An agent-based superalloy casting process design method and system

Abstract

The application discloses a high-temperature alloy casting process design method and system based on an agent, and the method comprises the following steps: structurally processing multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base; extracting and modeling entities and relationships in the multi-source process data base to obtain a process knowledge graph; inputting the multi-source process data base obtained through structural processing into a large language model based on a retrieval enhancement generation method, and combining a prompt word template and a thinking chain reasoning mechanism to verify and constrain the output of the large language model to obtain a process reasoning module; integrating the process knowledge graph and the process reasoning module with a professional tool set to establish an agent arrangement module to analyze user requirements and schedule tasks, and obtaining a casting high-temperature alloy process design agent; and outputting a final process scheme. The application improves the traceability and explainability of the designed process scheme and defect countermeasures.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided process planning technology, and in particular relates to a method and system for process design of high-temperature alloy castings based on intelligent agents. Background Technology

[0002] Cast high-temperature alloys are core materials for hot-end components of aero-engines and gas turbines. The level of their manufacturing process design directly determines the forming quality and service reliability of key hot-end components such as turbine blades, casings, and guide vanes. Currently, the manufacturing process design of cast high-temperature alloys still relies mainly on a combination of experience-based decisions and trial-and-error verification by process engineers. The selection of alloy grades for new castings, the formulation of process schemes, and the analysis of the causes of metallurgical defects usually require engineers to make judgments based on their long-accumulated professional knowledge and undergo several rounds of trial production verification. With the rapid improvement of the performance indicators of aero-engines and gas turbine equipment and the continuous acceleration of the development pace, the industry's demand for the evolution of manufacturing process design towards intelligence is becoming increasingly urgent.

[0003] In existing technologies, research has attempted to introduce artificial intelligence into the field of casting process design, mainly including defect prediction methods based on single deep learning models and conversational process assistants based on general large language models. However, due to the lack of deep, structured constraints on knowledge in the field of casting high-temperature alloys, the output results of these existing methods lack traceable knowledge basis when facing highly specialized problems such as alloy grade recommendation, process parameter determination, and defect cause diagnosis. This makes it impossible to effectively verify the reliability of the generated content, resulting in a high risk of misleading interpretations and making it difficult to directly apply to engineering practice. Summary of the Invention

[0004] The purpose of this invention is to provide a process design method and system for high-temperature alloy casting based on intelligent agents, in order to solve the problem that in the existing process design of casting high-temperature alloys, the model output results lack traceable knowledge basis, and the reliability of the generated content cannot be effectively verified, thus making it difficult to directly apply the generated results to engineering practice.

[0005] The technical solution of the present invention is as follows: In a first aspect, the present invention provides a process design method for high-temperature alloy castings based on intelligent agents, comprising:

[0006] S1. Structure the multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base.

[0007] S2. Extract and model the entities and relationships in the multi-source process data base to obtain a process knowledge graph;

[0008] S3. The multi-source process data base obtained by structured processing is connected to the large language model based on the retrieval enhancement generation method, and the output of the large language model is verified and constrained by the prompt word template and the thinking chain reasoning mechanism to obtain a process reasoning module with illusion suppression capability.

[0009] S4. Integrate the process knowledge graph and the process reasoning module with the professional toolset, and establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design.

[0010] S5. Use the intelligent agent for high-temperature alloy casting process design to design the high-temperature alloy casting process and obtain the final process scheme.

[0011] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S1 includes:

[0012] S11. The alloy composition data, physical property parameters, microstructure data, performance data and process window data of the cast high-temperature alloy are standardized and organized with the alloy grade as the index unit to obtain a basic materials science database.

[0013] S12. Classify and store the gating system layout data, mold shell process data, cooling process data, heat treatment system data and process operation data of historical castings according to product type, alloy grade, typical wall thickness and structural feature labels to obtain a process plan and process specification database.

[0014] S13. The descriptions of historical casting quality problems, their causes and solutions are structured and archived in the form of a problem, cause and countermeasure triplet to obtain a quality case database.

[0015] S14. After preprocessing the materials science database, the process scheme and process specification database, and the quality case database, they are uniformly stored in a hybrid storage database composed of vector database, relational database and index database to obtain a multi-source process data base; wherein, the preprocessing method includes at least cleaning processing and vectorization processing.

[0016] Optionally, in the above-described agent-based high-temperature alloy casting process design method,

[0017] In S11, five types of data, namely alloy composition data, physical property parameters, microstructure data, performance data and process window data, are uniformly classified into the material table of the relational database according to the alloy grade as the index unit. Each row of records corresponds to a complete physical property profile of an alloy grade, thus obtaining the basic database of materials science.

[0018] In S12, four types of data, namely, gating system layout data, mold shell process data, cooling process data, heat treatment system data and process operation data, are used to establish a composite index according to four levels of tags: product type, alloy grade, typical wall thickness and structural characteristics. The index is stored in the process table of the relational database. Each row of records corresponds to a complete process plan or operation tip entry, thus obtaining the process plan and process specification database.

[0019] In S13, each case of a historical casting quality problem is constructed into a data record in the form of a problem-cause-countermeasure triplet. Each field in the triplet data is uniquely identified by the case number and stored in the case table of the relational database. Each row of records corresponds to a complete quality handling case, thus obtaining the quality case database.

[0020] In step S14, the structured fields of the materials science database, process scheme and process specification database, and quality case database are stored in a relational database, the text vectors are stored in a vector database, and the metadata tags such as alloy grade, product type, process name, and defect type are stored in an index database in the form of key-value pairs. The three types of storage are associated through a unified data ID, thus obtaining a multi-source process data base.

[0021] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S2 includes:

[0022] S21. Taking product nodes, material nodes, structural feature nodes, quality requirement nodes and process nodes as the first basic node types, entity relationships between nodes are extracted from the multi-source process data base through entity recognition and ontology modeling technology to construct a product process design knowledge graph.

[0023] S22. Using defect type nodes, possible cause nodes, diagnostic feature nodes, and solution nodes as the second basic node types, the defect solution reasoning chain is extracted from the quality case database of the multi-source process data base through relation extraction technology, and priority weights are marked for each countermeasure based on historical case data to construct a defect solution knowledge graph.

[0024] S23. Connect the product process design knowledge graph and the defect resolution knowledge graph through unified metadata tags to obtain a process knowledge graph.

[0025] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S21 includes:

[0026] S21-1, scan the text in the materials science basic database and the process scheme and process specification database, identify the entities of the five types of nodes in the first basic node type, and then create nodes after classifying the identified entities according to node type and removing duplicates, and assign a unique node number to each node.

[0027] S21-2, perform relationship determination on entity pairs that co-occur within the same process solution text, extract requirement-satisfaction relationship between product node and material node, extract structure-risk-process compensation relationship between structural feature node and process node, and extract material-process window relationship between material node and process node. Each relationship is stored in the form of a triple, with the triple storage format as: <node A, relationship type, node B>.

[0028] S21-3. Import the extracted nodes and relationships into the graph database, remove duplicates, construct the graph structure, and build the product process design knowledge graph.

[0029] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S22 includes:

[0030] S22-1, parse each triplet record in the quality case database, extract the defect name in the problem field as the defect type node, extract the cause description in the cause field as the possible cause node, extract the countermeasure description in the solution field as the solution node, and extract the detection features implicit in the problem field and cause field as the diagnostic feature node.

[0031] S22-2, within the same case, establish directed edges for the four types of nodes of the second basic node type in the order of defect type node → possible cause node → diagnostic feature node → solution node, and statistically analyze the recurrence rate and first-pass rate improvement after each countermeasure is adopted in the historical case database. Store the statistical results as the priority attribute value of the solution node. The higher the value, the higher the priority of the countermeasure.

[0032] S22-3, import the nodes and directed edges into the graph database to construct a defect resolution knowledge graph.

[0033] Optionally, in the above-described agent-based high-temperature alloy casting process design method, step S23 includes:

[0034] S23-1: Traverse the material nodes and process nodes in the product process design knowledge graph, extract alloy grade attribute values, traverse the possible cause nodes and solution nodes in the defect resolution knowledge graph, extract the associated alloy grade attribute values, and establish relationship edges between the two graphs for nodes with the same alloy grade attribute values.

[0035] S23-2, traverse the product nodes in the product process design knowledge graph, extract the product type attribute values, traverse the defect type nodes in the defect resolution knowledge graph, extract the frequently associated product types, and establish relationship edges between the two graphs for nodes with the same product type attribute values;

[0036] S23-3, traverse the process nodes in the product process design knowledge graph, extract the process name attribute value, traverse the possible cause nodes in the defect resolution knowledge graph, extract the process name of the cause, and establish a relationship edge between the two graphs for nodes with the same process name attribute value.

[0037] S23-4, merge the two interconnected graphs and store them in the same graph database instance to obtain the process knowledge graph.

[0038] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S3 includes:

[0039] S31. Perform domain-customized semantic segmentation and embedding alignment on the multi-source process data base, and establish a multi-source knowledge index based on the retrieval enhancement generation method;

[0040] S32. Recall and reorder the user input based on the multi-source knowledge index to obtain a retrieval result set with source tracing tags;

[0041] S33. After merging the retrieval result set with the serialized subgraph description of the process knowledge graph, input the result set into the large language model to obtain a preliminary generated answer with explicit references.

[0042] S34. Perform multi-level prompt word template constraints and graph anchored thinking chain reasoning on the initially generated answer to obtain the graph anchored reasoning chain;

[0043] S35. Perform multi-dimensional hard constraint self-reflection verification and continuous optimization on the graph anchoring inference chain to obtain a process inference module with illusion suppression capability.

[0044] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S31 includes:

[0045] S31-1. Based on the process procedure boundary, the quality case phenomenon analysis and countermeasure paragraph boundary, and the alloy handbook grade item boundary, semantic block division is performed on the process scheme and process procedure database, quality case database, and materials science basic database in the multi-source process data base to obtain a domain semantic block set.

[0046] S31-2. Use the synonyms in the multi-source process data base to perform lightweight alignment on the embedding model, input the domain semantic block set into the aligned embedding model, and obtain the domain embedding index.

[0047] S31-3. Serialize the subgraphs in the process knowledge graph into natural language descriptions in the form of entity relationships, and merge the natural language descriptions with the domain embedding index into a hybrid storage database to obtain the multi-source knowledge index.

[0048] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S32 includes:

[0049] S32-1, Generate query vectors from the embedded model after aligning the received user input text, perform vector retrieval in the vector database, calculate the similarity between the query vector and all vectors in the index, obtain a preset number of candidate vectors, and query the original text and source identifier corresponding to each candidate vector from the relational database to obtain a semantic candidate data set;

[0050] S32-2, the user input text is segmented into a keyword sequence, the keyword matching algorithm is executed on all text records in the multi-source process data base, the matching score of each text record is calculated, and a preset number of candidate text records are obtained to obtain a keyword matching candidate data set;

[0051] S32-3, extract structured metadata from user input text, identify the specific content of alloy grade, product type and defect type, filter all records that match the identified specific content from the relational database, and obtain a metadata matching candidate data set;

[0052] S32-4: Take the union of the record IDs of the semantic candidate data set, the keyword matching candidate data set, and the metadata matching candidate data set, remove duplicate IDs, and summarize them into a total candidate data set;

[0053] S32-5: Concatenate the user input text with each candidate text in the candidate data set, obtain the relevance score of each concatenated text and sort them in descending order of the score, filter a preset number of candidate data, and add a source identifier to each candidate data to obtain a search result set with source tracing tags.

[0054] Optionally, in the above-described agent-based high-temperature alloy casting process design method, step S33 includes:

[0055] S33-1. Extract the original text and source identifier of each candidate data from the search results set with source tags, and concatenate them into an enhanced contextual text block;

[0056] S33-2. Parse the key entities in the user input and extract the alloy grade and defect type. Query the subgraph containing material nodes and defect type nodes in the process knowledge graph, extract all nodes and edges in the subgraph, and serialize them into a natural language description.

