Large model fine tuning method based on graphene industry chain application

Abstract

The invention discloses a large model fine tuning method based on graphene industry chain application. The method comprises the following steps: S1, obtaining a multi-dimensional data set of a graphene industry chain; s2, performing data preprocessing on the multi-dimensional data set; s3, constructing a graphene industry chain knowledge graph for supporting information sharing and cooperation of upstream and downstream enterprises on the basis of the multi-dimensional data set in combination with a large model and a knowledge graph technology; s4, performing preliminary fine adjustment on the pre-trained large model; s5, further finely tuning the large model based on the graphene industry chain knowledge graph and the self-supervised learning module; s6, applying the fine-tuned large model to graphene technology research and development, market prediction and industry chain optimization; s7, dynamically adjusting a fine adjustment strategy of the large model according to dynamic changes of technologies and markets in a graphene industry chain; and S8, outputting a fine-tuned large model analysis result. According to the invention, the application efficiency and popularization capability of the graphene technology in various fields are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of graphene technology, and in particular to a large model fine-tuning method based on the application of the graphene industry chain. Background Art

[0002] The graphene industry chain covers multiple links from basic research, technology development, production and manufacturing to market promotion. Accurate analysis and real-time processing of data are the key to ensuring the efficient operation of each link in the industry chain. However, existing technologies often show insufficient analytical capabilities when processing such complex and dynamic multi-dimensional data. It is difficult to effectively integrate data from research literature, technical patents, market data and production processes, making it difficult for information in each link of the graphene industry chain to be efficiently shared and collaborated, resulting in low efficiency in technology research and development and market application promotion.

[0003] Most existing analysis methods rely on traditional statistical models or simple machine learning algorithms, and lack the ability to comprehensively analyze multi-dimensional data. Especially when dealing with dynamic changes in the industrial chain, it is difficult to achieve real-time updates and dynamic optimization of the model. At the same time, cross-domain technology collaboration is also a major challenge in the graphene industry chain. Traditional technology applications are often limited to a single field and cannot fully utilize graphene-related technologies for collaborative optimization. Therefore, how to build an analysis system that can be updated in real time and dynamically adjusted to help each node in the graphene industry chain collaborate efficiently is an urgent problem that needs to be solved in current technology. Summary of the invention

[0004] One purpose of the present invention is to propose a large-model fine-tuning method based on the application of the graphene industry chain. The present invention greatly improves the application efficiency and promotion ability of graphene technology in various fields.

[0005] A large model fine-tuning method based on graphene industry chain application according to an embodiment of the present invention comprises the following steps:

[0006] S1. Obtain a multidimensional dataset of the graphene industry chain, including research literature, technology patents, market analysis, and production process data;

[0007] S2. performing data preprocessing on the multidimensional data set, including data cleaning, format conversion and data feature extraction;

[0008] S3. Based on multidimensional data sets combined with large models and knowledge graph technology, a graphene industry chain knowledge graph is constructed to support information sharing and collaborative cooperation between upstream and downstream enterprises. A dynamic graph update mechanism is introduced to dynamically update the node weights and relationship strengths in the graphene industry chain knowledge graph through upstream and downstream data interaction.

[0009] S4. Perform preliminary fine-tuning on the pre-trained large model, and combine the multi-dimensional collaborative attention mechanism to identify and focus on the key nodes in the technology research and development, market demand and production process in the graphene industry chain;

[0010] S5. Further fine-tune the large model based on the graphene industry chain knowledge graph and self-supervised learning module, dynamically learn and update new knowledge in the industry chain through the self-supervised learning module, and optimize the weights and parameters of the large model;

[0011] S6. Apply the fine-tuned large model to graphene technology research and development, market forecasting and industry chain optimization, analyze technology development trends, market demand and key nodes in the production process in real time, and achieve collaborative optimization of graphene and related technologies through cross-domain knowledge transfer technology;

[0012] S7. Introduce an adaptive feedback loop mechanism to dynamically adjust the fine-tuning strategy of the large model according to the technology and market dynamics in the graphene industry chain, so that the large model is synchronized with the needs of the industry chain;

[0013] S8. Output fine-tuned large model analysis results to assist enterprises in making decisions in technology research and development, market promotion and product application.

