Aerospace intelligent manufacturing large model construction method

Abstract

The invention discloses an aerospace intelligent manufacturing large model construction method, which comprises the steps of collecting original data, performing preprocessing and data association, and constructing an aerospace intelligent manufacturing database; establishing a knowledge acquisition and structured conversion assembly line, a multi-dimensional associated domain knowledge graph, a knowledge quality control system and a dynamic updating mechanism, and constructing a professional knowledge base; aligning the cross-modal manufacturing data to generate a corpus; combining base general large model pre-training, injecting terminology semantics and multi-modal association capability, and completing knowledge migration; based on the pre-trained aerospace intelligent manufacturing large model, constructing an aerospace manufacturing cognitive agent, and forming a complex engineering problem solving framework; professional ability is optimized through a two-stage progressive multi-task training strategy, and dynamic adaptation of a production environment is realized in combination with an online learning and incremental updating mechanism. The intelligent level of aerospace intelligent manufacturing is remarkably improved.

Description

Technical Field

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

[0002] In the aerospace field, traditional knowledge management and intelligent question-and-answer systems face significant technical bottlenecks and challenges. Existing systems often rely on isolated databases and expert rule bases, and it is difficult to effectively integrate diverse multi-modal data such as design drawings, sensor timings, and process logs, resulting in serious information islands. Although traditional general large language models perform excellently in open-domain tasks, they have serious deficiencies in professional domain knowledge coverage, lack a deep understanding of the complex term system and engineering logic in aerospace, and are difficult to accurately analyze complex manufacturing problems.

[0003] In addition, the knowledge update of traditional models usually relies on time-consuming manual annotation and retraining, resulting in lagging timeliness and being unable to meet the needs of the rapid iteration of aerospace technology and the dynamic changes in the production environment. Therefore, constructing a professional intelligent agent that can deeply understand domain knowledge, process multi-modal data, have logical reasoning, and continuous adaptation ability is the key challenge to promoting the upgrade of aerospace intelligent manufacturing. Summary of the Invention

[0004] Aiming at the above deficiencies of the existing technology, the present invention proposes a method for constructing an aerospace intelligent manufacturing large model, which solves the pain points such as term deviation, shallow knowledge, weak logic, and insufficient data processing ability existing in traditional general large models in professional scenarios, and improves the intelligent level of aerospace intelligent manufacturing.

[0005] To achieve the above technical objectives, the present invention provides the following technical solutions: A method for constructing an aerospace intelligent manufacturing large model specifically includes the following steps: S1. Build a professional data collection and management system, collect a large amount of original aerospace intelligent manufacturing real data from the actual scenarios of aerospace intelligent manufacturing, and perform preprocessing and data association to construct an aerospace intelligent manufacturing database; S2. Face the multi-modal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge collection and structured conversion pipeline, a multi-dimensional associated domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism, and construct an aerospace intelligent manufacturing professional knowledge base; S3. Align cross-modal manufacturing data to generate an aerospace intelligent manufacturing domain corpus; combine with a base general large model for pre-training, inject professional term semantics and multi-modal association capabilities through masked language modeling tasks and contrastive learning, complete the knowledge transfer of the aerospace domain of the base model, and obtain a pre-trained aerospace intelligent manufacturing large model; S4. Based on the pre-trained aerospace intelligent manufacturing large model, construct an aerospace manufacturing cognitive intelligent agent AMCA to form a complete, dynamic, and self-improving complex engineering problem-solving framework; S5. Through a two-stage progressive multi-task training strategy, perform supervised fine-tuning and reinforcement learning alignment on the large model to optimize the professional capabilities of the large model in the field of aerospace intelligent manufacturing. Combine online learning and incremental update mechanisms to enable the aerospace intelligent manufacturing large model to dynamically adapt to the needs of the actual production environment; finally, complete the construction of the aerospace intelligent manufacturing large model.

[0006] Further, step S1 specifically includes: S11. Establish a data aggregation channel. Through multi-source data collection and integration, obtain a large amount of real aerospace manufacturing data; for the scattered design, production, and inspection data sources in each type of real aerospace manufacturing scenario, define a unified data identification system, and adopt a heterogeneous protocol compatibility strategy to uniformly summarize each data source into the original database; S12. Design a hierarchical and classified data storage architecture, specifically including: first, uniformly store all the collected original data in a distributed database to build a data lake to retain all the original information; then, according to the data characteristics and application scenarios, classify and import the data into different databases: for the key business data in structured data, perform preprocessing including outlier filtering, unit standardization, and missing value imputation through a quality rule engine, and then import it into a columnar data warehouse; for the high-frequency device data with real-time changes in structured data, store it in a time series database; for semi-structured data, store its metadata and associated relationships in a document database; for unstructured data, manage file entities through object storage; S13. Establish a complete data governance system, specifically including: define a unified naming specification, build a global data catalog relying on the metadata management platform, clearly define the business meaning, data lineage, and access right attribution of each data entity; on this basis, implement automated data quality detection based on a dynamic rule engine to identify outliers and perform format verification on the data; finally, adopt a hierarchical encryption strategy to protect the data security and protect sensitive information, thereby forming a closed-loop data governance solution covering the entire process; S14. Construct an association network between data to provide support for the training and analysis of the large model; transform discrete data into an inferable knowledge network business entity network through an entity relationship model to complete the construction of the aerospace intelligent manufacturing database.

[0007] Further, step S2 specifically includes: S21. Construct a multi-modal knowledge acquisition and structured transformation pipeline for aerospace intelligent manufacturing, specifically as follows: directly collect structured documents in knowledge data; for semi-structured texts and unstructured data, parse key knowledge elements through natural language processing technology and transform them into structured triple forms; then for video-based unstructured data, use computer vision analysis technology to extract key frame sequences of the operation steps process and label the process key points; at the same time, establish an expert knowledge entry interface to support process engineers to contribute tacit experience in a form-based manner. S22. Construct a knowledge graph in the field of aerospace intelligent manufacturing with multi-dimensional associations, taking the product family tree, process flow, and fault tree as the core axes to organize the knowledge network; specifically, define a three-layer ontology model of "part - material - process", associate part design specifications with the material property library, and map them to machining parameter knowledge. At the same time, construct a process constraint rule library, describe the process dependency relationship through logical expressions, and link it to the inspection standard; in addition, integrate the historical fault case library, establish a causal chain of "abnormal phenomenon - root cause - solution measure", and form a traceable diagnostic knowledge network. And adopt a hybrid architecture of graph database Neo4j and vector database Milvus to meet the efficient storage and application of complex knowledge networks; the graph database Neo4j stores explicit structured relationships, while the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval; finally, establish a knowledge evidence base to associate each knowledge node with the original basis to ensure the two-way link between the knowledge node and the source evidence. S23. Construct a knowledge quality control system throughout the whole process; establish a dual knowledge verification mechanism: automatically discover and report conflict rules through logical consistency checks, trigger the manual review process, and verify the knowledge validity based on historical data backtracking; establish a version control system to record the revision history and effective time of knowledge items to ensure that the basis version of each process rule can be traced during airworthiness certification; at the access and usage level, adopt attribute-based access control permission management to dynamically control the knowledge access scope according to personnel roles and project confidentiality levels, and set an approval and unlocking process for core process parameters to ensure the security of knowledge assets. S24. Establish a dynamic update mechanism to continuously promote the optimization of the knowledge base; set the knowledge freshness monitoring index KFI and synchronize international standard updates regularly; automatically extract valuable potential knowledge patterns by analyzing newly generated production data through machine learning, and the extracted knowledge patterns need to be reviewed and confirmed by domain experts before being stored in the database; in addition, build an online collaboration platform to collect the tacit experience of front-line personnel to complete the construction of the aerospace intelligent manufacturing professional knowledge base.

