A method for constructing large models for aerospace intelligent manufacturing

Abstract

The present invention discloses a method for constructing a large-scale model for aerospace intelligent manufacturing, comprising: collecting raw data, performing preprocessing and data association, and constructing an aerospace intelligent manufacturing database; establishing a knowledge acquisition and structured conversion pipeline, a multi-dimensionally associated domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism to build a professional knowledge base; aligning cross-modal manufacturing data to generate a corpus; combining pre-training of a universal large-scale model with a base, injecting professional terminology semantics and multimodal association capabilities to complete knowledge transfer; constructing an aerospace manufacturing cognitive agent based on the pre-trained large-scale model for aerospace intelligent manufacturing, forming a framework for solving complex engineering problems; optimizing professional capabilities through a two-stage progressive multi-task training strategy, and combining online learning with an incremental update mechanism to achieve dynamic adaptation of the production environment. This invention significantly improves the intelligence level of aerospace intelligent manufacturing.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing technology, and specifically relates to a method for constructing a large model of aerospace intelligent manufacturing. Background Art

[0002] In the aerospace sector, traditional knowledge management and intelligent question-answering systems face significant technical bottlenecks and challenges. Existing systems often rely on isolated databases and expert rule bases, making it difficult to effectively integrate diverse multimodal data such as design drawings, sensor time series, and process logs, leading to severe information silos. While traditional general-purpose large language models excel in open-domain tasks, they lack a deep understanding of specialized domain knowledge and the complex aerospace terminology and engineering logic, making it difficult to accurately analyze complex manufacturing problems.

[0003] Furthermore, knowledge updates in traditional models often rely on time-consuming manual annotation and retraining, resulting in timeliness lags and an inability to adapt to the rapid iteration of aerospace technology and the dynamic changes in the production environment. Therefore, building a specialized intelligent agent capable of deeply understanding domain knowledge, processing multimodal data, and possessing logical reasoning and continuous adaptability is a key challenge in promoting the upgrade of intelligent aerospace manufacturing. Summary of the Invention

[0004] To address the shortcomings of the aforementioned existing technologies, this paper proposes a method for constructing a large-scale model for aerospace intelligent manufacturing. This method addresses the pain points of traditional general-purpose large-scale models in specialized scenarios, such as terminology errors, shallow knowledge, weak logic, and insufficient data processing capabilities, thereby enhancing the intelligence level of aerospace intelligent manufacturing.

[0005] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0006] A method for constructing a large model of aerospace intelligent manufacturing, which specifically includes the following steps:

[0007] S1. Build a professional data collection and management system to collect a large amount of raw aerospace intelligent manufacturing real data from actual aerospace intelligent manufacturing scenarios, perform pre-processing and data association, and build an aerospace intelligent manufacturing database;

[0008] S2. Targeting multimodal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge acquisition and structured conversion pipeline, a multi-dimensional domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism to build a professional knowledge base for aerospace intelligent manufacturing.

[0009] S3. Align cross-modal manufacturing data to generate a corpus in the field of aerospace intelligent manufacturing. Combined with the base general large model for pre-training, the system injects professional terminology semantics and multimodal association capabilities through masked language modeling tasks and contrastive learning, completing the aerospace domain knowledge transfer of the base model and obtaining the pre-trained aerospace intelligent manufacturing large model.

[0010] S4. Based on the pre-trained aerospace intelligent manufacturing big model, build the aerospace manufacturing cognitive agent AMCA to form a complete, dynamic, and self-improving framework for solving complex engineering problems;

[0011] S5. Through a two-stage progressive multi-task training strategy, supervised fine-tuning and reinforcement learning alignment are performed on the large model to optimize the professional capabilities of the large model in the field of aerospace intelligent manufacturing. Combined with online learning and incremental update mechanisms, the large model of aerospace intelligent manufacturing can dynamically adapt to the needs of the actual production environment; finally, the construction of the large model of aerospace intelligent manufacturing is completed.

[0012] Furthermore, step S1 specifically includes:

[0013] S11. Establish a data aggregation channel to acquire a large amount of real-world aerospace manufacturing data through multi-source data collection and integration. Define a unified data identification system for the dispersed design, production, and testing data sources in each type of real-world aerospace manufacturing scenario, and use a heterogeneous protocol compatibility strategy to aggregate each data source into the original database.

[0014] S12. Design a hierarchical and classified data storage architecture, specifically including: first, storing all collected raw data in a distributed database to build a data lake to retain the full amount of original information; then, classifying and importing the data into different databases based on data characteristics and application scenarios: For key business data in structured data, pre-process it through a quality rule engine, including outlier filtering, unit standardization, and missing value interpolation, and then import it into a columnar data warehouse; for high-frequency equipment data that changes in real time in structured data, store it in a time series database; for semi-structured data, store its metadata and associations in document data; for unstructured data, manage file entities through object storage;

[0015] S13. Establish a complete data governance system, including: defining unified naming conventions, building a global data catalog based on a metadata management platform, and clearly defining the business meaning, data lineage, and access rights of each data entity; on this basis, implementing automated data quality testing based on a dynamic rule engine to identify outliers and verify data formats; and finally, implementing a hierarchical encryption strategy to secure data and protect sensitive information, thereby forming a closed-loop data governance solution covering the entire process.

[0016] S14. Build an association network between data to provide support for the training and analysis of large models; transform discrete data into a reasonable knowledge network business entity network through the entity relationship model to complete the construction of the aerospace intelligent manufacturing database.

[0017] Furthermore, step S2 specifically includes:

[0018] S21. Construct a multimodal knowledge acquisition and structured conversion pipeline for intelligent aerospace manufacturing. Specifically, the pipeline involves directly acquiring structured documents from knowledge data, parsing key knowledge elements from semi-structured text and unstructured data using natural language processing techniques, and converting them into structured triples. Computer vision analysis techniques are then used for unstructured video data to extract key frame sequences from operational steps and annotate key process points. Furthermore, an expert knowledge entry interface is established to support process engineers in contributing implicit experience in a form-based format.

[0019] S22. Construct a multi-dimensional, interconnected knowledge graph for aerospace intelligent manufacturing, organizing the knowledge network around product family tree, process flow, and fault tree. Specifically, define a three-layer ontology model: "parts-materials-processes." Associate part design specifications with a material property library and map them to processing parameter knowledge. Simultaneously, construct a process constraint rule library, describe process dependencies through logical expressions, and link them to testing standards. Furthermore, integrate a historical failure case library to establish a causal chain of "anomaly-root cause-solution," forming a traceable diagnostic knowledge network.

[0020] A hybrid architecture combining graph database Neo4j and vector database Milvus is employed to efficiently store and apply complex knowledge networks. While the graph database Neo4j stores explicit structured relationships, the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval. Finally, a knowledge evidence library is established, linking each knowledge node to its original basis, ensuring a bidirectional link between knowledge nodes and source evidence.

