Intelligent material optimization method and system based on large model

Abstract

The invention discloses an intelligent material optimization method and system based on a large model, and belongs to the technical field of intelligent manufacturing and industrial internet, and the method comprises the steps: obtaining a natural language working condition description inputted by a user; inputting the data into a first round of big language model to generate structured demand data, and then inputting the structured demand data into a second round of big language model to carry out parameter slot extraction and standardized correction to generate standardized query data; triggering at least two independent retrieval agents, and obtaining a candidate material set from the knowledge graph, the structured knowledge base and the vector knowledge base; performing fusion, duplicate removal and conflict detection on the candidate material sets returned by the retrieval agents to generate a fusion candidate list; inputting to a fourth round of large language model for multi-source result refining and reordering, and generating a refining candidate list; inputting to the fifth round of large language model, generating an interpretable recommendation result, and outputting the interpretable recommendation result; the industrial pain points that traditional material type selection depends on personal experience, efficiency is low, and hidden risks are difficult to mine are solved.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent material selection based on a large model, belonging to the field of intelligent manufacturing and industrial internet technology. Background Technology

[0002] Material selection is a crucial step in industrial product R&D management. Traditional material selection relies on the manual experience of designers, requiring them to manually traverse material libraries, match operating parameters (such as temperature, pressure, and load), and compare product attributes (such as materials, dimensions, and certifications). With the increasing complexity of industrial products and the surge in material types (the number of materials for a single type of equipment can reach hundreds of thousands), existing methods reveal the following pain points: Material selection depends on experience; designers must manually match operating conditions and product parameters, lacking intelligent auxiliary tools, resulting in selection efficiency as low as hours, a material reuse rate of only 15%, leading to procurement redundancy and inventory backlog.

[0003] Inefficient: Manual screening requires comparing parameters one by one, and a single selection takes several hours to several days, which seriously affects the R&D cycle;

[0004] Insufficient accuracy: Relying on experience can easily lead to the omission of long-tail materials or emerging technology materials, and it cannot deeply explore the implicit relationships between parameters (such as the coupling relationship between material corrosion resistance and operating humidity).

[0005] Knowledge gap: Historical material data is stored in a scattered manner (such as documents, tables, and drawings), failing to form a structured knowledge system, resulting in high training costs for new employees;

[0006] Therefore, there is an urgent need for an intelligent screening solution that integrates large-scale model semantic understanding, knowledge graph association mining, and multi-strategy recall and ranking to achieve a leap from "parameter matching" to "demand understanding + knowledge reasoning".

[0007] In the existing technology, there are some solutions for manufacturing optimization and material selection that utilize artificial intelligence, large language models (LLM), and knowledge graphs. However, these solutions are usually limited to specific fields or partial technology integration and cannot achieve comprehensive end-to-end intelligent material selection.

[0008] For example, Chinese Patent Publication No. CN116010620B discloses a knowledge graph-based method for knowledge fusion and intelligent recommendation of sewing processes. This method designs a sewing knowledge ontology model, extracts entities using a BiLSTM-CRF model, and employs a stack-based pointer network based on a pre-trained language model to extract relationships. The knowledge is stored in the Neo4j graph database, and sewing parameters are recommended based on fabric mechanical properties, enabling rapid knowledge access and autonomous decision-making in sewing production.

[0009] Another existing technology is Chinese Patent Publication No. CN118014306A, which describes a unified representation method for customized furniture board processing data based on knowledge graphs. This method includes sorting out board processing data and processes, extracting knowledge elements, and constructing a knowledge graph dataset; defining entity, relation, and attribute types to complete ontology construction; performing knowledge extraction, fusion, and storage; using the Neo4j database to store triples; revealing the association relationships and mining potential factors of multi-source heterogeneous data; and supporting applications such as enterprise knowledge question answering systems.

[0010] In addition, Chinese Patent Publication No. CN118297275A proposes a data-driven intelligent process generation method and system. It identifies and cleans process-related data by connecting to a knowledge base, and fine-tunes the Large Language Model (LLM). It collects process business data, design data, and production data, assembles them at the process granularity, and stores them in an independent knowledge base. It uses LLM to generate intelligent process content and further fine-tunes the model through feedback data to support process optimization in the manufacturing process.

