A method, system, and program product for analyzing the performance degradation of electronic products based on large language models

Abstract

The present invention relates to the technical field of electronic product performance analysis, and in particular to a method, system and program product for analyzing the performance degradation of electronic products based on large language models. This method collects multi-source data, performs formatting and semantic parsing, uses large language models to pre-analyze the degradation mechanism, and combines multi-round scoring and weighted fusion by experts to construct a unified quantitative evaluation model for the degradation mechanism. At the same time, it correlates key performance parameters with internal and external independent variables, and outputs risk rankings and preventive optimization plans. This method can quickly identify potential degradation mechanisms of electronic products under complex environmental stress and working load conditions, and significantly improve the efficiency and accuracy of degradation identification through a quantitative analysis mechanism that combines large language models and experts.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic product performance analysis, and particularly to a method, system and program product for analyzing the performance degradation of electronic products based on large language models. Background Art

[0002] With the rapid development of electronic technology, the complexity of electronic products and the diversification of application scenarios, the reliability problem has become increasingly prominent. Electronic products are composed of a large number of electronic components, and the performance and reliability of the components directly determine the overall reliability of the product. Therefore, design engineers need to pay high attention to the problem of product performance degradation. To ensure that the product meets the long-term reliability requirements during the design stage, it is necessary to comprehensively analyze and identify the mechanisms that may cause performance degradation. However, performance degradation analysis involves knowledge in multiple fields. Especially under the action of complex environmental stresses and working loads, it is particularly difficult to identify and quantify the degradation process, and traditional analysis methods have significant deficiencies in terms of efficiency and accuracy.

[0003] In the traditional reliability analysis process, researchers usually use qualitative or semi-quantitative methods such as Failure Mode, Effects and Mechanisms Analysis (FMEA / FMMEA) to identify and evaluate the potential degradation mechanisms of electronic products. Although such methods can help engineers discover failure risks to a certain extent, they still face the following challenges or deficiencies:

[0004] 1. Dependence on expert experience and strong subjectivity in analysis: In the specific implementation of the FMEA or FMMEA method, a large amount of subjective judgment of experts on failure risks is required. Experts need to be familiar with the product design and manufacturing process, and also have a deep understanding of the component characteristics and the degradation phenomena that may occur under different stress environments. However, with the continuous complexity of the functions and structures of electronic products, it is difficult to comprehensively, accurately and consistently complete the identification and evaluation of potential failure mechanisms solely relying on expert experience.

[0005] 2. Multidisciplinary intersection and high difficulty in data acquisition and integration: The performance degradation of electronic products is often closely related to various factors such as material properties, structural design, thermal management, electromagnetic compatibility, and external stresses (temperature, humidity, vibration, radiation, etc.), and involves the dynamic performance in simulation, testing and large-scale production processes. Existing reliability analysis methods usually need to collect a large amount of data from multiple dimensions and perform statistics or modeling. Due to the scattered data, inconsistent formats and uneven credibility, it is often difficult to obtain the required information in a timely and efficient manner in engineering practice, resulting in the lack of integrity of the analysis results.

[0006] 3. Lack of intelligent quantitative evaluation, insufficient efficiency and accuracy: Traditional analysis processes often focus on qualitative descriptions of failure mechanisms, lacking quantitative evaluation of the degradation process and its impact on key performance parameters, making it difficult for designers to formulate targeted improvement strategies. At the same time, due to the large amount of manual modeling and manual calculations, the analysis efficiency is low, and it is prone to omissions or duplicate analyses, which further hinders large-scale and systematic performance degradation assessment activities.

[0007] 4. Scarcity of professional talents, difficult to meet the needs of high-difficulty failure mechanism analysis: The R & D and production processes of electronic products have become highly complex, requiring professional talents in multiple fields such as materials, design, technology, and environmental stress to deeply participate in the engineering team. However, in many enterprises or research institutions, the resources of senior reliability experts are relatively limited, making it difficult to comprehensively and accurately evaluate complex failure phenomena in a timely manner, and it is even more difficult to ensure the efficient implementation of time-consuming and laborious analyses such as FMMEA at the organizational level.