[0021] S23. Build a knowledge quality control system throughout the entire process; establish a dual knowledge verification mechanism: automatically discover and report conflicting rules through logical consistency checks, trigger a manual review process, and verify knowledge validity based on historical data backtracking; establish a version control system to record the revision history and effective time of knowledge items, ensuring that the reference version of each process rule can be traced during airworthiness certification; at the access and use level, adopt attribute-based access control permission management, dynamically control the scope of knowledge access based on personnel roles and project confidentiality levels, and set up an approval and unlocking process for core process parameters to ensure the security of knowledge assets;

[0022] S24. Establish a dynamic update mechanism to continuously optimize the knowledge base; set the knowledge freshness monitoring indicator KFI and regularly synchronize international standard updates; analyze newly generated production data through machine learning to automatically extract valuable potential knowledge patterns, and the extracted knowledge patterns must be reviewed and confirmed by domain experts before being stored; in addition, build an online collaboration platform to collect the implicit experience of front-line personnel and complete the construction of a professional knowledge base for aerospace intelligent manufacturing.

[0023] More specifically, KFI is obtained by weighting four core factors, namely:

[0024] ;

[0025] in, 、 、 、 is the weight of each factor, which is dynamically adjusted according to the knowledge type;

[0026] TDF is the time decay factor, and the formula is:

[0027] ;

[0028] Among them, T is the validity period of knowledge, is the current time, The time when the knowledge item was last verified and updated;

[0029] VAF is the version alignment factor, obtained by comparing with an external standard database;

[0030] UAF is the application activity factor, and the formula is:

[0031] ;

[0032] in, Indicates the frequency with which a piece of knowledge was called by the big model, retrieved, and adopted in knowledge question answering in the past cycle; Indicates the average frequency of all knowledge being called by the big model, retrieved and adopted in knowledge question answering in the past cycle;

[0033] PVF is the performance verification factor, which is used to test the validity of knowledge.

[0034] Furthermore, step S3 specifically includes:

[0035] S31. Construct a cross-modal, semantically precisely aligned corpus for aerospace intelligent manufacturing: Extract diverse information, including structured, semi-structured, and unstructured types, from the aerospace intelligent manufacturing database. Subsequently, transform and align the heterogeneous data. On the one hand, some structured data is converted into natural language descriptions using automated templates. On the other hand, different types of data pointing to the same manufacturing entity are precisely correlated by timestamps to construct cross-modal correlation samples of the "signal-image-parameter-text" type to train the model to understand the inherent connections between multi-source data. Finally, utilize the established aerospace intelligent manufacturing professional knowledge base to enhance the text content in the corpus with domain knowledge.

[0036] S32. Based on the established aerospace intelligent manufacturing domain corpus and the base general large model, domain adaptation pre-training is performed to inject basic knowledge of the aerospace intelligent manufacturing field into the base general large model; domain general corpus and some multimodal alignment samples are used for mixed training, and masked language modeling (MLM) tasks are adopted, focusing on masked domain entities, forcing the model to rely on knowledge clues in the context for prediction; at the same time, comparative learning tasks are introduced, given different modal descriptions of the same process, the model is required to determine whether they match, in order to enhance cross-modal alignment capabilities.

[0037] S41. After receiving a user query, AMCA first initiates a deep traversal of the knowledge graph, following the "retrieval-reasoning-pruning" mechanism to extract knowledge fragments. Specifically, it uses vector retrieval technology to match user queries with entity nodes in the knowledge graph, quickly locating the initial node most relevant to the query semantics as the seed node for graph traversal. Starting from the seed node, driven by the reasoning capability of the large model, it conducts multi-hop logical exploration to determine whether the exploration along the relationship path of each node is conducive to solving the problem, thereby connecting dispersed but logically strongly related information to form a complete chain of evidence. During the exploration process, it tracks multiple reasoning paths in parallel, continuously evaluates the information gain of each path, and prunes paths that lead to irrelevant information and logical dead loops. Finally, it aggregates, deduplicates, and logically reorganizes all high-value knowledge fragments on the reasoning path to form a comprehensive knowledge context.

[0038] S42. After obtaining the comprehensive knowledge context, AMCA enters the cognitive stage and adopts the thinking tree exploration framework to transform vague user intentions into clear, executable, and structured solutions. Specifically, the user's high-level goals are first decomposed into a series of thinking nodes, and multiple different but feasible next actions and hypotheses are generated for each thinking node. Then, the self-assessment mechanism is activated, and the comprehensive knowledge context and its own domain knowledge are used to evaluate the value of each newly generated path branch, calculate the evaluation score, and dynamically search and prune the entire thinking tree. Finally, a complete path is constructed starting from the initial problem, running through all necessary analysis steps, and with a high-confidence conclusion. This path is used as the final execution plan and output in a structured form.

[0039] S43. Entering the action phase, based on the collaborative scheduling-instant creation mode, each step in the execution plan is implemented and candidate solutions are generated. Specifically, the constructed database, knowledge base, and professional small models are first uniformly encapsulated and abstracted into an expert tool set; after parsing the execution plan, the scheduler agent delegates clearly defined and domain-specific subtasks to the most suitable expert agent in the expert tool set. The scheduler agent is responsible for managing the data flow and dependencies between different expert agents, ensuring that the steps in the plan are executed in the correct logical order and using the output of the previous task as the input of the next task to form a coherent automated workflow; when the scheduler agent delegates tasks and encounters subtasks that cannot be directly handled by expert agents and existing tools, the tool synthesis module is activated to generate subtasks that cannot be directly handled by new tools by writing Python functions on the fly. After sandbox verification, the output is integrated into the automated workflow;

[0040] S44, AMCA enters the reflection and verification stage, and conducts a review and correction cycle of the solution. Specifically: First, the scheduler agent summarizes and integrates the execution results of each expert agent and dynamic tool to form a preliminary solution; then AMCA internally criticizes the preliminary solution, generates a reflection text, raises relevant questions, and then switches to the evaluator role, using the expert tool set and knowledge graph to externally verify the solution, including logical consistency verification, compliance review, and historical case comparison; if problems arise during internal and external reviews, the relevant critical opinions and verification results are recorded and saved, the correction program optimization plan is started, and an iterative feedback cycle is performed; if all internal and external reviews are successfully passed, the solution is marked as verified and presented to the user as the final, highly reliable output solution.

[0041] Furthermore, step S5 specifically includes:

[0042] S51. Based on the established aerospace intelligent manufacturing domain corpus, a two-stage progressive multi-task training strategy is used to train the aerospace intelligent manufacturing large model after pre-training and construction of the aerospace manufacturing cognitive agent AMCA, enhancing the model's adaptability to domain data distribution and knowledge logic.

[0043] S52. Establish a multi-level evaluation and iteration mechanism to conduct performance evaluation and continuous dynamic update of the final constructed aerospace intelligent manufacturing large model.

[0044] More specifically, step S51 includes:

[0045] S511. The first phase involves multi-task joint supervised fine-tuning, aiming to comprehensively enhance the model's expertise in aerospace manufacturing. The model concurrently optimizes multiple heterogeneous downstream tasks, including process parameter prediction, text-image matching, quality inspection report generation, knowledge question answering, and anomaly root cause reasoning. The GradNorm algorithm is used to adaptively assign loss weights, dynamically monitoring the gradient norm and loss function of each task relative to the initial state and automatically adjusting the corresponding loss weights.