[0057] S33-3. Concatenate the enhanced contextual text block with the natural language description of the graph subgraph to construct a complete prompt word. Input the prompt word into the large language model. The response text generated by the large language model contains citation annotations of the source of the searched information, thus obtaining a preliminary generated answer with explicit citations.

[0058] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S34 includes:

[0059] S34-1. Construct a multi-level prompt word template based on role boundary setting, rejection rules, task input and output field constraints, few sample example guidance, and structured format output requirements. Constrain the preliminary generated answer within the preset expression range of casting high-temperature alloy process design to obtain a preliminary process reasoning result with standardized format.

[0060] S34-2. For the task of recommending process solutions, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the product process design knowledge graph of the process knowledge graph. The graph anchoring thinking chain reasoning is performed according to the fixed step sequence of requirement understanding, key variable identification, alloy grade and casting method determination, structural processability assessment, preliminary gating system scheme, mold shell and cooling process scheme, heat treatment and hot isostatic pressing scheme, risk point identification and comprehensive scheme to obtain the process solution reasoning chain.

[0061] S34-3. For the defect diagnosis task, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the defect solution knowledge graph of the process knowledge graph. The graph anchoring thinking chain reasoning is carried out according to a fixed sequence of steps: phenomenon analysis, key diagnostic feature location, traversal of possible causes along the defect solution knowledge graph, similar case matching, cause priority ranking, countermeasure generation and countermeasure verification condition evaluation. For complex and difficult defect problems, multiple candidate branches are deduced on the same reasoning node, and then summarized and compared. The branch with the most sufficient evidence is selected to obtain the defect diagnosis reasoning chain.

[0062] S34-4. Merge the process scheme reasoning chain with the defect diagnosis reasoning chain to obtain the map anchoring reasoning chain.

[0063] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S35 includes:

[0064] S35-1. Force the reflection process to be triggered for high-risk tasks, and trigger the reflection process for low-risk tasks when the confidence level is lower than the set threshold or when there is a lack of reference materials, and obtain the reflection trigger judgment result.

[0065] S35-2. According to the hard constraint verification rules of consistency between alloy grade and process window, correspondence between defects and countermeasures, compliance of numerical parameter range, relevance of referenced materials, and integrity of inference chain, the inference chain anchoring chain of the graph is verified one by one to obtain the verification results.

[0066] S35-3. Based on the verification results, the unqualified answers are rewritten and continuously optimized using the preliminary answer, reasoning chain, and verification results as new inputs. When the maximum number of iterations is reached and the verification is still not passed, the answer is explicitly marked as exceeding the range of reliable answers, and the optimization results are obtained.

[0067] S35-4. For the answers that pass the verification results, a weighted calculation is performed based on the relevance score of the retrieved data, the number of items that pass the knowledge graph verification, and the consistency index of the reasoning chain. A multidimensional confidence score and uncertainty description are added to obtain the process reasoning module.

[0068] Optionally, in the above-described agent-based high-temperature alloy casting process design method, in step S4, the professional toolset includes: drawing OCR recognition tool, gravity casting simulation tool, defect prediction deep learning tool, and process document generation tool.

[0069] Optionally, in the agent-based high-temperature alloy casting process design method described above, step S5 includes:

[0070] S51. Based on the intelligent agent orchestration module, the user's natural language input is parsed into parsing results including task type, key entities, and constraints;

[0071] S52. Based on the process knowledge graph and prompt word template, determine the tool call combination and execution order required for the current task;

[0072] S53. Execute multiple tool calls in the tool call combination in sequence according to the execution order, and merge the structured output results of each tool and send them back to the large language model;

[0073] S54. Generate the final process scheme based on the process knowledge graph, the search result set and the graph anchoring inference chain, and attach the cited sources, confidence score and uncertainty description.

[0074] In a second aspect, the present invention also provides an agent-based high-temperature alloy casting process design system for executing an agent-based high-temperature alloy casting process design method as described above, comprising:

[0075] Structured unit: Used to perform structured processing on multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base;

[0076] Modeling unit: used to extract and model entities and relationships in the multi-source process data base to obtain a process knowledge graph;

[0077] Constraint Unit: Used to connect the multi-source process data base obtained by structured processing into the large language model based on the retrieval enhancement generation method, and to combine the prompt word template and the thinking chain reasoning mechanism to verify and constrain the output of the large language model, so as to obtain a process reasoning module with illusion suppression capability;

[0078] Integration Unit: Used to integrate the process knowledge graph and the process reasoning module with a professional toolset, and to establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design;

[0079] The output unit is equipped with a casting high-temperature alloy process design intelligent agent obtained by the integrated unit, which is used to apply the configured casting high-temperature alloy process design intelligent agent to the high-temperature alloy casting process design and output the final process scheme.

[0080] The beneficial effects of this invention are as follows: The high-temperature alloy casting process design method and system based on intelligent agents provided by this invention, on the one hand, by constructing a multi-source process data foundation composed of a materials science fundamentals database, a process scheme and process specification database, a quality case database, and a dual-type process knowledge graph, it transforms the implicit experience long scattered in alloy handbooks, process documents, and operating tips into searchable and reasonable structured knowledge assets. This allows process engineers to quickly acquire domain knowledge accumulated by the system without relying on personal experience, significantly shortening the process scheme formulation cycle for new castings. On the other hand, through the introduction of domain-customized semantic segmentation and synonym term embedding alignment, the accuracy of retrieval and recall is significantly improved. The recall and cross-encoder re-ranking mechanism further ensures a high degree of matching between retrieval results and user needs. Combined with graph-anchored thought chain reasoning and a five-dimensional hard constraint self-reflection verification mechanism, it effectively suppresses the illusion risk of large language models in professional fields, ensuring that every... Both the process plan and defect countermeasures can be traced back to specific alloy manual numbers, process specification numbers, or historical case numbers, significantly ensuring the credibility and interpretability of the output results. Furthermore, by integrating drawing OCR recognition tools, gravity casting simulation tools, defect prediction deep learning tools, and process document generation tools, and with autonomous scheduling by the intelligent agent orchestration module, an end-to-end engineering closed loop is achieved, from product drawing input to process plan derivation, metallurgical defect risk pre-screening, simulation verification feedback, and automatic generation of process specification documents. This significantly reduces the time engineers spend on non-core tasks such as formatting and chart layout, effectively improving the delivery efficiency of the entire process design process. In addition, the model base in this invention maintains a decoupled design, adaptable to open-source models, supervised fine-tuning domain models, and private domain models deployed on the intranet, meeting the requirements for safe implementation in production environments. Through continuous accumulation of supervised fine-tuning samples via user feedback, the capabilities of the intelligent agent evolve continuously with the deepening of engineering practice. Attached Figure Description

[0081] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of the present invention and do not constitute a limitation on the technical solutions of the present invention.

[0082] Figure 1 A schematic diagram of a process design method for high-temperature alloy casting based on an intelligent agent is provided in an embodiment of the present invention.

[0083] Figure 2 This is a schematic diagram of a high-temperature alloy casting process design system based on an intelligent agent, provided as an embodiment of the present invention. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

[0085] The present invention provides the following specific embodiments, which can be combined with each other. For the same or similar concepts or processes, they may not be described again in some embodiments.

[0086] like Figure 1 The diagram shows a flowchart of a high-temperature alloy casting process design method based on an intelligent agent, provided by the present invention. The high-temperature alloy casting process design method based on an intelligent agent provided by this invention includes the following steps:

[0087] S1. Structure the multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base.

[0088] S2. Extract and model the entities and relationships in the multi-source process data base to obtain a process knowledge graph;

[0089] S3. The multi-source process data base obtained by structured processing is connected to the large language model based on the retrieval enhancement generation method, and the output of the large language model is verified and constrained by the prompt word template and the thinking chain reasoning mechanism to obtain a process reasoning module with illusion suppression capability.

[0090] S4. Integrate the process knowledge graph and the process reasoning module with the professional toolset, and establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design.

[0091] S5. Use the intelligent agent for high-temperature alloy casting process design to design the high-temperature alloy casting process and obtain the final process scheme.

[0092] The following is a detailed description of the specific implementation methods of each step in the intelligent agent-based high-temperature alloy casting process design method provided in the embodiments of the present invention.

[0093] S1. Structure the multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base.

[0094] In one implementation of this invention, the process of S1 includes:

[0095] S11. The original data of the cast high-temperature alloy, namely alloy composition data, physical property parameters, microstructure data, performance data and process window data, are standardized and organized with alloy grade as the index unit to obtain a basic materials science database.

[0096] In one implementation scheme of this method, the method for standardizing and organizing the raw data of the cast high-temperature alloy in S11 is illustrated below:

[0097] For alloy composition data, such as extracting the main element content, trace element content and typical impurity upper limit of different alloy grades such as K4169, K424 and K418, a data table is established with the alloy grade as the primary key. Each grade corresponds to a complete record, which contains the mass percentage value of each element.

[0098] For physical properties, such as density, specific heat, thermal conductivity, surface tension, viscosity, and latent heat of fusion, the curves showing how these properties change with temperature are stored as temperature-value pairs. Each physical property curve stored as a temperature-value pair is associated with a corresponding alloy grade.

[0099] For the microstructure data, such as equiaxed grain morphology, γ' phase size and volume fraction, carbide morphology, eutectic structure, and the evolution of microstructure in the as-cast and heat-treated states, each microstructure feature is recorded separately according to the alloy grade. Each microstructure feature is described in text and stored together with the corresponding metallographic image path.

[0100] For performance data, such as room temperature and high temperature tensile, creep, low cycle and high cycle fatigue, oxidation resistance and thermal corrosion, thermal expansion and thermal conductivity, etc., performance parameters are stored as numerical fields, and each performance index corresponds to a test temperature and test condition label.

[0101] For process window data, such as liquidus and solidus temperatures, suitable casting temperature range, mold preheating temperature range, and cooling medium parameters, extract them as numerical range fields.

[0102] Subsequently, in this embodiment of the invention, the above five types of data are uniformly categorized into the material table (e.g., the material_properties table) of a relational database using alloy grades as the index unit. Each row of records corresponds to a complete physical property profile of an alloy grade, thus obtaining a basic materials science database.

[0103] S12. Classify and store the process operation data of historical castings, namely gating system layout data, mold shell process data, cooling process data, heat treatment system data and process operation data, according to product type, alloy grade, typical wall thickness and structural feature labels, to obtain a process scheme and process specification database.

[0104] In one implementation scheme of this method, the method for classifying and storing the process operation data of historical castings in S12 is illustrated below:

[0105] For the gating system layout data, such as the location of the sprue, the number of runners, the cross-sectional area of ​​the ingate, and the design parameters of the runner guides, extract them into structured fields and label them with the corresponding product type tags, such as turbine blades, casings, and guides.

[0106] For mold shell process data, such as the number of coating layers, the ratio of each coating layer, the particle size of the sand, the dewaxing temperature curve, and the calcination temperature curve, process parameter sequences are extracted. Each process parameter sequence is associated with a dual label of alloy grade and product type.

[0107] For cooling process data, such as cooling medium type, cooling rate, cooling start temperature and cooling end temperature, extract them as numerical fields.

[0108] For heat treatment regime data, such as solution temperature, solution holding time, aging temperature, aging holding time, and cooling method, extract them as a set of process regime parameters.

[0109] For process operation data, such as the experience and tips in annotations, memos, and job operation instructions, extract them in text form, and label each tip with the process name, applicable product type, applicable alloy grade, typical wall thickness range, and structural feature keywords; among them, experience and tips include, for example, the empirical value of the opening angle of the gating for a certain thick part, the design of the duct for easily cold shut parts, and the correction coefficient of the remelting material ratio for a specific alloy pouring temperature, etc.