[0014] Optionally, the S1 step specifically includes:

[0015] S11. Constructing multi-dimensional data of graphene industry chain:

[0016] Research Literature Dataset D r , by collecting academic papers, technical reports and experimental data related to graphene technology, recording the keywords, citation times, research fields and publication time of research literature;

[0017] Technology Patent Dataset D p , collect graphene-related patent information from all over the world, including patent name, technical field, application date, patent applicant and patent validity status;

[0018] Market Analysis Dataset D m , market research reports and industry white papers based on the graphene industry chain record market share, sales volume, market trends, technical bottlenecks and technical competitor information;

[0019] Production process data set D s , collect the production process data of graphene materials, record the production process, capacity utilization, equipment operating parameters, material cost and production cycle data;

[0020] S12. Integrate the research literature dataset, technology patent dataset, market analysis dataset and production process dataset to construct a multidimensional dataset:

[0021] D={D r ,D p ,D m ,D s}.

[0022] Optionally, the S3 step specifically includes:

[0023] S31. Based on the multidimensional data set D, a knowledge graph of the graphene industry chain is constructed using knowledge graph technology. The knowledge graph of the graphene industry chain includes a node set N and an edge set E. The node set represents the research institutions, enterprises, technical patents and market demand in the graphene industry chain, and the edge set represents the association relationship between the node sets.

[0024] S32. Based on the features extracted from the multidimensional dataset, the research literature dataset D r Keywords and research fields in the r , the technical patent dataset D p The patent applicant and technical field in the p , the market analysis dataset D m Market trends and technological competitors in the m , the production process data set D s The equipment operating parameters and production process in the node N s ,Right now:

[0025] N={N r ,N p ,N m ,N s};

[0026] S33. Dynamically update the weight W(N) and relationship strength R(E) of nodes in the graphene industry chain knowledge graph through data interaction between upstream and downstream enterprises. The node weight W(N) indicates the importance of the node in the graphene industry chain, and the relationship strength R(E) indicates the data correlation between upstream and downstream enterprises:

[0027]

[0028] Among them, W(N x ) is node N x The importance weight in the graphene industry chain, γ is the global adjustment coefficient, is the timestamp weight, is the category weight, combining the domain weights of different categories of graphene-related data. is the inter-node association strength, indicating the relationship strength with other nodes in the graphene industry chain, T x is the time dimension feature, which is used to measure the trend of graphene technology related data evolving over time.x For node N x The characteristic mean of , which represents the average performance of the class data in the graphene industry chain, λ x is the adjustment coefficient, κ d is the dynamic weight, θ m is the market demand impact coefficient, C x is the category feature, σ x is the standard deviation of the node data feature;

[0029]

[0030] Among them, R(E xy ) represents node N x and node N y The strength of the relationship between j is the dynamic adjustment coefficient, For node N x and node N y The feature vector representation of is the interaction frequency between nodes, indicating that node N x and node N y The interaction intensity between them in the graphene industry chain, η xy is the structural relationship adjustment coefficient between nodes, ρ x , y For node N x and node N y The characteristic variance, φ xy The impact of external market fluctuations on the relationship between nodes;

[0031] S34, introduce a dynamic graph update mechanism to adjust the node weight W(N) in real time when the data of upstream and downstream enterprises in the graphene industry chain changes. x ) and the relationship strength R(E xy ).