[0046] S512. In the second stage, the model is deeply aligned through reinforcement learning to make its decision recommendations more in line with complex real-world needs: first, a reward model based on knowledge graphs and domain rules is trained, and the decision recommendations generated by the large model are scored from multiple dimensions in terms of technical feasibility, compliance, and economy. The multi-dimensional scores are integrated into a single scalar reward signal through weighted summation; at the same time, a logical constraint reward function is set to ensure that the model does not violate the predefined immutable rules, forcing the model to converge quickly to the compliant solution space during training, and forcing alignment of domain knowledge; then an artificial feedback loop is introduced to further improve the alignment accuracy, and the sampling results are regularly scored by domain experts; finally, the system will weightedly fuse the scaled scalar reward signal with the high-quality manual score to form the final reward signal, which is then combined with the proximal policy optimization PPO algorithm to fine-tune the large model.

[0047] More specifically, step S52 includes:

[0048] S521. First, we build a high-quality offline test set. We select the latest six months of fault cases and their associated multimodal data from the database as the test set. We strictly isolate the time window of the training data to simulate unknown problems encountered after the model is launched. We also inject artificially constructed adversarial examples to comprehensively test the model's anti-interference ability and logical consistency.

[0049] S522. Construct a multi-dimensional evaluation index system. Specifically, at the general text generation level, use the general generation indicators BLEU and ROUGE to compare with reference text to evaluate the fluency and information fidelity of the model-generated content. Secondly, define domain-specific key indicators, including process parameter deviation and knowledge compliance rate. Finally, to measure the real business value brought by the model, deploy a large-scale aerospace intelligent manufacturing model in a digital twin environment and evaluate the actual improvement effect of the decision recommendations provided by the model on core indicators such as yield rate and production efficiency.

[0050] S523. Establish an online learning and incremental update mechanism; use the decision recommendations given by the large model in the production environment and the results after actual execution as feedback data to build a feedback closed-loop data set; successful execution cases serve as positive samples to reinforce the correct behavior of the model, and failure cases trigger analysis and serve as negative samples to correct the model, realizing continuous learning based on actual feedback; use parameter efficient fine-tuning PEFT technology for LoRA low-rank adaptation, only updating a small number of parameters, without affecting general capabilities, to quickly realize monthly incremental updates of the model, ensuring that the model continues to adapt to changes in the production environment without interrupting service.

[0051] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0052] The method for constructing a large model of aerospace intelligent manufacturing proposed in this invention realizes multi-level and full-chain optimization in view of the characteristics of complex data, knowledge-intensive, rigorous logic and dynamic environment in the aerospace field. By building a high-quality multimodal data ecosystem and professional knowledge base, a solid foundation is provided for the model; the domain-based architecture transformation enables the model to deeply integrate and understand multiple modal information such as text, numerical, and visual, and has the decoding ability of knowledge perception, effectively avoiding physically infeasible output; the phased progressive training strategy combined with multi-task learning and reinforcement learning enables the model to fully master various professional skills of aerospace manufacturing, from basic parameter prediction to complex root cause reasoning; closed-loop evaluation and continuous iteration mechanism ensure the robustness and continuous adaptability of the model in the actual production environment, overcoming the problems of traditional model knowledge lag and difficulty in adapting to new scenarios; engineering deployment and human-machine collaboration mechanism ensure the efficiency, safety and reliability of the model in practical applications. The present invention effectively solves the pain points of traditional general large models in professional scenarios, such as terminology deviation, shallow knowledge, weak logic, and insufficient data processing capabilities, and significantly improves the intelligence level of aerospace intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for constructing a large model for intelligent manufacturing of aerospace proposed by the present invention;

[0054] Figure 2 This is a schematic diagram of the workflow of the aerospace manufacturing cognitive agent AMCA constructed by the method proposed in this invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] Although the steps in the present invention are arranged with numbers, they are not intended to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0057] like Figure 1-2 As shown, the present invention proposes a method for constructing a large model of aerospace intelligent manufacturing, which specifically includes the following steps:

[0058] S1. Build a professional data collection and management system to collect a large amount of raw aerospace intelligent manufacturing real data from actual aerospace intelligent manufacturing scenarios, perform pre-processing and data association, and build an aerospace intelligent manufacturing database;

[0059] As a preferred embodiment, step S1 specifically includes:

[0060] S11. Establish a data aggregation channel to obtain a large amount of real aerospace manufacturing data through multi-source data collection and integration; define a unified data identification system (such as process number, equipment ID, material batch) for the scattered design, production, and inspection data sources in each type of real aerospace manufacturing scenario, and adopt a heterogeneous protocol compatibility strategy to aggregate various data sources into the original database; in this embodiment, the heterogeneous protocol compatibility strategy is specifically as follows: real-time access to the time series data stream of CNC machine tools and sensors through the Industrial Internet of Things protocol, synchronizing the three-dimensional model and process files of the PLM system with the API interface, and using ETL tools to extract work order execution records and quality inspection results in the MES. At the same time, unstructured data such as laboratory material test reports and X-ray detection images are included in the collection scope and aggregated into the original database.

[0061] S12. To meet the storage and analysis needs of different types of data, this application designs a hierarchical and classified data storage architecture, which specifically includes: first, all collected raw data is uniformly stored in a distributed database to build a data lake to retain the full amount of original information; then, based on the data characteristics and application scenarios, the data is classified and imported into different databases: for key business data in structured data (such as BOM tables and quality inspection indicators), pre-processing including outlier filtering, unit standardization, and missing value interpolation is performed through the quality rule engine, and then imported into a columnar data warehouse to efficiently support cross-process online analytical processing (OLAP); for high-frequency equipment data that changes in real time in structured data, it is stored in a time series database to optimize high-speed reading and writing and time series query performance; for semi-structured data, its metadata and association relationships are stored in document data to facilitate rapid retrieval and call; for unstructured data, file entities are managed through object storage;

[0062] S13. To ensure the consistency, accuracy, and security of data throughout its entire lifecycle of collection, storage, and application, this application establishes a complete data governance system, specifically including: defining unified naming conventions, building a global data catalog based on a metadata management platform, and clearly defining the business meaning, data lineage, and access rights of each data entity; on this basis, implementing automated data quality testing based on a dynamic rule engine, identifying outliers and formatting data to ensure data reliability; and finally, adopting a hierarchical encryption strategy to secure data and protect sensitive information, thereby forming a closed-loop data governance solution covering the entire process;

[0063] In this embodiment, the hierarchical encryption strategy is as follows: the transport layer uses TLS 1.3 to ensure the security of the data channel, the storage layer encrypts sensitive design drawings with AES-256, and uses the RBAC model to limit the access granularity of different roles to the process library and test reports, meeting the AS9100 standard requirements for data traceability and permission isolation.

[0064] S14. Build an association network between data to provide support for the training and analysis of large models; transform discrete data into a reasonable knowledge network business entity network through the entity-relationship model, that is, use the graph database to build a process knowledge graph, and transform assembly constraints and failure modes into computable logical rules; design a feature engineering pipeline for time series data, extract time domain statistics (mean, variance) and frequency domain features (FFT main frequency amplitude), provide input features for the predictive maintenance model, and finally transform the scattered data into a reasonable knowledge network to complete the construction of the aerospace intelligent manufacturing database.

[0065] S2. Targeting multimodal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge acquisition and structured conversion pipeline, a multi-dimensional domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism to build a professional knowledge base for aerospace intelligent manufacturing.

