Reliability scheme generation method fusing knowledge graph and large model

Abstract

The application discloses a reliability scheme generation method fusing a knowledge graph and a large model, belongs to the field of electronic product reliability engineering, and comprises the following steps: S1, constructing an electronic product reliability field knowledge graph containing entity attributes, relationships and dynamic constraints; S2, fusing synthetic data and real data to obtain an electronic product reliability enhanced data set; S3, introducing knowledge consistency loss for joint training to obtain an ER-KG-RAG-LLM fusion model having field knowledge anchoring and reliability scheme generation capabilities; and S4, receiving an electronic product reliability requirement and inputting the ER-KG-RAG-LLM fusion model to output a reliability scheme. The reliability scheme generation method fusing the knowledge graph and the large model realizes the full-process credibility and high efficiency of electronic product reliability schemes from knowledge retrieval to intelligent generation, and takes into account engineering accuracy and scene adaptability.

Description

A reliable scheme generation method integrating knowledge graphs and large models Technical Field

[0001] This invention relates to the field of electronic product reliability engineering technology, and in particular to a method for generating reliability solutions that integrates knowledge graphs and large models. Background Technology

[0002] As electronic products develop towards higher integration, longer lifespan, and adaptability to complex environments (such as automotive MCUs needing to withstand a wide temperature range of -40℃ to 125℃, and industrial routers needing to meet an average mean time between failures of over 50,000 hours), the design difficulty and precision requirements of reliability solutions have significantly increased.

[0003] Traditional reliability solution generation relies on the accumulated experience of domain experts, requiring manual completion of processes such as failure mode identification, environmental impact analysis, and reliability solution feasibility verification. This process is not only inefficient (e.g., a complete FMEA (Failure Mode and Effects Analysis) report requires several weeks of preparation by an expert team), but also suffers from poor knowledge reusability and insufficient consistency of reliability solutions. Summary of the Invention

[0004] The purpose of this invention is to provide a reliable scheme generation method that integrates knowledge graphs and large models, thereby solving the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides a method for generating reliable schemes that integrates knowledge graphs and large models, comprising the following steps:

[0006] S1. Integrate structured data, unstructured data and experimental data related to fault-reliability solutions after analysis and testing in the field of electronic product reliability to form the original electronic product reliability data. Use knowledge graph to extract entities and relationships of components-fault-environment-indicators-reliability solutions, and add dynamic engineering constraints to construct a knowledge graph of electronic product reliability containing entity attributes, relationships and dynamic constraints.

[0007] S2. Preprocess the original electronic product reliability data to obtain real data. Based on the real data, use WGAN-digital twin generation technology to generate synthetic data, and then merge the synthetic data with the real data to obtain the electronic product reliability enhancement dataset.

[0008] S3. The LLM model is pre-trained using corpus on the reliability of electronic products. Based on the TransE model, the entities and relations of the knowledge graph on the reliability of electronic products are transformed into low-dimensional vectors, and an ER-KG retrieval enhancement module is constructed. Then, using the enhanced dataset on the reliability of electronic products as training samples, the subgraphs of the knowledge graph on the reliability of electronic products retrieved by the ER-KG retrieval enhancement module are used as context inputs to pre-train the LLM model. Knowledge consistency loss is introduced for joint training to obtain an ER-KG-RAG-LLM fusion model with domain knowledge anchoring and reliability scheme generation capabilities.

[0009] S4. Receive the reliability requirements of electronic products, transform them into structured requirements using the NER model, and input them into the ER-KG-RAG-LLM fusion model. Combine the component-fault-reliability solution related subgraphs retrieved by the ER-KG retrieval enhancement module to generate an FMEA report containing fault modes and risk priorities. After iterative optimization and multi-dimensional verification, output the reliability solution.

[0010] Therefore, the reliability generation method that integrates knowledge graphs and large models, as described above, has the following beneficial effects:

[0011] 1. Improve the accuracy and credibility of the solution: By leveraging the structured knowledge constraints of the Electronic Product Reliability Knowledge Graph (ER-KG) and the precise retrieval of the ER-KG retrieval enhancement module (RAG module), the "knowledge illusion" of the large model is effectively reduced, ensuring that the core content such as reliability indicators and fault solutions in the generated solution are consistent with the domain knowledge graph and industry standards, and the accuracy of key item matching is improved to over 85%.