[0110] Subsequently, in this embodiment of the invention, the above-mentioned gating system layout data, mold shell process data, cooling process data, heat treatment system data, and process operation data are used to establish a composite index according to four levels of tags: product type, alloy grade, typical wall thickness, and structural characteristics. This index is then stored in the process table (e.g., the process_knowledge table) of a relational database. Each row of records corresponds to a complete process plan or operation tip entry, thus obtaining a database of process plans and process specifications.

[0111] S13. The descriptions of historical casting quality problems, their causes, and solutions are archived in a structured format as a triplet of problem, cause, and countermeasure to obtain a quality case database.

[0112] In one implementation of this method, the method for archiving historical casting quality issues using a ternary structure in S13 is illustrated below:

[0113] For the description of quality problems, such as dimensional deviations, surface defects, internal porosity, inclusions, impurities, recrystallization, hot cracks, cold shuts, undercasting, arcing, abnormal grain size, and other defect types, their X-ray, fluorescence, metallographic, and CT test results are extracted into standardized defect names and detection feature fields.

[0114] For the analysis of the causes of quality problems, the conclusions of causes obtained by means such as fracture analysis, metallographic analysis, energy dispersive spectroscopy analysis, and process parameter review are classified and labeled into four categories: material factors, process factors, equipment factors, and operational factors. Each cause is stored in structured text and associated with the defect type.

[0115] Solutions to quality problems, such as process adjustments, equipment modifications, and revised operating procedures, are extracted into countermeasure texts, and the improved recurrence rate and first-pass yield after verification are stored as numerical fields.

[0116] Subsequently, in this embodiment of the invention, each historical casting quality problem case is constructed into a data record in the form of a problem-cause-countermeasure triple. Each field in the triple data is uniquely identified by the case number (case_id) and stored in the case table of the relational database (e.g., the quality_case table). Each row of records corresponds to a complete quality handling case, thus obtaining a quality case database.

[0117] S14. After preprocessing the materials science database, the process scheme and process specification database, and the quality case database, including at least cleaning and vectorization, they are uniformly stored in a hybrid storage database composed of vector database, relational database and graph database to obtain a multi-source process data base.

[0118] In one implementation of this method, the preprocessing and mixed storage of the three types of databases obtained in S11 to S13 in S14 is carried out as follows:

[0119] First, a data cleaning process is performed. Duplicate records in each database are deduplicated using a combination of alloy grade and process parameter keys. Missing values ​​are filled with the average of similar products for numerical fields and with a "not recorded" label for text fields. Outliers are filtered out by setting a reasonable range for physical properties. Next, data anonymization is performed. Fields involving specific product model codes are replaced with general product type labels, and fields involving specific company names are replaced with anonymized identifiers. Then, the anonymized data is vectorized. A pre-trained text embedding model is used to encode process specification text, operation tips text, and quality case description text, mapping each text segment to a dense vector. The vectors are stored with the original text ID and source identifier. Finally, the structured fields from the materials science database, process scheme and specification database, and quality case database are stored in a PostgreSQL relational database. The text vectors are stored in a Milvus vector database. Metadata tags such as alloy grade, product type, process name, and defect type are stored as key-value pairs in a Redis index database. These three types of storage are linked through a unified data ID, creating a multi-source process data foundation.

[0120] It should be noted that the anonymization process in S14 above is not a necessary step in the design method provided by the embodiments of the present invention, but rather a data processing measure flexibly selected according to the security requirements of the actual deployment environment. For example, in cloud deployment or cross-enterprise data sharing scenarios, to protect user trade secrets and product model information security, anonymization is implemented by replacing fields involving specific product model codes with general product type tags and replacing fields involving specific company names with anonymized identifiers. As another example, in intranet local deployment scenarios, since the data itself is within the protection scope of the enterprise firewall and does not circulate externally, users can selectively perform anonymization or directly retain the original sensitive data to maintain data integrity according to their own management requirements. Regardless of whether anonymization is performed, it does not affect the core function of the multi-source process data base, which is to structure, clean, and vectorize multi-source heterogeneous data and then uniformly store it as a searchable knowledge asset. Therefore, the specific execution of anonymization depends on the characteristics of the deployment environment and the user's security strategy, and does not constitute a limitation on the technical solution of the present invention.

[0121] S2. Extract and model the entities and relationships in the multi-source process data base to obtain a process knowledge graph.

[0122] In one implementation of this invention, the above-mentioned S2 process includes:

[0123] S21. Using product nodes, material nodes, structural feature nodes, quality requirement nodes, and process nodes as the first basic node types, entity relationships between nodes are extracted from the multi-source process data base through entity recognition and ontology modeling technology to construct a product process design knowledge graph.

[0124] In one implementation of this method, the method for extracting entity relationships between nodes from the multi-source process data base and constructing a product process design knowledge graph in step S21 is described below.

[0125] First, we define five basic node types, with the following definitions: Product nodes include attributes such as product name, product type (e.g., turbine blade or casing), application location, and performance requirement level; Material nodes include attributes such as alloy grade (e.g., K4169), alloy category (e.g., equiaxed crystal), key physical properties (e.g., liquidus temperature, casting characteristics); Structural feature nodes include complex feature attributes such as wall thickness distribution, hot spot location, weak zone location, internal cavity characteristics, tenon edge plate damping platform air film holes; Quality requirement nodes include attributes such as dimensional accuracy, surface roughness, internal metallurgical quality grade, and grain size requirements; and Process nodes include attributes such as mold shell process, gating system type, gating temperature, mold shell preheating temperature, heat treatment regime, and hot isostatic pressing parameters.

[0126] Secondly, a named entity recognition model is used to scan the text in the materials science basic database and the process scheme and process specification database to identify the entities of the five types of nodes in the first basic node type mentioned above. Then, the identified entities are classified according to node type and deduplicated before creating nodes, and each node is assigned a unique node_id.

[0127] Subsequently, a relation extraction model was used to determine the relationships between co-occurring entity pairs in the same process scheme text. The requirement-satisfaction relationship was extracted between product nodes and material nodes, the structure-risk-process compensation relationship was extracted between structural feature nodes and process nodes, and the material-process window relationship was extracted between material nodes and process nodes. Each relationship was stored in the form of a triple, with the triple storage format being: <node A, relation type, node B>.

[0128] Finally, in this embodiment of the invention, the extracted nodes and relationships are imported into the Neo4j graph database, and after deduplication using the MERGE statement, a graph structure is constructed to create a product process design knowledge graph.

[0129] S22. Using defect type nodes, possible cause nodes, diagnostic feature nodes, and solution nodes as basic node types, the defect solution reasoning chain is extracted from the quality case database of the multi-source process data base through relation extraction technology, and priority weights are marked for each countermeasure based on historical case data to construct a defect solution knowledge graph.

[0130] In one implementation of this method, the method of extracting the defect solution reasoning chain from the quality case database and constructing the defect solution knowledge graph in step S22 is described below.

[0131] First, four basic node types are defined. For example, the definition content is as follows: Defect type node includes defect name such as porosity, inclusion, impurities, hot crack, etc.; Possible cause node includes cause category such as material factors, mold shell factors, smelting factors, casting factors, heat treatment factors, equipment factors, operation factors, and specific cause description attributes; Diagnostic feature node includes metallographic morphology, energy dispersive spectroscopy characteristics, location, process time, etc.; Solution node includes specific process adjustment actions such as increasing the casting temperature by 10℃, adding ducts, etc.

[0132] In addition, each triplet record in the quality case database is parsed, and the defect name in the problem field is extracted as a defect type node, the cause description in the cause field is extracted as a possible cause node, the countermeasure description in the solution field is extracted as a solution node, and the detection features implicit in the problem field and cause field are extracted as diagnostic feature nodes.

[0133] Subsequently, in this embodiment of the invention, directed edges are established for the four types of nodes within the same case in the order of defect type node → possible cause node → diagnostic feature node → solution node. The recurrence rate and first-pass rate improvement after each countermeasure is adopted in the historical case database are statistically analyzed. The statistical results are stored as the priority_weight attribute value of the solution node. The higher the value, the higher the priority of the countermeasure.

[0134] Finally, in this embodiment of the invention, the above-mentioned nodes and directed edges are imported into the Neo4j graph database to construct a defect resolution knowledge graph.

[0135] S23. Connect the product process design knowledge graph and the defect resolution knowledge graph through unified metadata tags to obtain a process knowledge graph.

[0136] In one implementation of this method, the specific implementation method for establishing the interconnection between the product process design knowledge graph and the defect resolution knowledge graph through unified metadata tags in step S23 is described below.

[0137] On one hand, the material and process nodes in the product process design knowledge graph are traversed to extract their alloy grade attribute values, such as K4169. On the other hand, the possible cause and solution nodes in the defect resolution knowledge graph are traversed to extract their associated alloy grade attribute values. Subsequently, nodes with the same alloy grade attribute value are linked by an inter-graph relationship edge between the two graphs, with the relationship type labeled as related_by_alloy.

[0138] On the other hand, the product nodes in the product process design knowledge graph are traversed to extract their product type attribute values, such as turbine blades. Similarly, the defect type nodes in the defect resolution knowledge graph are traversed to extract their frequently associated product types. Subsequently, nodes with the same product type attribute value are linked together via inter-graph relationships, labeled as related_by_product.

[0139] On the other hand, the process nodes in the product process design knowledge graph are traversed to extract their process name attribute values, such as the casting process. Similarly, the possible cause nodes in the defect resolution knowledge graph are traversed to extract the process names of the causes. Subsequently, nodes with the same process name attribute value are linked together via inter-graph relationships, labeled as related_by_process.

[0140] Ultimately, in this embodiment of the invention, the two interconnected graphs are merged and stored in the same Neo4j graph database instance, and a process knowledge graph is obtained through a unified node_id namespace and relation type tag system.

[0141] S3. The multi-source process data base obtained by structured processing is connected to the large language model based on the retrieval enhancement generation method, and the output of the large language model is verified and constrained by the prompt word template and the thinking chain reasoning mechanism to obtain a process reasoning module with illusion suppression capability.

[0142] In one implementation of this invention, the above-mentioned S3 process includes:

[0143] S31. Perform domain-customized semantic segmentation and embedding alignment on the multi-source process data base, and establish a multi-source knowledge index based on the retrieval enhancement generation method.

[0144] It should be noted that the embodiments of the present invention do not impose restrictive constraints on the large language model base and its training method. The selected model type, training method and deployment method can be flexibly determined according to the performance requirements, security requirements and cost constraints of the actual application scenario.

[0145] Regarding model selection, large language models can be open-source, general-purpose models such as the LLaMA series, Qwen series, DeepSeek series, GLM series, Baichuan series, and ChatGLM series. These models can be directly used in the RAG pipeline and thought chain inference process of this invention without additional training, and can achieve good domain adaptability through prompt word engineering and retrieval enhancement. Alternatively, large language models can be commercial closed-source models such as the GPT series and Claude series, which can be accessed through API calls to the technical architecture of this invention. Furthermore, large language models can also be private domain models deployed in an enterprise intranet environment. These models are deployed locally within the user's firewall protection range, and all data processing and inference calculations are completed within the intranet, meeting data security requirements in a production environment.

[0146] In terms of training methods, large language models can employ various domain-adaptive training techniques, including: continuous pre-training, which involves further self-supervised pre-training on top of a general large model using unlabeled text corpora in the field of casting high-temperature alloys, such as alloy handbooks, process specifications, technical reports, and academic papers, enabling the model to learn domain-specific terminology and knowledge expression patterns; supervised fine-tuning, which uses manually labeled high-quality domain data, such as historical process question-and-answer pairs, defect diagnosis cases, and process specification interpretation samples, to conduct supervised training on the model, improving the model's understanding and execution capabilities for professional tasks; instruction fine-tuning, which involves fine-tuning the model by constructing an instruction-response dataset in the field of casting high-temperature alloys, enhancing the model's ability to follow domain instructions and task descriptions; efficient parameter fine-tuning techniques, such as LoRA, Adapter, Prefix-Tuning, and P-Tuning, which train only a small number of model parameters or insert a small number of trainable modules, achieving domain adaptation while keeping the main parameters of the base model frozen, significantly reducing training costs and storage overhead; and feedback-based reinforcement learning, which uses reinforcement learning algorithms to optimize the professionalism and reliability of the model's output by providing expert feedback on the model's output process schemes and defect diagnosis results. In addition, a hybrid training strategy can be adopted, such as first performing continuous pre-training to enhance domain knowledge reserves, then performing supervised fine-tuning to improve task execution capabilities, and finally optimizing output quality through RLHF.

