Intelligent manufacturing quality optimization method based on large model, privacy calculation and knowledge graph

Abstract

The invention discloses an intelligent manufacturing quality optimization method based on a large model, privacy calculation and a knowledge graph. The intelligent manufacturing quality optimization method comprises the steps of privacy acquisition and preprocessing of multi-source heterogeneous quality data, dynamic construction of the quality experience knowledge graph, knowledge enhancement quality reasoning of a generative large model and closed-loop feedback and adaptive evolution. According to the intelligent manufacturing quality optimization method based on the large model, the privacy calculation and the knowledge graph, the quality control problems of contradiction between data privacy and collaborative analysis, insufficient knowledge association depth, limited reasoning ability and the like in intelligent manufacturing are solved; by constructing a full-link technical system of'data privacy processing-knowledge graph construction and retrieval enhancement-multi-modal hybrid reasoning-closed loop evolution ', high-quality, safe and efficient intelligent manufacturing is realized, accurate diagnosis of quality problems, experience reuse and efficient optimization are realized, and the intelligent manufacturing quality control level and efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to an intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs. Background Technology

[0002] Currently, quality control in the field of intelligent manufacturing faces three core pain points that existing technologies cannot effectively address, as follows:

[0003] (1) Insufficient data privacy protection: Multi-source heterogeneous quality data contains sensitive information, and direct use can easily lead to privacy leaks. Existing privacy computing technologies are mostly applied to the data transmission stage and do not cover the data collection stage, so they cannot achieve "collection is privacy protection". Moreover, a single encryption scheme is difficult to adapt to the privacy requirements of different types of data such as real-time data and text documents.

[0004] (2) Weak knowledge association and updating capabilities: Quality experience is mostly stored in fragmented documents, lacking structured association; Traditional knowledge graph construction is mostly static, without dynamic updating of entity relationships through causal analysis, which makes traditional association analysis prone to false associations and low root cause localization accuracy.

[0005] (3) Poor reasoning efficiency and evolutionary ability: Existing quality reasoning models rely on a single data or knowledge source. When faced with the mixed requirements of "data query + knowledge association", they cannot take into account both data support and knowledge association, resulting in insufficient reasoning efficiency and accuracy. Furthermore, they lack a closed-loop optimization mechanism and cannot continuously evolve with the accumulation of quality experience.

[0006] In summary, existing technologies have not formed a complete solution encompassing "privacy protection, knowledge construction, hybrid reasoning, and closed-loop evolution." There is an urgent need for a quality optimization method that integrates multiple technologies and includes hybrid retrieval to address the aforementioned issues. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for optimizing intelligent manufacturing quality based on large models, privacy computing, and knowledge graphs. By constructing a full-link technology system of "data privacy processing - knowledge graph construction and retrieval enhancement - multimodal hybrid reasoning - closed-loop evolution", this invention effectively solves the pain points and problems in quality control of intelligent manufacturing and achieves high-quality, safe, and efficient intelligent manufacturing.

[0008] The technical solution to achieve the above objectives is: a smart manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs, comprising the following steps:

[0009] S1, Privacy-prepared collection and preprocessing steps for multi-source heterogeneous quality data: Construct a layered privacy protection architecture of "edge-cloud collaboration", put privacy protection in advance at the data collection stage, and use local differential privacy technology and dynamic masking technology to protect privacy for real-time device operation data and text data one by one, so as to realize privacy protection during data collection.

[0010] S2, the dynamic construction steps of the quality experience knowledge graph: combining causal analysis and vector model secondary fine-tuning, improving the accuracy of association through Granger causal model, and strengthening entity semantic capture through secondary encoding fine-tuning, to build a dynamically updated knowledge graph that can be automatically updated every 4 hours;

[0011] S3, Knowledge Enhancement Quality Inference Steps of Generative Large Model: Construct a knowledge-enhanced generative large model of "multimodal fusion + hybrid retrieval", classify intent through multimodal features, trigger hybrid retrieval, and integrate visual and text features to solve the hybrid needs of "data + knowledge";

[0012] S4, Closed-loop feedback and adaptive evolution steps: Establish a "feedback-evaluation-optimization" closed loop to achieve continuous evolution of quality experience reasoning.