[0012] 2. Enhanced reasoning capabilities in complex scenarios: Through subgraph multi-hop relationship retrieval (such as the "environmental parameters-component failure-solution" association chain) and hybrid reasoning mechanism, it supports collaborative analysis of multi-dimensional requirements (such as temperature and humidity, power consumption, cost constraints). The success rate of multi-hop reasoning for complex reliability solutions is over 90%, which is significantly better than pure text retrieval or single knowledge graph methods.

[0013] 3. Improve solution generation efficiency and dynamic adaptability: Millisecond-level retrieval based on FAISS vector index can shorten the solution generation cycle by more than 50%; at the same time, through dynamic updating of knowledge graph and linear dimensionality reduction to adapt to different demand vectors, it can flexibly respond to scenarios such as new product types and changes in environmental parameters.

[0014] 4. Enhance the interpretability and engineering feasibility of the solutions: All generated solutions are accompanied by the ER-KG triplet source (e.g., "Heating element solution originates from the 'Low Temperature Failure - Solution' subgraph"), which supports quantitative verification of indicator deviations, and the text output format conforms to engineering document specifications, reducing the cost of manual review;

[0015] 5. Lower the threshold for domain knowledge application: Through the semantic parsing of natural language requirements and automatic textification of triples by the BERT model, users do not need to master professional knowledge graph query language to generate solutions that conform to engineering logic. At the same time, it reduces the dependence on fine-tuning of large model domains and improves the stability of adapting to different LLM models by 40%.

[0016] The technical reliability of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 is a flowchart of the reliability scheme generation method that integrates knowledge graph and large model according to the present invention. Detailed Implementation

[0018] To make the objectives, technical reliability, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0019] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] As shown in Figure 1, the reliable scheme generation method integrating knowledge graphs and large models includes the following steps:

[0022] S1. Integrate structured data, unstructured data and experimental data related to fault-reliability solutions after analysis and testing in the field of electronic product reliability to form the original electronic product reliability data. Use knowledge graph to extract entities and relationships of components-fault-environment-indicators-reliability solutions, and add dynamic engineering constraints to construct a knowledge graph of electronic product reliability containing entity attributes, relationships and dynamic constraints.

[0023] The structured data mentioned in step S1 includes reliability standards and an enterprise electronic product failure database. The enterprise electronic product failure database contains component models, failure times, environmental parameters, failure modes, component characteristics, and environmental impact patterns. The unstructured data includes expert-annotated FMEA tables and maintenance record text. The FMEA tables contain failure modes, severity S, and probability of occurrence. Detection difficulty ;

[0024] Step S1 specifically includes the following steps:

[0025] S11. Convert the raw electronic product reliability data into triplet format:

[0026] Extract entities and relationships from structured data; extract entities and relationships from unstructured data using the BERT-NER model; and transform the fault-reliability scheme association experimental data labeled after expert analysis and testing into expert constraint rule condition text using rule templates.

[0027] S12. Define the entities and relationships in the knowledge graph of electronic product reliability: Use the causal structure ontology to define entity subclasses, relationship directions and attribute data types to obtain the ontology dictionary, which includes an entity type dictionary, a relationship type dictionary and attribute rules.

[0028] S13. Transform the expert constraint rule condition text into dynamic constraint conditions, and add the dynamic constraint conditions and entity attributes to the ontology dictionary to obtain a knowledge graph of electronic product reliability containing entity attributes, relationships, and dynamic constraints.

[0029] S14. Perform consistency and completeness checks on the electronic product reliability domain knowledge graph obtained in step S13, and verify it through expert review, then output the verified electronic product reliability domain knowledge graph.

[0030] S2. Preprocess the original electronic product reliability data to obtain real data. Based on the real data, use WGAN-digital twin generation technology to generate synthetic data, and then merge the synthetic data with the real data to obtain the electronic product reliability enhancement dataset.