[0032] Optionally, the S4 step specifically includes:

[0033] S41, based on the pre-trained large model M0 combined with the graphene industry chain knowledge map G (N, E), the large model is initially fine-tuned, and the multi-dimensional collaborative attention mechanism A is used. c Analyze key nodes in multidimensional datasets;

[0034] S42. Multi-dimensional collaborative attention mechanism A c The key nodes in the graphene industry chain are identified by using upstream and downstream association features, category features and time features, and a key node set K = {K1, K2, ..., K n}, calculate the attention weight of the key node set:

[0035]

[0036] in, Represents the key node K i The attention weight, For node K i Score, based on the upstream and downstream association strength R(E) of the node and the category feature C x and time feature T x Calculated;

[0037] S43, adjust the parameters related to the key node set K in the pre-trained large model M0, and update the attention mechanism parameters of the large model in combination with the node feature weight W(N) and the relationship strength R(E) in the graphene industry chain knowledge graph. The key node attention mechanism of the optimized large model M1 is updated as follows:

[0038]

[0039] in, are the parameters of the fine-tuned large model M1, is the parameter of the pre-trained model, η is the learning rate, is the gradient of the large model parameters, is the key node K i The attention weight, f(K i ) is the key node K i The characteristic function captures the specific characteristics of the node in technology research and development, market demand and production process. For key nodes The weight of is the strength of the relationship between nodes, is the time dimension weight, For node K i The time characteristics of t is the time weight adjustment coefficient, is the category feature weight, For node K i The category feature, γ c is the adjustment coefficient of the category feature, and Δt is the time change.

[0040] Optionally, the step S5 specifically includes:

[0041] S51, based on the graphene industry chain knowledge graph G(N,E) and self-supervised learning module S s Further fine-tuning the large model M1, the self-supervised learning module dynamically learns and updates new knowledge in the graphene industry chain by constructing pseudo labels and self-supervised tasks;

[0042] S52, in the self-supervised learning module S sIn the graphene industry chain knowledge graph, feature extraction is performed on node N, and a pseudo-label dataset is constructed through the pseudo-label generation function L(N):

[0043]

[0044] Among them, L(N x ) represents node N x Pseudo-label of , P(y|N x ,W(N x ),R(E x,y )) is node N x In the feature weight W(N x ) and the relationship strength R(E x,y ) classification probability under the condition, Y is the category set;

[0045] S53, combine the pseudo-label dataset to train the large model M1 for self-supervised tasks, and train the loss function It is expressed as:

[0046]

[0047] in, Represents the model M1 for the node The predicted probability of For Node Pseudo labels of

[0048] S54, through the dynamic learning process of the self-supervised learning module, the new knowledge in the graphene industry chain is updated in real time to optimize the parameters of the large model M1 The updated large model parameters are:

[0049]

[0050] Where η is the learning rate, is the gradient of the loss function, indicating the adjustment direction of the large model parameters.

[0051] Optionally, the step S6 specifically includes:

[0052] S61. Apply the fine-tuned large model M1 to graphene technology research and development, market forecasting and industry chain optimization, and use multidimensional data sets to analyze technology development trends, market demand and key nodes in the production process in real time;

[0053] S62. Use the large model M1 to predict the development trend of technology and build a trend prediction model P t :

[0054]

[0055] Among them, P t(K i ) represents the key node K i The predicted value of the technology development trend, σ is the activation function, For Node The feature weights of is the relationship strength between nodes, is the time feature, For node K j The characteristic gradient of

[0056] S63. Predict market demand based on large model M1 and build market demand model P m :

[0057]

[0058] Among them, P m (N x ) represents node N x The value in the market demand forecast, α k is the market demand forecast coefficient, For Node Market demand parameters in the production process, is the relationship strength between nodes, is the time dimension feature, λ m is the market demand adjustment coefficient, is the category feature;

[0059] S64. Through cross-domain knowledge transfer technology T k Realize the collaborative optimization of graphene and related technologies and build a collaborative optimization model C o :

[0060]

[0061] Among them, C o (N x ,N y ) represents node N x With node N y The collaborative optimization value between k is the cross-domain knowledge transfer adjustment coefficient, P t (N x ) is the predicted value of technology development trend, P m (N y ) is the market demand forecast value, R(E xy ) is the relationship strength between nodes, ∈ is a small constant to avoid the denominator being zero;

[0062] S65. Optimize the application efficiency of graphene technology in related industrial chains by analyzing technology development trends, market demands and key nodes in the production process in real time, and realizing the coordinated optimization of graphene and other fields through cross-domain knowledge transfer technology.