[0008] Based on the above industry pain points, the rapid rise of large language models (LLMs) provides a new idea for the performance degradation analysis of electronic products. Through deep neural networks and trained on a large amount of text data, large language models have powerful natural language understanding and semantic association capabilities, and can connect and reason about interdisciplinary knowledge when facing multi-source information. As a natural language processing technology based on deep learning, large language models can efficiently integrate multi-dimensional information, imitate the reasoning and learning processes of humans, and solve complex problems through powerful semantic understanding capabilities. In performance degradation analysis, large language models can quickly integrate usage scenarios, component design information, and historical degradation data, and intelligently identify potential degradation mechanisms and their impact on key performance parameters. This method not only significantly improves the analysis efficiency and accuracy, but also provides scientific and reliable decision-making basis for designers through automated quantitative analysis, thus injecting new technical impetus into the reliability design of electronic products. Summary of the Invention

[0009] To solve the above technical problems, the present invention proposes a method for analyzing the performance degradation of electronic products based on a large language model. By combining multi-source data collection, semantic association analysis, and expert collaborative scoring, a unified degradation mechanism identification and quantitative evaluation framework is constructed, which can not only significantly improve the accuracy and efficiency of degradation analysis, but also output actionable decision-making suggestions for design improvement, laying a technical foundation for the reliability design and full life cycle management of electronic products.

[0010] To achieve the above objectives, the present invention adopts the following technical solutions:

[0011] A method for analyzing the performance degradation of electronic products based on large language models. The method identifies and quantitatively evaluates the performance degradation mechanism and its influencing factors of electronic products in actual usage scenarios by iteratively combining large language models with an expert collaborative scoring mechanism multiple times. The method includes the following steps:

[0012] 1) Multi-source data acquisition and preprocessing:

[0013] Obtain and organize the multi-source data of electronic products. The multi-source data includes the usage scenario information of electronic products, product functions, design margin analysis reports, performance degradation models, and supplier component performance degradation experimental data; format, denoise, and semantically parse the multi-source data to obtain the key performance parameter set, internal independent variable set, and external independent variable set to be analyzed;

[0014] 2) Preliminary analysis by large language models:

[0015] Input the organized data into a pre-trained large language model to generate a preliminary list of performance degradation causes and degradation mechanisms, including mechanism names, degradation process descriptions, and key performance parameters potentially affected; and the large language model automatically infers a preliminary score of the importance of the degradation mechanism based on the product usage scenario and historical data, including a harmfulness scoring standard and a possibility scoring standard;

[0016] 3) Expert collaborative multi-round iterative scoring:

[0017] 3.1) Provide the large language model with descriptions of the electronic product degradation mechanism, harmfulness scoring standard, possibility scoring standard, usage scenario information, and historical data, and the large language model automatically generates a score for each degradation mechanism;

[0018] 3.2) The expert group independently scores the degradation mechanism based on the harmfulness scoring standard and the possibility scoring standard in combination with their understanding of product functions, usage scenario information, and failure risks;

[0019] 3.3) For the differences between the large language model scores and the expert scores, during the scoring process, comprehensively evaluate the degradation mechanism through the collaborative mechanism of the large language model and expert scores;

[0020] For the differences between the large language model scores and the expert scores, adopt a multi-round iterative collaborative scoring mechanism: a. If the difference is within the preset threshold, take the average of the two or directly adopt the large language model score; b. If the difference exceeds the preset threshold, introduce expert secondary evaluation or weighted fusion algorithms to obtain more accurate harmfulness and possibility scores;

[0021] After each round of iteration, update the analysis parameters or model weights of the large language model and recalculate the importance of the degradation mechanism;

[0022] 4) Degradation factor association and model correction:

[0023] Based on the degradation mechanism and importance determined through multiple rounds of collaborative scoring, combined with the correlation between key performance parameters and internal and external variables, construct or correct the performance degradation analysis model; through the semantic association ability of the large language model, map each degradation mechanism to corresponding information such as material properties, structural parameters, and environmental stress to form a complete degradation factor network;

[0024] 5) Output analysis results and prevention strategies:

[0025] Output the final hazard score, likelihood score, and importance of each degradation mechanism; combined with expert experience and the inference results of the large language model, output the influence mechanism of internal and external variables on the degradation of key performance parameters, and give corresponding prevention strategies and reliability optimization plans.

[0026] Preferably, in the multi-source data acquisition and preprocessing step 1), it further includes hierarchical annotation of the environmental stress involved in the usage scenario information, including temperature, humidity, radiation, mechanical vibration; and the working load, including voltage, current, high-frequency operation, to form a multi-level label data structure that can be parsed by the large language model.

[0027] Preferably, in the large language model pre-analysis step 2), the large language model uses a deep learning network to perform vector embedding on the text-based degradation mechanism description to automatically identify possible degradation modes under different environmental stresses, and the generated list of degradation mechanisms includes the corresponding degradation process description and its correlation with key performance parameters.

[0028] Preferably, the specific algorithm for the multi-round iterative scoring by experts in step 3) is as follows:

[0029] ,

[0030] where, is the absolute error between the two scores, is the score of the large language model, is the expert score;

[0031] Due to the subjective deviation and misunderstanding in expert scoring, the median method is used to obtain the expert scoring results; the processing of the scoring results is specifically divided into the following situations:

[0032] a. When, directly use the scoring result for importance calculation;

[0033] b. When, take the average of the two scores as the final score;

[0034] c, When , the designer conducts the final scoring through the weight allocation method, and the scoring algorithm is as follows:

[0035] ;

[0036] Among them, is the final scoring result, is the weight of the large language model scoring, is the weight of the expert scoring, .