[0031] Step S2 specifically includes the following steps:

[0032] S21, use in sequence The criteria involve handling outliers, filling missing values ​​using neighbor-to-neighbor interpolation, and then cleaning the data to obtain the true data. ;

[0033] S22. Introduce dynamic constraints and train the WGAN network using the electronic product reliability data processed in step S21 until convergence; wherein, the generator loss function and discriminator loss function of the WGAN network are... The expression is as follows:

[0034] ;

[0035] ;

[0036] In the formula, This represents the loss value of the generator; This represents the synthesized data output by the discriminator to the generator. Average rating Represents a random noise vector. Represents a random noise vector The probability distribution it follows This represents the discriminator network. This represents the synthesized data output by the generator; Indicates constraint weights; This represents the synthesized data output by the generator. Penalties for violating dynamic constraints, and , Represents a dynamic constraint set. This represents the synthesized data output by the generator. Under constraints The actual value below, Representing constraints The corresponding compliance value; This indicates that the discriminator is effective against real data. Average rating; Representing real data The probability distribution;

[0037] S23. Input real data into the pre-trained WGAN network to output synthetic data. Then, fuse the real data and synthetic data, standardize them, and output the electronic product reliability enhancement dataset. .

[0038] S3. The LLM model is pre-trained using corpus on the reliability of electronic products. Based on the TransE model, the entities and relations of the knowledge graph on the reliability of electronic products are transformed into low-dimensional vectors, and an ER-KG retrieval enhancement module is constructed. Then, using the enhanced dataset on the reliability of electronic products as training samples, the subgraphs of the knowledge graph on the reliability of electronic products retrieved by the ER-KG retrieval enhancement module are used as context inputs to pre-train the LLM model. Knowledge consistency loss is introduced for joint training to obtain an ER-KG-RAG-LLM fusion model with domain knowledge anchoring and reliability scheme generation capabilities.

[0039] Step S3 specifically includes the following steps:

[0040] S31. Convert the triplet data obtained in step S11 into natural language sentences, and combine them with the fault report and FMEA table to convert them into UTF-8 text format to obtain the electronic product reliability corpus.

[0041] S32. Set the task type of the LLM model as causal language modeling, and use perplexity... To determine whether the LLM model is suitable for the field of electronic product reliability, when... At that time, the pre-trained LLM model is output, where the perplexity is... The expression is as follows:

[0042] ;

[0043] In the formula, Represents an exponential function; The sequence length of the corpus representing the reliability of electronic products; This represents the token index in a corpus sequence related to the reliability of electronic products. This indicates that the LLM model predicts the first... tokens The conditional probability;

[0044] S33. Based on the TransE model, transform the entity and relation vectors of the knowledge graph in the field of electronic product reliability to obtain an entity vector library and a relation vector library, so that the semantic association between entities and relations can be reflected through the distance in the vector space.

[0045] S34. Construct an ER-KG retrieval enhancement module;

[0046] S341. Input the natural language requirements into the BERT-base model to generate a reliability requirement vector. ;

[0047] S342. Extract each entity from the knowledge graph of electronic product reliability. The triples centered on the center form candidate subgraphs. ; where each entity All include head entities Tail-end entity ;

[0048] S343, Candidate Subgraphs All included entity vectors With relation vector Take the average to obtain the subgraph vector. Among them, entity vector Including head entity vectors Tail entity vector ;

[0049] S344, Calculate the reliability requirement vector and subgraph vectors Cosine similarity score between :

[0050] ;

[0051] In the formula, Represents the cosine similarity function; Represents a linearly reduced-dimensional matrix;

[0052] S345. Sort the candidate subgraphs in descending order of similarity, and the first predetermined number of candidate subgraphs are the search results. and the search results The triples in the text are transformed into natural sentences as context text for the LLM model;

[0053] S35. Construct training samples:

[0054] S351, From Electronic Product Reliability Enhancement Dataset Extracting Reliability Requirements for Electronic Products - Sample Reliability Solutions , This indicates the reliability solutions corresponding to the reliability requirements of electronic products.