[0008] More specifically, KFI is obtained by weighting four core factors, specifically: ; Among them, , , , are the weights of each factor, which are dynamically adjusted according to the knowledge type; TDF is the time decay factor, and the formula is expressed as: ; Among them, T is the knowledge validity period, is the current time, is 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; UAF is the application activity factor, and the formula is expressed as: ; Among them, represents the frequency at which a certain piece of knowledge is called by the large model and retrieved and adopted in knowledge Q&A within the past cycle; represents the average frequency at which all knowledge is called by the large model and retrieved and adopted in knowledge Q&A within the past cycle; PVF is the performance verification factor, which is used to test the validity of knowledge.

[0009] Furthermore, step S3 specifically includes: S31. Construct a corpus in the field of aerospace intelligent manufacturing with cross-modal and semantic precise alignment: Extract diverse information covering structured, semi-structured, and unstructured types from the aerospace intelligent manufacturing database; subsequently, perform transformation and alignment on heterogeneous data. On the one hand, convert some structured data into natural language descriptions through automated templates, and on the other hand, finely associate different types of data pointing to the same manufacturing entity according to timestamps to construct cross-modal association samples in the form of "signal-image-parameter-text" to train the model to understand the internal relationships between multi-source data; finally, use the constructed aerospace intelligent manufacturing professional knowledge base to enhance the domain knowledge of the text content in the corpus. S32. Perform domain adaptation pre-training based on the constructed aerospace intelligent manufacturing field corpus and the base general large model to inject basic knowledge in the field of aerospace intelligent manufacturing into the base general large model; use domain general corpus and part of the multi-modal alignment samples for mixed training, adopt the masked language modeling MLM task, focusing on masking domain entities, forcing the model to rely on knowledge clues in the context for prediction; at the same time, introduce a contrastive learning task, given different modal descriptions of the same process, and require the model to judge whether they match to enhance the cross-modal alignment ability.

[0010] 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: Using vector retrieval technology, it matches the user query with the entity nodes in the knowledge graph to quickly locate the initial node most relevant to the query semantics 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 each node's relationship path is conducive to solving the problem, thereby connecting dispersed but logically strongly related information to form a complete evidence chain. During the exploration process, it parallelly tracks multiple reasoning paths, continuously evaluates the information gain of each path, and prunes the paths leading to irrelevant information and logical deadlocks. Finally, it summarizes, deduplicates, and logically reorganizes all high-value knowledge fragments on the reasoning paths to form a comprehensive knowledge context. S42. After obtaining the comprehensive knowledge context, AMCA enters the cognitive stage and adopts a thought tree exploration framework to transform the vague user intention into a clear, executable, and structured solution. Specifically: First, it decomposes the user's high-level goal into a series of thinking nodes, generates multiple different but feasible next steps and hypotheses for each thinking node. Then, it activates the self-evaluation mechanism, uses the comprehensive knowledge context and its own domain knowledge to evaluate the value of each newly generated path branch, calculates the evaluation score, and dynamically searches and prunes the entire thought tree. Finally, it constructs a complete path starting from the initial problem, passing through all necessary analysis steps, and having a high-confidence conclusion, takes this path as the final execution plan, and outputs it in a structured form. S43. Entering the action stage, based on the collaborative scheduling - instant creation mode, it implements each step in the execution plan and generates candidate solutions. Specifically: First, it uniformly encapsulates and abstracts the constructed database, knowledge base, and professional small models into an expert toolset. After parsing the execution plan, through the scheduler agent, it delegates well-defined and domain-specific subtasks to the most matching expert agent in the expert toolset. 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 there are subtasks that no expert agent and existing tools can directly handle during the process of the scheduler agent delegating tasks, it activates the tool synthesis module, generates new tools by instantaneously writing Python functions for the subtasks that cannot be directly handled at all, and after sandbox verification, integrates the output 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 plan.

[0011] Furthermore, step S5 specifically includes: S51. Based on the established aerospace intelligent manufacturing domain corpus, a two-stage progressive multi-task training strategy is adopted to train the aerospace intelligent manufacturing large model after pre-training and construction of the aerospace manufacturing cognitive agent AMCA, to enhance 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.

[0012] More specifically, step S51 specifically includes: S511. In the first phase, multi-task joint supervised fine-tuning is carried out to comprehensively improve the model's professional skills in the field of aerospace manufacturing: multiple heterogeneous downstream tasks are optimized in parallel, including process parameter prediction, text-image matching, quality inspection report generation, knowledge question answering, and abnormal root cause reasoning; the GradNorm algorithm is used to achieve adaptive allocation of loss weights, dynamically monitor the gradient norm and loss function of each task relative to the initial state, and automatically adjust the corresponding loss weights; S512. In the second stage, the model is deeply aligned through reinforcement learning to make its decision-making suggestions more in line with complex real-world requirements: First, a reward model based on a knowledge graph and domain rules is trained to score the decision-making suggestions generated by the large model from multiple dimensions including technical feasibility, compliance, and economy, and 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 predefined immutable rules, forcing the model to quickly converge to the compliance solution space during training and aligning with domain knowledge. Then, an artificial feedback loop is introduced to further improve the alignment accuracy by having domain experts regularly score the sampling results. Finally, the system will perform weighted fusion of the large-scale scalar reward signal and high-quality artificial scores to form the final reward signal, and based on this, the proximal policy optimization (PPO) algorithm is combined to fine-tune the large model.