[0013] The above-mentioned intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs, wherein the specific steps of S1 are as follows:

[0014] S11, Real-time Data Acquisition Steps: Collect real-time operating data of temperature and resistance of temperature sensing elements through edge nodes, add Gaussian noise using local differential privacy technology, and only upload the desensitized feature vector to obtain privacy-protected real-time data of the device.

[0015] S12, Quality Inspection Document Data Collection Steps: Quality inspection documents include inspection reports and non-conforming item processing forms. Key information in each text is extracted using OCR recognition, entity alignment, and multimodal technology. Sensitive fields are hidden using dynamic masking technology to obtain privacy-protected quality inspection document data. The key information includes defect type and / or handling measures, and the sensitive fields include supplier name and batch number.

[0016] S13, Data Standardization Preprocessing Step: The privacy-protected real-time device data collected in step S11 is transmitted via the MQTT protocol, and the privacy-protected quality inspection document data collected in step S12 is transmitted via the RESTful interface. After cleaning and removing outliers, standardizing and unifying the units, and associating features and labels, a structured quality dataset containing "device identifier, feature vector, defect label, and processing record" is formed.

[0017] The above-mentioned intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs, wherein the specific steps of S2 are as follows:

[0018] S21, Core Entity and Relationship Definition Steps: Define 8 types of core entities and 6 types of key relationships to clarify the basic framework of the knowledge graph; the 8 types of core entities include defect type, equipment model, process parameters, material batch, handling measures, operator / de-identification label, production environment, and quality indicators; the 6 types of key relationships include "Defect type - Cause of - Abnormal process parameters", "Handling measures - Applicable scenario - Equipment model", "Defect type - Associated material - Material batch", "Production environment - Impact - Stability of process parameters", "Quality indicators - Associated detection - Defect type", and "Operator - Execution - Handling measures";

[0019] S22, the knowledge extraction and dynamic update steps, are as follows:

[0020] S221, Causal Association Extraction: Based on the privacy-protected real-time equipment data collected in step S11, the causal association between "time-series changes in process parameters" and "defect occurrence" is examined using the Granger causal model to generate implicit association knowledge.

[0021] S222, Secondary Encoding Fine-tuning: Using Milvus as the vector database, the privacy-protected quality inspection document data collected in step S12 is used to generate an initial semantic vector through the BGE vector model; the position and type codes of "defects and process parameters" entities are extracted by Named Entity Recognition and merged with the original BGE semantic code to form a secondary encoding of "semantics + entity attributes"; based on this secondary encoding, the BGE vector model is modified and fine-tuned, and the entity relationship extraction logic is optimized by combining the Langgraph framework, so that the F1-score of the "defect-cause-measure" triplet extraction is ≥0.92 and can be directly mapped to the neo4j graph database;

[0022] S223, Real-time Dynamic Updates: Receives the structured quality dataset from step S13 every 4 hours, automatically adds entity nodes; merges duplicate nodes using cosine similarity with a threshold of 0.25 to ensure graph consistency; synchronously updates the Milvus vector database and neo4j graph database to guarantee knowledge timeliness.

[0023] The above-mentioned intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs includes the following specific steps in S3:

[0024] S31, Feature extraction step, specifically:

[0025] S311, Semantic Feature Extraction: Extracts 1024-dimensional semantic vectors from user input using a BERT fine-tuning model, and uses the Milvus vector database to pre-vectorize and index the knowledge graph text, supporting fast semantic retrieval;

[0026] S312, Entity Feature Extraction: Reuse the named entity recognition capability in step S2 to extract entities of "time, device ID, and defect type". If it is an entity of "time and value", the weight of "structured query" will be enhanced. If it is an entity of "defect type", the weight of "knowledge graph retrieval" will be enhanced.

[0027] S313, Sentence Structure Feature Extraction: Differentiating structured requests such as "query / statistics" through syntactic analysis.

[0028] The demand for knowledge graphs related to "root cause tracing";

[0029] S32, Intent Classification and Hybrid Retrieval Steps, are as follows:

[0030] S321, concatenates the "semantic + entity + sentence structure" features into a multimodal vector, inputs it into a lightweight CNN model, and outputs the structured query probability P1 and the knowledge graph retrieval probability P2;

[0031] S322, When the difference between the structured query probability P1 and the knowledge graph retrieval probability P2 is <0.2, and the user input contains a "data + knowledge" expression or a mixed input of "image + text", the mixed retrieval of structured query and knowledge graph retrieval is triggered.