[0055] S352, Compare the electronic product reliability requirements - reliability solution samples Inputting the ER-KG search enhancement module yields results related to the reliability requirements of electronic products. Corresponding retrieval subgraph ;

[0056] S36. Construct the following multi-objective loss function and train the ER-KG-RAG-LLM fusion model:

[0057] ;

[0058] in,

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, Indicates the total loss; , and These represent prediction loss, knowledge consistency loss, and recall loss, respectively. and Both represent loss weights; The total number of tokens representing the reliability scheme text; This indicates that the ER-KG-RAG-LLM fusion model is in context Reliability requirements of electronic products and the Reliability scheme text The Probability of predicting a reliable text token; This indicates the confidence level of the LLM model in determining that the generated reliability scheme meets the dynamic constraints. Represents a retrieval subgraph The number of triples related to the reliability of electronic products; Representing a knowledge graph in the field of electronic product reliability The number of triples related to the reliability of electronic products;

[0063] S37. In each round of training, an accuracy verification ER-KG-RAG-LLM fusion model is generated based on the knowledge consistency and reliability scheme. If the verification fails, training continues; otherwise, the trained ER-KG-RAG-LLM fusion model is output.

[0064] Step S33 specifically includes the following steps:

[0065] S331. Extract all valid positive triples from the knowledge graph of electronic product reliability constructed in step S1. ,in, , and These represent the head entity, predefined relation, and tail entity in the knowledge graph of electronic product reliability, respectively.

[0066] S332. Configure the TransE model: Set the vector dimensions of entities and relationships. And assume that both entity vectors and relation vectors follow a normal distribution; and simultaneously for each positive triplet Random replacement or Generate negative triples for other entities of the same type. ,in, and These represent the head and tail entities after the replacement, respectively.

[0067] S333. The TransE model is trained using the SGD optimizer with a learning rate of 0.01, a batch size of 128, and 50 training epochs. The loss function is calculated in each epoch. If the loss does not decrease for 3 consecutive epochs, the training stops early. The low-dimensional vectors of all entities and relations are output and stored in the FAISS index. The entity ID-vector and relation ID-vector mapping tables are constructed.

[0068] Among them, the loss function of the TransE model The expression is as follows:

[0069] ;

[0070] In the formula, Represents the set of positive triples; Represents the set of negative triples; and These represent the head entity vector and the tail entity vector in a positive triple, respectively. Represents a relation vector; and These represent the head entity vector and the tail entity vector in the negative triplet, respectively. This represents the marginal parameter.

[0071] S4. Receive the reliability requirements of electronic products, transform them into structured requirements using the NER model, and input them into the ER-KG-RAG-LLM fusion model. Combine the component-fault-reliability solution related subgraphs retrieved by the ER-KG retrieval enhancement module to generate an FMEA report containing fault modes and risk priorities. After iterative optimization and multi-dimensional verification, output the reliability solution.

[0072] Step S4 specifically includes the following steps:

[0073] S41. Transform the received natural language requirements into structured reliability requirements using the NER model. And based on a knowledge graph in the field of electronic product reliability, the feasibility of receiving requirements is assessed:

[0074] ;

[0075] In the formula, This indicates the structured reliability requirements for reception. The results of the feasibility assessment; Indicates the requirement for structured reliability All reliability metrics target value Perform a traversal and obtain the feasibility indicator function for each indicator. Perform a series of multiplications; This represents the corresponding reliability indicators in the knowledge graph of electronic product reliability. The current maximum achievable value for engineering;

[0076] S42, After feasibility assessment Input the ER-KG-RAG-LLM fusion model to generate an FMEA report that includes failure modes and risk priorities;

[0077] S43. Verify the pass rate of the FMEA report generated in step S42. If the pass rate is greater than the set threshold, return to step S41; otherwise, proceed to step S44. The calculation formula is as follows:

[0078] ;

[0079] In the formula, Indicates the risk priority in the FMEA report The number of valid entries in the calculation result, and ; This indicates the total number of risk priority entries in the FMEA report;

[0080] S44. Set the following optimization objective function to iteratively optimize the FMEA report and output the optimized reliability solution:

[0081] ;