[0013] More specifically, step S52 specifically includes: S521. First, construct a high-quality offline test set. Select the failure cases and their associated multi-modal data in the latest 6 months from the database as the test set, strictly isolate the time window of the training data, and simulate the unknown problems encountered after the model goes online. At the same time, inject artificially constructed adversarial samples to comprehensively test the anti-interference ability and logical consistency of the model. S522. Construct a multi-dimensional evaluation index system, specifically: At the general text generation level, use the general generation metrics BLEU and ROUGE to compare with the reference text to evaluate the fluency and information fidelity of the content generated by the model. Secondly, define key metrics specific to the domain, including process parameter deviation and knowledge compliance rate. Finally, to measure the real business value brought by the model, deploy the aerospace intelligent manufacturing large model in the digital twin environment and evaluate the actual improvement effect of the decision-making suggestions given by the model on core metrics including the good product rate and production efficiency. S523. Establish an online learning and incremental update mechanism; use the decision-making suggestions given by the large model in the production environment and the results after actual execution as feedback data to flow back, and construct a feedback closed-loop data set. Successful execution cases are used as positive samples to strengthen the correct behavior of the model, and failed cases trigger analysis and are used as negative samples to correct the model, realizing continuous learning based on actual feedback. The parameter-efficient fine-tuning (PEFT) technology is used for LoRA low-rank adaptation, only updating a small number of parameters, and quickly realizing the monthly incremental update of the model without affecting the general capabilities, ensuring that the model continuously adapts to changes in the production environment without interrupting the service.

[0014] Based on the above technical solutions, the present invention has at least the following beneficial effects: A method for constructing a large model for aerospace intelligent manufacturing proposed by the present 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 constructing a high-quality multi-modal data ecosystem and a professional knowledge base, a solid foundation is provided for the model; the domain-specific architecture transformation enables the model to deeply integrate and understand various modal information such as text, numerical values, and vision, and has the decoding ability of knowledge perception, effectively avoiding physically infeasible outputs; the phased progressive training strategy combines multi-task learning and reinforcement learning, enabling the model to comprehensively master various professional skills in aerospace manufacturing, from basic parameter prediction to complex root cause reasoning; the 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; the engineering deployment and human-machine collaboration mechanism ensure the efficiency, safety, and reliability of the model in actual applications. The present invention effectively solves the pain points such as term deviation, shallow knowledge, weak logic, and insufficient data processing ability existing in traditional general large models in professional scenarios, and significantly improves the intelligent level of aerospace intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flow chart of a method for constructing a large model for aerospace intelligent manufacturing proposed by the present invention; Figure 2 is a schematic work flow chart of the aerospace manufacturing cognitive intelligent agent AMCA constructed by the method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the attached Figure 1-2 drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0017] Although the steps in the present invention are numbered, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It can be understood that the term "and / or" used herein relates to and encompasses any and all possible combinations of one or more of the associated listed items.

[0018] As Figure 1-2 shown, a method for constructing a large model for aerospace intelligent manufacturing proposed by the present invention specifically includes the following steps: S1. Build a professional data collection and management system, collect a large amount of original real data of aerospace intelligent manufacturing from the actual scenarios of aerospace intelligent manufacturing, and perform preprocessing and data association to construct an aerospace intelligent manufacturing database; As a preferred implementation manner, step S1 specifically includes: S11. Establish a data aggregation channel, and obtain a large amount of real aerospace manufacturing data through multi-source data collection and integration; for the scattered design, production, and inspection data sources in each type of real aerospace manufacturing scenario, define a unified data identification system (such as process number, equipment ID, material batch), and adopt a heterogeneous protocol compatibility strategy to uniformly aggregate each data source into the original database; in this embodiment, the heterogeneous protocol compatibility strategy is specifically: real-time access the time-series data streams of numerically controlled machine tools and sensors through industrial Internet of Things protocols, synchronize the 3D models and process files of the PLM system through API interfaces, extract the work order execution records and quality inspection results in the MES using ETL tools, and at the same time include unstructured data such as laboratory material test reports and X-ray detection images in the collection scope, and uniformly aggregate them into the original database.

[0019] S12. To meet the storage and analysis requirements of different types of data, this application designs a hierarchical and classified data storage architecture, which specifically includes: first, uniformly store all the collected original data in a distributed database to build a data lake to retain the full amount of original information; then, according to the data characteristics and application scenarios, classify and import the data into different databases: for the key business data (such as BOM tables, quality inspection indicators) in structured data, perform preprocessing including outlier filtering, unit standardization, and missing value imputation through a quality rule engine, and then import it into a columnar data warehouse to efficiently support online analytical processing (OLAP) across processes; for the high-frequency equipment data with real-time changes in structured data, store it in a time-series database to optimize high-speed read and write and time-series query performance; for semi-structured data, store its metadata and association relationships in document data for convenient and fast retrieval and invocation; for unstructured data, manage the file entities through object storage; S13. To ensure the consistency, accuracy, and security of data throughout the entire life cycle of data collection, storage, and application, this application establishes a complete data governance system, which specifically includes: defining a unified naming specification, constructing a global data directory relying on the metadata management platform, clearly defining the business meaning, data lineage, and access permission ownership of each data entity; on this basis, implement automated data quality detection based on a dynamic rule engine to identify outliers and perform format verification on the data to ensure the reliability of the data; finally, adopt a hierarchical encryption strategy to protect the data security and protect sensitive information, thereby forming a closed-loop data governance solution covering the entire process; In this embodiment, the hierarchical encryption policy is specifically as follows: The transport layer ensures the security of the data channel through TLS 1.3. The storage layer encrypts sensitive design drawings with AES-256 and restricts the access granularity of different roles to the process library and inspection reports through the RBAC model, meeting the requirements of the AS9100 standard for data traceability and permission isolation.

[0020] S14. Construct an association network between data to provide support for the training and analysis of the large model; transform discrete data into an inferable knowledge network business entity network through an entity relationship model, that is, use a graph database to construct 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 an inferable knowledge network to complete the construction of the aerospace intelligent manufacturing database.