[0037] Preferably, in the degradation factor association and model correction step 4), it further includes using the large language model to perform automatic interpretability analysis on the identified degradation mechanism, and outputting the multi-dimensional association relationship between the degradation mechanism and key performance parameters, internal variables, and external variables based on knowledge graph technology, enabling the designer to intuitively obtain the degradation source and influence path.

[0038] Preferably, in the output analysis result and prevention strategy step 5), the large language model combines expert feedback information to automatically generate a list of ranked degradation risks and corresponding preventive improvement measures, and visually displays the degradation trend of key performance parameters, the sensitivity of influencing factors, and the ranking of mechanism importance, providing a decision-making basis for subsequent design optimization.

[0039] Furthermore, the present invention also provides an electronic product performance degradation analysis system based on a large language model, which implements the described method, including:

[0040] A data management module for performing multi-source data acquisition and preprocessing, cleaning, annotating, and archiving usage scenario data, product design information, and component degradation experiment data;

[0041] A model analysis module for performing semantic reasoning, mining, and ranking of degradation mechanisms based on a large language model, and automatically associating internal variables and external variables;

[0042] An expert collaboration module for realizing multi-round iterative fusion of large language model scoring and expert scoring, including a scoring difference comparison and weighted calculation mechanism;

[0043] A result output and visualization module for outputting the hazard score, possibility score, and importance of the degradation mechanism, and generating a reliability optimization report and visualization chart.

[0044] Preferably, the expert collaboration module further includes a multi-round interaction interface for the expert to supplement and correct the degradation mechanism, scoring basis, and interpretability conclusion initially identified by the large language model in each round, so that the model analysis result can fit the actual usage environment and product requirements.

[0045] Preferably, the system adopts a deployment method combining cloud and local servers. The cloud server is responsible for the training and inference of large language models, and the local server is responsible for the offline management and secure calculation of sensitive data, and exchanges data with the cloud server through a secure interface;

[0046] And / or, the system is further configured with a knowledge graph engine for correlating degradation mechanisms, key performance parameters and their internal and external variables into a multi-dimensional knowledge graph, enabling design engineers to retrieve and view the degradation source, mechanism reasoning path and corresponding improvement measures based on the graph.

[0047] Furthermore, the present invention also provides a computer-readable storage medium having a computer program or instruction stored thereon, and when the computer program or instruction is executed by a processor, the method is implemented.

[0048] Furthermore, the present invention also provides a computer program product including a computer program or instruction, and when the computer program or instruction is executed by a processor, the method is implemented.

[0049] Due to the adoption of the above technical solution, the present invention combines large language models (LLMs) with traditional fault mechanism analysis processes such as FMMEA, and makes full use of the ability of large language models to understand natural language semantics and correlate multi-dimensional information, and can establish a unified association among multi-source data (usage scenarios, component design information, historical degradation data, expert knowledge, etc.), making the identification and evaluation of degradation mechanisms more systematic. On the one hand, for complex or interdisciplinary degradation mechanisms, the large language model can automatically identify potential failure sources through semantic reasoning and pattern matching, reducing human dependence and blind spots; on the other hand, through multiple rounds of scoring in collaboration with experts, the subjective bias brought by a single method can be effectively corrected, ensuring the accuracy and consistency of the analysis results under different environmental stresses and process conditions.

[0050] Traditional degradation analysis of electronic products requires a large amount of manpower for data collection, fault mode troubleshooting and qualitative judgment. The process is cumbersome and highly dependent on the professional experience of senior reliability experts. The present invention uses a large language model to batch semantically parse and automatically associate product characteristics and degradation data, and can complete the screening and evaluation of degradation mechanisms under various stress environments and workload combinations in a short time, thereby realizing the rapid diagnosis of the performance degradation of electronic products. Since the preliminary diagnosis work is undertaken by an intelligent model, experts only need to conduct multiple rounds of iterative review and scoring in key links or for difficult degradation mechanisms, greatly reducing the input requirements of enterprises or R & D institutions for high-level reliability experts.

[0051] Furthermore, the present invention introduces a quantitative evaluation method for degradation mechanisms that combines hazard and probability scoring, and adopts a collaborative mechanism of automatic scoring by large language models and expert scoring, significantly improving the operability and objectivity of quantitative scoring. The large language model performs semantic reasoning based on pre-set scoring criteria and existing degradation data, automatically gives preliminary values and attaches explanations of the scoring logic, and then experts make corrections or confirmations based on actual engineering experience, finally generating an authoritative score that conforms to the actual usage conditions of the product. Through this scoring mechanism, designers can not only quickly distinguish the importance of each degradation mechanism, but also formulate targeted prevention or improvement plans based on these quantitative results, providing a scientific basis for optimizing the reliability of the product.