[0066] As a preferred embodiment, step S2 specifically includes:

[0067] S21. Construct a multimodal knowledge acquisition and structured conversion pipeline for aerospace intelligent manufacturing. Specifically, the pipeline involves: directly acquiring structured documents (such as design manuals, process specifications, and quality inspection standards) from knowledge data; parsing key knowledge elements from semi-structured text (such as R&D reports and failure analysis records) and unstructured data (such as expert experience and assembly operation videos) through natural language processing technology, and converting them into structured triples; then, using computer vision analysis technology for video-based unstructured data, extracting key frame sequences of operation steps and annotating the key process points; and establishing an expert knowledge entry interface to support process engineers in contributing implicit experience in a form-based manner.

[0068] S22. Construct a multi-dimensional, interconnected knowledge graph for aerospace intelligent manufacturing, organizing the knowledge network around product family tree, process flow, and fault tree. Specifically, define a three-layer ontology model: "parts-materials-processes." Associate part design specifications with a material property library and map them to processing parameter knowledge. Simultaneously, construct a process constraint rule library, describe process dependencies through logical expressions, and link them to testing standards. Furthermore, integrate a historical failure case library to establish a causal chain of "anomaly-root cause-solution," forming a traceable diagnostic knowledge network.

[0069] A hybrid architecture combining graph database Neo4j and vector database Milvus is employed to efficiently store and apply complex knowledge networks. While the graph database Neo4j stores explicit structured relationships, the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval. Finally, a knowledge evidence library is established, linking each knowledge node to its original basis, ensuring a bidirectional link between knowledge nodes and source evidence.

[0070] S23. To ensure the authority and reliability of the knowledge base throughout its lifecycle, this application constructs a knowledge quality control system that runs through the entire process; establishes a dual knowledge verification mechanism: automatically discovers and reports conflicting rules through logical consistency checks, triggers a manual review process, and verifies the validity of knowledge based on historical data backtracking; establishes a version control system to record the revision history and effective time of knowledge items, ensuring that the version of each process rule can be traced during airworthiness certification; at the access and use level, adopts attribute-based access control permission management, dynamically controls the scope of knowledge access based on personnel roles and project confidentiality levels, and sets an approval and unlocking process for core process parameters to ensure the security of knowledge assets;

[0071] S24. Establish a dynamic update mechanism to continuously optimize the knowledge base; set up a knowledge freshness monitoring indicator (KFI), and regularly synchronize with international standards to ensure the cutting-edge and compliance of knowledge content;

[0072] In this application, KFI is obtained by weighting four core factors, specifically:

[0073] ;

[0074] in, 、 、 、 The weight of each factor is dynamically adjusted according to the type of knowledge (such as safety regulations, process parameters, fault plans, etc.);

[0075] TDF is the time decay factor, which is used to measure the risk of knowledge aging and becoming obsolete over time. The formula is:

[0076] ;

[0077] Among them, T is the validity period of knowledge, is the current time, The time when the knowledge item was last verified and updated. The longer the knowledge item has not been updated, the lower the TDF score.

[0078] VAF is a version alignment factor, obtained by comparing with an external standard database. It is used to ensure that the knowledge is consistent with the version of external authoritative standards (such as international / industry standards and design specification manuals). In this embodiment, it is specifically set in the form of a discrete factor: 1.0 (Latest): The standard version based on which the knowledge is based is the latest valid version; 0.5 (Updated): The standard version based on which the knowledge is based has an updated version, but the old version has not been abolished; 0.0 (Abolished): The standard version based on which the knowledge is based has been explicitly abolished or replaced; N / A (Not Applicable): The knowledge is not associated with any external standard.

[0079] UAF is the application activity factor, which evaluates the effectiveness of knowledge through its actual application frequency. Its formula is expressed as:

[0080] ;

[0081] in, Indicates the frequency with which a piece of knowledge was called by the big model, retrieved, and adopted in knowledge question answering in the past cycle; Indicates the average frequency of all knowledge being called by the big model, retrieved and adopted in knowledge question answering in the past cycle;

[0082] PVF stands for Performance Verification Factor, which measures the performance of knowledge in guiding actual production or solving problems. It is the gold standard for testing knowledge validity. In this embodiment, the specific design can be as follows depending on the knowledge type: For process parameter knowledge: regularly apply the parameter in the digital twin environment or small-batch trial production, and PVF = product yield or key performance indicator achievement rate; For fault diagnosis knowledge: when new fault cases arise, use this knowledge for reasoning, and PVF = diagnosis accuracy or success rate of recommended solutions; For rule constraint knowledge: analyze the latest production log, and PVF = 1 - (number of rule violations / total number of production batches);

[0083] After that, newly generated production data is analyzed through machine learning to automatically extract valuable potential knowledge patterns. The extracted knowledge patterns must be reviewed and confirmed by domain experts before they are stored, forming a verification closed loop from production practice to expert knowledge. In addition, an online collaboration platform is built to gather the implicit experience of front-line personnel and build an online collaboration platform that supports the participation of multiple teams such as design, manufacturing, and operation and maintenance. Personnel from all parties are allowed to initiate knowledge revision proposals. These proposals will go through a rigorous cross-departmental review process to ensure that the final updated knowledge is both authoritative and applicable on the field. This completes the construction of the aerospace intelligent manufacturing professional knowledge base.

[0084] S3. Align cross-modal manufacturing data to generate a corpus in the field of aerospace intelligent manufacturing. Combined with the base general large model for pre-training, the system injects professional terminology semantics and multimodal association capabilities through masked language modeling tasks and contrastive learning, completing the aerospace domain knowledge transfer of the base model and obtaining the pre-trained aerospace intelligent manufacturing large model.

[0085] As a preferred embodiment, step S3 specifically includes:

[0086] S31. Construct a cross-modal, semantically precisely aligned corpus in the field of aerospace intelligent manufacturing: Extract diverse information covering structured, semi-structured, and unstructured types from the aerospace intelligent manufacturing database; then, to achieve cross-modal understanding, transform and align the heterogeneous data. On the one hand, some structured data (such as work order execution records and minute-level time series data collected by sensors) are converted into natural language descriptions through automated templates; on the other hand, different types of data pointing to the same manufacturing entity, such as unstructured information (detection images), structured data (sensor parameters), and semi-structured data (process logs, quality inspection reports) are converted into natural language descriptions. ), and finely correlate by timestamp to construct a cross-modal correlation sample of "signal-image-parameter-text" to train the model to understand the intrinsic connection between multi-source data; finally, the established aerospace intelligent manufacturing professional knowledge base is used to enhance the domain knowledge of the text content in the corpus; in this embodiment, the specific measures for domain knowledge enhancement include: injecting domain entities and rules into the text context, replacing common terms with industry standard terms, and inserting knowledge triples as context supplements, so as to improve the professional semantic accuracy and logical depth of the corpus, so that the model can accurately learn the precise correspondence between professional terms and logical constraints.

[0087] S32. Based on the established aerospace intelligent manufacturing domain corpus and the base general large model, domain adaptation pre-training is performed to inject basic knowledge of the aerospace intelligent manufacturing field into the base general large model; domain general corpus (such as aerospace journal articles, technical manuals) and some multimodal alignment samples are used for mixed training, and masked language modeling (MLM) tasks are adopted, focusing on masked domain entities, forcing the model to rely on knowledge clues in the context for prediction; at the same time, comparative learning tasks are introduced, given different modal descriptions of the same process, the model is required to determine whether they match, in order to enhance cross-modal alignment capabilities.