[0021] S2. For multi-modal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge acquisition and structured transformation pipeline, a multi-dimensional associated domain knowledge graph, a knowledge quality control system, and a dynamic update mechanism to construct a professional knowledge base for aerospace intelligent manufacturing; As a preferred implementation method, step S2 specifically includes: S21. Construct a multi-modal knowledge acquisition and structured transformation pipeline for aerospace intelligent manufacturing, specifically: directly collect structured documents (such as design manuals, process specifications, and quality inspection standards) in knowledge data, and for semi-structured texts (such as R & D reports and failure analysis records) and unstructured data (such as expert experience and assembly operation videos), parse key knowledge elements through natural language processing technology and transform them into structured triple forms; then for video-based unstructured data, use computer vision analysis technology to extract key frame sequences of the operation steps and label the process key points; at the same time, establish an expert knowledge entry interface to support process engineers to contribute implicit experience in a form-based manner; S22. Construct a multi-dimensional associated knowledge graph in the field of aerospace intelligent manufacturing, organize the knowledge network with the product family tree, process flow, and fault tree as the core axes; specifically, define a three-layer ontology model of "part - material - process", associate part design specifications with the material property library, and map them to machining parameter knowledge. At the same time, construct a process constraint rule library, describe the process dependency relationship through logical expressions, and link it to the inspection standard; in addition, integrate the historical fault case library, establish a causal chain of "abnormal phenomenon - root cause - solution measure" to form a traceable diagnostic knowledge network; A hybrid architecture of the graph database Neo4j and the vector database Milvus is adopted to meet the efficient storage and application of complex knowledge networks; the graph database Neo4j stores explicit structured relationships, while the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval; finally, a knowledge evidence base is established to associate each knowledge node with the original basis, ensuring two-way links between knowledge nodes and source evidence; S23. To ensure the authority and reliability of the knowledge base throughout its life cycle, this application constructs a knowledge quality control system that runs through the entire process; a dual knowledge verification mechanism is established: automatically discover and report conflicting rules through logical consistency checks, trigger the manual review process, and verify the 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 basis version of each process rule can be traced during airworthiness certification; at the access and usage level, adopt attribute-based access control permission management to dynamically control the knowledge access scope according to personnel roles and project confidentiality levels, and set up an approval unlocking process for core process parameters to ensure the security of knowledge assets; S24. Establish a dynamic update mechanism to continuously promote the optimization of the knowledge base; set the knowledge freshness monitoring index KFI, and regularly synchronize international standard updates to ensure the forefront and compliance of knowledge content; In this application, KFI is obtained by weighting four core factors, specifically: ; Among them, 、 、 、 are the weights of each factor, which are dynamically adjusted according to knowledge types (such as safety specifications, process parameters, fault plans, etc.); TDF is the time decay factor, which is used to measure the risk that knowledge naturally ages and may become obsolete over time. The formula is expressed as: ; Among them, T is the knowledge validity period, is the current time, is the time when the knowledge item was last verified and updated; the longer the knowledge has not been updated, the lower the TDF score; VAF is the version alignment factor, which is obtained by comparing with an external standard database and is used to ensure that the knowledge is consistent with the versions of external authoritative standards (such as international / industry standards, design specification manuals). In this embodiment, it is specifically set in the form of a discrete factor: 1.0 (latest): the standard version on which the knowledge is based is the current latest and valid version; 0.5 (update available): there is an updated version of the standard on which the knowledge is based, but the old version has not been repealed; 0.0 (repealed): the standard version on which the knowledge is based has been explicitly repealed or replaced; N / A (not applicable): the knowledge is not associated with any external standard; UAF is the application activity factor, which indirectly evaluates its effectiveness through the actual application frequency of the knowledge. Its formula is expressed as: ; where, represents the frequency of a certain piece of knowledge being called by the large model, retrieved and adopted in knowledge Q&A within the past cycle; represents the average frequency of all knowledge being called by the large model, retrieved and adopted in knowledge Q&A within the past cycle; PVF is the performance verification factor, which is used to measure the performance of knowledge in guiding actual production or solving problems and is the golden standard for testing the effectiveness of knowledge. In this embodiment, it can be specifically designed as follows according to different knowledge categories: For process parameter knowledge: regularly apply this parameter in the digital twin environment or small-batch trial production, PVF = the yield rate of the product or the achievement rate of key performance indicators; For fault diagnosis knowledge: when a new fault case appears, use this knowledge for reasoning, PVF = the accuracy rate of diagnosis or the success rate of recommended solutions; For rule constraint knowledge: analyze the latest production logs, PVF = 1 - (the number of times the rule is violated / the total number of production batches); After that, machine learning is used to analyze the newly generated production data to automatically extract valuable potential knowledge patterns, and the extracted knowledge patterns need to be reviewed and confirmed by domain experts before being stored in the database, forming a verification closed-loop from production practice to expert knowledge. In addition, an online collaboration platform is built to collect the tacit experience of front-line personnel, and an online collaboration platform that supports the participation of multiple teams such as design, manufacturing, and operation and maintenance is built, allowing all parties to initiate knowledge revision proposals, and these proposals will go through a rigorous cross-departmental review process to ensure that the finally updated knowledge has both authority and on-site applicability. Thus, the construction of the aerospace intelligent manufacturing professional knowledge base is completed.

[0022] S3. Align cross-modal manufacturing data to generate an aerospace intelligent manufacturing domain corpus; perform pre-training in combination with the base general large model, inject professional term semantics and multi-modal association capabilities through the masked language modeling task and contrastive learning, complete the knowledge transfer of the base model in the aerospace field, and obtain the pre-trained aerospace intelligent manufacturing large model; As a preferred embodiment, step S3 specifically includes: S31. Construct a corpus in the field of aerospace intelligent manufacturing with cross-modal and semantically precise alignment: Extract diverse information covering structured, semi-structured, and unstructured types from the aerospace intelligent manufacturing database; Subsequently, for cross-modal understanding, transform and align heterogeneous data. On the one hand, convert some structured data (such as work order execution records, minute-level time series data collected by sensors) into natural language descriptions through automated templates. On the other hand, finely associate different types of data pointing to the same manufacturing entity, such as unstructured information (inspection images), structured data (sensor parameters), and semi-structured data (process logs, quality inspection reports), according to timestamps to construct cross-modal association samples in the form of "signal-image-parameter-text" to train the model to understand the internal relationships between multi-source data; Finally, use the constructed professional knowledge base of aerospace intelligent manufacturing 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 general terms with industry-standard terms, and inserting knowledge triples as context supplements to improve the professional semantic accuracy and logical depth of the corpus, enabling the model to accurately learn the precise corresponding relationships between professional terms and logical constraints S32. Conduct domain adaptation pre-training based on the constructed aerospace intelligent manufacturing field corpus and the base general large model to inject basic knowledge in the field of aerospace intelligent manufacturing into the base general large model; Use domain-general corpus (such as aerospace journal papers, technical manuals) and some multi-modal alignment samples for mixed training, adopt the masked language modeling MLM task, focusing on masking domain entities, forcing the model to rely on knowledge clues in the context for prediction; At the same time, introduce a contrastive learning task, given different modal descriptions of the same process, and require the model to judge whether they match to enhance cross-modal alignment ability.