[0011] These existing technologies utilize AI, LLM, and knowledge graphs to some extent to improve manufacturing efficiency, but they mainly focus on specific areas such as sewing, furniture panel processing, process generation, or process planning, and the degree of integration is limited.

[0012] The aforementioned existing technologies suffer from the following drawbacks, failing to comprehensively address the pain points in industrial material screening, such as low efficiency, insufficient accuracy, and knowledge gaps. Only this invention, through the deep integration of knowledge bases, multimodal knowledge graphs, multi-path recall strategies, and LLM, can achieve an end-to-end closed loop from natural language condition descriptions to accurate material recommendations, overcoming the limitations of "parameter matching" to "demand understanding + knowledge reasoning."

[0013] While the method in CN116010620B achieves the application of knowledge graphs in sewing processes, it heavily relies on manual annotation of high-quality corpora, resulting in a complex extraction process. Furthermore, it is limited to recommendations based on fabric mechanical properties and cannot handle flexible parsing of natural language descriptions of working conditions. It also lacks the integration of multimodal knowledge bases and multi-path recall, leading to insufficient breadth and depth of recall in broader industrial material screening scenarios. Only the combination of knowledge base + knowledge graph + multi-path recall + LLM in this invention can comprehensively uncover implicit relationships and optimize ranking, achieving efficient material reuse.

[0014] The solution in CN118014306A emphasizes unified representation and knowledge extraction from multi-source heterogeneous data, but it does not integrate a large language model for semantic understanding. Knowledge fusion relies on specific AI technologies such as Siamese networks and convolutional neural networks, lacking multi-path recall and hybrid ranking strategies. It cannot deeply mine the implicit relationships between parameters and does not directly target precise recommendations for material selection. Only the multi-strategy fusion (knowledge base + knowledge graph + multi-path recall + LLM) of this invention can perform multi-dimensional analysis from generalized working condition information, avoiding knowledge gaps.

[0015] CN118297275A utilizes LLM for process generation but fails to integrate knowledge graphs for association mining. Its knowledge base construction relies on data vectorization and RAG technology, resulting in a high dependence on data quality. Furthermore, the feedback mechanism may introduce biases, failing to achieve an end-to-end closed loop from operational description to material recommendation. Only by integrating a knowledge base, knowledge graph, multi-path recall, and LLM in this invention can the recall depth be enhanced, results validated, and accuracy and efficiency improved.

[0016] These shortcomings indicate that while existing technologies may have some local innovations, they lack the irreplaceable nature of the multi-strategy integration of this invention, which makes it impossible to comprehensively improve efficiency, accuracy, and knowledge continuity when dealing with the screening of materials for highly complex industrial products. Summary of the Invention

[0017] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an intelligent material selection method and system based on a large model, realizing a complete capability from natural language to second-level accuracy, traceability, and closed-loop optimization, and completely solving the industry pain points of traditional material selection relying on personal experience, low efficiency, and difficulty in uncovering hidden risks.

[0018] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0019] In a first aspect, the present invention provides a smart material selection method based on a large model, comprising:

[0020] Obtain the user's input in natural language to describe the working conditions;

[0021] The natural language description of the working condition is input into the first round of large language model for intent recognition and key constraint extraction, generating structured requirement data;

[0022] The structured demand data is input into the second round of the large language model for parameter slot extraction and standardization correction to generate standardized query data;

[0023] Based on standardized query data, at least two independent retrieval agents are triggered in parallel to obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively;

[0024] The candidate material sets returned by each retrieval agent are merged, deduplicated, and conflict detected to generate a merged candidate list;

[0025] The fusion candidate list is input into the fourth round of the large language model to refine and reorder the multi-source results, generating a refined candidate list.

[0026] The refined candidate list is input into the fifth round of the large language model, which combines material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

[0027] Furthermore, the step of inputting the natural language working condition description into the first-round large language model for intent recognition and key constraint extraction to generate structured requirement data includes:

[0028] The user intent is identified through the first round of large language modeling, and the intent includes at least one of new material selection, alternative material search, or BOM optimization.