[0082] In the formula, This represents the optimization objective function value of the reliability scheme in the FMEA report; and All represent weighting coefficients; Indicates an indicator function; This indicates whether the reliability scheme in the FMEA report meets the requirements. All metrics; This represents the reliability solution cost factor in the FMEA report;

[0083] S45. Perform the following multi-dimensional verification and output the reliability scheme that passes the verification:

[0084] ;

[0085] in,

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, This indicates the overall score; , and These represent the compliance rate of the optimized reliability scheme with dynamic constraints, the compliance rate with reliability standards, and the feasibility rate, respectively. This represents the number of reliability scheme entries in the optimized reliability scheme that conform to the dynamic constraints of the knowledge graph in the field of electronic product reliability. This indicates the total number of entries in the optimized reliability scheme; This indicates the number of reliability scheme entries in the optimized reliability scheme that conform to industry reliability standards. This indicates the number of reliability scheme entries in the optimized reliability scheme that meet the product engineering design constraints.

[0090] Step S4 is followed by: S5, using built-in sensors in the electronic product and a cloud monitoring platform, collecting reliability data after the deployment of the electronic reliability solution output in step S4 at a fixed frequency, and analyzing the deviation between actual and target indicators; then, based on the deviation analysis results, updating the entity attributes and dynamic constraints of the electronic product reliability domain knowledge graph in step S1, and adjusting the ER-KG-RAG-LLM fusion model in step S3 using deployment monitoring data; then calling the updated ER-KG-RAG-LLM fusion model and the electronic product reliability domain knowledge graph, re-inputting the electronic product reliability requirements, generating an iteratively optimized reliability solution, and comparing the indicator improvement effects before and after iteration to form a closed-loop optimization.

[0091] The actual indicators described in step S5 With target indicators Deviation between The calculation formula is as follows:

[0092] ;

[0093] The loss function of the ER-KG-RAG-LLM fusion model is adjusted as follows:

[0094] ;

[0095] In the formula, This represents the adjusted loss function value; Indicates the bias weight; This represents the function for calculating the reliability index deviation of the output reliability scheme of the ER-KG-RAG-LLM fusion model; This indicates the deviation of the reliability index of the adjusted ER-KG-RAG-LLM fusion model output.

[0096] Simulation Experiment

[0097] Experimental subject: Selected vehicle communication module (including core components such as MCU, electrolytic capacitor, and RF chip); Requirements: "Mean time between failures (MTBF) ≥ 30,000 h in an environment of -40℃ to 85℃, and solution cost ≤ 200 yuan".

[0098] The following comparison methods are set: Method 1 (Manual Design): Engineers manually design reliability schemes based on the "Automotive Electronics FMEA Manual," the "ISO26262 Functional Safety Standard," and their experience. Method 2 (Pure Large Model Method): Using a GPT-3.5 model finely tuned with an "augmented dataset," natural language requirements are directly input to generate the scheme.

[0099] The following data was collected and integrated: Real data: A 5-year fault database of an automotive communication module from a certain automaker (containing 2000+ fault records, environmental parameters, and repair solutions). Synthetic data: 10,000 virtual "fault-environment-solution" data points were generated using WGAN (combined with ER-KG constraints) to address the scarcity of real data. Knowledge graph (ER-KG): An ER-KG for the automotive communication field was constructed, containing 5 types of entities (components, fault modes, environmental parameters, reliability indicators, and solutions) and 4 types of core relationships (causing, affecting, having, and resolving), totaling 100,000+ triples.

[0100] Table 1 Comparison of Simulation Results

[0101]

[0102] As shown in Table 1, this application achieves a strict match between the solution and the domain triplet (e.g., when selecting an industrial-grade MCU of model XX, its ER-KG records 'MTBF=35000h at -40℃'), by using knowledge anchoring in the ER-KG and precise retrieval in the RAG. The accuracy of the solution is improved to 92%, and the prediction bias and cost are reduced to 8% and 6% respectively, which is close to the accuracy of human experience and more stable (human experience depends on individual experience, and the results of different engineers vary greatly), thus proving the effectiveness of this application.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical reliability of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical reliability of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical reliability to deviate from the spirit and scope of the technical reliability of the present invention.

