, but also provides a reusable technical path for the digitalization and intelligentization of process in the field of aviation equipment manufacturing. SUMMARY

[0004] In view of the problems in the prior art, the present application proposes a large language model guided aircraft manufacturing process instruction automatic compilation method, which combines Figure 1 The specific method, process and content implemented by the present application are as follows:

[0005] S1, structured knowledge is extracted from unstructured process text. In view of the limitations of traditional natural language processing model training in small sample scenarios, a strategy of knowledge driving and deep learning cooperation is adopted to break through the relationship extraction and knowledge fusion of complex semantics in small samples;

[0006] S2, construction of a graph structure expert library based on a large language model. Through a pre-training model, process documents, historical data and domain experience are analyzed, multi-level rules are defined, and an artificial feedback mechanism is designed to iteratively optimize the rule library. Finally, knowledge storage and dynamic updating are realized through a graph database to form a standardized knowledge base;

[0007] S3, based on component object model and parameterized driving non-geometric manufacturing information intelligent extraction, adopts the strategy of automatic extraction and multi-modal verification fusion, accurately extracts non-geometric manufacturing information through component object model technology and parameterized rules;

[0008] S4, combined with point cloud sampling and surface reconstruction technology to complete the digital mapping of geometric features, form multi-version structured comparison data, drive the rapid verification of process changes;

[0009] S5, based on semantic retrieval enhanced intelligent interaction paradigm, through vectorization technology fusion process knowledge graph and software interface, establish functional semantic index library;

[0010] S6, combined with retrieval enhancement generation technology, realize the logical verification and instruction optimization of multi-tool collaborative scheduling.

[0011] In step S1, focus on extracting structured knowledge from unstructured process text, aiming at the limitations of traditional natural language processing model training in small sample scenarios, study the strategy of knowledge driven and deep learning cooperation, break through the relationship extraction and knowledge fusion of complex semantics in small sample, such as Figure 2 As shown, the following steps are included:

[0012] S11, extract the text in paper files and electronic files through optical character recognition and electronic text recognition methods;

[0013] S12, pre-process the data, such as removing garbled characters, replacing non-standard expressions and wrong characters, etc;

[0014] S13, use pre-trained models to generate and context supplement to further enrich text information and reduce understanding difficulty;

[0015] S14, based on semantic or entity extraction, the complete text is segmented, that is, the text is cut into standard length text blocks;

[0016] S15, vectorize the text, compress the semantic information in the text into vector form, and assist the model to understand the similarity and relationship between words and sentences through these vectors.

[0017] In step S2, the pre-trained model is used to analyze the process documents, historical data and domain experience, define multi-level rules, and design artificial feedback mechanism to iteratively optimize the rule library. Finally, through the graph database, the knowledge storage and dynamic update are realized, and the standardized knowledge base is formed, which provides core data support for intelligent process design. The following steps are included:

[0018] S21, use large language model to extract entities and their relationships from text;

[0019] S22 transforms the identified entities and relationships into a graph structure, where entities become nodes and relationships become edges. Nodes can have attribute information attached to them, and edges represent the relationships between entities.

[0020] S23. The generated graph structure is stored in the graph database Neo4j, forming an expert database. Within this expert database, information can be retrieved using a graph query language, and graph reasoning can be performed using graph algorithms.

[0021] Steps S3 and S4 focus on extracting geometric and non-geometric manufacturing features from the model-based definition model, enabling multi-version modal comparison and providing data support for process instruction generation and process change verification. Addressing the issues of low efficiency and weak feature correlation in traditional manual analysis, this study employs a strategy that integrates automated extraction and multi-modal verification. Non-geometric manufacturing information is accurately extracted using component object modeling technology and parametric rules, while point cloud sampling and surface reconstruction techniques are combined to complete the digital mapping of geometric features. This ultimately forms multi-version structured comparison data, driving rapid verification of process changes. Specifically, the following steps are included:

[0022] S31 enhances system integration capabilities through component object models, builds data interaction between design software and manufacturing systems, and simplifies the process of calling and integrating non-geometric manufacturing information.