[0029] Extract the key constraints from the natural language working condition description and output them in structured JSON format.

[0030] Furthermore, the step of inputting structured demand data into the second-round large language model for parameter slot extraction and standardization correction to generate standardized query data includes:

[0031] The key constraints are normalized and implicit parameters are completed using a second round of large language modeling.

[0032] Outputs standardized query data that conforms to the preset JSON Schema.

[0033] Furthermore, the method of triggering at least two independent retrieval agents in parallel based on standardized query data includes:

[0034] The knowledge graph retrieval agent is triggered, and based on the entity and relation constraints in the standardized query data, multi-hop graph reasoning is performed to return the first candidate material set and reasoning path;

[0035] Trigger the structured query agent to convert the standardized query data into an SQL query statement, execute an exact or range query from the relational database, and return the second candidate material set;

[0036] The vector retrieval agent is triggered to perform semantic expansion on the natural language working condition description and standardized query data, and then retrieves similar cases and document paragraphs from the vector database to return a third candidate material set.

[0037] Furthermore, the process of merging, deduplicating, and detecting conflicts in the candidate material sets returned by each retrieval agent to generate a fused candidate list includes:

[0038] The candidate material sets returned by each retrieval agent are deduplicated;

[0039] Detect conflicts between candidate materials from different sources and record the conflict type and confidence level;

[0040] The candidate materials are initially merged and sorted based on the weighted Reciprocal Rank Fusion algorithm to generate a fusion candidate list.

[0041] Furthermore, the step of inputting the fusion candidate list into the fourth round of the large language model for multi-source result refinement and re-ranking to generate a refined candidate list includes:

[0042] The fusion candidate list is input into the fourth round of the large language model, and secondary completion and verification are performed in combination with the conflict detection results.

[0043] The candidate materials are cross-coded and scored using a re-ranking model, and then sorted using a rule engine to generate a refined candidate list.

[0044] Furthermore, the refined candidate list is input into the fifth round of the large language model, and combined with material attributes, historical usage data, and risk knowledge, an interpretable recommendation result is generated, including:

[0045] The fifth round of large language model generates recommendation reasons for each candidate material. The recommendation reasons include matching parameters, graph reasoning path, historical usage statistics and risk warnings.

[0046] The output includes structured recommendation results containing material model, supplier, matching score, recommendation reason, and alternative solutions.

[0047] Furthermore, after outputting the recommendation results, it also includes:

[0048] If the number of candidate materials is less than the preset threshold or the highest matching score is lower than the preset score, the abnormal logic routing process will be automatically triggered, and at least one of the following operations will be performed: parameter relaxation, cross-category substitution retrieval, or external directory query.

[0049] Secondly, the present invention provides an intelligent material selection system based on a large model, used to implement the intelligent material selection method based on a large model as described in any of the preceding claims, comprising:

[0050] The user input module is used to obtain the natural language description of the working conditions input by the user;

[0051] The requirement understanding module is used to input natural language descriptions of working conditions into the first-round large language model for intent recognition and key constraint extraction, generating structured requirement data;

[0052] The data correction module is used to input structured requirement data into the second-round large language model for parameter slot extraction and standardization correction, generating standardized query data;

[0053] The parallel retrieval module is used to trigger at least two independent retrieval agents in parallel based on standardized query data, and obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively;

[0054] The result fusion module is used to merge, deduplicate, and detect conflicts in the candidate material sets returned by each retrieval agent to generate a fused candidate list.

[0055] The refinement and reordering module is used to input the fusion candidate list into the fourth round of the large language model to refine and reorder the multi-source results and generate a refined candidate list.

[0056] The recommendation generation module is used to input the refined candidate list into the fifth round of the large language model, and combine material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

[0057] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0058] Fourthly, the present invention provides an electronic device, comprising:

[0059] Memory, used to store computer programs / instructions;

[0060] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0061] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0062] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0063] 1. This invention provides an intelligent material selection method and system based on a large model. By receiving natural language descriptions of working conditions and performing automated demand analysis, multi-source knowledge retrieval, and result refinement through multiple rounds of large language model, it directly outputs interpretable material recommendations, shortening the traditional selection process that relies on human experience and takes hours or even days to minutes.