[0147] In the technical architecture of this invention, all the different types of large language models and different training methods mentioned above are accessed through a unified API interface or a local inference engine. Regardless of the type of large model used, whether domain training is performed, or what training technology is used, the integrity and feasibility of the technical solution of this invention are not affected. The technical effect of this invention mainly comes from the structured organization of domain knowledge, the reasoning guidance of graph constraints, the multi-level illusion suppression mechanism, and the collaborative scheduling of professional tools, rather than relying on the inherent capabilities of a specific large model. Therefore, the specific selection of the large language model base, the specific adoption of the training method, and the specific configuration of the deployment environment do not constitute a limitation on the scope of protection of this invention.

[0148] In one implementation of this method, the above-described S31 includes:

[0149] S31-1. Based on the process procedure boundary, the quality case phenomenon analysis and countermeasure paragraph boundary, and the alloy handbook grade item boundary, semantic segmentation is performed on the process scheme and process procedure database, quality case database, and materials science basic database in the multi-source process data base to obtain a domain semantic segmentation set.

[0150] In S31-1, the specific implementation method for domain-customized semantic segmentation of the multi-source process data base is as follows: For the process scheme and process specification database, identify the process identification words in the text, such as "wax mold preparation", "modeling", "slurry coating", "dewaxing", "calcination", "melting", "casting", "heat treatment", etc., and segment each process identification word and its subsequent paragraph as a complete semantic unit. Each semantic block contains the process name, operation step description, key parameter values, and precaution text. Each segmented semantic block is identified by chunk_id and the process specification number to which it belongs is recorded.

[0151] For the quality case database, the paragraph markers "problem phenomenon", "cause analysis" and "solution" in the text are identified. Each case is divided into three paragraphs: phenomenon paragraph - analysis paragraph - countermeasure paragraph. Each paragraph maintains the complete three-tuple structure without being destroyed. Each paragraph block after segmentation records the case number and paragraph type label.

[0152] For the materials science database, the alloy grade identifiers such as "K4169", "K424" and "K418" in the text are identified. Each alloy grade entry is segmented as an independent semantic block. Each semantic block contains a complete descriptive text of the grade's composition data, microstructure data, property data and process window data.

[0153] Finally, in this embodiment of the invention, all the segmented semantic blocks are aggregated, and each semantic block is accompanied by the source database identifier, the original record ID, and the segmentation boundary location information to obtain a domain semantic block set.

[0154] S31-2. Use the synonyms in the multi-source process data base to perform lightweight alignment on the embedding model, input the domain semantic block set into the aligned embedding model, and obtain the domain embedding index.

[0155] In S31-2, the specific implementation method for lightweight alignment of the embedding model and generation of domain embedding index is as follows: First, synonym pairs are extracted from the multi-source process data base, including synonyms in the professional field such as "white spot" and "mixture", "Treated" and "heat-treated state", "EQ" and "equiaxed crystal". Then, a pre-trained text embedding model is loaded, and each pair of synonyms is input into the model to obtain its initial embedding vector, and the cosine distance between the two vectors is calculated. Then, a contrastive learning loss function is constructed to bring the embedding vectors of synonym pairs closer to the target distance threshold and push the embedding vectors of non-synonymous pairs further away from the target distance threshold. The weights of the last two layers of the embedding model are updated through gradient descent, and iterative training is performed until the loss function converges to obtain the aligned embedding model. After the embedding model is established, the text of each semantic block in the domain semantic block set is input into the aligned embedding model, and the vector of the model output layer is extracted as the embedding representation of the semantic block. Each embedding vector is accompanied by the chunk_id, source identifier, and original text of the corresponding semantic block. Finally, all embedded vectors and their metadata are inserted into the Milvus vector database, and an HNSW index structure is established with index parameters set to M=16 and efConstruction=200 to obtain the domain embedding index.

[0156] S31-3. Serialize the subgraphs in the process knowledge graph into natural language descriptions in the form of entity relationships, and merge the natural language descriptions with the domain embedding index into a hybrid storage database to obtain the multi-source knowledge index.

[0157] In S31-3, the specific implementation method for serializing the subgraph in the process knowledge graph and merging it with the domain embedding index is as follows: First, traverse each node in the process knowledge graph, extract the node_id, node type, and attribute key-value pairs of the node, and traverse all edges originating from that node, extracting the relation type of the edge and the target node_id. Then, concatenate the node and edge information into a natural language sentence, in the format "<Entity A's name> (type: <Entity A's type>) is connected to <Entity B's name> (type: <Entity B's type>) through the <relation type> relation," for example, "K4169 alloy (type: material node) is connected to casting temperature 1480-1520℃ (type: process node) through the constraints relation." Then, serialize each node and its associated edges into several natural language descriptions, each description including the original node_id and relation path identifier, and generate an embedding vector from the aligned embedding model of all natural language description text inputs. After obtaining the embedding vectors, the serialized embedding vectors from the process knowledge graph are inserted into the Milvus vector database, sharing the same index space as the domain embedding index in S31-2. The vector records are appended with a `graph_source` identifier field set to "knowledge_graph" to distinguish the source. Finally, a `metadata_index` table is created in a PostgreSQL relational database to store metadata fields for each vector, including `chunk_id` or `node_id`, source type, alloy grade label, product type label, defect type label, and process name label. The `chunk_id` or `node_id` is used to establish a connection with the vector records in the vector database, thus obtaining a multi-source knowledge index.

[0158] S32. Recall and reorder the user input based on the multi-source knowledge index to obtain a retrieval result set with source tracing tags.

[0159] In one implementation scheme of this method, the implementation method of performing three-way recall and reordering of user input based on multi-source knowledge index in S32 above is described below.

[0160] First, the system receives user input text, such as "How to handle the loosening defect in K4169 alloy turbine blades." The embedded model, aligned with this input text, generates a query vector. Next, based on the query vector, a vector retrieval is performed in the Milvus vector database. The cosine similarity between the query vector and all vectors in the index is calculated, and the Top-50 candidate vectors are returned in descending order of similarity. The corresponding chunk_id or node_id is extracted from the candidate vectors, and the corresponding original text and source identifier are retrieved from the PostgreSQL relational database to obtain a semantic candidate dataset. After the retrieval, the user input text is segmented into a keyword sequence ["K4169", "alloy", "turbine blade", "loosening", "defect", "handling"]. The BM25 keyword matching algorithm is executed on all text records in the multi-source process data base, calculating the BM25 score for each record. The Top-50 records are returned in descending order of score to obtain the keyword matching candidate dataset.

[0161] Subsequently, this embodiment of the invention extracts structured metadata from the user-input text, identifies alloy grade "K4169", product type "turbine blade", and defect type "loose". An SQL query is executed in the metadata_index table of the PostgreSQL relational database to filter all records where the alloy grade field value is "K4169", the product type field value contains "turbine blade", and the defect type field value is "loose". The filtering results are returned, obtaining a metadata matching candidate data set. After obtaining the metadata matching candidate data set, the record IDs of the semantic candidate data set, the keyword matching candidate data set, and the metadata matching candidate data set are combined, duplicate IDs are removed, and the results are summarized into a total candidate data set.

[0162] Based on the candidate data set, a Cross-Encoder model is loaded. The user input text is concatenated with each candidate text in the candidate data set into the format "[CLS]User Input[SEP]Candidate Text[SEP]". This concatenation is then input into the model to obtain a relevance score. The scores are sorted in descending order to select the Top-10 candidate data. Finally, a source identifier is added to each Top-10 candidate data. If the source is a process specification, the process specification number is added; if the source is a quality case, the case number is added; and if the source is an alloy handbook, the handbook number is added, resulting in a search result set with traceability tags.

[0163] S33. After merging the retrieval result set with the serialized subgraph description of the process knowledge graph, input the result set into the large language model to obtain a preliminary generated answer carrying explicit references.

[0164] In one implementation scheme of this method, the implementation of S33 is described as follows: The original text and source identifier of each candidate data item are extracted from the retrieval result set with traceability tags and concatenated into an enhanced context text block, formatted as "[Source: Process Specification GY-2023-045] The gating system adopts top-pouring type, with a sprue diameter of 30mm...". After extraction, key entities in the user input are parsed, and the alloy grade "K4169" and defect type "porosity" are further extracted. Simultaneously, a subgraph containing the "K4169" material node and the "porosity" defect type node is queried in the process knowledge graph. All nodes and edges in the subgraph are extracted and serialized into a natural language description, such as "The suitable pouring temperature for K4169 alloy is 1480-1520℃; possible causes of porosity defects include excessively fast cooling rates; a typical countermeasure for excessively fast cooling rates is to reduce the cooling rate to 5℃ / min". Then, the contextual text block and the natural language description of the graph subgraph will be concatenated to construct a complete prompt input, formatted as "As an assistant to a casting high-temperature alloy process engineer. The following are the relevant retrieved materials: [Source: Process Specification GY-2023-045]... The following are the relevant knowledge in the knowledge graph: The suitable pouring temperature for K4169 alloy is 1480-1520℃... User question: How to deal with the porosity defect in K4169 alloy turbine blades; Please answer based on the above information and indicate the source in your answer."

[0165] Finally, in this embodiment of the invention, the prompt word is input into a large language model, which generates an answer text. The answer text includes a reference to the source of the searched information, such as "According to process specification GY-2023-045, the answer is...", thus obtaining a preliminary generated answer with explicit references.

[0166] S34. Perform multi-level prompt word template constraints and graph anchored thinking chain reasoning on the initially generated answer to obtain the graph anchored reasoning chain.

[0167] In one implementation of this method, the implementation of S34 is described below.

[0168] S34-1. Construct a multi-level prompt word template based on role boundary settings, rejection rules, task input and output field constraints, few-sample example guidance, and structured format output requirements. Constrain the preliminary generated answer within the preset expression range of casting high-temperature alloy process design to obtain a preliminary process reasoning result with standardized format.

[0169] In S34-1, the specific method for constructing multi-level prompt word templates and constraining the initial generation of answers is as follows: First, define role boundaries at the system prompt word level and set the text content as "As an assistant to the casting high-temperature alloy process engineer, the answer must be based on the retrieved field information. It is forbidden to fabricate alloy grades, process parameters and technical conditions. When the user asks a question that is beyond the scope of casting high-temperature alloys, a prompt that it cannot be answered should be clearly given."

[0170] Subsequently, at the task template level, input and output fields are defined for different task types. For defect diagnosis tasks, the input fields include "alloy grade," "product type," "defect description," "defect location," and "detection method," while the output fields include "diagnostic conclusion," "confidence level," "ranking of typical causes," "recommended countermeasures," "similar case number," and "referenced materials." The output format is constrained to be JSON. Simultaneously, this invention selects high-quality cases reviewed by experts from historical question-and-answer records, extracting three as a small sample example. Each example contains complete input and output field values ​​and is embedded in the prompt word template as example guidance text.

[0171] Subsequently, the system prompts, task templates, few-sample examples, user input, search data, and map descriptions are concatenated in a fixed order to construct a complete multi-level prompt. This is then input into the large language model, and the model's output automatically follows JSON format constraints. It includes structured content such as the "diagnostic conclusion" field value "This loose defect is shrinkage-type looseness", the "confidence" field value 0.85, and the "typical cause ranking" field value a list of causes, thus obtaining a preliminary process reasoning result with a standardized format.