[0052] Furthermore, in the degradation analysis process, the present invention makes full use of the integration ability of large language models for large-scale historical data and in-service information, helping design engineers and decision-makers more accurately locate possible degradation risk points and pre-plan preventive measures or improvement strategies. Since it reduces the waste of resources caused by repeated trial and error and repeated testing, enterprises can complete the reliability assessment from R & D to mass production more quickly. At the same time, with the help of the model automatic analysis and visualization report functions of the present invention, the R & D team can identify key degradation risks and make timely adjustments in the early design stage of the product, achieving effective control of the reliability of the product throughout its life cycle.

[0053] Furthermore, the degradation analysis framework of the present invention can continuously absorb new scenario information, component design parameters, and failure history data, realizing the generalization of degradation mechanism analysis for various types of electronic products. Whether it is consumer electronics, communication equipment, or high-demand application fields such as aerospace devices, the present invention can quickly adapt and conduct degradation analysis according to actual stress conditions and performance parameter configurations, providing an intelligent and automated solution for the long-term reliability design and management of various electronic products. Specific embodiments

[0054] Next, in combination with the embodiments of the present invention, taking a CMOS inverter as an example, the technical content in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall be regarded as the protection scope of the present invention.

[0055] The present invention provides a method for analyzing the performance degradation of electronic products based on a large language model. The specific steps of this method are as follows:

[0056] Step 1: Extract relevant information of the electronic product, mainly including:

[0057] a. Collect and organize the usage scenario analysis reports of electronic products, including information such as working modes, environmental stresses, and working loads, to clarify the working conditions of the products in actual applications. The detailed process is as follows:

[0058] 1) The CMOS inverter mainly serves as the basic unit of the logic circuit and is widely used in the processing and conversion of digital signals. In actual operation, its typical working modes include high-speed switching mode, low-power mode, and long-term stable working mode;

[0059] 2) The CMOS inverter may be affected by various environmental stresses during operation, mainly including: temperature stress, humidity stress, electrical stress, mechanical stress, radiation environmental stress, etc.;

[0060] 3) The CMOS inverter will be affected by working load electrical stress and high-frequency switching load during actual operation, and these loads may directly affect its performance. Among them, the working load electrical stress directly leads to an increase in the dynamic power consumption, deterioration of signal delay, and attenuation of long-term reliability of the CMOS inverter by exacerbating the risk of gate oxide layer breakdown, metal interconnect electromigration, and hot carrier injection effects; the high-frequency switching load directly causes a sharp increase in dynamic power consumption and signal noise coupling.

[0061] b. Collect relevant information of similar products, including the FPMA reports of similar products and performance degradation models, and draw on their degradation characteristics and analysis conclusions;

[0062] c. Collect the component performance degradation analysis and performance degradation experimental data provided by the supplier;

[0063] d. Collect and organize the product function, performance, and margin analysis (FPMA) reports, extract the functional principles, key performance parameters, and related internal and external independent variables of the products, and the analysis results are shown in Table 1;

[0064] Table 1 Analysis Table of Key Performance Parameters and Associated Internal and External Independent Variables of CMOS Inverters

[0065]

[0066] Format, denoise, and semantically parse the multi-source data to obtain the key performance parameter set, internal independent variable set, and external independent variable set to be analyzed.

[0067] Step 2: Pre-analysis by the large language model

[0068] Input the organized data into a pre-trained large language model to generate a preliminary list of performance degradation causes and degradation mechanisms; and the large language model automatically infers the preliminary importance scores of the degradation mechanisms based on the product usage scenarios and historical data, including the hazard score Table 2 and the possibility score Table 3.

[0069] a. Hazard rating criteria: According to the degree of influence of performance degradation on the key performance of the product, it is divided into four levels: mild hazard, medium hazard, severe hazard, and major hazard. The scoring criteria clearly indicate the typical descriptions corresponding to different hazard levels. For example, mild hazard mainly shows a slight impact on product performance, while major hazard may lead to product destruction or personal injury.

[0070] Table 2 Hazard rating criteria table for performance degradation mechanisms

[0071]

[0072] b. Possibility rating criteria: Score according to the occurrence ratio of the performance degradation mechanism in the usage scenario, with a range of 1 to 10 points. Specifically: 0 < p ≤ 10% corresponds to 1 point, 10% < p ≤ 20% corresponds to 2 points, and so on, until 90% < p ≤ 100% corresponds to 10 points. The scoring is graded according to the coverage ratio of the usage scenario, which can clearly reflect the occurrence probability of the degradation mechanism in product use.