[0064] 2. This invention utilizes a multi-round, progressive refinement of a large language model to ensure accurate understanding of requirements, and triggers parallel heterogeneous retrieval of knowledge graphs, structured databases, and vector libraries. After fusion sorting and conflict detection, it achieves deep coverage and accurate matching of complex working conditions, implicit relationships, and long-tail materials. Attached Figure Description

[0065] Figure 1 This is an overall execution flowchart of the intelligent material selection system based on a large model provided in this embodiment of the invention;

[0066] Figure 2 This is a flowchart illustrating the intelligent material selection method based on a large model provided in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0068] Example 1: This example introduces a smart material selection method based on a large model, including:

[0069] Obtain the user's input in natural language to describe the working conditions;

[0070] The natural language description of the working condition is input into the first round of large language model for intent recognition and key constraint extraction, generating structured requirement data;

[0071] The structured demand data is input into the second round of the large language model for parameter slot extraction and standardization correction to generate standardized query data;

[0072] Based on standardized query data, at least two independent retrieval agents are triggered in parallel to obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively;

[0073] The candidate material sets returned by each retrieval agent are merged, deduplicated, and conflict detected to generate a merged candidate list;

[0074] The fusion candidate list is input into the fourth round of the large language model to refine and reorder the multi-source results, generating a refined candidate list.

[0075] The refined candidate list is input into the fifth round of the large language model, which combines material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

[0076] This embodiment provides an intelligent material selection system based on multi-round large language model-driven, multi-agent parallel execution, and hybrid knowledge storage. The system achieves a closed-loop end-to-end process from natural language description of operating conditions to highly interpretable material recommendation. Its core innovation lies in the complete technical framework of "five-round LLM progressive refinement + three types of agent parallel recall + anomaly logic routing + feedback closed loop." The overall execution flow is as follows: Figure 1 As shown.

[0077] 1. Hybrid Knowledge Storage Layer

[0078] PostgreSQL Structured Knowledge Base (including pgvector extension): Stores material master data, parameter tables, supplier information, and historical selection cases, supporting exact / range / fuzzy SQL queries.

[0079] Neo4j graph database: Stores material entities, operating condition entities, parameter entities, and more than 30 types of relationships, including compatibility, substitutability, robust coupling, historical sharing, and failure cases. It supports multi-hop Cypher / GQL inference.

[0080] Vector knowledge base (PG pgvector or Milvus): Stores multimodal embedded vectors (text uses BGE-large-zh, images use CLIP-ViT-L-14) of material manual text paragraphs, drawing instructions, and inspection reports, and supports Top-K semantic similarity retrieval.

[0081] 2. Automatic Knowledge Construction and Update Module

[0082] Information is extracted periodically or incrementally from PDF manuals, CAD drawings, inspection reports, and historical ERP BOMs.

[0083] Use LLM (Qwen-Max or Llama3-70B) in conjunction with LayoutLMv3 for OCR and table structure restoration.

[0084] A two-stage entity-relation extraction Prompt is used to automatically generate triples and write them to Neo4j and PostgreSQL.

[0085] Once new materials are added or user feedback is confirmed, an incremental update of the knowledge graph is automatically triggered.

[0086] 3. User Interaction and Input

[0087] It supports three input methods: natural language input, voice input, and BOM table upload.

[0088] 4. Five-round LLM progressive refining process

[0089] Round 1: Intent Recognition and Requirement Understanding

[0090] The Prompt template guides the LLM to recognize user intent (new material selection / alternative material search / BOM optimization) and outputs structured JSON.

[0091] Output example:

[0092] The system, through the first round of large language model processing, identified the user's intent as "new material selection", determined the material category as "sealing ring", and extracted key constraints, including: temperature resistance not lower than 150℃, acid and alkali resistance, working pressure not lower than 10MPa, and the medium as seawater.