[0088] S4. Build an Aerospace Manufacturing Cognitive Agent (AMCA) based on the pre-trained aerospace intelligent manufacturing model to form a complete, dynamic, and self-improving framework for solving complex engineering problems.

[0089] As a preferred embodiment, Figure 2 As shown, step S4 specifically includes:

[0090] S41. After receiving the user query, AMCA first initiates a deep traversal of the knowledge graph, following the "retrieval-reasoning-pruning" mechanism to extract knowledge fragments. Specifically, it uses vector retrieval technology to match the user query with the entity nodes in the knowledge graph, quickly locates the initial node most relevant to the query semantics, and uses it as the seed node for graph traversal. Then, starting from the seed node, driven by the reasoning ability of the large model, it conducts multi-hop logical exploration to determine whether the exploration along the relationship path of each node is conducive to solving the problem. This judgment-guided reasoning enables the traversal process to cross the boundaries of multiple knowledge documents, thereby connecting dispersed but logically strongly related information to form a complete chain of evidence. During the exploration process, multiple reasoning paths are tracked in parallel, the information gain of each path is continuously evaluated, and paths that lead to irrelevant information and logical dead loops are pruned. Finally, all high-value knowledge fragments on the reasoning path are aggregated, deduplicated, and logically reorganized to form a comprehensive knowledge context.

[0091] S42. After obtaining the comprehensive knowledge context, AMCA enters the cognitive stage and adopts the thinking tree exploration framework to transform vague user intentions into clear, executable, and structured solutions. Specifically, it first decomposes the user's high-level goals into a series of thinking nodes, and generates multiple different but feasible next actions and hypotheses for each thinking node; then it starts the self-assessment mechanism, using the comprehensive knowledge context and its own domain knowledge to evaluate the value of each newly generated path branch (for example, whether the path has data support, whether it conflicts with known engineering principles or safety constraints, and how high the probability is that the path leads to the final deterministic conclusion), and calculates the evaluation score to evaluate the entire thinking process. The tree is dynamically searched and pruned. This design allows AMCA to prioritize in-depth exploration of paths with the highest assessed value, while also reserving a certain amount of computing resources to explore new paths that appear less optimal but have potential. When a path proves to be a dead end or of extremely low value, AMCA can perform a "backtracking" operation, returning to the previous decision node and investing resources in other more promising paths. Ultimately, a complete path is constructed that starts from the initial problem, runs through all necessary analysis steps, and has a high-confidence conclusion. This path is used as the final execution plan and output in a structured form (such as a step list or dependency diagram) to provide clear guidance for the next stage of action.

[0092] S43. After formulating a detailed execution plan, AMCA enters the action phase. Based on the collaborative scheduling-instant creation mode, it implements each step in the execution plan and generates candidate solutions. Specifically, the constructed database, knowledge base and professional small models are first encapsulated and abstracted into an expert tool set; the professional small models are trained differently for different tasks; after parsing the execution plan, the scheduler agent delegates the subtasks with clear definitions and domain expertise to the most matching expert agent in the expert tool set. The scheduler agent is responsible for managing the data flow and dependencies between different expert agents, ensuring that the steps in the plan are executed in the correct logical order and using the output of the previous task as the input of the next task to form a coherent automated workflow; when the scheduler agent delegates tasks and encounters subtasks that cannot be directly handled by expert agents and existing tools, the tool synthesis module is activated to generate subtasks that cannot be directly handled by new tools by writing Python functions on the fly, and after sandbox verification, the output is integrated into the automated workflow;

[0093] Specifically, within the tool synthesis module, the scheduler agent leverages its powerful reasoning and code generation capabilities to instantly write a standalone Python function capable of completing a specific task. This function is designed to have a single responsibility and a clear interface, and automatically generates a documentation string containing input parameters, output format, and functional descriptions to facilitate subsequent calls and understanding. The newly generated code is not directly executed in the main production environment. Instead, the system deploys it to an isolated, resource-constrained "sandbox" environment. Simultaneously, the scheduler agent automatically generates a set of unit test cases based on the problem context and runs these tests in the sandbox to verify the logical correctness, computational accuracy, and operational safety of the new tool, ensuring that it contains no malicious behavior or serious defects. Only tools that pass the sandbox verification will be officially called by the scheduler to complete the corresponding subtasks in the plan, and their output results will be integrated into the workflow. To improve efficiency, this dynamically synthesized and verified tool can be temporarily added to the scheduler's tool set during the current task session. If the same computational or processing requirements reappear in subsequent planning steps, the scheduler can directly reuse the new tool without regenerating it.

[0094] S44. After that, AMCA enters the reflection and verification phase, conducting a review and correction cycle of the solution. Specifically, the scheduler agent summarizes and integrates the execution results of each expert agent and dynamic tool to form a preliminary solution. Then, AMCA internally criticizes the preliminary solution, generates a reflection text, and raises relevant questions. Then, it switches to the evaluator role and uses the expert tool set and knowledge graph to externally verify the solution, including logical consistency verification, compliance review, and historical case comparison.

[0095] In this embodiment, during logical consistency verification, AMCA proactively queries the knowledge graph to verify whether its recommendations contradict known causal relationships or physical constraints. During compliance review, AMCA invokes a specialized compliance checking tool to compare its solutions with industry standards, internal enterprise process specifications, and relevant safety regulations stored in the database to ensure full compliance. During historical case comparison, AMCA searches the historical case database for similar scenarios, analyzes the success and failure rates of similar measures taken in the past, and provides data-driven risk assessments for current decisions.

[0096] If any issues are identified during internal or external reviews, the relevant critical comments and verification results are recorded and saved, and a correction process is initiated: for minor issues, only the recommended parameters are fine-tuned; for major defects, a complete re-planning is triggered, bringing the newly discovered constraints back to the mind tree exploration stage to regenerate a better strategic execution plan, thereby planning a plan that is better than the initial plan and conducting an iterative feedback loop;

[0097] If all internal and external reviews are successfully passed, the solution will be marked as verified and presented to the user as the final, high-confidence output solution.

[0098] S5. Through a two-stage progressive multi-task training strategy, supervised fine-tuning and reinforcement learning alignment are performed on the large model to optimize its professional capabilities in the field of aerospace intelligent manufacturing. Combined with online learning and incremental update mechanisms, the large model of aerospace intelligent manufacturing can dynamically adapt to the needs of the actual production environment. Ultimately, the construction of the large model of aerospace intelligent manufacturing is completed.

[0099] As a preferred embodiment, step S5 specifically includes:

[0100] S51. Based on the established aerospace intelligent manufacturing domain corpus, a two-stage progressive multi-task training strategy is used to train the aerospace intelligent manufacturing large model after pre-training and construction of the aerospace manufacturing cognitive agent AMCA, enhancing the model's adaptability to domain data distribution and knowledge logic.

[0101] More specifically, step S51 includes:

[0102] S511. The first phase involves multi-task joint supervised fine-tuning, aiming to comprehensively enhance the model's expertise in aerospace manufacturing. This includes parallel optimization of multiple heterogeneous downstream tasks, including process parameter prediction (taking sensor data from the previous N time steps as input and predicting key parameters for the next M time steps, evaluated using a mean squared error loss function); text-image matching (determining the correspondence between process descriptions and X-ray images); quality inspection report generation (given inspection images and equipment status data, generating AS9100-compliant quality inspection conclusion text, evaluated using a cross-entropy loss function); knowledge question answering (answering complex questions based on a process knowledge base, using a hybrid span prediction and generative loss function); and anomaly root cause reasoning (taking an anomaly description as input and outputting possible failure chains, using a graph structure matching loss function).