[0023] S32, based on the parameterized rule matching method, uses feature type and name as indexes to associate hierarchical manufacturing structures;

[0024] S33, quickly locates and accurately extracts manufacturing semantic information, and establishes logical mapping relationships between information;

[0025] S34 combines manual review and dynamic feedback mechanisms to verify the extraction results from multiple dimensions. Through the experience correction of process personnel and the iterative optimization of algorithms, the extracted digital model information is stored in an Excel spreadsheet.

[0026] S35, by comparing the information of different versions, the differences are highlighted in the table.

[0027] S41 uses the point cloud sampling tool in CATIA to generate a uniformly distributed set of points on the model surface, adjusting the spacing and number, or manually using the mesh point placement function to cover key areas to ensure the capture of geometric details;

[0028] S42, combining geometric feature analysis and parametric processing, identifies high curvature regions and locations of abrupt changes in normals, extracts contour boundaries through cross-sectional analysis, and establishes association rules between features;

[0029] S43 maps the analysis results into parametric geometric entities, generates editable curves and surfaces based on digital reconstruction, automatically compares the reconstructed digital features with historical versions, and outputs a list of geometric differences. It also outputs information such as part numbers and names to an Excel spreadsheet for process engineers to check and verify.

[0030] In steps S5 and S6, a dynamic interaction framework between AI and the process knowledge base is constructed to achieve accurate conversion of natural language instructions into executable process parameters. Addressing the issues of traditional process generation relying on human experience and having low response efficiency, a semantic retrieval-enhanced intelligent interaction paradigm is proposed: a functional semantic index is established by integrating the process knowledge graph and software interfaces through vectorization technology. A dynamic programming algorithm is designed, combined with retrieval-enhanced generation technology, to achieve logical verification and instruction optimization for multi-tool collaborative scheduling.

[0031] S51. Construct a process information retrieval framework for natural language semantic understanding. After modeling process entities, relationships, attributes, etc. in graph structure, they are vectorized and encoded. Using graph embedding and semantic representation methods, node feature vectors that support high-dimensional matching are generated to realize the construction of semantic index of knowledge nodes.

[0032] S52, design a hybrid retrieval mechanism that integrates graph traversal and semantic search. Based on the natural language instructions input by the user, relevant nodes are located through semantic similarity calculation to form a candidate set of results with context awareness.

[0033] S53 constructs a joint retrieval engine based on semantic vector recall and graph structure verification, supporting fuzzy queries, keyword expansion, and combined invocation of process fragments, significantly enhancing the coverage and accuracy of retrieval results, and providing structured and traceable knowledge support for the AI ​​interaction module.

[0034] S61, Construct a language understanding module based on a retrieval-enhanced generation framework. Obtain structured information such as process parameters and equipment characteristics from the graph database through semantic retrieval. Combined with context feature alignment technology, convert natural language instructions into preliminary process execution plans.

[0035] S62, based on the MCP (Model Context Protocol) interface specification, encapsulates core functions such as process optimization and parameter verification into standardized service modules. Through semantic matching and intent understanding, it enables the language model to accurately call process functions, forming a closed-loop link from requirement parsing to function execution.

[0036] S63 employs a dynamic programming algorithm to optimize the execution order of multi-step instructions, and combines the context management capabilities of MCP with the process logic verification of the graph database to ensure that the final instruction set meets the requirements of engineering feasibility and interpretability.

[0037] Compared with the prior art, the significant advantages of the above-described specific content of the present invention are as follows:

[0038] (1) Significantly improves the efficiency of process planning;

[0039] (2) Reduce reliance on human experience and enhance standardization and consistency;

[0040] (3) Supports comparison of multiple model versions and rapid process iteration;

[0041] (4) Implement human-machine collaborative programming to free process personnel from repetitive labor;

[0042] (5) A complete digital link from design information to executable process instructions has been formed, which is universal and portable, providing a reference paradigm for the intelligent manufacturing transformation of the equipment manufacturing industry. Attached Figure Description

[0043] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0044] Figure 1 The flowchart of the manufacturing process instruction compilation method guided by the large language model of the present invention is shown;