[0073] Table 3 Possibility rating criteria table for performance degradation mechanisms

[0074]

[0075] As shown in Table 4, taking the CMOS inverter as an example, the analysis results are as follows, and the specific details of the process are as follows. The analysis method for other types of electronic products is the same:

[0076] Table 4 Analysis table of performance degradation causes and mechanisms for CMOS inverters

[0077]

[0078] a. Sort out the types of environmental stresses and working loads that act continuously or intermittently in the long term in the product usage scenario, clarify the working mode of the product, and analyze the impact of these stresses and loads on the product structure in combination with the product function principle. Determine the possible causes of performance parameter degradation, and generate the "Performance degradation causes" part in Table 4 according to the analysis results;

[0079] b. According to the degradation causes, analyze the process of performance degradation, identify the performance degradation mechanism, and determine the possible degradation mechanism in combination with the identification methods of common degradation mechanisms, and generate the "Performance degradation mechanism" part in Table 4;

[0080] c. Based on the description of the performance degradation process and the analysis results of the degradation mechanism, combine the key performance parameters in the FPMA report to determine the key performance parameters that each degradation mechanism may affect, and generate the "Key performance parameters affected" part in Table 4.

[0081] Step 3: Expert collaborative multi-round iterative scoring:

[0082] Based on the professional understanding of product functions, failure risks, and environmental stresses by the expert group, the list of degradation mechanisms initially given for the large language model is verified and scored independently or collectively. For the importance scoring of each degradation mechanism, the present invention improves the efficiency and accuracy of degradation mechanism evaluation through a collaborative scoring method that combines the large language model (ChatGPT) and expert evaluation. In the scoring process, the importance of the performance degradation mechanism is evaluated by calculating the product of the hazard score and the likelihood score, generating the "importance" part of Table 4.

[0083] Specifically, it includes the following steps:

[0084] 1) Provide CHATGPT with descriptions of CMOS inverter degradation mechanisms, hazard scoring criteria, likelihood scoring criteria, usage scenarios, and historical data. Taking the information in item 1 as an example, CHATGPT automatically generates scores for each degradation mechanism.

[0085] The relevant information before each degradation mechanism is shown in Table 4. Subsequently, the corresponding information is uniformly represented by serial numbers from top to bottom, and the scoring results are shown in Table 5.

[0086] Table 5 CHATGPT Scoring Results Table

[0087]

[0088] 2) The expert group combines its in-depth understanding of product functions, actual usage scenarios, and failure risks, and independently scores the degradation mechanisms according to the hazard and likelihood scoring criteria. The scoring results are shown in Table 6.

[0089] Table 6 Hazard and Likelihood Scoring Table of Expert Group for CMOS Inverter Degradation Mechanisms

[0090]

[0091] Regarding the differences between the scores of the large language model and the expert scores, in the scoring process, a comprehensive evaluation of the degradation mechanisms is carried out through the collaborative mechanism of CHATGPT and expert scores. Among them is the CHATGPT score, is the expert score, is the absolute error between the two scores, is the final scoring result. At the same time, considering the situation of subjective deviation in expert scores, the median method is used to obtain the values of expert score results. The processing of scoring results is specifically divided into the following situations:

[0092] When, directly use the scoring result for importance calculation;

[0093] When it comes to this, take the average of the scores of the two as the final score;

[0094] When it comes to this, the designer can conduct the final scoring through the weight allocation method, and the scoring algorithm is as follows:

[0095] ;

[0096] The parameter definitions in the formula are as follows:

[0097] : The weight of the CHATGPT score;

[0098] : The weight of the expert score;

[0099] .

[0100] Taking the CMOS inverter as an example, assuming that the weights are all 0.5, calculate the content in Table 3 and Table 4 according to the above calculation method, and fill in the "importance" part of Table 2.

[0101] Step 4: Degradation factor association and model correction:

[0102] Based on the degradation mechanism and importance determined by multi-round collaborative scoring, combined with the correlation between key performance parameters and internal and external variables, construct or correct the performance degradation analysis model; through the semantic association ability of the large language model, map each degradation mechanism to the corresponding material properties, structural parameters, environmental stress and other information to form a complete degradation factor network; generate the content in Table 7. Taking the CMOS inverter as an example, the analysis results are shown in Table 8.