[0103] The GradNorm algorithm is also used to adaptively allocate loss weights. It dynamically monitors the gradient norm and loss function of each task relative to the initial state, and automatically adjusts the corresponding loss weights. Specifically, it reduces the weight of tasks that learn too quickly and have excessively large gradient norms. Conversely, it increases the weight of tasks that learn more slowly, thereby dynamically balancing the gradient norms of each task. This ensures that all specialized skills are fully and evenly optimized, making the final fine-tuned model more comprehensive and reliable.

[0104] In this example, the gradient loss function Calculate the gradient norm of each task at the current time t The target gradient norm that it expects to achieve The L1 distance (absolute value difference) between them is used to guide the update of the loss weight of each task; the formula of the gradient loss function is expressed as:

[0105] ;

[0106] Relative reverse training rate It is used to measure the learning progress of each task, and its formula is expressed as:

[0107] ;

[0108] in, is the initial loss of task i at training step t, The closer the value of is to 1, the slower the loss of the task decreases, that is, the slower the learning progress;

[0109] If the learning progress of task i is slow ( larger), its target gradient norm will be set higher, prompting it to obtain greater learning intensity in subsequent training; on the contrary, if the task learning progress is fast ( is smaller), its goal will be set lower to slow down its learning speed;

[0110] Gradient norm of each task Calculate the weighted loss of the i-th task versus the shared layer parameters The L2 norm of the gradient (i.e., the length of the vector) is used to measure the learning intensity of each task at the current step; the larger this value is, the greater the impact of the task on the update of the model's shared parameters; The formula is expressed as:

[0111] ;

[0112] Among them, the average gradient norm Calculate the arithmetic mean of the learning strengths of all tasks at the current step.

[0113] S512. In the second stage, the model is deeply aligned through reinforcement learning to make its decision recommendations more in line with complex real-world needs: first, a reward model based on knowledge graphs and domain rules is trained, and the decision recommendations generated by the large model are scored from multiple dimensions in terms of technical feasibility, compliance, and economy. The multi-dimensional scores are integrated into a single scalar reward signal through weighted summation. At the same time, in order to force the model to comply with the insurmountable red lines in the domain, a logical constraint reward function is set to ensure that the model does not violate the predefined immutable rules. During training, the model is forced to converge quickly to the compliant solution space, and domain knowledge is forced to be aligned. An artificial feedback loop is then introduced to further improve the alignment accuracy, and the sampling results are regularly scored by domain experts. Finally, the system will weightedly fuse the scaled scalar reward signal with the high-quality manual score to form the final reward signal, which is then combined with the proximal policy optimization PPO algorithm to fine-tune the large model.

[0114] In this embodiment, the calculation formula of the scalar reward signal is as follows: For a given decision suggestion, the reward model outputs a vector consisting of scores of multiple dimensions. Since different dimensions have different importance in the final decision, the weight of each dimension is set in advance by domain experts and project managers. By multiplying the score of each dimension by its corresponding weight and then adding all the results, a single, comprehensive scalar reward signal is obtained. , the formula is:

[0115] ;

[0116] in, represents the compliance weight, represents the compliance score, represents the technical feasibility weight, represents the technical feasibility score, represents the economic weight, Indicates the economic score;

[0117] The workflow of the logical constraint reward function is as follows: the decision generated by the large model will first be sent to the logical constraint reward function. The parser inside the function will extract the key information related to the rules, check them one by one, and perform a strict logical comparison with each rule in the "immutable rule library" (including the hard constraint rules defined and stored in the corpus of the aerospace intelligent manufacturing field). Based on the verification results, the function will output a reward value: as long as any rule is found to be violated, the function will immediately stop checking and return -100 as a penalty signal; only when the decision recommendation has completely passed the verification of all immutable rules will the function return a neutral value of 0.

[0118] S52. Establish a multi-level evaluation and iteration mechanism to conduct performance evaluation and continuous dynamic update of the final aerospace intelligent manufacturing large model;

[0119] More specifically, step S52 includes:

[0120] S521. First, we build a high-quality offline test set. We select the latest six months of fault cases and their associated multimodal data from the database as the test set. We strictly isolate the time window of the training data to simulate unknown problems encountered after the model is launched. We also inject artificially constructed adversarial examples to comprehensively test the model's anti-interference ability and logical consistency.

[0121] S522. Construct a multi-dimensional evaluation index system. Specifically, at the general text generation level, use the common generation indicators BLEU and ROUGE to compare with reference text to evaluate the fluency and information fidelity of the model-generated content. Secondly, evaluate key indicators specific to its professional definition field, including process parameter deviation (used to quantify the gap between the parameter value predicted by the model and the engineering standard or optimal value) and knowledge compliance rate (used to calculate the proportion of decision recommendations generated by the model that conform to the established rules of the knowledge base).

[0122] In this embodiment, the mean absolute percentage error (MAPE) is calculated as the process parameter deviation, and the formula is expressed as follows:

[0123] ;

[0124] The specific evaluation process of knowledge compliance rate is as follows: Design a A test set of questions is specifically designed to stimulate the model to answer in scenarios where specific rules must be followed; Answer each question one by one, and then judge whether each answer violates the relevant rules in the knowledge base, and count the number of answers that do not violate any rules at all. , and thus calculate the compliance rate:

[0125] ;

[0126] Ultimately, to measure the true commercial value of the model, a large-scale aerospace intelligent manufacturing model was deployed in a digital twin environment. The impact of the model's decision-making recommendations on core indicators, including yield and production efficiency, was evaluated. The process parameters or scheduling plans generated by the model were used as input, and simulations were performed on a virtual production line. The resulting data was analyzed to quantify its true impact on production efficiency, thereby achieving a closed-loop, quantitative, and risk-free assessment of the model's application value.

[0127] S523. Establish an online learning and incremental update mechanism; use the decision recommendations given by the large model in the production environment and the results after actual execution as feedback data to build a feedback closed-loop data set; successful execution cases serve as positive samples to reinforce the correct behavior of the model, and failure cases trigger analysis and serve as negative samples to correct the model, realizing continuous learning based on actual feedback; use parameter efficient fine-tuning PEFT technology for LoRA low-rank adaptation, only updating a small number of parameters, without affecting general capabilities, to quickly realize monthly incremental updates of the model, ensuring that the model continues to adapt to changes in the production environment without interrupting service.

[0128] In summary, the construction and training of aerospace intelligent manufacturing large models based on the universal large model of the base has been realized. The present invention transforms the universal large model into a professional intelligent entity that deeply understands the industry logic and accurately responds to manufacturing needs by building a high-quality multimodal data ecosystem, domain-based architecture transformation, phased multi-task joint training, integrated retrieval enhancement and function call, closed-loop evaluation and continuous iteration and other key technologies. This method effectively solves the challenges faced by traditional large models in the aerospace field, such as multimodal data processing, deep understanding of domain knowledge, logical constraint satisfaction and dynamic environment adaptation, and provides a feasible and high-performance technical path for the digital and intelligent upgrade of aerospace intelligent manufacturing.