[0023] S4. Based on the pre-trained aerospace intelligent manufacturing large model, construct an Aerospace Manufacturing Cognitive Agent (AMCA) to form a complete, dynamic, and self-improving complex engineering problem-solving framework; As a preferred embodiment, as Figure 2 shown, step S4 specifically includes: S41. After receiving a user query, AMCA first initiates a depth-first traversal of the knowledge graph, following the "retrieval - reasoning - pruning" mechanism to extract knowledge fragments. Specifically: Using vector retrieval technology, it matches the user query with the entity nodes in the knowledge graph to quickly locate the initial node most relevant to the query semantics 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 each node's relationship path is conducive to problem-solving. This reasoning guided by such judgments enables the traversal process to cross the boundaries of multiple knowledge documents, thereby concatenating scattered but logically strongly related information to form a complete evidence chain. During the exploration process, it parallelly tracks multiple reasoning paths, continuously evaluates the information gain of each path, and prunes the paths leading to irrelevant information and logical deadlocks. Finally, it summarizes, deduplicates, and logically reorganizes all high-value knowledge fragments on the reasoning paths to form a comprehensive knowledge context. S42. After obtaining the comprehensive knowledge context, AMCA enters the cognitive stage and adopts a thought tree exploration framework to transform the vague user intention into a clear, executable, and structured solution. Specifically: First, it decomposes the user's high-level goal into a series of thinking nodes, generates multiple different but feasible next actions and hypotheses for each thinking node. Then, it initiates a self-evaluation mechanism, uses the comprehensive knowledge context and its own domain knowledge to evaluate the value of each newly generated path branch (such as evaluating whether the path has data support, whether it conflicts with known engineering principles or safety constraints, and the probability of the path leading to a final definite conclusion), calculates the evaluation score, and dynamically searches and prunes the entire thought tree. Such a design will make AMCA preferentially explore the paths with the highest evaluation value in depth, and at the same time reserve some computing resources to explore some seemingly less optimal but potentially possible new paths. And when a path is proven to be a dead end or of extremely low value, AMCA can perform a "backtracking" operation, retreat to the previous decision node, and instead invest resources in other more promising paths. Finally, it constructs a complete path starting from the initial problem, passing through all necessary analysis steps, and having a high-confidence conclusion, takes this path as the final execution plan, and outputs it in a structured form (such as a step list, a dependency graph) to provide clear guidance for the actions in the next stage. 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 as follows: First, the constructed database, knowledge base, and professional small models are uniformly encapsulated and abstracted into an expert toolset; the professional small models are trained differently for different tasks; after parsing the execution plan, the scheduler agent delegates well-defined and domain-specific subtasks to the most matching expert agents in the expert toolset. 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 there are subtasks that no expert agent and existing tools can directly handle during the process of the scheduler agent delegating tasks, the tool synthesis module is activated to generate new tools that can completely handle the subtasks that cannot be directly processed by instantaneously writing Python functions. After sandbox verification, the output is integrated into the automated workflow; Specifically, within the tool synthesis module, the scheduler agent utilizes its powerful reasoning and code generation capabilities to instantaneously write an independent Python function that can complete the specific task. The function will be designed with a single responsibility and clear interfaces, and automatically generate a docstring containing input parameters, output formats, and function descriptions for subsequent calls and understanding; the newly generated code will not be directly executed in the main production environment. The system will deploy it to an isolated and resource-constrained "sandbox" environment. At the same time, the scheduler agent will also automatically generate a set of unit test cases according to the problem context and run these tests in the sandbox to verify the logical correctness, computational accuracy, and running security of the new tool, ensuring that it has no malicious behavior or serious defects; only the tools that have passed 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 toolset during the current task session. If the same calculation or processing requirements appear again in subsequent plan steps, the scheduler can directly reuse this new tool without having to regenerate it; S44. Then AMCA enters the reflection and verification phase to conduct a review and correction cycle of the solution, specifically as follows: First, the scheduler agent summarizes and synthesizes the execution results of each expert agent and dynamic tool to form a preliminary solution; then AMCA first conducts an internal critique of the preliminary solution to generate reflection text and raise relevant questions, and then switches to the role of an evaluator to externally verify the solution using the expert toolset and knowledge graph, including logical consistency verification, compliance review, and comparison with historical cases; In this embodiment, during logical consistency verification, AMCA actively queries the knowledge graph to verify whether its suggestions are contrary to known causal relationships or physical constraints; during compliance review, AMCA calls a dedicated compliance check tool to compare its solutions with industry standards, enterprise internal process specifications, and relevant safety regulations stored in the database to ensure full compliance; during historical case comparison, AMCA retrieves similar scenarios in the historical case database, analyzes the success and failure rates after taking similar measures in history, and provides data-driven risk assessment for the current decision-making; If problems occur in internal and external reviews, record and save relevant critical opinions and verification results, and initiate the correction procedure: for minor problems, only fine-tune the suggested parameters; for major defects, trigger a complete replanning, bring the newly discovered constraints back to the thinking tree exploration stage, regenerate a better strategic execution plan, thus planning a better plan than the preliminary plan, and perform an iterative feedback loop; If all internal and external reviews are successfully passed, mark the solution as verified and present it to the user as the final, highly reliable output solution.

[0024] S5. Through a two-stage progressive multi-task training strategy, perform supervised fine-tuning and reinforcement learning alignment on the large model, optimize the professional capabilities of the large model in the field of aerospace intelligent manufacturing, and combine online learning and incremental update mechanisms to achieve the dynamic adaptation of the aerospace intelligent manufacturing large model to the actual production environment requirements; finally complete the construction of the aerospace intelligent manufacturing large model; As a preferred implementation method, step S5 specifically includes: S51. Based on the established corpus in the field of aerospace intelligent manufacturing, adopt a two-stage progressive multi-task training strategy to train the aerospace intelligent manufacturing large model after pre-training and constructing the aerospace manufacturing cognitive intelligent agent AMCA, and strengthen the adaptability of the model to the domain data distribution and knowledge logic; More specifically, step S51 specifically includes: S511. In the first stage, multi-task joint supervised fine-tuning is carried out to comprehensively improve the model's various professional skills in the aerospace manufacturing field: multiple heterogeneous downstream tasks are optimized in parallel, specifically including process parameter prediction (inputting sensor data of the first N time steps to predict key parameters in the next M steps, evaluated using the mean squared error loss function), text-image matching (judging the correspondence between process descriptions and X-ray images), quality inspection report generation (given inspection images and equipment status data, generating quality inspection conclusion texts that comply with the AS9100 standard, evaluated using the cross-entropy loss function), knowledge Q&A (answering complex questions based on a process knowledge base, using a hybrid loss function of span prediction and generation), and root cause inference of anomalies (inputting descriptions of abnormal phenomena and outputting possible fault chains, using a graph structure matching loss); Meanwhile, the GradNorm algorithm is adopted to achieve adaptive allocation of loss weights, dynamically monitor the gradient norms of each task and the degree of decline of the loss function relative to the initial state, and automatically adjust their corresponding loss weights; specifically: for tasks with too fast learning and too large gradient norms, their weights are correspondingly reduced; conversely, the weights of tasks with slower learning are increased, thereby dynamically balancing the gradient norms of each task; ensuring that all professional skills can be fully and evenly optimized, making the finally fine-tuned model more comprehensive and reliable; In this example, through the gradient loss function Calculate the gradient norm of each task at the current moment t And its expected target gradient norm The L1 distance (absolute difference) between them is used to guide the update of the loss weights of each task; the formula of the gradient loss function is expressed as: ; Relative reverse training rate Is used to measure the learning progress of each task, and its formula is expressed as: ; Among them, Is the initial loss of task i at training step t, The closer the value of Is to 1, it indicates that the loss of this task decreases more slowly, that is, the learning progress is slower; If task i has a slow learning progress ( Is larger), its target gradient norm will be set higher to prompt it to obtain a greater learning intensity in subsequent training; conversely, if the task has a fast learning progress ( The gradient norms of each task Calculate the weighted loss of the i-th task with respect to the shared layer parameters The L2 norm of the gradient (i.e., the length of the vector), which is used to measure the learning intensity of each task at the current step; the larger this value, the greater the impact of the task on the update of the model's shared parameters; The formula is expressed as: ; Among them, the average gradient norm calculates the arithmetic mean of the learning intensities of all tasks at the current step.