[0103] Table 7 Analysis Table of Performance Degradation Influence Factors

[0104]

[0105] Table 8 Analysis Table of Performance Degradation Influence Factors for CMOS Inverters

[0106]

[0107] The specific process is as follows:

[0108] a. Combining the functional principle and performance requirements of the CMOS inverter, gradually decompose the top-level performance parameters to the bottom-level performance parameters, list all the key performance parameters that may be affected by degradation according to the hierarchical relationship, including the first-level performance, second-level performance and the lowest-level performance parameters, form a complete performance parameter hierarchical relationship table, and generate the "Key Performance Parameters with Degradation" part of Table 8;

[0109] b. Qualitatively analyze how the associated internal independent variables of the key performance parameters of the CMOS inverter affect the corresponding structural characteristics and working process after changes, and comprehensively judge by combining the functional principle and relevant performance parameters to generate the "Internal Independent Variables" part in Table 8;

[0110] c. Qualitatively analyze how the associated external independent variables of the key performance parameters of the CMOS inverter change the rate or trend of performance degradation by affecting material characteristics, structural characteristics, or internal independent variables after changes, and comprehensively judge by combining the actual usage scenario of the product and the degradation mechanism to generate the "External Independent Variables" part in Table 8.

[0111] Furthermore, after clarifying the degradation mechanism and its importance score, it is necessary to further analyze and clarify the multi-dimensional correlation relationships among the degradation mechanism, key performance parameters, internal independent variables, and external independent variables. In this embodiment, by using the semantic understanding and automatic reasoning capabilities of the large language model (LLM), the automated interpretable analysis of the degradation mechanism is realized, and then the multi-dimensional network relationship of the degradation factors is constructed by combining the knowledge graph technology, so that designers can intuitively and accurately identify the source, influence path, and correlation degree of performance degradation. The specific steps are as follows:

[0112] (I) Input data preparation and preprocessing

[0113] 1. Degradation mechanism set:

[0114] Extract the mechanisms with high importance from the degradation mechanisms determined in the expert collaborative scoring stage, such as: threshold voltage drift, carrier mobility reduction, leakage current increase.

[0115] 2. Related factor set:

[0116] Sort out the correlation information of key performance parameters (such as delay time, power consumption, signal amplitude, etc.), internal independent variables (such as carrier mobility, threshold voltage), and external independent variables (such as temperature, power supply voltage) to form a clear initial list.

[0117] Taking the CMOS inverter as an example, the initial structure of the input data is shown in Table 9:

[0118] Table 9 List of Degradation Mechanisms and Related Factors of CMOS Inverter

[0119]

[0120] (II) Automated interpretable analysis of the large language model

[0121] Based on the input data sorted out in step (I), use the pre-trained large language model to automatically reason and interpret the semantic relationships among each degradation mechanism, performance parameters, internal and external independent variables. The implementation steps are as follows:

[0122] 1. Semantic Analysis Input Construction

[0123] Design clear semantic analysis instructions for each degradation mechanism, for example:

[0124] "Given that the degradation mechanism is 'threshold voltage drift', involving the key performance parameter 'delay time', the internal variable is 'gate oxide layer defect density', and the external variables are 'temperature, voltage fluctuation', please automatically generate the logical correlation relationships between these factors and explain them item by item in a simple and easy-to-understand manner."

[0125] 2. Output Results of the Large Language Model

[0126] The large language model automatically returns the logical correlation relationships and concise explanation text, as shown in the following table:

[0127] Table 10 Association Relationship and Explanation Table Automatically Generated by the Large Language Model

[0128]

[0129] (3) Construction of a Multidimensional Association Relationship Network Based on a Knowledge Graph

[0130] Using the semantic association analysis results generated in step (2), further construct a knowledge graph of degradation factors to clearly show the multidimensional association relationships between these factors. The specific process is as follows:

[0131] 1. Definition of Knowledge Graph Elements

[0132] The knowledge graph consists of nodes and edges, and the definitions are as follows:

[0133] 1) Node classification:

[0134] Degradation mechanism nodes (such as 'threshold voltage drift');

[0135] Performance parameter nodes (such as 'delay time', 'power consumption');

[0136] Internal variable nodes (such as 'gate oxide layer defect density');

[0137] External variable nodes (such as 'temperature', 'voltage fluctuation').

[0138] 2) Edges are defined as clear semantic association relationships:

[0139] "Cause" (directly generate);

[0140] "Aggravate" (promoting effect);

[0141] "Determine" (fundamental influence);

[0142] "Indirectly aggravate" (indirect promotion).

[0143] 2. Atlas Construction Method

[0144] Integrate the automatically generated relationships and explanatory texts into the atlas database to form a complete multi-dimensional network structure. For example:

[0145] "Temperature increase" → (intensify) → "Threshold voltage drift" → (result in) → "Increase in delay time"

[0146] "Gate oxide layer defect density" → (determine) → "Threshold voltage drift" → (result in) → "Increase in delay time".