Claims

1. A reliable scheme generation method integrating knowledge graphs and large models, characterized by: Includes the following steps: S1. Integrate structured data, unstructured data and experimental data related to fault-reliability solutions after analysis and testing in the field of electronic product reliability to form the original electronic product reliability data. Use knowledge graph to extract entities and relationships of components-fault-environment-indicators-reliability solutions, and add dynamic engineering constraints to construct a knowledge graph of electronic product reliability containing entity attributes, relationships and dynamic constraints. S2. Preprocess the original electronic product reliability data to obtain real data. Based on the real data, use WGAN-digital twin generation technology to generate synthetic data, and then fuse the synthetic data with the real data to obtain the electronic product reliability enhancement dataset. S3. Pre-train the LLM model using electronic product reliability domain corpus, and based on the TransE model, transform the entities and relations of the electronic product reliability domain knowledge graph into low-dimensional vectors to construct an ER-KG retrieval enhancement module. Then, using the electronic product reliability enhancement dataset as training samples, the ER-KG retrieval enhancement module retrieves electronic product reliability domain data. The subgraphs of the domain knowledge graph are used as context inputs to pre-train the LLM model, and knowledge consistency loss is introduced for joint training to obtain the ER-KG-RAG-LLM fusion model with domain knowledge anchoring and reliability solution generation capabilities; S4, receive the reliability requirements of electronic products, use the NER model to transform them into structured requirements, and input them into the ER-KG-RAG-LLM fusion model. Combined with the component-fault-reliability solution related subgraphs retrieved by the ER-KG retrieval enhancement module, an FMEA report containing fault modes and risk priorities is generated; and after iterative optimization and multi-dimensional verification, the reliability solution is output.

2. The reliability scheme generation method integrating knowledge graphs and large models according to claim 1, characterized in that: The structured data mentioned in step S1 includes reliability standards and an enterprise electronic product failure database. The enterprise electronic product failure database contains component models, failure times, environmental parameters, failure modes, component characteristics, and environmental impact patterns. The unstructured data includes expert-annotated FMEA tables and maintenance record text. The FMEA tables contain failure modes, severity S, and probability of occurrence. Detection difficulty Step S1 specifically includes the following steps: S11, converting the original electronic product reliability data into a triplet format: extracting entities and relations from structured data; extracting entities and relations from unstructured data using the BERT-NER model; converting the fault-reliability scheme association experimental data labeled after expert analysis and testing into expert constraint rule condition text using rule templates; S12, defining entities and relations in the electronic product reliability domain knowledge graph: defining entity subclasses, relation directions, and attribute data types using a causal structure ontology to obtain an ontology dictionary, which includes an entity type dictionary, a relation type dictionary, and attribute rules; S13, converting the expert constraint rule condition text into dynamic constraint conditions, and adding the dynamic constraint conditions and entity attributes to the ontology dictionary to obtain an electronic product reliability domain knowledge graph containing entity attributes, relations, and dynamic constraints; S14, performing consistency and completeness checks on the electronic product reliability domain knowledge graph obtained in step S13, and verifying it through expert review, outputting the verified electronic product reliability domain knowledge graph.