[0172] S34-2. For the task of recommending process solutions, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the product process design knowledge graph of the process knowledge graph. The reasoning chain is obtained by graph anchoring thinking chain reasoning according to the fixed step sequence of requirement understanding, key variable identification, alloy grade and casting method determination, structural processability assessment, preliminary gating system scheme, mold shell and cooling process scheme, heat treatment and hot isostatic pressing scheme, risk point identification and comprehensive scheme.

[0173] In S34-2, the specific implementation method for graph-anchored thinking chain reasoning for the process scheme recommendation task is as follows: extract the task type field value "process scheme recommendation" from the preliminary process reasoning results in the format specification, and identify that the task requires an eight-step fixed reasoning sequence.

[0174] In this embodiment of the invention, for the eight-step fixed inference sequence in S34-2, the first step, "requirement understanding," is executed. This involves extracting the product type "turbine blade," alloy grade "K4169," and performance requirement "X-ray level 1" from the user input. The product type node "turbine blade" is queried in the product process design knowledge graph to obtain the attributes and associated edges of this node. The current inference focus is then anchored to this node, and the node ID is recorded as anchor_node_1. Next, the second step, "key variable identification," is executed. Starting from the product node anchor_node_1, the structural feature nodes connected to it are traversed to extract the wall thickness distribution attributes "edge plate 3-5mm, blade body 2-3mm." The hot spot risk location is identified as "the connection between the edge plate and the blade body." The inference focus is then anchored to the structural feature node, and the node ID is recorded as anchor_node_2. Next, execute the third step, "Alloy grade and casting method determination". Starting from the material node "K4169", traverse its process window relationship edges, extract the liquidus temperature of 1380℃, solidus temperature of 1320℃, and suitable pouring temperature range of 1480-1520℃, determine that the casting method is equiaxed crystal casting, anchor the inference focus to the material node, and record the node ID as anchor_node_3. Subsequently, steps four through eight are executed sequentially. In each step of the reasoning, the node attribute values ​​are extracted and filled into the output fields of the corresponding reasoning step, starting from the current anchor node and traversing along the graph edge to the next relevant node. Step four generates the structural processability assessment result "The edge plate is prone to thermal cracking, so increase the fillet radius". Step five generates the preliminary plan for the gating system "Use top injection type, 30mm straight runner". Step six generates the mold shell and cooling process plan "Mold shell preheating 1050℃, directional cooling of the edge plate area". Step seven generates the heat treatment plan "Solution treatment 1180℃ / 4h, aging 850℃ / 16h". Step eight generates the risk point identification "High risk of thermal shrinkage and loosening of the edge plate, add drainage ribs".

[0175] Finally, in this embodiment of the invention, the anchored node sequence [anchor_node_1, anchor_node_2, anchor_node_3, ...] in the above eight-step reasoning process is concatenated with the output result of each step to form a structured reasoning chain, in the format "{step:1, node:anchor_node_1, reasoning:'Product is turbine blade', output:'Extract wall thickness and thermal node information'}, {step:2, node:anchor_node_2, ...}", thus obtaining the process scheme reasoning chain.

[0176] S34-3. For the defect diagnosis task, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the defect solution knowledge graph of the process knowledge graph. The graph anchoring thinking chain reasoning is carried out according to a fixed sequence of steps: phenomenon analysis, key diagnostic feature localization, traversal of possible causes along the defect solution knowledge graph, similar case matching, cause priority ranking, countermeasure generation, and countermeasure verification condition evaluation. For complex and difficult defect problems, multiple candidate branches are deduced on the same reasoning node, and then summarized and compared. The branch with the most sufficient evidence is selected to obtain the defect diagnosis reasoning chain.

[0177] In S34-3, the specific implementation method for graph-anchored thinking chain reasoning for defect diagnosis tasks is as follows: identify the task type as "defect diagnosis" from the preliminary process reasoning results in the format specification, and execute a seven-step fixed reasoning sequence.

[0178] In the specific implementation, the first step, "phenomenon analysis," is executed. The defect description "holes with a diameter of 1-2 mm appear in the blade edge region" is extracted from the user input. The defect type node "loose" is queried in the defect resolution knowledge graph, and the reasoning focus is anchored to this node, recording its ID as `defect_node_1`. Next, the second step, "key diagnostic feature localization," is executed. The detection method "X-ray detection" is extracted from the user input, and the defect location "edge region" is identified. Starting from the defect type node `defect_node_1`, the edges of the `diagnosed_by` type are traversed to extract the metallographic morphology attribute "circular hole, smooth edges" of the diagnostic feature node. This is matched with the defect features described by the user, and the reasoning focus is anchored to the diagnostic feature node, recording its ID as `feature_node_1`. Next, the third step, "Traversing possible causes along the defect resolution knowledge graph," is executed. Starting from the defect type node `defect_node_1`, edges of type `caused_by` are traversed to extract all possible cause nodes, including cause nodes such as "excessive cooling rate," "excessive casting temperature," and "poor mold permeability." The IDs of each cause node are recorded as `cause_node_1`, `cause_node_2`, and `cause_node_3`. Following this, the fourth step, "Similar case matching," is executed. Historical cases with alloy grade "K4169," product type "turbine blade," and defect type "porosity" are retrieved from the quality case database. These cases are sorted by text similarity between the case description and the current defect description, and the top-3 similar cases are extracted, with case numbers QA-2022-078, QA-2021-134, and QA-2023-009, respectively. Next, the fifth step, "Cause Priority Sorting," is executed. The final cause conclusions annotated in each similar case are extracted. The frequency of "excessive cooling rate" is 2 times, and the frequency of "excessive pouring temperature" is 1 time. Combining the priority_weight attribute values ​​of each cause node in the defect resolution knowledge graph, a weighted priority score is calculated. The cause list is arranged in descending order of score as ["excessive cooling rate (score 0.78)", "excessive pouring temperature (score 0.52)", "poor mold permeability (score 0.31)"]. Following this, the sixth step, "Countermeasure Generation," is executed. Starting from the cause node "excessive cooling rate" (cause_node_1), edges of type `solved_by` are traversed to extract the countermeasure description of the solution node: "reduce the cooling rate to 5℃ / min, extend the solidification time." The recurrence rate of this countermeasure in historical cases is 0.08, and the first-pass yield improvement is 15%. Finally, perform the seventh step, "Countermeasure Validation Condition Assessment," and extract the validation condition attribute "Need to monitor the solidification process temperature curve and confirm that the cooling rate drops to the target range" from the solution node.

[0179] Finally, the embodiments of the present invention record the anchored node sequence and output results in the seven-step reasoning process, construct a structured reasoning chain, and obtain a defect diagnosis reasoning chain.

[0180] It should be noted that, in this embodiment of the invention, the detection also includes whether the user input contains the keyword "difficult" or whether the complexity score of the defect description is higher than the threshold of 0.7. If it is determined to be a difficult defect, in the third step "traversing possible causes", an independent reasoning branch is developed for each cause node. Branch A assumes the cause is "cooling rate is too fast" and deduces countermeasures and expected effects along this cause path. Branch B assumes the cause is "poor casting temperature" and deduces countermeasures. Branch C assumes the cause is "poor mold permeability" and deduces countermeasures. In addition, in the fifth step of performing branch summary comparison, the evidence support degree of each branch is calculated. The evidence support degree is obtained by weighting three indicators: the number of similar case matches, the completeness of the graph path, and the success rate of historical countermeasures. Branch A with the highest evidence support degree is selected as the final diagnostic conclusion, and a defect diagnosis reasoning chain with multi-branch reasoning records is obtained.

[0181] The tree-like thinking mechanism described above is particularly suitable for diagnosing complex quality problems that are rare in history and difficult to judge with a single clue. It effectively avoids the risk of falling into local errors or missing key causes by single linear reasoning, and further enhances the illusion suppression ability and decision reliability of the process reasoning module in complex scenarios.

[0182] S34-4. Merge the process scheme reasoning chain with the defect diagnosis reasoning chain to obtain the map anchoring reasoning chain.

[0183] In S34-4, the specific implementation of merging the process solution reasoning chain and the defect diagnosis reasoning chain is as follows: check whether the current task involves both process solution recommendation and defect diagnosis. If the user input includes "formulate process solution and analyze possible defects", then both the process solution reasoning chain and the defect diagnosis reasoning chain exist.

[0184] After determining that all conditions exist, the eight-step reasoning records of the process solution reasoning chain are arranged by step number as chain_part_1, and the seven-step reasoning records of the defect diagnosis reasoning chain are arranged as chain_part_2. These are then concatenated into a complete reasoning chain [chain_part_1, chain_part_2] according to the task execution sequence. Each step record retains its anchored node ID, reasoning content, and output result. Finally, a task type identifier is appended to the end of the reasoning chain to obtain the graph-anchored reasoning chain.

[0185] S35. Perform multi-dimensional hard constraint self-reflection verification and continuous optimization on the graph anchoring inference chain to obtain a process inference module with illusion suppression capability.

[0186] In one implementation of this method, the implementation process of S35 includes:

[0187] S35-1. For high-risk tasks, a reflection process is forcibly triggered. For low-risk tasks, a reflection process is triggered when the confidence level is lower than a set threshold or when there is a lack of reference materials, and the reflection trigger judgment result is obtained.

[0188] In S35-1, the specific implementation method of triggering the reflection process based on the task risk level is as follows: First, extract the task attributes from the user input and check whether it contains product model keywords such as "a certain type of engine" or high-risk keywords such as "critical defect" or "product model process plan". If the above keywords are detected, the task is marked as a high-risk task, the reflection process is forcibly triggered, and the trigger type is recorded as "mandatory".

[0189] Subsequently, the confidence field value in the graph anchoring inference chain is checked. If the confidence of the process scheme inference chain or the defect diagnosis inference chain is lower than the set threshold of 0.7, the task is marked as a low-risk task and the reflection process is triggered, and the trigger type is recorded as "conditional_low_confidence".

[0190] Subsequently, the reference data field in the graph anchoring inference chain is checked, and the number of reference data is counted. If the number of reference data is 0, or only general knowledge is referenced without specific case numbers or process specification numbers, the reflection process is triggered, and the trigger type is recorded as "conditional_no_citation". This invention summarizes the trigger determination results, and the output field includes a Boolean value of "whether reflection is triggered" and a "trigger type" identifier to obtain the reflection trigger determination result.

[0191] S35-2. Based on the hard constraint verification rules of consistency between alloy grade and process window, correspondence between defects and countermeasures, compliance of numerical parameter range, relevance of referenced materials, and integrity of inference chain, the inference chain anchored by the graph is verified one by one to obtain the verification results.

[0192] In S35-2, the specific implementation method for verifying the graph anchoring inference chain according to the five hard constraint rules is as follows: For the five hard constraint rules, the first check "consistency between alloy grade and process window" is executed first. For example, this invention extracts the recommended casting temperature value of 1510℃ from the graph anchoring inference chain, queries the appropriate casting temperature range of K4169 alloy 1480-1520℃ from the process knowledge graph, and determines whether 1510℃ falls within the range. If it falls within the range, the check passes and the check result "check_1:pass" is recorded. If it exceeds the range, "check_1:fail, recommended temperature 1510℃ exceeds the reasonable range of alloy K4169" is recorded.

[0193] Secondly, perform the second check, "the correspondence between defects and countermeasures." For example, extract the recommended countermeasure "reduce the cooling rate to 5℃ / min" from the defect diagnosis reasoning chain, extract the diagnosed defect cause "cooling rate is too fast," and query the solution node connected to the solved_by edge of the cause node "cooling rate is too fast" in the defect solution knowledge graph. Check whether there is a solution node containing the description "reduce the cooling rate." If there is, the check passes and "check_2: pass" is recorded; otherwise, "check_2: fail, the recommended countermeasure does not have a corresponding relationship in the graph" is recorded.