[0093] Second round: Slot extraction and standardization calibration

[0094] The JSON Schema is used to force the output, normalize the user's ambiguous expression, such as "resistant to seawater corrosion" → "resistant to chloride ion corrosion + resistant to damp heat", and complete the implicit parameters.

[0095] Third round: Multi-Agent task allocation and parallel recall

[0096] LLM acts as a central scheduler, dynamically launching the following agents (up to 3 simultaneously):

[0097] KG Agent: Input standardized entities and relational constraints, perform up to 5 hops of graph reasoning, and return a candidate list and reasoning path.

[0098] SQL Agent: Transforms slots into SQL conditions, executes exact / range queries, and returns Top-50 candidates.

[0099] Vector Agent (optional): After expanding the user's original description and standardized slots using HyDE assumptions, it performs vector retrieval and returns historical success cases and manual paragraphs.

[0100] Each Agent type independently returns a list of candidate material IDs, a matching score, and a preliminary explanation.

[0101] Round Four: Multi-Source Result Fusion and Conflict Refinement

[0102] After merging and deduplicating the results from the three agents, input them into the LLM function:

[0103] Conflict detection (e.g., SQL hit but KG inference excludes it).

[0104] Secondary data completion for slots, such as automatically triggering external supplier manual searches to complete missing data.

[0105] Re-evaluate (0~100 points) and output the merged Top-20 candidates.

[0106] Fifth round: Generate interpretable recommendations

[0107] Prompt requires each recommendation to include:

[0108] Material model, supplier, unit price, inventory, and historical reuse count.

[0109] Match score and detailed explanation (e.g., "Exact match from SQL Agent for temperature and pressure resistance; inference from KG Agent for seawater corrosion resistance").

[0110] Risk warning (e.g., "This material has a long service life of only 8000 hours at 180℃, so we recommend paying attention").

[0111] Top-3 alternatives.

[0112] 5. Exception Logic Routing Module

[0113] If there are fewer than 3 candidates or the highest score is less than 70 points after the fourth round, the following fallback strategy will be automatically triggered (in order of priority):

[0114] Parameters relaxed by 10%~20% and then recalled.

[0115] Cross-category substitution reasoning (e.g., no solution for sealing rings → recommend metal C-rings).

[0116] Call the external vendor catalog (the latest version cached locally).

[0117] If no solution is found, the system will output "Suggestion for New Material Development" and generate a draft technical requirements document.

[0118] 6. Results Display and Feedback Closed-Loop Module

[0119] The front end displays the recommendation results in the form of cards, including a matching progress bar, an expandable reasoning path, and a one-click access to the original manual PDF page.

[0120] Users can provide five-star ratings or point out errors. Feedback data is written to the experience base in real time and automatically triggers knowledge graph updates and prompts for fine-tuning the following day.

[0121] 7. System Deployment and Security Module

[0122] End-to-end local deployment (K8s + Ollama / Qwen-Max quantization model).

[0123] All enterprise data remains within the domain, supporting operation in an air-isolated environment.

[0124] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0125] like Figure 2 As shown in the figure, the application process of the intelligent material selection method based on a large model provided in this embodiment includes the following steps:

[0126] 1. User Scenarios and Input

[0127] Scenario: Selection of dynamic seal for the main shaft of ultra-high pressure seawater injection pump on a deep-water platform in the South China Sea.

[0128] Operating conditions: pressure 28MPa, medium is high chloride ion filtered seawater + trace amount of associated crude oil, peak temperature 158℃, speed about 3000rpm, shaft diameter 90mm.

[0129] Natural language input: "Urgent! Help me select a mechanical seal for the main shaft. The operating conditions are a seawater injection pump, 28 MPa, maximum temperature of 158 degrees Celsius, with a little bit of crude oil mixed in, speed of around 3000 rpm, and a shaft diameter of 90 mm. The imported John Crane seal used to crack frequently, and the domestic one kept leaking. The lifespan should be at least 3 years. I need the drawing by tomorrow morning!"

[0130] 2. System's Five-Wheel LLM Core Processing Details

[0131] Round-1: Intent recognition and activation of invalid knowledge;

[0132] The intent classification was completed using the locally deployed Qwen-Max-72B model, combined with the industry-specific System Prompt (including the group's 28MPa seawater pump failure mode knowledge base).