[0032] Structured queries retrieve statistical results from structured databases, and Elasticsearch fuzzy search quickly filters irrelevant text noise, along with Milvus semantic search knowledge.

[0033] Knowledge graph retrieval uses a combination of dense and sparse semantic recall to retrieve knowledge from the neo4j graph database. After being sorted by RRF and rearranged by BGE-Reranker, it matches knowledge related to "defect-cause-measure".

[0034] S323, a generative large model integrates the hybrid retrieval results of structured queries and knowledge graph retrieval to output a solution;

[0035] S33, the multimodal knowledge-enhanced reasoning steps are as follows:

[0036] S331: After the user inputs a question, similar cases are retrieved from the knowledge graph, and the cases are combined with the original question to form an input sequence.

[0037] S332 extracts visual features from defect / equipment status images, fuses them with text features through a cross-modal attention mechanism, and inputs them into the fusion model of "Qwen3-7B language model + SigLIP visual model + feature mapping network".

[0038] S333: During training, it is first pre-trained with "defect image-text" and "equipment image-process text", then the visual layer parameters are frozen, and the language model is fine-tuned with quality-specific data; during inference, it generates results containing "root cause location, treatment measures, prevention plan, and effect prediction", and the confidence level of root cause location is ≥0.85.

[0039] The above-mentioned intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs includes the following specific steps in S4:

[0040] S41, Multi-dimensional feedback collection steps: Operators use terminals to give the solution a "1-5 star quantitative rating + qualitative annotation"; compare with production data to generate a "0-100 point measure effectiveness score", forming subjective + objective feedback;

[0041] S42, Knowledge Graph Optimization Steps: Adjust entity relationship weights based on subjective and objective feedback, add new causal relationships, and update the neo4j graph database synchronously;

[0042] S43, Generative Large Model Optimization Steps: Incremental fine-tuning is triggered every 100 accumulated valid feedbacks. Drawing on the "semantic + entity attribute" secondary encoding approach in step S2, the generative large model's understanding of entity semantics is enhanced, and high-accuracy output is rewarded through reinforcement learning.

[0043] S44, Privacy and Security Audit Steps: Record privacy computing operation logs, knowledge graph change logs, and model training logs, retain them for 2 years to ensure compliance.

[0044] The above-mentioned intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs includes, in step S44, the privacy computing operation log including encryption algorithms, parameters, and executors; the knowledge graph change log including node additions, deletions, and relationship adjustments; and the model training log including dataset sources and iteration metrics.

[0045] The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs in this invention has the following advantages compared with existing technologies:

[0046] (1) Synergy between privacy protection and data utilization: Through the "collection is privacy protection" architecture, a balance is achieved between the privacy and security of multi-source data and its effective utilization, solving the pain point of "not daring to use data";

[0047] (2) Improved knowledge association and accuracy: The dynamic knowledge graph, combined with causal analysis and secondary coding fine-tuning, makes the F1-score of triple extraction ≥ 0.92 and the root cause location reliability ≥ 0.85, thus solving the pain point of "knowledge not being connected";

[0048] (3) Hybrid reasoning and efficiency optimization: The multimodal hybrid retrieval mechanism meets the needs of "data query + knowledge association", and the reasoning response speed is improved by more than 30% compared with the traditional model, solving the pain point of "inaccurate reasoning";

[0049] (4) Closed-loop evolution and experience reuse: Continuous evolution is achieved through feedback optimization, reducing the recurrence rate of quality problems by more than 25%, and realizing the whole process management of quality experience "generation-accumulation-reuse-evolution". Attached Figure Description

[0050] Figure 1 This is a flowchart of the intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to the present invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand the technical solution of the present invention, its specific embodiments are described in detail below with reference to the accompanying drawings:

[0052] Please see Figure 1 The preferred embodiment of the present invention, a method for optimizing intelligent manufacturing quality based on large models, privacy computing, and knowledge graphs, includes the following steps:

[0053] S1, Privacy-preserving collection and preprocessing steps for multi-source heterogeneous quality data: Addressing the limitation of existing privacy computing applications only in the transmission stage, a layered privacy protection architecture based on "edge-cloud collaboration" is constructed to achieve "collection equals privacy protection." Specific steps include:

[0054] S11, Real-time data acquisition steps: Real-time operating data of the device, such as temperature and resistance of the temperature measuring element, are collected through edge nodes. Gaussian noise is added using local differential privacy (LDP) technology, and only the desensitized feature vector is uploaded to obtain privacy-protected real-time device data.