[0129] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0130] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for constructing a large model for aerospace intelligent manufacturing, characterized in that: The specific steps include: S1. Build a professional data collection and management system to collect a large amount of raw aerospace intelligent manufacturing real data from actual aerospace intelligent manufacturing scenarios, perform pre-processing and data association, and build an aerospace intelligent manufacturing database; S2. Targeting multimodal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge acquisition and structured conversion pipeline, a multi-dimensional domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism to build a professional knowledge base for aerospace intelligent manufacturing. Step S2 specifically includes: S21. Construct a multimodal knowledge acquisition and structured conversion pipeline for intelligent aerospace manufacturing. Specifically, the pipeline involves directly acquiring structured documents from knowledge data, parsing key knowledge elements from semi-structured text and unstructured data using natural language processing techniques, and converting them into structured triples. Computer vision analysis techniques are then used for unstructured video data to extract key frame sequences from operational steps and annotate key process points. Furthermore, an expert knowledge entry interface is established to support process engineers in contributing implicit experience in a form-based format. S22. Construct a multi-dimensional, interconnected knowledge graph for aerospace intelligent manufacturing, organizing the knowledge network around product family trees, process flows, and fault trees. Specifically, define a three-layer ontology model: "parts-materials-processes." Associate part design specifications with a material property library and map them to processing parameter knowledge. Simultaneously, construct a process constraint rule library, describe process dependencies through logical expressions, and link them to testing standards. Furthermore, integrate a historical failure case library to establish a causal chain from "anomaly-root cause-solution," forming a traceable diagnostic knowledge network. A hybrid architecture combining graph database Neo4j and vector database Milvus is employed to efficiently store and apply complex knowledge networks. While the graph database Neo4j stores explicit structured relationships, the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval. Finally, a knowledge evidence library is established, linking each knowledge node to its original basis, ensuring a bidirectional link between knowledge nodes and source evidence. S23. Build a knowledge quality control system throughout the entire process; establish a dual knowledge verification mechanism: automatically discover and report conflicting rules through logical consistency checks, trigger a manual review process, and verify knowledge validity based on historical data backtracking; establish a version control system to record the revision history and effective time of knowledge items, ensuring that the reference version of each process rule can be traced during airworthiness certification; at the access and use level, adopt attribute-based access control permission management, dynamically control the scope of knowledge access based on personnel roles and project confidentiality levels, and set up an approval and unlocking process for core process parameters to ensure the security of knowledge assets; S24. Establish a dynamic update mechanism to continuously optimize the knowledge base. Set a knowledge freshness monitoring indicator (KFI) and regularly synchronize with international standard updates. Use machine learning to analyze newly generated production data to automatically extract valuable potential knowledge patterns. Extracted knowledge patterns must be reviewed and confirmed by domain experts before being stored. In addition, build an online collaboration platform to pool the implicit experience of frontline personnel to complete the construction of a professional knowledge base for aerospace intelligent manufacturing. S3. Align cross-modal manufacturing data to generate a corpus in the field of aerospace intelligent manufacturing. Combined with the base general large model for pre-training, the system injects professional terminology semantics and multimodal association capabilities through masked language modeling tasks and contrastive learning, completing the aerospace domain knowledge transfer of the base model and obtaining the pre-trained aerospace intelligent manufacturing large model. S4. Based on the pre-trained aerospace intelligent manufacturing big model, build the aerospace manufacturing cognitive agent AMCA to form a complete, dynamic, and self-improving framework for solving complex engineering problems; S5. Through a two-stage progressive multi-task training strategy, supervised fine-tuning and reinforcement learning alignment are performed on the large model to optimize the professional capabilities of the large model in the field of aerospace intelligent manufacturing. Combined with online learning and incremental update mechanisms, the large model of aerospace intelligent manufacturing can dynamically adapt to the needs of the actual production environment; finally, the construction of the large model of aerospace intelligent manufacturing is completed.

2. The method for constructing a large model of aerospace intelligent manufacturing according to claim 1, characterized in that: Step S1 specifically includes: S11. Establish a data aggregation channel to acquire a large amount of real-world aerospace manufacturing data through multi-source data collection and integration. Define a unified data identification system for the dispersed design, production, and testing data sources in each type of real-world aerospace manufacturing scenario, and use a heterogeneous protocol compatibility strategy to aggregate each data source into the original database. S12. Design a hierarchical and classified data storage architecture, specifically including: first, storing all collected raw data in a distributed database to build a data lake to retain the full amount of original information; then, classifying and importing the data into different databases based on data characteristics and application scenarios: For key business data in structured data, pre-process it through a quality rule engine, including outlier filtering, unit standardization, and missing value interpolation, and then import it into a columnar data warehouse; for high-frequency equipment data that changes in real time in structured data, store it in a time series database; for semi-structured data, store its metadata and associations in document data; for unstructured data, manage file entities through object storage; S13. Establish a complete data governance system, including: defining unified naming conventions, building a global data catalog based on a metadata management platform, and clearly defining the business meaning, data lineage, and access rights of each data entity; on this basis, implementing automated data quality testing based on a dynamic rule engine to identify outliers and verify data formats; and finally, implementing a hierarchical encryption strategy to secure data and protect sensitive information, thereby forming a closed-loop data governance solution covering the entire process. S14. Build an association network between data to provide support for the training and analysis of large models; transform discrete data into a reasonable knowledge network business entity network through the entity relationship model to complete the construction of the aerospace intelligent manufacturing database.

3. The method for constructing a large model for intelligent aerospace manufacturing according to claim 1, characterized in that: KFI is obtained by weighting four core factors, specifically: ； in, is the weight of each factor, which is dynamically adjusted according to the knowledge type; TDF is the time decay factor, and the formula is: ； Among them, T is the validity period of knowledge, is the current time, The time when the knowledge item was last verified and updated; VAF is the version alignment factor, which is obtained by comparing with an external standard database; it is specifically set in the form of a discrete factor: 1.0: The standard version on which the knowledge is based is the latest valid version; 0.5: The knowledge is based on a newer version of the standard, but the old version has not been abolished; 0.0: The knowledge is based on a standard version that has been explicitly obsoleted or superseded; N / A: Knowledge is not linked to any external criteria; UAF is the application activity factor, and the formula is: ； in, Indicates the frequency with which a piece of knowledge was called by the big model, retrieved, and adopted in knowledge question answering in the past cycle; Indicates the average frequency of all knowledge being called by the big model, retrieved and adopted in knowledge question answering in the past cycle; PVF is the performance verification factor, which is used to test the validity of knowledge.

4. The method for constructing a large model for intelligent aerospace manufacturing according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31. Construct a cross-modal, semantically precisely aligned corpus for aerospace intelligent manufacturing: Extract diverse information, encompassing structured, semi-structured, and unstructured types, from the aerospace intelligent manufacturing database. Subsequently, transform and align the heterogeneous data. On the one hand, some structured data is converted into natural language descriptions using automated templates. On the other hand, different types of data referring to the same manufacturing entity are precisely correlated by timestamp, constructing cross-modal correlation samples of the "signal-image-parameter-text" model to train the model to understand the inherent connections between multi-source data. Finally, using the established aerospace intelligent manufacturing professional knowledge base, the text content in the corpus is enhanced with domain knowledge. S32. Based on the established aerospace intelligent manufacturing domain corpus and the base general large model, domain adaptation pre-training is performed to inject basic knowledge of the aerospace intelligent manufacturing field into the base general large model; domain general corpus and some multimodal alignment samples are used for mixed training, and masked language modeling (MLM) tasks are adopted, focusing on masked domain entities, forcing the model to rely on knowledge clues in the context for prediction; at the same time, comparative learning tasks are introduced, given different modal descriptions of the same process, the model is required to determine whether they match, in order to enhance cross-modal alignment capabilities.