[0045] Figure 2 The flowchart illustrating the present invention for extracting structured knowledge from unstructured process text is shown. Detailed Implementation

[0046] To better understand the technical content of this invention, the technical solution of this invention will be clearly and completely described below with reference to the accompanying drawings and specific examples. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] A preferred embodiment of this invention is the initial hole drilling instruction for the front section of the right wing spars and ribs of a certain type of aircraft. The instruction includes processes such as part preparation, hole location determination, initial hole drilling, deburring, cleaning, and quality inspection. The supporting design data is based on the CATIA 3D model and accompanying drawings. Process engineers aim to generate an assembly instruction document that is clearly structured, accurate in content, and meets the requirements of actual operation.

[0048] Specifically Figure 1 The flowchart of an automatic command generation method for aircraft manufacturing processes guided by a large language model is presented, including the following steps:

[0049] Step 1: Transform unstructured information in traditional paper and electronic process documents into structured knowledge, laying the foundation for subsequent graph construction and knowledge extraction; Step 2: Use Deepseek V3 as the engine to automatically complete the extraction, structuring, and graph construction of process knowledge; Step 3: Analyze and compare complex structured information in the model-based definition 3D model; Step 4: Intelligent conversion from natural language instructions to structured process parameters.

[0050] Step 11: Extract text content from historical process cards, operating instructions, design specifications, and other documents using optical character recognition technology (such as Tesseract) and electronic document parsing interfaces. Preprocess the extracted results to remove garbled characters, standardize the format, and correct typos to ensure data quality.

[0051] Step 12: Use the pre-trained Deepseek to perform context completion and semantic reconstruction to solve problems such as abbreviations and polysemous words in the original process language, and improve the model's ability to understand complex semantics.

[0052] Step 13: Use a semantic boundary recognition algorithm to cut the text into semantic units of controllable length (such as "riveting fitter", "precautions" and "remove excess materials"), and use Word2Vec to vectorize the text to construct vectorized semantic features for relation extraction.

[0053] Step 21: Identify key entities and their logical relationships through a large language model (the entities are two independent parts: "Process A" and "Instruction A", and their relationship should be that "Process A" should use "Instruction A").

[0054] Step 22: Construct the above entities and relationships into a directed graph structure, attach semantic attributes to the nodes (such as “process hole on beam”, “task number” and “special industrial cotton cloth”), and label the edges with logical descriptions (such as “use”, “strictly prohibited” and “install”).

[0055] Step 23: Construct the Neo4j database, import the graph structure, and enable fast entity indexing, relationship querying, and multi-hop reasoning capabilities to support subsequent graph-based knowledge retrieval and recommendation.

[0056] Step 31: Integrate CATIA V5 via the Component Object Model interface. Locate the component level by parsing the assembly structure tree layer by layer, confirming that beams and ribs are adjacent assembly units of the same level. Load their associated model-based definition structure model and version number, and bind them to the current instruction.

[0057] Step 32: Using the defined rule dictionary (such as "assembly feature = hole positioning + rivet + mating surface", "process attribute = surface quality + treatment method"), the system matches model information based on the parametric semantic template and extracts the relevant information, including: hole feature type (such as "initial hole", "final hole"), allowable deviation, connection method (such as "wet assembly", "dry assembly"), etc. Semantic matching is completed by combining part name + connection method + hole attribute.

[0058] Step 33: Based on the "region name", "hole attribute", and "preprocessing requirements" corresponding to the hole position and the connection boundary, the system establishes a three-segment logical mapping from hole position → process statement → process attribute. The "Φ2.6 initial hole" is bound to the standardized semantics "preprocessing before connection → test cut → scribing → hole making → deburring".

[0059] Step 34: Extract hole location attribute information and confirm the accuracy of system labeling. Mark outliers in red and provide suggested corrections. Incorporate the correction results into the drawing database for subsequent instruction adjustments. For assembly processes involving multiple digital model versions, perform multi-version comparison based on structured manufacturing information tags (e.g., compare hole diameter, tolerance, and surface annotations in versions v1.0 and v2.0, output a list of differences, and highlight them in an Excel template).