[0055] S12, Quality Inspection Document Data Collection Steps: This includes semi-structured texts such as inspection reports and non-conformance handling sheets. Key information such as "defect type and handling measures" is extracted using OCR recognition, entity alignment, and multimodal technology. Sensitive fields such as supplier name and batch number are hidden using dynamic masking technology to obtain privacy-protected quality inspection document data.

[0056] S13, Data Standardization Preprocessing Step: The privacy-protected real-time device data (such as temperature / resistance of temperature measuring elements) collected in step S11 is transmitted via the MQTT protocol. The privacy-protected quality inspection document data collected in step S12 is transmitted via the RESTful interface. After cleaning and removing outliers, standardizing and unifying the units, and associating features and labels, a structured quality dataset containing "device identifier, feature vector, defect label, and processing record" is formed.

[0057] Compared with existing technologies: Existing technologies only encrypt data during the data transmission stage. This step brings privacy protection to the collection stage, using LDP and Masking technologies for real-time device data and text data respectively, to achieve "collection is protection" and adapt to the privacy needs of multiple types of data.

[0058] S2: Dynamic Construction Steps of Quality Experience Knowledge Graph: Addressing the shortcomings of existing knowledge graphs, such as static nature and low accuracy of associations, a dynamically updated knowledge graph is constructed by combining causal analysis and secondary fine-tuning of the vector model. The specific steps are as follows:

[0059] S21, Core Entity and Relationship Definition Steps: Define 8 types of core entities and 6 types of key relationships to clarify the basic framework of the knowledge graph; the 8 types of core entities include defect type, equipment model, process parameters, material batch, handling measures, operator / de-identification label, production environment, and quality indicators; the 6 types of key relationships include "defect type - cause - abnormal process parameters", "handling measures - applicable scenario - equipment model", "defect type - associated material - material batch", "production environment - impact - process parameter stability", "quality indicators - associated detection - defect type", and "operator - execution - handling measures";

[0060] S22, the knowledge extraction and dynamic update steps, are as follows:

[0061] S221, Causal Association Extraction: Based on the privacy-protected real-time equipment data collected in step S11, the causal association between "time-series changes in process parameters" and "defect occurrence" is examined using the Granger causal model to generate implicit association knowledge.

[0062] S222, Secondary Encoding Fine-tuning: Using Milvus as the vector database, the privacy-protected quality inspection document data collected in step S12 is used to generate initial semantic vectors through the BGE vector model; combined with Named Entity Recognition (NER) to extract the position and type codes of "defects and process parameters" entities, these are fused with the original BGE semantic codes to form a secondary encoding of "semantics + entity attributes"; based on this secondary encoding, the BGE vector model is modified and fine-tuned, and the entity relationship extraction logic is optimized using the Langgraph framework, so that the F1-score of the "defect-cause-measure" triple extraction is ≥0.92, and

[0063] It can be directly mapped to the neo4j graph database;

[0064] S223, Real-time Dynamic Updates: Receives the structured quality dataset from step S13 every 4 hours, automatically adds entity nodes; merges duplicate nodes using cosine similarity (threshold 0.25) to ensure graph consistency; synchronously updates the Milvus vector database and neo4j graph database to guarantee knowledge timeliness.

[0065] Compared with existing technologies: Most existing knowledge graphs are statically constructed. This step improves the accuracy of associations through the Granger causal model, strengthens the capture of entity semantics through secondary encoding fine-tuning, and achieves automatic updates every 4 hours, thus solving the problem of "knowledge lag" in static graphs.