5. The method for constructing a large model for intelligent aerospace manufacturing according to claim 1, characterized in that: Step S4 specifically includes: S41. After receiving a user query, AMCA first initiates a deep traversal of the knowledge graph, following the "retrieval-inference-pruning" mechanism to extract knowledge fragments. Specifically, it uses vector retrieval technology to match user queries with entity nodes in the knowledge graph, quickly locating the initial node most relevant to the query semantics as the seed node for graph traversal. Starting from the seed node, driven by the reasoning capability of the large model, it conducts multi-hop logical exploration to determine whether the exploration along the relationship path of each node is conducive to solving the problem, thereby connecting dispersed but logically strongly related information to form a complete chain of evidence. During the exploration process, it tracks multiple reasoning paths in parallel, continuously evaluates the information gain of each path, and prunes paths that lead to irrelevant information and logical dead loops. Finally, it aggregates, deduplicates, and logically reorganizes all high-value knowledge fragments on the reasoning path to form a comprehensive knowledge context. S42. After obtaining the comprehensive knowledge context, AMCA enters the cognitive stage and adopts the thinking tree exploration framework to transform vague user intentions into clear, executable, and structured solutions. Specifically, the user's high-level goals are first decomposed into a series of thinking nodes, and multiple different but feasible next steps and hypotheses are generated for each thinking node. Then, the self-assessment mechanism is activated, and the comprehensive knowledge context and its own domain knowledge are used to evaluate the value of each newly generated path branch, calculate the evaluation score, and dynamically search and prune the entire thinking tree. Finally, a complete path is constructed starting from the initial problem, running through all analysis steps, and with a high-confidence conclusion. This path is used as the final execution plan and output in a structured form. S43. Entering the action phase, based on the collaborative scheduling-instant creation mode, each step in the execution plan is implemented and candidate solutions are generated. Specifically, the constructed database, knowledge base, and professional small models are first uniformly encapsulated and abstracted into an expert tool set; after parsing the execution plan, the scheduler agent delegates clearly defined and domain-specific subtasks to the most suitable expert agent in the expert tool set. The scheduler agent is responsible for managing the data flow and dependencies between different expert agents, ensuring that the steps in the plan are executed in the correct logical order and using the output of the previous task as the input of the next task to form a coherent automated workflow; when the scheduler agent delegates tasks and encounters subtasks that cannot be directly handled by expert agents and existing tools, the tool synthesis module is activated to generate subtasks that cannot be directly handled by new tools by writing Python functions on the fly. After sandbox verification, the output is integrated into the automated workflow; S44, AMCA enters the reflection and verification stage, and conducts a review and correction cycle of the solution. Specifically: First, the scheduler agent summarizes and integrates the execution results of each expert agent and dynamic tool to form a preliminary solution; then AMCA internally criticizes the preliminary solution, generates a reflection text, raises relevant questions, and then switches to the evaluator role, using the expert tool set and knowledge graph to externally verify the solution, including logical consistency verification, compliance review, and historical case comparison; if problems arise during internal and external reviews, the relevant critical opinions and verification results are recorded and saved, the correction program optimization plan is started, and an iterative feedback cycle is performed; if all internal and external reviews are successfully passed, the solution is marked as verified and presented to the user as the final, highly reliable output solution.

6. The method for constructing a large model for intelligent aerospace manufacturing according to claim 1, characterized in that: Step S5 specifically includes: S51. Based on the established aerospace intelligent manufacturing domain corpus, a two-stage progressive multi-task training strategy is used to train the aerospace intelligent manufacturing large model after pre-training and construction of the aerospace manufacturing cognitive agent AMCA, enhancing the model's adaptability to domain data distribution and knowledge logic. S52. Establish a multi-level evaluation and iteration mechanism to conduct performance evaluation and continuous dynamic update of the final constructed aerospace intelligent manufacturing large model.

7. The method for constructing a large model for intelligent aerospace manufacturing according to claim 6, characterized in that: Step S51 specifically includes: S511. The first phase involves multi-task joint supervised fine-tuning, aiming to comprehensively enhance the model's expertise in aerospace manufacturing. The model concurrently optimizes multiple heterogeneous downstream tasks, including process parameter prediction, text-image matching, quality inspection report generation, knowledge question answering, and anomaly root cause reasoning. The GradNorm algorithm is used to adaptively assign loss weights, dynamically monitoring the gradient norm and loss function of each task relative to the initial state and automatically adjusting the corresponding loss weights. S512. In the second stage, the model is deeply aligned through reinforcement learning to make its decision recommendations more in line with complex real-world needs: first, a reward model based on knowledge graphs and domain rules is trained, and the decision recommendations generated by the large model are scored from multiple dimensions in terms of technical feasibility, compliance, and economy. The multi-dimensional scores are integrated into a single scalar reward signal through weighted summation; at the same time, a logical constraint reward function is set to ensure that the model does not violate the predefined immutable rules, forcing the model to converge quickly to the compliant solution space during training, and forcing alignment of domain knowledge; then an artificial feedback loop is introduced to further improve the alignment accuracy, and the sampling results are regularly scored by domain experts; finally, the system will weightedly fuse the scaled scalar reward signal with the high-quality manual score to form the final reward signal, which is then combined with the proximal policy optimization PPO algorithm to fine-tune the large model.

8. The method for constructing a large model for intelligent aerospace manufacturing according to claim 6, characterized in that: Step S52 specifically includes: S521. First, we build a high-quality offline test set. We select the latest six months of fault cases and their associated multimodal data from the database as the test set. We strictly isolate the time window of the training data to simulate unknown problems encountered after the model is launched. We also inject artificially constructed adversarial examples to comprehensively test the model's anti-interference ability and logical consistency. S522. Construct a multi-dimensional evaluation index system. Specifically, at the general text generation level, use the general generation indicators BLEU and ROUGE to compare with reference text to evaluate the fluency and information fidelity of the model-generated content. Secondly, define domain-specific key indicators, including process parameter deviation and knowledge compliance rate. Finally, to measure the real business value brought by the model, deploy a large-scale aerospace intelligent manufacturing model in a digital twin environment and evaluate the actual improvement effect of the decision recommendations provided by the model on core indicators such as yield rate and production efficiency. S523. Establish an online learning and incremental update mechanism; use the decision recommendations given by the large model in the production environment and the results after actual execution as feedback data to build a feedback closed-loop data set; successful execution cases serve as positive samples to reinforce the correct behavior of the model, and failure cases trigger analysis and serve as negative samples to correct the model, realizing continuous learning based on actual feedback; use parameter efficient fine-tuning PEFT technology for LoRA low-rank adaptation, only updating a small number of parameters, without affecting general capabilities, to quickly realize monthly incremental updates of the model, ensuring that the model continues to adapt to changes in the production environment without interrupting service.

Images