3. The reliability scheme generation method integrating knowledge graphs and large models according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21, sequentially using The criteria involve handling outliers, filling missing values ​​using neighbor-to-neighbor interpolation, and then cleaning the data to obtain the true data. S22. Introduce dynamic constraints and train the WGAN network using the electronic product reliability data processed in step S21 until convergence; whereby the generator loss function and discriminator loss function of the WGAN network are... The expression is as follows: ； In the formula, This represents the loss value of the generator; This represents the synthesized data output by the discriminator to the generator. Average rating Represents a random noise vector. Represents a random noise vector The probability distribution it follows This represents the discriminator network. This represents the synthesized data output by the generator; Indicates constraint weights; This represents the synthesized data output by the generator. Penalties for violating dynamic constraints, and ， Represents a dynamic constraint set. This represents the synthesized data output by the generator. Under constraints The actual value below, Representing constraints The corresponding compliance value; This indicates that the discriminator is effective against real data. Average rating; Representing real data The probability distribution; S23, input real data into the pre-trained WGAN network, output synthetic data, then fuse the real data and synthetic data, standardize, and output the electronic product reliability enhancement dataset. 。 4. The reliability scheme generation method integrating knowledge graphs and large models according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31, converting the triplet data obtained in step S11 into natural language sentences, and combining it with the fault report and FMEA table, then converting it into UTF-8 text format to obtain electronic product reliability corpus; S32, setting the task type of the LLM model as causal language modeling, and using perplexity... To determine whether the LLM model is suitable for the field of electronic product reliability, when... At that time, the pre-trained LLM model is output, where the perplexity is... The expression is as follows: In the formula, Represents an exponential function; The sequence length of the corpus representing the reliability of electronic products; This represents the token index in a corpus sequence related to the reliability of electronic products. This indicates that the LLM model predicts the first... tokens Conditional probability; S33. Based on the TransE model, transform entity and relation vectors in the knowledge graph of electronic product reliability domain to obtain entity vector library and relation vector library, so that the semantic association between entities and relations can be reflected through vector space distance; S34. Construct ER-KG retrieval enhancement module; S341. Input natural language requirements into BERT-base model to generate reliability requirement vector. S342. Extract each entity from the knowledge graph of electronic product reliability. The triples centered on the center form candidate subgraphs. ; where each entity All include head entities Tail-end entity S343, Candidate Subgraphs All included entity vectors With relation vector Take the average to obtain the subgraph vector. Among them, entity vector Including head entity vectors Tail entity vector S344, Calculate the reliability requirement vector and subgraph vectors Cosine similarity score between ： In the formula, Represents the cosine similarity function; S345 represents a linear dimensionality reduction matrix; Candidate subgraphs are sorted in descending order of similarity, with the first predetermined number of candidate subgraphs representing the search results. and the search results The triples in the dataset are transformed into natural language sentences to serve as the context text for the LLM model; S35, Constructing training samples: S351, From the Electronic Product Reliability Enhancement Dataset Extracting Reliability Requirements for Electronic Products - Sample Reliability Solutions ， This indicates the reliability solutions corresponding to the reliability requirements of electronic products; S352, compare the electronic product reliability requirements with the reliability solution samples. Inputting the ER-KG search enhancement module yields results related to the reliability requirements of electronic products. Corresponding retrieval subgraph S36. Construct the following multi-objective loss function and train the ER-KG-RAG-LLM fusion model: ;in, ； ； In the formula, Indicates the total loss; 、 and These represent prediction loss, knowledge consistency loss, and recall loss, respectively. and Both represent loss weights; The total number of tokens representing the reliability scheme text; This indicates that the ER-KG-RAG-LLM fusion model is in context Reliability requirements of electronic products and the Reliability scheme text The Probability of predicting a reliable text token; This indicates the confidence level of the LLM model in determining that the generated reliability scheme meets the dynamic constraints. Represents a retrieval subgraph The number of triples related to the reliability of electronic products; Representing a knowledge graph in the field of electronic product reliability The number of triples related to the reliability of electronic products; S37. In each round of training, an accuracy verification ER-KG-RAG-LLM fusion model is generated based on the knowledge consistency and reliability scheme. If the verification fails, training continues; otherwise, the trained ER-KG-RAG-LLM fusion model is output.

5. The reliability scheme generation method integrating knowledge graphs and large models according to claim 4, characterized in that: Step S33 specifically includes the following steps: S331, extracting all valid positive triples from the electronic product reliability domain knowledge graph constructed in step S1. ,in, 、 and These represent the head entity, predefined relation, and tail entity in the knowledge graph of electronic product reliability, respectively; S332, Configure the TransE model: Set the vector dimensions of entities and relations. And assume that both entity vectors and relation vectors follow a normal distribution; and simultaneously for each positive triplet Random replacement or Generate negative triples for other entities of the same type. ,in, and These represent the replaced head and tail entities, respectively; S333, using the SGD optimizer with a learning rate of 0.01, a batch size of 128, and 50 training epochs, the TransE model is trained. The loss function is calculated in each epoch, and training stops early if the loss does not decrease for three consecutive epochs. Low-dimensional vectors of all entities and relations are output and stored in the FAISS index, constructing entity ID-vector and relation ID-vector mapping tables; where the loss function of the TransE model... The expression is as follows: In the formula, Represents the set of positive triples; Represents the set of negative triples; and These represent the head entity vector and the tail entity vector in a positive triple, respectively. Represents a relation vector; and These represent the head entity vector and the tail entity vector in the negative triplet, respectively. This represents the marginal parameter.