[0147] (4) Quantitative Analysis Model for Association Relationship Strength

[0148] To better guide designers to focus on critical paths, an association strength index S is introduced to quantify the importance of each association path. The defined formula is as follows:

[0149] Let there be a path path The association strength is:

[0150] ,

[0151] Where:

[0152] S

[0154] , ,

[0152] ,

[0157] ,

[0153] , , , rel,i ,

[0155] ,

[0151] ,

[0156] , , , , , exp,i ,

[0158] , , path , ,

[0160] , , ,

[0159] : Overall strength of the path;

[0153] W rel,i : Weight of the i-th association relationship type (for example: the weight of "result in" is 1.0, "intensify" is 0.7, "determine" is 1.2, "indirectly intensify" is 0.5);

[0154] I exp,i : Expert importance score of the i-th association relationship (based on the scoring results of step three).

[0155] The calculation example is shown in Table 11 below:

[0156] Table 11 Quantitative Table of Association Path Strength

[0157]

[0158] Through the above quantitative analysis, designers can quickly determine the high-risk degradation paths that need to be focused on and intervened first.

[0159] (5) Visualization and Interactive Interface of the Multi-Dimensional Association Relationship Network

[0160] Finally, use a graphical interactive interface to visually display the knowledge atlas, support node classification, association path, and association strength visualization. Designers can conveniently identify the source of performance degradation and key impact paths through interactive operations, enhancing the efficiency of problem-solving.

[0161] Step 5: Conduct multiple rounds of collaborative scoring on the degradation mechanism and output the results

[0162] Output the final hazard score, probability score, and importance of each degradation mechanism; combine expert experience with the inference results of the large language model to output the internal and external variable influence mechanisms of the degradation of key performance parameters, and give corresponding prevention strategies and reliability optimization plans. After completing the discrimination and quantitative evaluation of the degradation mechanism, the system will generate the following output information:

[0163] 1. List of degradation mechanisms and their importance

[0164] Display the hazard score, probability score, and final importance of each mechanism under different environmental stresses and working loads to assist design engineers in quickly locating the degradation factors with the highest risk.

[0165] 2. Influence path diagram of internal and external variables

[0166] If a visualization module or knowledge graph engine is deployed, a multi-dimensional association diagram of key performance parameters, internal variables, and external variables can be output, clearly showing logical relationships such as how "threshold voltage drift" leads to an increase in delay time and a decrease in device noise tolerance under high-temperature stress.

[0167] 3. Prevention strategies and optimization plans

[0168] Based on the inference results of the large language model of the present invention and collaborative evaluation by experts, improvement suggestions are given, such as optimizing the package design, controlling the working temperature range, and alleviating mechanical vibration; it is recommended to conduct targeted tests or simulations in the early stage of product development to avoid large-scale rework and higher failure rates in the later stage.

[0169] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented 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.

[0170] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0171] The memory may include non-permanent memory in the computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

[0172] A computer-readable medium includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer-readable medium does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0173] The foregoing is a description of embodiments of the present invention. Through the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for analyzing the performance degradation of electronic products based on large language models, characterized in that, The method identifies and quantitatively evaluates the performance degradation mechanism and its influencing factors of electronic products in actual usage scenarios by iteratively combining a large language model with an expert collaborative scoring mechanism for multiple rounds. The method includes the following steps: 1) Multi-source data acquisition and preprocessing: Acquire and organize multi-source data of electronic products. The multi-source data includes the usage scenario information of electronic products, product functions, design margin analysis reports, performance degradation models, and supplier component performance degradation experiment data. The environmental stresses involved in the usage scenario information include temperature, humidity, radiation, and mechanical vibration. Format, denoise, and semantically parse the multi-source data to obtain a set of key performance parameters, internal independent variables, and external independent variables to be analyzed; 2) Preliminary analysis by the large language model: Input the organized data into a pre-trained large language model to generate a preliminary list of performance degradation causes and degradation mechanisms, including mechanism names, degradation process descriptions, and key performance parameters potentially affected; and the large language model automatically infers a preliminary score for the importance of the degradation mechanism based on the product usage scenario and historical data, including a harmfulness scoring standard and a possibility scoring standard; 3) Expert collaborative multi-round iterative scoring: 3.1) Provide the large language model with descriptions of the degradation mechanism of electronic products, harmfulness scoring standards, possibility scoring standards, usage scenario information, and historical data, and the large language model automatically generates a score for each degradation mechanism; 3.2) The expert group independently scores the degradation mechanism based on their understanding of product functions, usage scenario information, and failure risks, according to the harmfulness scoring standard and the possibility scoring standard; 3.3) For the differences between the scores of the large language model and the expert scores, during the scoring process, comprehensively evaluate the degradation mechanism through the collaborative mechanism of the large language model and the expert scores; For the differences between the scores of the large language model and the expert scores, adopt a multi-round iterative collaborative scoring mechanism: a. If the difference is within the preset threshold, take the average of the two or directly adopt the score of the large language model; b. If the difference exceeds the preset threshold, introduce expert secondary evaluation or weighted fusion algorithm to obtain more accurate harmfulness and possibility scores; After each round of iteration ends, update the analysis parameters or model weights of the large language model, and recalculate the importance of the degradation mechanism; 4) Degradation factor association and model correction: Based on the degradation mechanism and importance determined by multi-round collaborative scoring, combine the correlation relationships between key performance parameters and internal and external independent variables to construct or correct a performance degradation analysis model; through the semantic association ability of the large language model, map each degradation mechanism to the information of corresponding material properties, structural parameters, and environmental stresses to form a complete degradation factor network; 5) Output analysis results and prevention strategies: Output the final harmfulness score, possibility score, and importance of each degradation mechanism; combine expert experience and the inference results of the large language model to output the influencing mechanism of internal and external independent variables on the degradation of key performance parameters, and give corresponding prevention strategies and reliability optimization solutions.