[0060] Step 41: Open the CATIA model, use the point cloud tool to evenly distribute points in the contact area at a spacing of 0.5mm, and manually specify denser distribution points in high curvature areas of the model (such as the corners of wing beams) to ensure that no geometric details are missed.

[0061] Step 42: Using a geometric mutation detection algorithm, the platform identifies characteristic areas such as depressions on the side margins of the web and pore distribution areas. Local cross-sections are generated using the section extraction function, and boundary segments are extracted from them for parametric modeling.

[0062] Step 43: Generate a non-uniform rational B-spline surface based on the reconstructed point cloud data. The output is an editable CAD structure, which is then mapped to the original design surface for deviation analysis. If the curvature error is within a given threshold, the reconstruction is considered successful. Subsequently, different surfaces are compared. If the differences between the different surfaces are within a given threshold, they are considered indistinguishable.

[0063] Step 51: Based on the existing nodes such as “hole making”, “initial hole”, “Φ2.6mm”, and “rivet connection” in the graph, retrieve the associated process instruction keywords such as “hole position scribing” and “digital model 571L0600101” as query input and convert them into vector query statements.

[0064] Step 52: Locate the path node of the combination "Φ2.6 + rivet + aluminum alloy + wet assembly" by traversing the graph structure, and then use semantic matching to find typical case instructions containing the process of "pre-marking → test cutting → hole making → deburring → quality inspection" as candidate references.

[0065] Step 53: Taking into account factors such as semantic score, path rationality, and context overlap, three instruction fragments are output for the generation module to call. Finally, the combination path with the highest score is selected, matched to the current structure instance, and the generation process begins.

[0066] Step 61: Taking "making holes in the front section of the rib and the beam" as the assembly intention, extract relevant entities and process parameters from the drawing, such as "Φ2.6 hole", "wet assembly", "aluminum alloy material", "manual scribing", etc., as the context input for generating the model, and construct the prompt: "Please generate the process flow for making holes between 571L0600701 and 571L0060197 based on the digital model 571L0060121, 571L0600101 and R_571L0600101".

[0067] Step 62: Convert the "hole position marking, scribing method, hole making tool, material matching" in the process generation results into a structured parameter package defined in the MCP protocol, such as: {D: 2.6, tool: electric drill, material: HB6298-5x12, process: wet assembly, cleaning method: acetone wiping}, for use on the downstream platform.

[0068] Step 63: The system sorts the instruction steps according to the predefined template and the logical dependency graph, and verifies whether there are execution dependency errors in the instruction set (such as "no test cut before hole making" or "no scribing"). It automatically adjusts the order to ensure rationality and consistency.

[0069] The final structure is as follows:

[0070] 010 Hole Position Determination: Mark the connection hole positions (8 + 2 Φ2.6);

[0071] 015 Check the lines and spacing;

[0072] 020 Hole making: Complete 10-Φ2.6 initial holes;

[0073] 025 Deburring;

[0074] 030 Cleaning;

[0075] 035 Inspection: Hole quality + surface finish + cleanliness.

[0076] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should be within the protection scope of the present invention.