6. The reliability scheme generation method integrating knowledge graphs and large models according to claim 5, characterized in that: Step S4 Specifically, it includes the following steps: S41, transforming the received natural language requirements into structured reliability requirements using the NER model. And based on a knowledge graph in the field of electronic product reliability, the feasibility of receiving requirements is assessed: In the formula, This indicates the structured reliability requirements for reception. The results of the feasibility assessment; Indicates the requirement for structured reliability All reliability metrics Target value Perform a traversal and obtain the feasibility indicator function for each indicator. Perform a series of multiplications; This represents the corresponding reliability indicators in the knowledge graph of electronic product reliability. The current maximum achievable value of the project; S42, after feasibility assessment Input the ER-KG-RAG-LLM fusion model to generate an FMEA report containing failure modes and risk priorities; S43, verify the pass rate of the FMEA report generated in step S42. If the pass rate is greater than the set threshold, return to step S41 if not, otherwise proceed to step S44; The calculation formula is as follows: In the formula, Indicates the risk priority in the FMEA report The number of valid entries in the calculation result, and ； This indicates the total number of risk priority entries in the FMEA report; S44. Set the following optimization objective function to iteratively optimize the FMEA report and output the optimized reliability solution: In the formula, This represents the optimization objective function value of the reliability scheme in the FMEA report; and All represent weighting coefficients; Indicates an indicator function; This indicates whether the reliability scheme in the FMEA report meets the requirements. All metrics; This represents the reliability solution cost factor in the FMEA report; S45, perform the following multi-dimensional verifications and output the verified reliability solutions: ;in, ； ； In the formula, This indicates the overall score; 、 and These represent the compliance rate of the optimized reliability scheme with dynamic constraints, the compliance rate with reliability standards, and the feasibility rate, respectively. This represents the number of reliability scheme entries in the optimized reliability scheme that conform to the dynamic constraints of the knowledge graph in the field of electronic product reliability. This indicates the total number of entries in the optimized reliability scheme; This indicates the number of reliability scheme entries in the optimized reliability scheme that conform to industry reliability standards. This indicates the number of reliability scheme entries in the optimized reliability scheme that meet the product engineering design constraints.

7. The reliability scheme generation method integrating knowledge graphs and large models according to claim 6, characterized in that: Step S4 is followed by: S5, using built-in sensors in the electronic product and a cloud monitoring platform, collecting reliability data after the deployment of the electronic reliability solution output in step S4 at a fixed frequency, and analyzing the deviation between actual and target indicators; then, based on the deviation analysis results, updating the entity attributes and dynamic constraints of the electronic product reliability domain knowledge graph in step S1, and adjusting the ER-KG-RAG-LLM fusion model in step S3 using deployment monitoring data; then calling the updated ER-KG-RAG-LLM fusion model and the electronic product reliability domain knowledge graph, re-inputting the electronic product reliability requirements, generating an iteratively optimized reliability solution, and comparing the indicator improvement effects before and after iteration to form a closed-loop optimization.

8. The reliability scheme generation method integrating knowledge graphs and large models according to claim 7, characterized in that: The actual indicators described in step S5 With target indicators Deviation between The calculation formula is as follows: The loss function of the ER-KG-RAG-LLM fusion model is adjusted as follows: In the formula, This represents the adjusted loss function value; Indicates the bias weight; This represents the function for calculating the reliability index deviation of the reliability scheme output by the ER-KG-RAG-LLM fusion model. This indicates the deviation of the reliability index of the adjusted ER-KG-RAG-LLM fusion model output.

Images