[0025] S512. In the second stage, the model is deeply aligned through reinforcement learning to make its decision suggestions more in line with complex real-world requirements: First, a reward model based on knowledge graphs and domain rules is trained to score the decision suggestions generated by the large model from multiple dimensions including technical feasibility, compliance, and economy, and through weighted summation, the multi-dimensional scores are integrated into a single scalar reward signal; at the same time, 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, forcing the model to quickly converge to the compliant solution space during training and forcing alignment with domain knowledge; then an artificial feedback loop is introduced to further improve the alignment accuracy by having domain experts regularly score the sampling results; finally, the system will perform weighted fusion of the scaled scalar reward signal and the high-quality artificial scores to form the final reward signal, and in combination with the proximal policy optimization PPO algorithm, fine-tune the large model.

[0026] In this embodiment, the formula for the scalar reward signal is as follows: For a given decision suggestion, the reward model outputs a vector composed of scores in multiple dimensions. Since the importance of different dimensions in the final decision is different, the weights of each dimension are set in advance by domain experts and project managers. By multiplying the score of each dimension by its corresponding weight and then summing all the results, a single, comprehensive scalar reward signal is obtained , and the formula is expressed as: ; Among them, represents the compliance weight, represents the compliance score, represents the technical feasibility weight, represents the technical feasibility score, represents the economy weight, represents the economy score; The workflow of the logical constraint reward function is as follows: The decisions generated by the large model are first fed into the logical constraint reward function. The parser inside the function extracts the key information related to the rules, checks them one by one, and conducts a strict logical comparison with each rule in the "immutable rule library" (which contains the hard constraint rules defined and stored in the corpus of the aerospace intelligent manufacturing field). According to the verification results, the function outputs a reward value: As long as any rule is found to be violated, the function immediately stops checking and returns -100 as a penalty signal; only when the decision recommendation completely passes the verification of all immutable rules will the function return a neutral value of 0.

[0027] S52. Establish a multi-level evaluation and iteration mechanism to evaluate the performance of the finally constructed aerospace intelligent manufacturing large model and perform continuous dynamic updates; More specifically, step S52 specifically includes: S521. First, construct a high-quality offline test set. Select the fault cases and their associated multi-modal data in the latest 6 months from the database as the test set, strictly isolate the time window of the training data, and simulate the unknown problems encountered after the model goes online; at the same time, inject artificially constructed adversarial samples to comprehensively test the anti-interference ability and logical consistency of the model; S522. Construct a multi-dimensional evaluation index system. Specifically: at the general text generation level, adopt the general generation indexes BLEU and ROUGE to compare with the reference text to evaluate the fluency and information fidelity of the content generated by the model; secondly, to assess its professionalism, define key indexes specific to the field, including the process parameter deviation (used to quantify the gap between the parameter values predicted by the model and the engineering standards or optimal values), and the knowledge compliance rate (used to calculate the proportion of the decision recommendations generated by the model that conform to the established rules in the knowledge base); In this embodiment, the mean absolute percentage error MAPE is calculated as the process parameter deviation, and the formula is expressed as: ; The specific evaluation process of the knowledge compliance rate is as follows: Design a test set containing questions, specifically used to stimulate the model to answer in scenarios where specific rules need to be followed; let the model answer these questions one by one, and then judge whether each answer violates the relevant rules in the knowledge base item by item, and count the number of answers that do not violate any rules , so as to calculate the compliance rate: ; Finally, to measure the real business value brought by the model, the aerospace intelligent manufacturing large model is deployed in the digital twin environment, and the actual improvement effect of the decision-making suggestions given by the model on the core indicators including the good product rate and production efficiency is evaluated. That is, the process parameters or scheduling plans generated by the model are used as inputs and simulated on the virtual production line, and the real impact on production benefits is quantified by analyzing the data output from the simulation, so as to realize the closed-loop, quantitative, and risk-free evaluation of the application value of the model. S523. Establish an online learning and incremental update mechanism; use the decision-making suggestions given by the large model in the production environment and the results after actual execution as feedback data to flow back, and construct a feedback closed-loop dataset; successful execution cases are used as positive samples to strengthen the correct behavior of the model, while failed cases trigger analysis and are used as negative samples to correct the model, so as to achieve continuous learning based on actual feedback; adopt the parameter-efficient fine-tuning PEFT technology for LoRA low-rank adaptation, and only update a small number of parameters. Without affecting the general capabilities, quickly realize the monthly incremental update of the model to ensure that the model continuously adapts to the changes in the production environment without interrupting the service.

[0028] In summary, the construction and training of the aerospace intelligent manufacturing large model based on the base general large model have been realized. Through key technologies such as constructing a high-quality multi-modal data ecosystem, domain architecture transformation, phased multi-task joint training, integrating retrieval enhancement and function call, closed-loop evaluation and continuous iteration, the general large model is transformed into a professional intelligent agent that deeply understands the industry logic and accurately responds to manufacturing needs. This method effectively solves the challenges faced by traditional large models in the aerospace field, such as multi-modal data processing, in-depth understanding of domain knowledge, satisfaction of logical constraints, and adaptation to dynamic environments, and provides a feasible and high-performance technical path for the digital and intelligent upgrade of aerospace intelligent manufacturing.