Claims

1. A method for automatically generating aircraft manufacturing process instructions guided by a large language model, characterized in that, Includes the following steps: S1 extracts structured knowledge from unstructured process texts. To address the limitations of training traditional natural language processing models in small sample scenarios, it adopts a strategy that combines knowledge-driven learning with deep learning to overcome the challenges of complex semantic relationship extraction and knowledge fusion in small sample scenarios. S2 is based on the construction of a graph-structured expert database using a large language model. It analyzes process documents, historical data, and domain experience through a pre-trained model, defines multi-level rules, and designs a human feedback mechanism to iteratively optimize the rule base. Finally, it uses a graph database to achieve knowledge storage and dynamic updates, forming a standardized knowledge base. S3 enhances system integration capabilities through the component object model, builds data interaction between design software and manufacturing systems, and simplifies the process of calling and integrating non-geometric manufacturing information. Based on the parameterized rule matching method, a hierarchical structure is generated by using feature type and name as indexes. Quickly locate and accurately extract manufacturing semantic information, and establish logical mapping relationships between information; By combining manual review and dynamic feedback mechanisms, the extraction results are verified from multiple dimensions. Through the experience of process engineers and the iterative optimization of algorithms, the extracted mathematical model information is stored in an Excel spreadsheet. By comparing the mathematical model information of different versions, the differences are highlighted in the spreadsheet. S4 combines point cloud sampling and surface reconstruction technology to complete the digital mapping of geometric features, forming multi-version structured comparison data to drive rapid verification of process changes; S5, based on a semantic retrieval-enhanced intelligent interaction paradigm, integrates process knowledge graphs and software interfaces through vectorization technology to establish a functional semantic index library; S6, construct a language understanding module based on a retrieval-enhanced generation framework, obtain process parameters and equipment characteristics from the graph database through semantic retrieval, and combine context feature alignment technology to convert natural language instructions into preliminary process execution plans; Based on the MCP (Model Context Protocol) interface specification, process optimization and parameter verification are encapsulated into standardized service modules. Through semantic matching and intent understanding, the language model can accurately call process functions, forming a closed-loop link from requirement parsing to function execution. Dynamic programming algorithm is used to optimize the execution order of multi-step instructions, and the context management capabilities of MCP and the process logic of graph database are used for verification.

2. The method for automatically compiling aircraft manufacturing process instructions guided by a large language model according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11 extracts text from paper and electronic documents using optical character recognition and electronic text recognition methods; S12, preprocess the data, remove garbled characters, replace non-standard expressions and typos; S13 uses a pre-trained model to generate and supplement textual information with contextual information to further enrich the text and reduce the difficulty of understanding it; S14, segment the complete text based on semantic or entity extraction, that is, cut the text into text blocks of standard length; S15 vectorizes the text, compressing the semantic information in the text into vector form, which helps the model understand the similarity and relationship between words and sentences.

3. The method for automatically compiling aircraft manufacturing process instructions guided by a large language model according to claim 1, characterized in that, Step S2 includes the following steps: S21, using a large language model to extract entities and the relationships between them from the text; S22, the identified entities and relationships are converted into a graph structure, where entities become nodes in the graph and relationships become edges; nodes can be attached with attribute information of entities, and edges represent the relationships between entities; S23. Store the generated graph structure in the graph database Neo4j to form an expert database. In the established expert database, information retrieval is performed using graph query language and graph reasoning is performed using graph algorithms.

4. The method for automatically compiling aircraft manufacturing process instructions guided by a large language model according to claim 1, characterized in that, Step S4 includes the following steps: S41 uses the point cloud sampling tool in CATIA to generate a uniformly distributed set of points on the model surface, adjusting the spacing and number, or manually using the mesh point placement function to cover key areas to ensure the capture of geometric details; S42, combining geometric feature analysis and parametric processing, identifies high curvature regions and locations of abrupt changes in normals, extracts contour boundaries through cross-sectional analysis, and establishes association rules between features; S43 maps the analysis results into parametric geometric entities, generates editable curves and surfaces based on digital reconstruction, automatically compares the reconstructed digital features with historical versions, outputs a list of geometric differences, and outputs part numbers and names to an Excel spreadsheet for process engineers to check and verify.

5. The large language model-guided aircraft according to claim 1 The method for automatically generating manufacturing process instructions is characterized by, Step S5 includes the following steps: S51. Construct a process information retrieval framework for natural language semantic understanding. After modeling process entities, relationships, and attributes in a graph structure, they are vectorized and encoded. Using graph embedding and semantic representation methods, node feature vectors that support high-dimensional matching are generated to realize the construction of semantic indexes for knowledge nodes. S52, design a hybrid retrieval mechanism that integrates graph traversal and semantic search. Based on the natural language instructions input by the user, relevant nodes are located through semantic similarity calculation to form a candidate set of results with context awareness. S53 constructs a joint retrieval engine based on semantic vector recall and graph structure verification, supporting fuzzy queries, keyword expansion, and combined invocation of process fragments, significantly enhancing the coverage and accuracy of retrieval results, and providing structured and traceable knowledge support for the AI ​​interaction module.

Images