[0133] Round-2: Slot depth analysis and implicit constraint auto-completion;

[0134] The Function Calling + Pydantic JSON Schema is used to force the output of 57 slots.

[0135] The core algorithm formulas include:

[0136] Real-time calculation of PV value:

[0137] ;

[0138] in, This represents the pressure-velocity product of the sealing system, used to assess the workload of the sealing surface. or Indicates system workload. Indicates shaft diameter, Indicates shaft speed. Pi is the constant value of a circle, and 60 × 1000 in the denominator is the conversion factor between rotational speed and unit. =28 MPa =90 mm, =2980 rpm → PV ≈ 243.6 bar·m / s >230 bar·m / s threshold, forcibly activate the balanced constraint.

[0139] Material corrosion resistance scoring function:

[0140] ;

[0141] in, This indicates the overall corrosion resistance score of a material under a specific corrosive environment. The rating indicates the material's resistance to chloride ion corrosion. The score indicates the material's resistance to hydrogen sulfide corrosion. The score indicates the material's resistance to corrosion from oily media. The weighting coefficients for the three capabilities mentioned above are respectively, satisfying... The adjustment is made dynamically based on the composition of the medium under operating conditions.

[0142] Only Hastelloy C-276 and Inconel 718 with scores >0.95 are whitelisted.

[0143] Negative sample weighting for failure risk:

[0144] ;

[0145] in, Indicates candidate materials The failure risk assessment value is determined by the value; a higher value indicates a higher risk. This represents the total number of historical failure cases. Indicates the first The severity level of each failure case, Indicates the current operating condition and the first The similarity of individual failure cases.

[0146] Round-3: Multi-Agent Parallel Heterogeneous Recall

[0147] Technical framework: LangChain + LangGraph (directed acyclic graph orchestration) + Ray distributed parallel execution.

[0148] KG Agent: Employs a weighted improved Dijkstra + Beam Search (beam=10) for a maximum of 5-hop inference. Path score formula:

[0149] ;

[0150] in, This represents the overall credibility score of a reasoning path within a knowledge graph. This represents the total number of entity relationships in the path. Indicates the first in the path The weight of each negative relationship; Indicates the first in the path The weight of positive relationships;

[0151] SQL Agent: Dynamically generates SQL with window functions based on Llama3-70B-8bit + SQLGlot parser.

[0152] Vector Agent: Employs Milvus 2.4 + BGE-m3 multilingual encoder + HyDE hypothesis expansion + Cohere Rerank, rearranging the formula:

[0153] ;

[0154] in, This represents the final fusion score of the vector retrieval results. This represents the similarity score of dense vectors. This represents the sparse retrieval score based on the BM25 algorithm. This represents the score after secondary sorting using the rearrangement model. The fusion weight coefficients for the above three scores are respectively, satisfying the following conditions: .

[0155] Round-4: Multi-source result fusion and automatic conflict resolution

[0156] The fusion framework consists of LightGBM Ranker (offline training) + online bge-reranker-v2 + the rules engine Drools.

[0157] Three-stage fusion algorithm:

[0158] 1. Weighted RRF (Reciprocal Rank Fusion):

[0159] ;

[0160] in, Indicates material The fusion ranking score in multi-way recall results It is a smoothing constant. =60; Indicates material In the Ranking in the recall results; This represents the set of all call circuits.

[0161] 2. bge-reranker cross-coding for secondary reordering.

[0162] 3. Rules + LLM Final Approval: Load 86 Drools rules to generate the final ruling and interpretation.

[0163] Round-5: Explainable Recommendation Generation and Visualization of Multi-Source Inference Chains

[0164] The front-end uses React 18 + AntV G6 (relationship graph) + Monaco Editor (SQL highlighting) + PDF.js (manual location).

[0165] Each recommendation is required to include a four-layer evidence structure:

[0166] SQL exact match (highlighting parameters that exceed the limit).

[0167] Complete visualization of the KG path.