[0066] S3: Knowledge-Enhanced Quality Inference Steps for Generative Large Models: Addressing the limitation of existing models in simultaneously handling "data query + knowledge association," a knowledge-enhanced generative large model based on "multimodal fusion + hybrid retrieval" is constructed. The specific steps are as follows:

[0067] S31, Feature extraction step, specifically:

[0068] S311, Semantic Feature Extraction: Extracts 1024-dimensional semantic vectors from user input using a BERT fine-tuning model, and uses the Milvus vector database to pre-vectorize and index the knowledge graph text, supporting fast semantic retrieval;

[0069] S312, Entity Feature Extraction: Reuse the NER (Named Entity Recognition) capability in step S2 to extract entities of "time, device ID, and defect type". If the entity is "time" or "value", the weight of "structured query" will be enhanced. If the entity is "defect type", the weight of "knowledge graph retrieval" will be enhanced.

[0070] S313, Sentence Structure Feature Extraction: Differentiating structured requests such as "query / statistics" through syntactic analysis.

[0071] The demand for knowledge graphs related to "root cause tracing";

[0072] S32, Intent Classification and Hybrid Retrieval Steps, are as follows:

[0073] S321, concatenates the "semantic + entity + sentence structure" features into a multimodal vector, inputs it into a lightweight CNN model, and outputs the structured query probability P1 and the knowledge graph retrieval probability P2;

[0074] S322, When the difference between the structured query probability P1 and the knowledge graph retrieval probability P2 is <0.2, and the user input contains a "data + knowledge" expression or a mixed input of "image + text", the mixed retrieval of structured query and knowledge graph retrieval is triggered.

[0075] Structured queries retrieve statistical results from structured databases, and Elasticsearch fuzzy search quickly filters irrelevant text noise, along with Milvus semantic search knowledge.

[0076] Knowledge graph retrieval uses a combination of dense and sparse semantic recall to retrieve knowledge from the neo4j graph database. After being sorted by RRF and rearranged by BGE-Reranker, it matches knowledge related to "defect-cause-measure".

[0077] S323, a generative large model integrates the hybrid retrieval results of structured queries and knowledge graph retrieval to output a solution;

[0078] S33, the multimodal knowledge-enhanced reasoning steps are as follows:

[0079] S331: After the user inputs a question, similar cases are retrieved from the knowledge graph, and the cases are combined with the original question to form an input sequence.

[0080] S332 extracts visual features from defect / equipment status images, fuses them with text features through a cross-modal attention mechanism, and inputs them into the fusion model of "Qwen3-7B language model + SigLIP visual model + feature mapping network".

[0081] S333: During training, it is first pre-trained with "defect image-text" and "equipment image-process text", then the visual layer parameters are frozen, and the language model is fine-tuned with quality-specific data; during inference, it generates results containing "root cause location, treatment measures, prevention plan, and effect prediction", and the confidence level of root cause location is ≥0.85.

[0082] Compared with existing technologies: Existing models rely on single retrieval, while this step classifies intent through multimodal features, triggers hybrid retrieval, and integrates visual and textual features to solve the hybrid needs of "data + knowledge" and improve the accuracy of reasoning.

[0083] S4: Closed-Loop Feedback and Adaptive Evolution Steps: Establish a "feedback-evaluation-optimization" closed loop to achieve continuous evolution of quality experience-based reasoning. The specific steps are as follows:

[0084] S41, Multi-dimensional feedback collection steps: Operators use terminals to give the solution a "1-5 star quantitative rating + qualitative annotation"; compare with production data to generate a "0-100 point measure effectiveness score", forming subjective + objective feedback;

[0085] S42, Knowledge Graph Optimization Steps: Adjust entity relationship weights based on subjective and objective feedback, add new causal relationships, and update the neo4j graph database synchronously;

[0086] S43, Generative Large Model Optimization Steps: Incremental fine-tuning is triggered every 100 accumulated valid feedbacks. Drawing on the "semantic + entity attribute" secondary encoding approach in step S2, the generative large model's understanding of entity semantics is enhanced, and high-accuracy output is rewarded through reinforcement learning.

[0087] S44, Privacy and Security Audit Steps: Record privacy computing operation logs (encryption algorithms, parameters, executors), knowledge graph change logs (node ​​additions and deletions, relationship adjustments), and model training logs (dataset sources, iteration metrics), with a retention period of 2 years to ensure compliance.

[0088] Compared with existing technologies, this step optimizes the knowledge graph and generative large model through feedback synchronization, realizing a positive cycle of "experience-model-effect", thereby solving the pain point of "inability to evolve".