[0194] Subsequently, the third check, "Numerical parameter range compliance," is performed. This involves extracting all numerical process parameters from the inference chain, such as the mold shell preheating temperature of 1050℃ and the solution temperature of 1180℃. The parameter value ranges of historical process schemes are queried from the multi-source process data base. It is determined whether each parameter falls between the historical minimum and maximum values. If all parameters fall within a reasonable range, the check passes and "check_3:pass" is recorded. If any parameter exceeds the historical range, "check_3:fail, solution temperature 1180℃ exceeds the historical range of 1100-1170℃" is recorded.

[0195] Then, the fourth check, "Relevance of Referenced Materials," is performed. This involves extracting all referenced material source identifiers from the inference chain, such as "Process Specification GY-2023-045" and "Case Number QA-2022-078." The original text of the corresponding materials is extracted from the search results set with traceability tags. The original text is then matched with the key entities in the user input, and the text coverage is calculated. If the referenced materials do indeed involve the same alloy grade, the same product type, or the same defect type, the check passes, and "check_4:pass" is recorded. If the referenced materials are irrelevant to the user's question, "check_4:fail, the referenced materials involve K424 alloy instead of the K4169 alloy that the user inquired about" is recorded.

[0196] Next, the fifth check, "Inference Link Integrity," is performed. Specifically, in this embodiment of the invention, all step records in the graph anchored inference chain are traversed. For all step records, it is checked whether the input data of each step comes from the output of the previous step or from the graph node attributes, and whether the inference target node of each step is connected to the previous node through graph edges. If there are no logical jumps or missing evidence in the inference link, the check passes and "check_5: pass" is recorded. If it is found that step 3 jumps directly to step 6 and there are no corresponding nodes to support intermediate steps 4-5, it is recorded "check_5: fail, logical jump exists in the inference chain."

[0197] Finally, this embodiment of the invention summarizes the five verification results, counts the number of items that pass, and outputs "overall:pass" if all five items pass, otherwise it outputs "overall:fail" and a list of specific failed items to obtain the verification result.

[0198] S35-3. Based on the verification results, the unqualified answers are rewritten and continuously optimized using the preliminary answer, reasoning chain, and verification results as new inputs. When the maximum number of iterations is reached and the verification is still not passed, the answer is explicitly marked as exceeding the range of reliable answers, and the optimization results are obtained.

[0199] In S35-3, the specific implementation method for triggering rewriting and continuous optimization based on the verification result is as follows: Check the overall field value of the verification result. If it is "fail", extract the list of failed items. Concatenate the initially generated answer, the graph anchoring inference chain, and the description of the failed items in the verification result into a new input prompt word, in the format of "The previous answer has the following problems: [check_1:fail, recommended temperature is outside the reasonable range][check_3:fail, solution temperature is outside the historical range]. Please regenerate the answer based on the following information and ensure that the recommended parameters fall within the reasonable range. [Original search information]...[Original graph description]...[Original user question]...".

[0200] After obtaining a new input prompt, the new input prompt is input into the large language model again. The model generates a revised answer. The prompt constraint, graph anchoring inference, and verification process of S34-1 to S35-2 are re-executed on the revised answer. The current iteration round is recorded as iteration_count=2.

[0201] Subsequently, the verification result of the revised answer is checked. If it is still "fail", the next round of rewriting is triggered, and iteration_count is incremented to 3, repeating the above process.

[0202] In this embodiment of the invention, the maximum number of iterations, max_iterations = 3, is set in the above iterations. When iteration_count reaches max_iterations and the verification result is still "fail", rewriting stops, and an explicit annotation text, "This problem exceeds the reliable answer range of the current process intelligence agent and requires human intervention," is generated. This annotation text is used as the final output to obtain the optimization result. If the verification result changes to "pass" within the maximum number of iterations, the revised answer that passes the verification is output as the optimization result.

[0203] S35-4. For the answers that pass the verification results, a weighted calculation is performed based on the relevance score of the retrieved data, the number of items that pass the knowledge graph verification, and the consistency index of the reasoning chain. A multidimensional confidence score and uncertainty description are added to obtain the process reasoning module.

[0204] In section S35-4, the specific implementation method for adding confidence scores and uncertainty explanations to the validated answers is as follows: First, extract the cross-encoder relevance scores of the Top-10 candidate data from the retrieval result set with source tracing tags, and calculate the average score as the relevance score of the retrieved data. Then, count the number of items that passed the five hard constraint validations in S35-2, divide the number of passed items by the total number of items (5), and obtain the knowledge graph validation pass rate. Next, traverse all step records of the knowledge graph anchoring inference chain, check the consistency between the output result of each step and the input data of the next step, calculate the percentage of steps where data field values ​​match successfully, and obtain the inference chain consistency index.

[0205] Subsequently, the three indicators were calculated by weighting them according to their respective weights. The weights were set as follows: the relevance score of the retrieved data was 0.4, the pass rate of the knowledge graph verification was 0.4, and the consistency of the reasoning chain was 0.2. Finally, the overall confidence score was calculated by weighting the three indicators according to their respective weights.

[0206] In addition, in this embodiment of the invention, the steps with lower confidence in the inference chain are extracted from the generated uncertainty description text, such as "the confidence level of alloy grade determination is 0.65. This recommendation is valid when the wall thickness of the flange is less than 2 mm. It is effective within ±10℃ for casting temperature sensitivity. The potential failure mode is thermal cracking caused by improper cooling rate control".

[0207] Finally, the confidence score and uncertainty description text are appended to the end of the answer, and the formatted output is "[Confidence score: 0.82][Uncertainty description: This recommendation is valid when the edge plate wall thickness is less than 2mm...]", which is the final output of the process reasoning module.

[0208] S4. Integrate the process knowledge graph and the process reasoning module with the professional toolset, and establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design.

[0209] In this embodiment of the invention, the professional toolset in S4 above includes, for example: drawing OCR recognition tool, gravity casting simulation tool, defect prediction deep learning tool, and process document generation tool.

[0210] In one implementation of this invention, the professional toolset includes four types of tools. The specific implementation method for integrating the process knowledge graph and process reasoning module with the professional toolset in S4 is described below.

[0211] This drawing OCR recognition tool receives PDF or TIFF format drawing files and outputs the drawing parsing results in structured JSON format. The JSON fields include alloy grade, technical specifications, dimensional tolerances, metallurgical quality grade requirements, non-destructive testing requirements, etc.

[0212] The gravity casting simulation tool accepts the geometric parameters of the gating system, the preheating temperature of the mold shell, the pouring temperature, and the alloy physical properties as inputs. It calls a commercial or self-developed finite element solver to perform filling and solidification simulations and outputs results such as shrinkage rate distribution cloud map, filling integrity percentage, temperature gradient field, and hot spot location coordinates.

[0213] The defect prediction deep learning tool receives the coordinates of three-dimensional mesh nodes, the process parameter vectors of each node such as local cooling rate, and the alloy composition vectors such as the mass percentage of each element. It inputs these into a pre-trained convolutional neural network or graph neural network and outputs numerical values ​​such as the probability of porosity risk, inclusion risk, and hot cracking risk for each mesh node.

[0214] The process document generation tool receives structured process design data, such as a list of process names, a table of parameter values, and text describing risk points. It calls the python-docx library to automatically fill in the chapters of the process card, the parameter table, and the precaution paragraphs according to the company's internal Word template format. It calls the python-pptx library to automatically format the title page, process plan page, simulation result page, and defect analysis page according to the reporting scenario template, and outputs complete docx and pptx files.

[0215] In this embodiment of the invention, a standardized calling interface is defined for each type of tool. The interface adopts the FunctionCalling protocol, and each tool registers its function name, input parameters, and output parameters to the tool registry.

[0216] In addition, the professional toolset also includes: an extensible tool interface, which reserves a standardized tool calling interface based on the Function Calling protocol, without limiting the types and number of external tools. In the future, it can be easily extended to access other professional tools such as heat treatment simulation, hot isostatic pressing simulation, crystal growth phase field simulation, online visual defect recognition, PLM system integration, and MES system integration, forming a sustainable and evolving tool ecosystem, enabling the intelligent agent to have the expansion capability to meet new future needs.

[0217] In addition, this embodiment of the invention establishes an orchestration module for intelligent agents. This orchestration module is used to maintain the query interface of the tool registry and the process knowledge graph. During operation, the orchestration module receives user requirement text, calls a large language model to parse the requirement, and breaks it down into task types such as "process scheme generation," "defect diagnosis," and "simulation verification." It extracts key entities such as alloy grade, product type, and defect type, and extracts constraints such as time constraints and cost constraints. Simultaneously, the orchestration module is also used to match tool combinations according to the task type. If the task type is "process scheme generation," the tool combination is [drawing OCR recognition tool, process inference module, gravity casting simulation tool, process document generation tool]; if the task type is "defect diagnosis," the tool combination is [process inference module, defect prediction deep learning tool, process document generation tool]. In addition, the orchestration module is used to determine the execution order of tools, sorting them according to dependencies into a directed acyclic graph. For example, the output of the drawing OCR recognition tool is the input of the process inference module, the output of the process inference module is the input of the gravity casting simulation tool, the output of the gravity casting simulation tool is the feedback input of the process inference module, and the final output of the process inference module is the input of the process document generation tool. Furthermore, the orchestration module also maintains a task state machine (task_state_machine), recording the calling status of each tool as pending, running, completed, or failed, scheduling tools sequentially according to the execution order, starting the subsequent tool after the preceding tool has completed.

[0218] Ultimately, this embodiment of the invention encapsulates the process knowledge graph, process reasoning module, professional toolset, and intelligent agent orchestration module into a unified service interface, providing RESTful API or local function call entry points to obtain an intelligent agent for casting high-temperature alloy process design.

[0219] S5. Use the intelligent agent for high-temperature alloy casting process design to design the high-temperature alloy casting process and obtain the final process scheme.

[0220] In one implementation of this invention, the above-mentioned S5 process includes:

[0221] S51. Based on the intelligent agent orchestration module, the user's natural language input is parsed into parsing results including task type, key entities, and constraints.

[0222] In S51, the implementation method of parsing the user's natural language input based on the intelligent agent orchestration module is described below. For example, if the user inputs the text "Develop a process plan for the first-stage guide of a certain type of aero-engine power turbine, alloy K4169, flange surface roughness Ra3.2, X-ray first stage", the intelligent agent orchestration module calls the large language model to perform intent recognition and slot filling on the text. Subsequently, the user intent is recognized as "develop a process plan", and the task type field is assigned the value "process_design".

[0223] Based on the extracted intent, key entities are extracted from the text, identifying the product type as "power turbine first-stage guide", the alloy grade as "K4169", and the quality requirements as "surface roughness Ra3.2" and "X-ray grade 1". The extracted entities are stored in the entity dictionary entity_dict={product_type:"guide", alloy:"K4169", surface_roughness:"Ra3.2", xray_grade:"grade 1"}.

[0224] Next, constraints are extracted from the text, and implicit constraints such as "a formal process plan document needs to be generated" are identified. The constraints are then stored in constraints_list=["generate_formal_document"].

[0225] Finally, this invention packages the task type, entity dictionary, and constraint list into a parsing result: parsed_input={task_type:"process_design", entities:entity_dict, constraints:constraints_list}, thus obtaining the parsing result.

[0226] S52. Based on the process knowledge graph and prompt word template, determine the tool call combination and execution order required for the current task.

[0227] In S52, the implementation method for determining the tool call combination and execution order based on the process knowledge graph and prompt word template is as follows: extract the task type "process_design" from the parsing result of S51, query the recommended tool chain corresponding to the task type in the process knowledge graph, and the knowledge rules stored in the graph indicate that "the process_design task needs to call the drawing OCR recognition tool → process reasoning module → defect prediction deep learning tool → gravity casting simulation tool → process reasoning module (feedback correction) → process document generation tool in sequence".

[0228] Subsequently, the system determines whether to call the drawing OCR recognition tool based on whether the user has uploaded a drawing file. If the user has uploaded a drawing PDF file, the drawing OCR recognition tool is added to the tool call combination tools_to_call; otherwise, the tool is skipped.