[0168] Vector Case Study: MTBF Curve (ECharts Dynamic Rendering)

[0169] Failure probability Bayesian update analysis:

[0170] ;

[0171] in, This represents the posterior probability of material failure given the evidence. This represents the likelihood probability of observing the current evidence in the event of material failure. This represents the prior probability of material failure. This represents the marginal probability of observing the current evidence.

[0172] 3. System's final recommendation results (Top-1 example)

[0173] Material Code: MAT-2024-11892

[0174] Model: HAST-CW (API 682 Category 3 Balanced Cartridge Mechanical Seal)

[0175] Key technological highlights:

[0176] Corrugated pipe: Hastelloy C-276 laser welded, with a chloride ion stress corrosion resistance life of >100,000h.

[0177] Hard surface combination: WC-10Ni self-lubricating coating vs. oil-resistant high-purity graphite, PV limit >350 bar·m / s.

[0178] Auxiliary seal: DuPont Kalrez 6375 low-temperature flexible FFKM.

[0179] Verified number of operational units: 73 (cumulative operation time exceeding 1.8 million hours, zero leakage).

[0180] 4. Feedback loop and instant evolution of knowledge graph

[0181] The knowledge graph update (Neo4j Cypher write) is completed in less than 2 seconds after user confirmation.

[0182] Relationship weights are dynamically updated (a PageRank variant): weight +0.18.

[0183] QLoRA fine-tuning started at 22:00 that night (only the reranker and SQL generation header were updated, and the dataset was updated with 1 positive sample).

[0184] The ERP interface pushes suggestions for maintaining inventory.

[0185] This embodiment fully demonstrates that under the most demanding working conditions in heavy industry, the present invention achieves complete capabilities from natural language to second-level accuracy, traceability, and closed-loop optimization through a technical solution of five rounds of progressive LLM refinement + multi-agent heterogeneous parallel recall (LangGraph + Ray) + knowledge graph deep reasoning + multi-source interpretable fusion (RRF + bge-reranker + Drools) + real-time feedback evolution (LoRA + graph dynamic weights). This completely solves the industry pain points of traditional material selection, which rely on personal experience, are inefficient, and have hidden risks that are difficult to uncover.

[0186] Example 2: This example provides an intelligent material selection system based on a large model, including:

[0187] The user input module is used to obtain the natural language description of the working conditions input by the user;

[0188] The requirement understanding module is used to input natural language descriptions of working conditions into the first-round large language model for intent recognition and key constraint extraction, generating structured requirement data;

[0189] The data correction module is used to input structured requirement data into the second-round large language model for parameter slot extraction and standardization correction, generating standardized query data;

[0190] The parallel retrieval module is used to trigger at least two independent retrieval agents in parallel based on standardized query data, and obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively;

[0191] The result fusion module is used to merge, deduplicate, and detect conflicts in the candidate material sets returned by each retrieval agent to generate a fused candidate list.

[0192] The refinement and reordering module is used to input the fusion candidate list into the fourth round of the large language model to refine and reorder the multi-source results and generate a refined candidate list.

[0193] The recommendation generation module is used to input the refined candidate list into the fifth round of the large language model, and combine material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

[0194] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0195] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0196] Example 4: This example provides an electronic device, including:

[0197] Memory, used to store computer programs / instructions;

[0198] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0199] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.

[0200] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0201] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0202] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0203] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0204] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A smart material selection method based on a large model, characterized in that, include: Obtain the user's input in natural language to describe the working conditions; The natural language description of the working condition is input into the first round of large language model for intent recognition and key constraint extraction, generating structured requirement data; The structured demand data is input into the second round of the large language model for parameter slot extraction and standardization correction to generate standardized query data; Based on standardized query data, at least two independent retrieval agents are triggered in parallel to obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively; The candidate material sets returned by each retrieval agent are merged, deduplicated, and conflict detected to generate a merged candidate list; The fusion candidate list is input into the fourth round of the large language model to refine and reorder the multi-source results, generating a refined candidate list. The refined candidate list is input into the fifth round of the large language model, which combines material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

2. The intelligent material selection method based on a large model according to claim 1, characterized in that, The process of inputting the natural language description into the first-round large language model for intent recognition and key constraint extraction to generate structured requirement data includes: The user intent is identified through the first round of large language modeling, and the intent includes at least one of new material selection, alternative material search, or BOM optimization. Extract the key constraints from the natural language working condition description and output them in structured JSON format.