[0089] In summary, the intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs in this invention addresses quality control challenges in intelligent manufacturing, such as the contradiction between data privacy and collaborative analysis, insufficient depth of knowledge association, and limited reasoning capabilities. By constructing a full-link technology system encompassing "data privacy processing - knowledge graph construction and retrieval enhancement - multimodal hybrid reasoning - closed-loop evolution," it effectively solves the pain points and problems in intelligent manufacturing quality control, achieving high-quality, safe, and efficient intelligent manufacturing. This enables accurate diagnosis of quality problems, reuse of experience, and efficient optimization, thereby improving the level and efficiency of intelligent manufacturing quality control.

[0090] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for optimizing intelligent manufacturing quality based on large models, privacy computing, and knowledge graphs, characterized in that, Includes the following steps: S1, Privacy-prepared collection and preprocessing steps for multi-source heterogeneous quality data: Construct a layered privacy protection architecture of "edge-cloud collaboration", put privacy protection in advance at the data collection stage, and use local differential privacy technology and dynamic masking technology to protect privacy for real-time device operation data and text data one by one, so as to realize privacy protection during data collection. S2, the dynamic construction steps of the quality experience knowledge graph: combining causal analysis and vector model secondary fine-tuning, improving the accuracy of association through Granger causal model, and strengthening entity semantic capture through secondary encoding fine-tuning, to build a dynamically updated knowledge graph that can be automatically updated every 4 hours; S3, Knowledge Enhancement Quality Inference Steps of Generative Large Model: Construct a knowledge-enhanced generative large model of "multimodal fusion + hybrid retrieval", classify intent through multimodal features, trigger hybrid retrieval, and integrate visual and text features to solve the hybrid needs of "data + knowledge"; S4, Closed-loop feedback and adaptive evolution steps: Establish a "feedback-evaluation-optimization" closed loop to achieve continuous evolution of quality experience reasoning.

2. The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Real-time Data Acquisition Steps: Collect real-time operating data of temperature and resistance of temperature sensing elements through edge nodes, add Gaussian noise using local differential privacy technology, and only upload the desensitized feature vector to obtain privacy-protected real-time data of the device. S12, Quality Inspection Document Data Collection Steps: Quality inspection documents include inspection reports and non-conforming item processing forms. Key information in each text is extracted using OCR recognition, entity alignment, and multimodal technology. Sensitive fields are hidden using dynamic masking technology to obtain privacy-protected quality inspection document data. The key information includes defect type and / or handling measures, and the sensitive fields include supplier name and batch number. S13, Data Standardization Preprocessing Step: The privacy-protected real-time device data collected in step S11 is transmitted via the MQTT protocol, and the privacy-protected quality inspection document data collected in step S12 is transmitted via the RESTful interface. After cleaning and removing outliers, standardizing and unifying the units, and associating features and labels, a structured quality dataset containing "device identifier, feature vector, defect label, and processing record" is formed.

3. The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to claim 2, characterized in that, The specific steps of S2 are as follows: S21, Core Entity and Relationship Definition Steps: Define 8 types of core entities and 6 types of key relationships to clarify the basic framework of the knowledge graph; the 8 types of core entities include defect type, equipment model, process parameters, material batch, handling measures, operator / desensitization label, production environment, and quality indicators; the 6 types of key relationships include "defect type - cause - abnormal process parameters", "handling measures - applicable scenario - equipment model", "defect type - associated material - material batch", "production environment - impact - process parameter stability", "quality indicators - associated detection - defect type", and "operator - execution - handling measures"; S22, the knowledge extraction and dynamic update steps, are as follows: S221, Causal Correlation Extraction: Based on the privacy-protected real-time device data collected in step S11, the causal relationship between "time-series changes in process parameters" and "defect occurrence" is examined using a Granger causal model. Generate implicit knowledge associations; S222, Secondary Encoding Fine-tuning: Using Milvus as the vector database, the privacy-protected quality inspection document data collected in step S12 is used to generate an initial semantic vector through the BGE vector model; the position and type codes of "defects and process parameters" entities are extracted by Named Entity Recognition and merged with the original BGE semantic code to form a secondary encoding of "semantics + entity attributes"; based on this secondary encoding, the BGE vector model is modified and fine-tuned, and the entity relationship extraction logic is optimized by combining the Langgraph framework, so that the F1-score of the "defect-cause-measure" triplet extraction is ≥0.92 and can be directly mapped to the neo4j graph database; S223, Real-time Dynamic Updates: Receives the structured quality dataset from step S13 every 4 hours, automatically adds entity nodes; merges duplicate nodes using cosine similarity with a threshold of 0.25 to ensure graph consistency; synchronously updates the Milvus vector database and neo4j graph database to guarantee knowledge timeliness.