[0029] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0030] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments 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, Specifically, it includes the following steps: S1. Build a professional data collection and management system. Collect a large amount of original real data of aerospace intelligent manufacturing from the actual scenarios of aerospace intelligent manufacturing, and perform preprocessing and data association to construct an aerospace intelligent manufacturing database; S2. For multi-modal knowledge sources in the field of aerospace intelligent manufacturing, establish a knowledge collection and structured transformation pipeline, a multi-dimensional associated domain knowledge graph, a knowledge quality control system and a dynamic update mechanism, and construct 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; Perform pre-training in combination with the base general large model. Inject the semantics of professional terms and multi-modal association capabilities through the masked language modeling task and contrastive learning, complete the knowledge transfer of the base model in the aerospace field, and obtain the pre-trained large model of aerospace intelligent manufacturing; S4. Based on the pre-trained large model of aerospace intelligent manufacturing, construct an aerospace manufacturing cognitive agent AMCA to form a complete, dynamic and self-improving complex engineering problem-solving framework; S5. Through a two-stage progressive multi-task training strategy, perform supervised fine-tuning and reinforcement learning alignment on the large model, optimize the professional capabilities of the large model in the field of aerospace intelligent manufacturing, and combine the online learning and incremental update mechanism to realize the dynamic adaptation of the large model of aerospace intelligent manufacturing to the requirements of the actual production environment; finally complete the construction of the large model of aerospace intelligent manufacturing.

2. The method for constructing a large model for aerospace intelligent manufacturing according to claim 1, characterized in that Step S1 specifically includes: S11. Establish a data aggregation channel. Through multi-source data collection and integration, obtain a large amount of real aerospace manufacturing data; for the scattered design, production, and inspection data sources in each type of real aerospace manufacturing scenario, define a unified data identification system, and adopt a heterogeneous protocol compatibility strategy to uniformly aggregate each data source into the original database; S12. Design a hierarchical and classified data storage architecture, specifically including: first, uniformly store all the collected original data in a distributed database to build a data lake to retain all the original information; then, according to the data characteristics and application scenarios, classify and import the data into different databases: for the key business data in the structured data, perform preprocessing including outlier filtering, unit standardization, and missing value imputation through a quality rule engine, and then import it into a columnar data warehouse; for the high-frequency device data with real-time changes in the structured data, store it in a time series database; for semi-structured data, store its metadata and association relationships in a document database; for unstructured data, manage the file entities through object storage; S13. Establish a complete data governance system, specifically including: define a unified naming specification, build a global data catalog relying on the metadata management platform, and clearly define the business meaning, data lineage, and access permission ownership of each data entity; on this basis, implement automated data quality detection based on a dynamic rule engine to identify outliers and perform format verification on the data; finally, adopt a hierarchical encryption strategy to protect the data security and protect sensitive information, so as to form a closed-loop data governance solution covering the whole process. S14. Build an association network between data to support the training and analysis of large models; transform discrete data into an inferable knowledge network business entity network through an entity relationship model to complete the construction of an aerospace intelligent manufacturing database.

3. A method for constructing an aerospace intelligent manufacturing large model according to claim 1, characterized in that, Step S2 specifically includes: S21. Build a multi-modal knowledge acquisition and structured transformation pipeline for aerospace intelligent manufacturing, specifically: directly collect structured documents in knowledge data, for semi-structured text and unstructured data, parse key knowledge elements through natural language processing technology and transform them into structured triple forms; then for video-like unstructured data, use computer vision analysis technology to extract key frame sequences of the operation steps process and label the process key points; at the same time, establish an expert knowledge entry interface to support process engineers to contribute implicit experience in a form-based manner; S22. Build a knowledge graph in the field of aerospace intelligent manufacturing with multi-dimensional associations, organize the knowledge network with the product family tree, process flow, and fault tree as the core axes; specifically, define a three-layer ontology model of "part - material - process", associate part design specifications with the material property library, and map them to machining parameter knowledge, and at the same time build a process constraint rule library, describe the process dependency relationship through logical expressions, and link to inspection standards; in addition, integrate the historical fault case library, establish a causal chain of "abnormal phenomenon - root cause - solution measure" to form a traceable diagnostic knowledge network; And adopt a hybrid architecture of graph database Neo4j and vector database Milvus to meet the efficient storage and application of complex knowledge networks; the graph database Neo4j stores explicit structured relationships, and the vector database Milvus generates embedded semantic vectors through deep learning models to support similarity retrieval; finally, establish a knowledge evidence base to associate each knowledge node with the original basis to ensure the two-way link between the knowledge node and the source evidence; S23. Build a knowledge quality control system throughout the whole process; establish a dual knowledge verification mechanism: automatically discover and report conflict rules through logical consistency checks, trigger the manual review process, and verify the knowledge validity based on historical data backtracking; establish a version control system to record the revision history and effective time of knowledge items to ensure that the basis 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 knowledge access scope according to 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 promote the optimization of the knowledge base; set a knowledge freshness monitoring index KFI and regularly synchronize international standard updates; automatically extract valuable potential knowledge patterns through machine learning analysis of newly generated production data, and the extracted knowledge patterns need to be reviewed and confirmed by domain experts before being put into the database; in addition, build an online collaboration platform to collect the implicit experience of front-line personnel to complete the construction of an aerospace intelligent manufacturing professional knowledge base.

4. A method for constructing an aerospace intelligent manufacturing large model according to claim 2, characterized in that KFI is obtained by weighting four core factors, specifically: ； Among them, , , , are the weights of each factor and are dynamically adjusted according to the knowledge type; TDF is the time decay factor, and the formula is expressed as: ； where T is the knowledge validity period, is the current time, is the time when the knowledge item was last verified and updated; The VAF is the version alignment factor, which is obtained by comparing with an external standard database; The UAF is the application activity factor, and its formula is expressed as: ； Among them, represents the frequency of a certain piece of knowledge being called by the large model and retrieved and adopted in knowledge Q&A within the past cycle; represents the average frequency of all knowledge being called by the large model and retrieved and adopted in knowledge Q&A within the past cycle; The PVF is the performance verification factor, which is used to verify the validity of knowledge.