2. The method according to claim 1, wherein: In the multi-source data acquisition and preprocessing step 1), it also includes hierarchical annotation of the environmental stresses involved in the usage scenario information, including temperature, humidity, radiation, and mechanical vibration; and the working loads, including voltage, current, and high-frequency operation, to form a multi-level tag data structure that can be parsed by the large language model.

3. The method according to claim 1, characterized in that: In the large language model pre-analysis step 2), the large language model uses a deep learning network to perform vector embedding on the texturized degradation mechanism description to automatically identify the possible degradation modes under different environmental stresses. The generated list of degradation mechanisms includes the corresponding degradation process descriptions and their correlation relationships with the key performance parameters.

4. The method according to claim 1, characterized in that: The specific algorithm for the multi-round iterative scoring by experts in step 3) is as follows: DS =｜ S GPT －S EXP ｜， Among them, DS is the absolute error of the scores of the two, S GPT is the score of the large language model, S EXP is the score of the expert; Due to the subjective bias in expert scoring, the median method is used to obtain the value of the expert scoring result. The processing of the scoring result is divided into the following specific situations: a, DS When it is 0, directly use the scoring result for importance calculation; b, DS When ≤ 1, take the average of the two scores as the final score; c、 DS When it is > 1, the designer performs the final scoring through the weight distribution method, and the scoring algorithm is as follows: S Final = w GPT ×S GPT + w EXP ×S EXP ； Among them, S Final is the final scoring result, w GPT is the weight of the large language model scoring, w EXP is the weight of the expert scoring, w GPT +w EXP = 1.

5. The method according to claim 1, wherein: In the degradation factor association and model correction step 4), it also includes using the large language model to perform automatic interpretability analysis on the identified degradation mechanisms, and outputting the multi-dimensional association relationships between the degradation mechanisms, key performance parameters, internal variables, and external variables based on the knowledge graph technology, enabling designers to intuitively obtain the degradation sources and influence paths.

6. The method according to claim 1, wherein: In the output analysis result and prevention strategy step 5), the large language model combines the expert feedback information to automatically generate a list of ranked degradation risks and corresponding preventive improvement measures, and visually displays the degradation trends of key performance parameters, the sensitivity of influencing factors, and the ranking of mechanism importance, providing a decision-making basis for subsequent design optimization.

7. An electronic product performance degradation analysis system based on a large language model, characterized in that, The system implements the method described in any one of claims 1-6, including: A data management module for performing multi-source data acquisition and preprocessing, and cleaning, annotating, and archiving the usage scenario data, product design information, and component degradation experiment data; A model analysis module for performing semantic reasoning, mining, and ranking of degradation mechanisms based on the large language model, and automatically associating internal variables and external variables; An expert collaboration module for realizing multi-round iterative fusion of large language model scoring and expert scoring, including a scoring difference comparison and weighted calculation mechanism; A result output and visualization module for outputting the hazard score, possibility score, and importance of degradation mechanisms, and generating a reliability optimization report and visualization charts.

8. The system for analyzing the performance degradation of electronic products based on a large language model according to claim 7, wherein: The expert collaboration module also includes a multi-round interaction interface for experts to supplement and correct the degradation mechanisms, scoring bases, and interpretability conclusions initially identified by the large language model in each round, so that the model analysis results can conform to the actual usage environment and product requirements.

9. The system for analyzing the performance degradation of electronic products based on a large language model according to claim 7, wherein: The system adopts a deployment method combining cloud and local servers, where the cloud server is responsible for the training and inference of the large language model, and the local server is responsible for offline management and secure calculation of sensitive data, and exchanges data with the cloud server through a secure interface; And / or, the system is further configured with a knowledge graph engine for correlating degradation mechanisms, key performance parameters, and their internal and external variables into a multi-dimensional knowledge graph, enabling design engineers to retrieve and view the degradation sources, mechanism inference paths, and corresponding improvement measures based on the graph.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instruction is executed by a processor, it implements the method according to any one of claims 1-6.

Images