[0229] Then, the template for the "process_design" task is loaded from the prompt word template library. The template indicates that the process reasoning module should first generate a preliminary process plan, then call the defect prediction tool to assess the risk, then call the simulation tool to verify the plan, and finally correct the plan based on the simulation results.

[0230] In this embodiment of the invention, based on the feedback correction scheme determined above, the execution order is determined according to the template instruction: execution_order=[OCR_tool, reasoning_module, defect_prediction_tool, simulation_tool, reasoning_module_refine, document_generation_tool]. Each tool in the execution order is marked with the source of the calling parameters, thus obtaining the tool calling combination and execution order.

[0231] S53. Execute multiple tool calls in the tool call combination in sequence according to the execution order, and merge the structured output results of each tool and send them back to the large language model.

[0232] In S53, the implementation method for executing multiple tool calls in the tool call combination according to the execution order is as follows: First, the first step is executed, calling the drawing OCR recognition tool, passing in the path of the user-uploaded drawing file / mnt / user-data / uploads / guide drawing.pdf, the tool internally calls the OCR engine to recognize text and symbols, and outputs structured JSON data ocr_result={alloy:"K4169", surface_roughness:"Ra3.2", dimension_tolerance:"±0.1mm", xray_grade:"Level 1", fluorescence_grade:"Level 2"}, and storing ocr_result in the context. Secondly, the second step is executed, calling the process reasoning module and passing in the parsed results parsed_input and ocr_result. The process reasoning module internally executes the search enhancement, prompt word constraint, graph anchoring reasoning, and self-reflection verification processes from S31 to S35, and outputs the preliminary process plan initial_process_plan={casting_system:"top casting type, straight runner 30mm", mold_preheat_temp:"1050℃", pouring_temp:"1510℃", cooling_strategy:"edge plate directional cooling 5℃ / min", heat_treatment:"solution 1180℃ / 4h + aging 850℃ / 16h"}, and stores the initial_process_plan in the context. Next, the third step involves calling a deep learning tool for defect prediction. This tool extracts the casting temperature (1510℃) and cooling rate (5℃ / min) from the initial process plan, queries the composition vector of K4169 alloy from a materials science database, constructs a three-dimensional mesh model, and inputs the local process parameters and alloy composition of each mesh node into a pre-trained neural network. The output defect risk prediction result is `defect_risk={shrinkage_risk_map:[node1:0.12, node2:0.68, ...], hot_crack_risk_map:[node1:0.05, ...]}`. The high-risk area is identified as node 2 at the connection between the flange and the blade, with a risk value of 0.68. The `defect_risk` value is then stored in the context.Subsequently, the fourth step is executed: the gravity casting simulation tool is invoked, and the gating system parameters, gating temperature, and mold preheating temperature in the initial_process_plan are combined with the product's 3D CAD model and input into the finite element solver to perform filling and solidification simulation. The simulation result is output as simulation_result={shrinkage_rate:0.032, filling_completeness:0.98, hot_spot_location:"center coordinates of the edge plate (x:50, y:20, z:10)", temperature_gradient_field:[...]}, and the simulation_result is stored in the context. After storage, the fifth step is executed, and the process reasoning module is called again for feedback correction. The initial_process_plan, defect_risk, and simulation_result are passed in. The process reasoning module identifies node 2 as high-risk based on defect risk prediction, and queries the defect resolution knowledge graph to find the countermeasure "adding ribs". Based on the simulation results, the shrinkage rate of 0.032 is slightly high. The process design knowledge graph is queried to adjust the cooling rate to 4℃ / min, and the revised process plan is generated: refined_process_plan={casting_system:"top casting type, 30mm straight runner, 2 additional ribs added to the rim plate", mold_preheat_temp:"1050℃", pouring_temp:"1510℃", cooling_strategy:"directional cooling of the rim plate 4℃ / min", heat_treatment:"solution 1180℃ / 4h + aging 850℃ / 16h"}. The refined_process_plan is then stored in the context.Finally, in step six, the process document generation tool is invoked, passing in refined_process_plan and all data from the context. Internally, the tool calls python-docx to load the company's Word template, filling the process card sections with process names ["wax mold preparation", "modeling", "slurry coating", "dewaxing", "calcination", "melting", "casting", "heat treatment"], filling the parameter values ​​into the parameter table, inserting the simulation cloud image simulation_result.temperature_gradient_field and converting it to PNG format, generating a complete Word document process_plan.docx, and calling python-pptx to generate presentation.pptx. The paths of process_plan.docx and presentation.pptx are stored in the context, obtaining a summary context of the structured output results of all tools.

[0233] S54. Generate the final process scheme based on the process knowledge graph, the search result set and the graph anchoring inference chain, and attach the cited sources, confidence score and uncertainty description.

[0234] In S54, the specific implementation method for generating the final process plan based on the process knowledge graph, the retrieval result set, and the graph anchored inference chain is as follows: extract refined_process_plan from context as the core process plan content, extract the confidence score generated by the process inference module in S35 (confidence_score=0.82), and extract the uncertainty description: "This plan is effective within the range of 3-5mm edge plate wall thickness and is stable within ±10℃ of casting temperature. The potential risk is that improper control of the cooling rate may lead to thermal cracking." Subsequently, the list of cited sources is extracted from the graph anchoring inference chain of the process inference module: citations=["Process Specification GY-2023-045", "Quality Case QA-2022-078", "Alloy Handbook K4169 Technical Conditions"]. The cited sources are appended to the end of the process scheme text in the format "References: [1] Process Specification GY-2023-045 [2] Quality Case QA-2022-078 [3] Alloy Handbook K4169 Technical Conditions". Then, the refined_process_plan, confidence_score, uncertainty_description, and citations are concatenated into the final process plan text, formatted as "{Gating system: top-injection, 30mm sprue, 2 additional ribs added to the rim plate; mold shell preheating temperature: 1050℃; pouring temperature: 1510℃; cooling strategy: directional cooling of the rim plate at 4℃ / min; heat treatment: solution treatment at 1180℃ / 4h + aging at 850℃ / 16h} [Confidence level: 0.82][Uncertainty description: This plan is effective within the rim plate wall thickness range of 3-5mm...][References:...]". Finally, the final process plan text, along with the generated process document process_plan.docx and presentation.pptx, are output together to obtain the final process plan deliverable.

[0235] Based on the above embodiments of the present invention, the high-temperature alloy casting process design method based on intelligent agents is provided, such as Figure 2 As shown, the present invention also provides a high-temperature alloy casting process design system based on an intelligent agent, comprising:

[0236] Structured Unit 100: Used to perform structured processing on multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base.

[0237] Modeling unit 200: used to extract and model entities and relationships in the multi-source process data base to obtain a process knowledge graph;

[0238] Constraint Unit 300: Used to connect the multi-source process data base obtained by structured processing to the large language model based on the retrieval enhancement generation method, and to verify and constrain the output of the large language model by combining prompt word templates and thinking chain reasoning mechanism, so as to obtain a process reasoning module with illusion suppression capability;

[0239] Integration unit 400: used to integrate the process knowledge graph and the process reasoning module with a professional toolset, and to establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design;

[0240] Output unit 500 is configured with a casting high-temperature alloy process design intelligent agent obtained by integration unit 400, which is used to apply the configured casting high-temperature alloy process design intelligent agent to the high-temperature alloy casting process design and output the final process scheme.

[0241] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0242] This invention provides a method and system for process design of high-temperature alloy castings based on intelligent agents. On the one hand, by constructing a multi-source process data foundation composed of a materials science fundamentals database, a process scheme and process specification database, a quality case database, and a dual-type process knowledge graph, implicit experience that has long been scattered in alloy manuals, process documents, and operating tips is transformed into searchable and reasonable structured knowledge assets. This allows process engineers to quickly acquire domain knowledge accumulated by the system without relying on personal experience, significantly shortening the process scheme formulation cycle for new castings. On the other hand, by introducing domain-customized semantic segmentation and synonym term embedding alignment, the accuracy of retrieval and recall is significantly improved. The recall and cross-encoder re-ranking mechanism further ensures a high degree of matching between retrieval results and user needs. Combined with graph-anchored thought chain reasoning and a five-dimensional hard constraint self-reflection verification mechanism, the illusion risk of large language models in professional fields is effectively suppressed, ensuring that each process scheme is consistent with the user's needs. Defect countermeasures can all be traced back to specific alloy manual numbers, process specification numbers, or historical case numbers, significantly ensuring the credibility and interpretability of the output results. Furthermore, by integrating drawing OCR recognition tools, gravity casting simulation tools, defect prediction deep learning tools, and process document generation tools, and with autonomous scheduling by the intelligent agent orchestration module, an end-to-end engineering closed loop is achieved, from product drawing input to process scheme deduction, metallurgical defect risk pre-screening, simulation verification feedback, and automatic generation of process specification documents. This significantly reduces the time engineers spend on non-core tasks such as formatting and chart layout, effectively improving the delivery efficiency of the entire process design process. In addition, the model base in this invention maintains a decoupled design, adaptable to open-source models, supervised fine-tuning domain models, and private domain models deployed on the intranet, meeting the requirements for safe implementation in production environments. Through continuous accumulation of supervised fine-tuning samples via user feedback, the capabilities of the intelligent agent evolve with the deepening of engineering practice.

[0243] In one specific embodiment, the present invention proposes an agent-based high-temperature alloy casting process design system that operates with an agent for high-temperature alloy casting process design. This system includes: a multi-source data base module for storing a materials science fundamentals database, a process scheme and procedure database, a quality case database, and a dual-type process knowledge graph and its vectorized index; a large-scale model RAG enhancement module for providing RAG and prompt word template management, thought chain reasoning process control, and a self-reflective verification pipeline; an external tool scheduling module, including a drawing OCR recognition tool, a gravity casting simulation interface, a defect prediction deep learning model interface, an office software document generation interface, and an extensible tool interface; an agent orchestration module responsible for requirement analysis, tool planning, result merging, confidence assessment, and interpretable output; and a human-computer interaction front-end supporting natural language dialogue, file upload, drawing preview, report download, and feedback collection.

[0244] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the agent-based high-temperature alloy casting process design method described in various embodiments or some parts of embodiments.

[0245] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A process design method for high-temperature alloy castings based on intelligent agents, characterized in that, include: S1. Structure the multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base. S2. Extract and model the entities and relationships in the multi-source process data base to obtain a process knowledge graph; S3. The multi-source process data base obtained by structured processing is connected to the large language model based on the retrieval enhancement generation method, and the output of the large language model is verified and constrained by the prompt word template and the thinking chain reasoning mechanism to obtain a process reasoning module with illusion suppression capability. S4. Integrate the process knowledge graph and the process reasoning module with the professional toolset, and establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design. S5. Use the intelligent agent for high-temperature alloy casting process design to design the high-temperature alloy casting process and obtain the final process scheme.

2. The method for designing high-temperature alloy casting processes based on intelligent agents according to claim 1, characterized in that, S1 includes: S11. The alloy composition data, physical property parameters, microstructure data, performance data and process window data of the cast high-temperature alloy are standardized and organized with the alloy grade as the index unit to obtain a basic materials science database. S12. Classify and store the gating system layout data, mold shell process data, cooling process data, heat treatment system data and process operation data of historical castings according to product type, alloy grade, typical wall thickness and structural feature labels to obtain a process plan and process specification database. S13. The descriptions of historical casting quality problems, their causes and solutions are structured and archived in the form of a problem, cause and countermeasure triplet to obtain a quality case database. S14. After preprocessing the materials science database, the process scheme and process specification database, and the quality case database, they are uniformly stored in a hybrid storage database composed of vector database, relational database and index database to obtain a multi-source process data base; wherein, the preprocessing method includes at least cleaning processing and vectorization processing.

3. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 2, characterized in that, In S11, five types of data, namely alloy composition data, physical property parameters, microstructure data, performance data and process window data, are uniformly classified into the material table of the relational database according to the alloy grade as the index unit. Each row of records corresponds to a complete physical property profile of an alloy grade, thus obtaining the basic database of materials science. In S12, four types of data, namely, gating system layout data, mold shell process data, cooling process data, heat treatment system data and process operation data, are used to establish a composite index according to four levels of tags: product type, alloy grade, typical wall thickness and structural characteristics. The index is stored in the process table of the relational database. Each row of records corresponds to a complete process plan or operation tip entry, thus obtaining the process plan and process specification database. In S13, each case of a historical casting quality problem is constructed into a data record in the form of a problem-cause-countermeasure triplet. Each field in the triplet data is uniquely identified by the case number and stored in the case table of the relational database. Each row of records corresponds to a complete quality handling case, thus obtaining the quality case database. In step S14, the structured fields of the materials science database, process scheme and process specification database, and quality case database are stored in a relational database, the text vectors are stored in a vector database, and the metadata tags such as alloy grade, product type, process name, and defect type are stored in an index database in the form of key-value pairs. The three types of storage are associated through a unified data ID, thus obtaining a multi-source process data base.

4. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 3, characterized in that, S2 includes: S21. Taking product nodes, material nodes, structural feature nodes, quality requirement nodes and process nodes as the first basic node types, entity relationships between nodes are extracted from the multi-source process data base through entity recognition and ontology modeling technology to construct a product process design knowledge graph. S22. Using defect type nodes, possible cause nodes, diagnostic feature nodes, and solution nodes as the second basic node types, the defect solution reasoning chain is extracted from the quality case database of the multi-source process data base through relation extraction technology, and priority weights are marked for each countermeasure based on historical case data to construct a defect solution knowledge graph. S23. Connect the product process design knowledge graph and the defect resolution knowledge graph through unified metadata tags to obtain a process knowledge graph.

5. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 4, characterized in that, S21 includes: S21-1, scan the text in the materials science basic database and the process scheme and process specification database, identify the entities of the five types of nodes in the first basic node type, and then create nodes after classifying the identified entities according to node type and removing duplicates, and assign a unique node number to each node. S21-2, perform relationship determination on entity pairs that co-occur within the same process solution text, extract requirement-satisfaction relationship between product node and material node, extract structure-risk-process compensation relationship between structural feature node and process node, and extract material-process window relationship between material node and process node. Each relationship is stored in the form of a triple, with the triple storage format as: <node A, relationship type, node B>. S21-3. Import the extracted nodes and relationships into the graph database, remove duplicates, construct the graph structure, and build the product process design knowledge graph.

6. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 5, characterized in that, S22 includes: S22-1, parse each triplet record in the quality case database, extract the defect name in the problem field as the defect type node, extract the cause description in the cause field as the possible cause node, extract the countermeasure description in the solution field as the solution node, and extract the detection features implicit in the problem field and cause field as the diagnostic feature node. S22-2, within the same case, establish directed edges for the four types of nodes of the second basic node type in the order of defect type node → possible cause node → diagnostic feature node → solution node, and statistically analyze the recurrence rate and first-pass rate improvement after each countermeasure is adopted in the historical case database. Store the statistical results as the priority attribute value of the solution node. The higher the value, the higher the priority of the countermeasure. S22-3, import the nodes and directed edges into the graph database to construct a defect resolution knowledge graph.

7. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 6, characterized in that, S23 includes: S23-1: Traverse the material nodes and process nodes in the product process design knowledge graph, extract alloy grade attribute values, traverse the possible cause nodes and solution nodes in the defect resolution knowledge graph, extract the associated alloy grade attribute values, and establish relationship edges between the two graphs for nodes with the same alloy grade attribute values. S23-2, traverse the product nodes in the product process design knowledge graph, extract the product type attribute values, traverse the defect type nodes in the defect resolution knowledge graph, extract the frequently associated product types, and establish relationship edges between the two graphs for nodes with the same product type attribute values; S23-3, traverse the process nodes in the product process design knowledge graph, extract the process name attribute value, traverse the possible cause nodes in the defect resolution knowledge graph, extract the process name of the cause, and establish a relationship edge between the two graphs for nodes with the same process name attribute value. S23-4, merge the two interconnected graphs and store them in the same graph database instance to obtain the process knowledge graph.

8. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 7, characterized in that, S3 includes: S31. Perform domain-customized semantic segmentation and embedding alignment on the multi-source process data base, and establish a multi-source knowledge index based on the retrieval enhancement generation method; S32. Recall and reorder the user input based on the multi-source knowledge index to obtain a retrieval result set with source tracing tags; S33. After merging the retrieval result set with the serialized subgraph description of the process knowledge graph, input the result set into the large language model to obtain a preliminary generated answer with explicit references. S34. Perform multi-level prompt word template constraints and graph anchored thinking chain reasoning on the initially generated answer to obtain the graph anchored reasoning chain; S35. Perform multi-dimensional hard constraint self-reflection verification and continuous optimization on the graph anchoring inference chain to obtain a process inference module with illusion suppression capability.

9. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 8, characterized in that, S31 includes: S31-1. Based on the process procedure boundary, the quality case phenomenon analysis and countermeasure paragraph boundary, and the alloy handbook grade item boundary, semantic block division is performed on the process scheme and process procedure database, quality case database, and materials science basic database in the multi-source process data base to obtain a domain semantic block set. S31-2. Use the synonyms in the multi-source process data base to perform lightweight alignment on the embedding model, input the domain semantic block set into the aligned embedding model, and obtain the domain embedding index. S31-3. Serialize the subgraphs in the process knowledge graph into natural language descriptions in the form of entity relationships, and merge the natural language descriptions with the domain embedding index into a hybrid storage database to obtain the multi-source knowledge index.

10. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 9, characterized in that, S32 includes: S32-1, Generate query vectors from the embedded model after aligning the received user input text, perform vector retrieval in the vector database, calculate the similarity between the query vector and all vectors in the index, obtain a preset number of candidate vectors, and query the original text and source identifier corresponding to each candidate vector from the relational database to obtain a semantic candidate data set; S32-2, the user input text is segmented into a keyword sequence, the keyword matching algorithm is executed on all text records in the multi-source process data base, the matching score of each text record is calculated, and a preset number of candidate text records are obtained to obtain a keyword matching candidate data set; S32-3, extract structured metadata from user input text, identify the specific content of alloy grade, product type and defect type, filter all records that match the identified specific content from the relational database, and obtain a metadata matching candidate data set; S32-4: Take the union of the record IDs of the semantic candidate data set, the keyword matching candidate data set, and the metadata matching candidate data set, remove duplicate IDs, and summarize them into a total candidate data set; S32-5: Concatenate the user input text with each candidate text in the candidate data set, obtain the relevance score of each concatenated text and sort them in descending order of the score, filter a preset number of candidate data, and add a source identifier to each candidate data to obtain a search result set with source tracing tags.

11. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 10, characterized in that, S33 includes: S33-1. Extract the original text and source identifier of each candidate data from the search results set with source tags, and concatenate them into an enhanced contextual text block; S33-2. Parse the key entities in the user input and extract the alloy grade and defect type. Query the subgraph containing material nodes and defect type nodes in the process knowledge graph, extract all nodes and edges in the subgraph, and serialize them into a natural language description. S33-3. Concatenate the enhanced contextual text block with the natural language description of the graph subgraph to construct a complete prompt word. Input the prompt word into the large language model. The response text generated by the large language model contains citation annotations of the source of the searched information, thus obtaining a preliminary generated answer with explicit citations.

12. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 11, characterized in that, S34 includes: S34-1. Construct a multi-level prompt word template based on role boundary setting, rejection rules, task input and output field constraints, few sample example guidance, and structured format output requirements. Constrain the preliminary generated answer within the preset expression range of casting high-temperature alloy process design to obtain a preliminary process reasoning result with standardized format. S34-2. For the task of recommending process solutions, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the product process design knowledge graph of the process knowledge graph. The graph anchoring thinking chain reasoning is performed according to the fixed step sequence of requirement understanding, key variable identification, alloy grade and casting method determination, structural processability assessment, preliminary gating system scheme, mold shell and cooling process scheme, heat treatment and hot isostatic pressing scheme, risk point identification and comprehensive scheme to obtain the process solution reasoning chain. S34-3. For the defect diagnosis task, the reasoning objects in the preliminary process reasoning results are sequentially anchored to the corresponding nodes in the defect solution knowledge graph of the process knowledge graph. The graph anchoring thinking chain reasoning is carried out according to a fixed sequence of steps: phenomenon analysis, key diagnostic feature location, traversal of possible causes along the defect solution knowledge graph, similar case matching, cause priority ranking, countermeasure generation and countermeasure verification condition evaluation. For complex and difficult defect problems, multiple candidate branches are deduced on the same reasoning node, and then summarized and compared. The branch with the most sufficient evidence is selected to obtain the defect diagnosis reasoning chain. S34-4. Merge the process scheme reasoning chain with the defect diagnosis reasoning chain to obtain the map anchoring reasoning chain.

13. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 12, characterized in that, The S35 includes: S35-1. Force the reflection process to be triggered for high-risk tasks, and trigger the reflection process for low-risk tasks when the confidence level is lower than the set threshold or when there is a lack of reference materials, and obtain the reflection trigger judgment result. S35-2. According to the hard constraint verification rules of consistency between alloy grade and process window, correspondence between defects and countermeasures, compliance of numerical parameter range, relevance of referenced materials, and integrity of inference chain, the inference chain anchoring chain of the graph is verified one by one to obtain the verification results. S35-3. Based on the verification results, the unqualified answers are rewritten and continuously optimized using the preliminary answer, reasoning chain, and verification results as new inputs. When the maximum number of iterations is reached and the verification is still not passed, the answer is explicitly marked as exceeding the range of reliable answers, and the optimization results are obtained. S35-4. For the answers that pass the verification results, a weighted calculation is performed based on the relevance score of the retrieved data, the number of items that pass the knowledge graph verification, and the consistency index of the reasoning chain. A multidimensional confidence score and uncertainty description are added to obtain the process reasoning module.

14. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 1, characterized in that, In S4, the professional toolset includes: drawing OCR recognition tool, gravity casting simulation tool, defect prediction deep learning tool, and process document generation tool.

15. The method for designing a high-temperature alloy casting process based on an intelligent agent according to claim 13, characterized in that, S5 includes: S51. Based on the intelligent agent orchestration module, the user's natural language input is parsed into parsing results including task type, key entities, and constraints; S52. Based on the process knowledge graph and prompt word template, determine the tool call combination and execution order required for the current task; S53. Execute multiple tool calls in the tool call combination in sequence according to the execution order, and merge the structured output results of each tool and send them back to the large language model; S54. Generate the final process scheme based on the process knowledge graph, the search result set and the graph anchoring inference chain, and attach the cited sources, confidence score and uncertainty description.

16. A high-temperature alloy casting process design system based on intelligent agents, used to execute the high-temperature alloy casting process design method based on intelligent agents as described in any one of claims 1 to 15, characterized in that, include: Structured unit: Used to perform structured processing on multi-source heterogeneous data in the field of casting high-temperature alloys to obtain a multi-source process data base; Modeling unit: used to extract and model entities and relationships in the multi-source process data base to obtain a process knowledge graph; Constraint Unit: Used to connect the multi-source process data base obtained by structured processing into the large language model based on the retrieval enhancement generation method, and to combine the prompt word template and the thinking chain reasoning mechanism to verify and constrain the output of the large language model, so as to obtain a process reasoning module with illusion suppression capability; Integration Unit: Used to integrate the process knowledge graph and the process reasoning module with a professional toolset, and to establish an intelligent agent orchestration module to analyze user needs and schedule tasks to obtain an intelligent agent for casting high-temperature alloy process design; The output unit is equipped with a casting high-temperature alloy process design intelligent agent obtained by the integrated unit, which is used to apply the configured casting high-temperature alloy process design intelligent agent to the high-temperature alloy casting process design and output the final process scheme.

Images