5. A method for constructing an aerospace intelligent manufacturing large model according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Construct a corpus in the field of aerospace intelligent manufacturing with cross-modal and semantically precise alignment: Extract diverse information covering structured, semi-structured, and unstructured types from the aerospace intelligent manufacturing database; subsequently, perform transformation and alignment on heterogeneous data. On the one hand, convert some structured data into natural language descriptions through automated templates, and on the other hand, finely correlate different types of data pointing to the same manufacturing entity according to timestamps to construct cross-modal association samples in the form of "signal-image-parameter-text" to train the model to understand the internal relationships between multi-source data; finally, use the constructed professional knowledge base of aerospace intelligent manufacturing to enhance the domain knowledge of the text content in the corpus. S32. Perform domain adaptation pre-training based on the constructed corpus in the field of aerospace intelligent manufacturing and the base general large model, and inject basic knowledge in the field of aerospace intelligent manufacturing into the base general large model; use domain general corpus and some multi-modal alignment samples for mixed training, adopt the masked language modeling MLM task, focusing on masking domain entities, forcing the model to rely on knowledge clues in the context for prediction; at the same time, introduce a contrastive learning task, given different modal descriptions of the same process, and require the model to judge whether they match to enhance the cross-modal alignment ability.

6. The method for constructing an aerospace intelligent manufacturing large model 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 - reasoning - pruning" mechanism to refine knowledge fragments; specifically: use vector retrieval technology to match the user query with the entity nodes in the knowledge graph, quickly locate the initial node most relevant to the query semantics as the seed node for graph traversal; then starting from the seed node, driven by the reasoning ability of the large model, perform multi-hop logical exploration, judge whether the exploration along each node's relationship path is conducive to solving the problem, so as to concatenate dispersed but logically strongly related information to form a complete evidence chain; during the exploration process, parallelly track multiple reasoning paths, continuously evaluate the information gain of each path, and prune the paths leading to irrelevant information and logical dead loops; finally, summarize, deduplicate, and logically reorganize 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. Using the thinking tree exploration framework, it transforms the vague user intention into a clear, executable, and structured solution. Specifically: First, it decomposes the user's high-level goal into a series of thinking nodes, and generates multiple different but feasible next actions and hypotheses for each thinking node. Then, it activates the self-evaluation mechanism, uses the comprehensive knowledge context and its own domain knowledge to evaluate the value of each newly generated path branch, calculates the evaluation score, and dynamically searches and prunes the entire thinking tree. Finally, it constructs a complete path starting from the initial problem, passing through all necessary analysis steps, and having a high-confidence conclusion, takes this path as the final execution plan, and outputs it in a structured form. S43. Enter the action stage. Based on the collaborative scheduling-instant creation mode, implement each step in the execution plan and generate candidate solutions. Specifically: First, uniformly package and abstract the constructed database, knowledge base, and professional small models into an expert toolset. After parsing the execution plan, the scheduler agent delegates the well-defined and domain-specific subtasks to the most matching expert agent in the expert toolset. 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 there are subtasks that no expert agent and existing tools can directly handle during the process of the scheduler agent delegating tasks, the tool synthesis module is activated to generate new tools by instantaneously writing Python functions for the subtasks that cannot be directly handled at all, and after sandbox verification, integrate the output 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 synthesizes the execution results of each expert agent and dynamic tool to form a preliminary solution. Then AMCA first conducts internal criticism on the preliminary solution, generates reflection text, and raises relevant questions. Then it switches to the evaluator role and uses the expert toolset and knowledge graph to conduct external verification of the solution, including logical consistency verification, compliance review, and comparison with historical cases. If problems occur in the internal and external reviews, record and save the relevant critical opinions and verification results, start the correction program to optimize the solution, and conduct an iterative feedback cycle. If it successfully passes all internal and external reviews, mark the solution as verified and present it to the user as the final, high-confidence output solution.

7. A method for constructing an aerospace intelligent manufacturing large model according to claim 1, characterized in that Step S5 specifically includes: S51. Based on the constructed corpus in the field of aerospace intelligent manufacturing, adopt a two-stage progressive multi-task training strategy to train the aerospace intelligent manufacturing large model after pre-training and constructing the aerospace manufacturing cognitive intelligent agent AMCA, and strengthen the adaptability of the model to the domain data distribution and knowledge logic. S52. Establish a multi-level evaluation and iteration mechanism to evaluate the performance of the finally constructed aerospace intelligent manufacturing large model and perform continuous dynamic updates.

8. A method for constructing an aerospace intelligent manufacturing large model according to claim 7, characterized in that, Step S51 specifically includes: S511. In the first stage, perform multi-task joint supervised fine-tuning to comprehensively improve the model's various professional skills in the aerospace manufacturing field: optimize multiple heterogeneous downstream tasks in parallel, specifically including process parameter prediction, text-image matching, quality inspection report generation, knowledge Q&A, and root cause reasoning of anomalies; use the GradNorm algorithm to achieve adaptive allocation of loss weights, dynamically monitor the gradient norms of each task and the degree of decline of the loss function relative to the initial state, and automatically adjust their corresponding loss weights; S512. In the second stage, deeply align the model through reinforcement learning to make its decision-making suggestions more in line with complex real-world requirements: first, train a reward model based on knowledge graphs and domain rules to score the decision-making suggestions generated by the large model from multiple dimensions of technical feasibility, compliance, and economy, and integrate the multi-dimensional scores into a single scalar reward signal through weighted summation; at the same time, set a logical constraint reward function to ensure that the model does not violate predefined immutable rules, forcing the model to quickly converge to the compliance solution space during training and forcefully aligning with domain knowledge; then introduce an artificial feedback loop to further improve the alignment accuracy by having domain experts regularly score the sampled results; finally, the system will perform weighted fusion of the large-scale scalar reward signal and the high-quality artificial scores to form the final reward signal, and use this to fine-tune the large model in combination with the proximal policy optimization (PPO) algorithm.

9. A method for constructing an aerospace intelligent manufacturing large model according to claim 7, characterized in that, Step S52 specifically includes: S521. First, construct a high-quality offline test set. Select the latest six months of failure cases and their associated multi-modal data from the database as the test set, strictly isolate the time window of the training data, and simulate unknown problems that the model may encounter after going online; at the same time, inject artificially constructed adversarial samples 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 metrics BLEU and ROUGE to compare with the reference text to evaluate the fluency and information fidelity of the content generated by the model; 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 the aerospace intelligent manufacturing large model in a digital twin environment and evaluate the actual improvement effect of the decision-making suggestions given by the model on core indicators including the good product rate and production efficiency; S523. Establish an online learning and incremental update mechanism; use the decision-making suggestions given by the large model in the production environment and the results after actual execution as feedback data to flow back, and build a feedback closed-loop dataset; use successful execution cases as positive samples to strengthen the correct behavior of the model, and failure cases trigger analysis and are used as negative samples to correct the model, realizing continuous learning based on actual feedback; adopt the parameter-efficient fine-tuning PEFT technology for LoRA low-rank adaptation, only update a small number of parameters, and quickly achieve monthly incremental updates of the model without affecting the general capabilities, ensuring that the model continuously adapts to changes in the production environment without interrupting the service.

Images