3. The intelligent material selection method based on a large model according to claim 2, characterized in that, The process of inputting structured demand data into the second-round large language model for parameter slot extraction and standardization correction to generate standardized query data includes: The key constraints are normalized and implicit parameters are completed using a second round of large language modeling. Outputs standardized query data that conforms to the preset JSON Schema.

4. The intelligent material selection method based on a large model according to claim 3, characterized in that, The method, based on standardized query data, triggers at least two independent retrieval agents in parallel, including: The knowledge graph retrieval agent is triggered, and based on the entity and relation constraints in the standardized query data, multi-hop graph reasoning is performed to return the first candidate material set and reasoning path; Trigger the structured query agent to convert the standardized query data into an SQL query statement, execute an exact or range query from the relational database, and return the second candidate material set; The vector retrieval agent is triggered to perform semantic expansion on the natural language working condition description and standardized query data, and then retrieves similar cases and document paragraphs from the vector database to return a third candidate material set.

5. The intelligent material selection method based on a large model according to claim 4, characterized in that, The process of merging, deduplicating, and detecting conflicts in the candidate material sets returned by each retrieval agent to generate a merged candidate list includes: The candidate material sets returned by each retrieval agent are deduplicated; Detect conflicts between candidate materials from different sources and record the conflict type and confidence level; The candidate materials are initially merged and sorted based on the weighted Reciprocal Rank Fusion algorithm to generate a fusion candidate list.

6. The intelligent material selection method based on a large model according to claim 5, characterized in that, The step of inputting the fusion candidate list into the fourth round of the large language model for multi-source result refinement and re-ranking to generate a refined candidate list includes: The fusion candidate list is input into the fourth round of the large language model, and secondary completion and verification are performed in combination with the conflict detection results. The candidate materials are cross-coded and scored using a re-ranking model, and then sorted using a rule engine to generate a refined candidate list.

7. The intelligent material selection method based on a large model according to claim 6, characterized in that, The process involves inputting the refined candidate list into the fifth-round large language model, combining material attributes, historical usage data, and risk knowledge to generate interpretable recommendation results, including: The fifth round of large language model generates recommendation reasons for each candidate material. The recommendation reasons include matching parameters, graph reasoning path, historical usage statistics and risk warnings. The output includes structured recommendation results containing material model, supplier, matching score, recommendation reason, and alternative solutions.

8. The intelligent material selection method based on a large model according to claim 7, characterized in that, After outputting the recommendation results, it also includes: If the number of candidate materials is less than the preset threshold or the highest matching score is lower than the preset score, the abnormal logic routing process will be automatically triggered, and at least one of the following operations will be performed: parameter relaxation, cross-category substitution retrieval, or external directory query.

9. A large-scale model-based intelligent material selection system, used to implement the large-scale model-based intelligent material selection method according to any one of claims 1-8, characterized in that, include: The user input module is used to obtain the natural language description of the working conditions input by the user; The requirement understanding module is used to input natural language descriptions of working conditions into the first-round large language model for intent recognition and key constraint extraction, generating structured requirement data; The data correction module is used to input structured requirement data into the second-round large language model for parameter slot extraction and standardization correction, generating standardized query data; The parallel retrieval module is used to trigger at least two independent retrieval agents in parallel based on standardized query data, and obtain candidate material sets from the knowledge graph, structured knowledge base and vector knowledge base respectively; The result fusion module is used to merge, deduplicate, and detect conflicts in the candidate material sets returned by each retrieval agent to generate a fused candidate list. The refinement and reordering module is used to input the fusion candidate list into the fourth round of the large language model to refine and reorder the multi-source results and generate a refined candidate list. The recommendation generation module is used to input the refined candidate list into the fifth round of the large language model, and combine material attributes, historical usage data and risk knowledge to generate interpretable recommendation results and output them.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-8.

Images