4. The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to claim 3, characterized in that, The specific steps for S3 are as follows: S31, Feature extraction step, specifically: S311, Semantic Feature Extraction: Extracts 1024-dimensional semantic vectors from user input using a BERT fine-tuning model, and uses the Milvus vector database to pre-vectorize and index the knowledge graph text, supporting fast semantic retrieval; S312, Entity Feature Extraction: Reuse the named entity recognition capability in step S2 to extract entities of "time, device ID, and defect type". If it is an entity of "time and value", the weight of "structured query" will be enhanced. If it is an entity of "defect type", the weight of "knowledge graph retrieval" will be enhanced. S313, Sentence Feature Extraction: Differentiate between "query / statistics" type structured requirements and "root cause tracing" type knowledge graph requirements through syntactic analysis; S32, Intent Classification and Hybrid Retrieval Steps, are as follows: S321, concatenates the "semantic + entity + sentence structure" features into a multimodal vector, inputs it into a lightweight CNN model, and outputs the structured query probability P1 and the knowledge graph retrieval probability P2; S322, When the difference between the structured query probability P1 and the knowledge graph retrieval probability P2 is <0.2, and the user input contains "data + knowledge" or "image + text" mixed input, the mixed retrieval of structured query and knowledge graph retrieval is triggered; Structured queries retrieve statistical results from structured databases, and Elasticsearch fuzzy search quickly filters irrelevant text noise, along with Milvus semantic search knowledge. Knowledge graph retrieval uses "dense + sparse semantic recall" to retrieve knowledge from the neo4j graph database, sorts it using RRF and rearranges it using BGE-Reranker, and matches "defect-cause-measure" related knowledge. S323, a generative large model integrates the hybrid retrieval results of structured queries and knowledge graph retrieval to output a solution; S33, the multimodal knowledge-enhanced reasoning steps are as follows: S331: After the user inputs a question, similar cases are retrieved from the knowledge graph, and the cases are combined with the original question to form an input sequence. S332 extracts visual features from defect / equipment status images, fuses them with text features through a cross-modal attention mechanism, and inputs them into the "Qwen3-7B language model + SigLIP visual model + feature mapping network" fusion model; S333: During training, it is first pre-trained with "defect image-text" and "equipment image-process text", then the visual layer parameters are frozen, and the language model is fine-tuned with quality-specific data; during inference, it generates results containing "root cause location, treatment measures, prevention plan, and effect prediction", and the confidence level of root cause location is ≥0.

85.

5. The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to claim 4, characterized in that, The specific steps of S4 are as follows: S41, Multi-dimensional feedback collection steps: Operators use terminals to give the solution a "1-5 star quantitative rating + qualitative annotation"; compare with production data to generate a "0-100 point measure effectiveness score", forming subjective + objective feedback; S42, Knowledge Graph Optimization Steps: Adjust entity relationship weights based on subjective and objective feedback, add new causal relationships, and update the neo4j graph database synchronously; S43, Generative Large Model Optimization Steps: Trigger incremental fine-tuning every 100 accumulated valid feedbacks. Borrowing the "semantic + entity attribute" secondary encoding approach from step S2, strengthen the generative large model's understanding of entity semantics and reward high-accuracy outputs through reinforcement learning. S44, Privacy and Security Audit Steps: Record privacy computing operation logs, knowledge graph change logs, and model training logs, retain them for 2 years to ensure compliance.

6. The intelligent manufacturing quality optimization method based on large models, privacy computing, and knowledge graphs according to claim 5, characterized in that, In S44, the privacy computing operation log includes the encryption algorithm, parameters, and executor; the knowledge graph change log includes node additions, deletions, and relationship adjustments; and the model training log includes the dataset source and iteration metrics.

Images