[0063] Optionally, the step S7 specifically includes:

[0064] S71. Introducing an adaptive feedback loop mechanism F ad Real-time monitoring of technology and market dynamics in the graphene industry chain, adaptive feedback loop mechanism generates feedback signals by monitoring key indicators in multidimensional data sets s To dynamically adjust the fine-tuning strategy of large models:

[0065]

[0066] Among them, α i is the adjustment factor, is the gradient value of the technology trend forecast, is the gradient value of market demand forecast, T i is the time feature, λ t is the adjustment coefficient, C i is the category feature;

[0067] S72, according to the feedback signal F s Adjust the fine-tuning strategy of the large model in real time and update the parameters of the fine-tuned large model for:

[0068]

[0069] in, is the parameter of the current large model, η is the learning rate, F s It is a feedback signal, indicating the update adjustment direction of the large model parameters;

[0070] S74. When the technology or market dynamics of the graphene industry chain change, the adaptive feedback loop mechanism automatically adjusts the fine-tuning strategy by continuously monitoring and analyzing new data to keep the big model in sync with the actual needs of the industry chain.

[0071] The beneficial effects of the present invention are:

[0072] (1) The present invention uses a multi-dimensional collaborative attention mechanism combined with the graphene industry chain knowledge graph to perform preliminary fine-tuning on the large model, which can identify and focus on the key nodes in the graphene industry chain in real time. When processing complex graphene technology research and development data, market demand information and production process data, it can dynamically adapt to data changes, accurately locate technical breakthroughs and market bottlenecks, and effectively improve the response speed and accuracy of the model in multi-dimensional data analysis.

[0073] (2) The present invention introduces a self-supervised learning module, which further fine-tunes the model by dynamically constructing pseudo-labels and self-supervised tasks. The large model can continuously learn and update new knowledge in the graphene industry chain, perform real-time optimization in the dynamic changes of technology, and automatically capture the latest technological development trends and market demands in the industry chain, thereby enhancing the adaptability of the large model and its ability to predict future trends and reducing dependence on human intervention.

[0074] (3) The present invention realizes the coordinated optimization of graphene technology and other related technologies through cross-domain knowledge transfer technology, and uses the large model of cross-domain knowledge transfer technology to fully explore the potential relationship between graphene and other new materials and technology fields, thereby realizing the global optimization of the industrial chain. In market applications, the model strategy can be dynamically adjusted according to the needs of different industries, which greatly improves the application efficiency and promotion ability of graphene technology in various fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0076] Figure 1 A flow chart of a large model fine-tuning method based on graphene industry chain application proposed by the present invention;

[0077] Figure 2 This is a flowchart of key node identification based on graphene industry chain knowledge graph construction and multi-dimensional collaborative attention mechanism in a large model fine-tuning method based on graphene industry chain application proposed in the present invention. DETAILED DESCRIPTION

[0078] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0079] refer to Figure 1-2 , a large model fine-tuning method based on graphene industry chain application, comprising the following steps:

[0080] S1. Obtain a multidimensional dataset of the graphene industry chain, including research literature, technology patents, market analysis, and production process data;

[0081] S2, preprocessing the multidimensional data set, including data cleaning, format conversion and data feature extraction;

[0082] S3. Based on multidimensional data sets combined with large models and knowledge graph technology, a graphene industry chain knowledge graph is constructed to support information sharing and collaborative cooperation between upstream and downstream enterprises. A dynamic graph update mechanism is introduced to dynamically update the node weights and relationship strengths in the graphene industry chain knowledge graph through upstream and downstream data interaction.

[0083] S4. Perform preliminary fine-tuning on the pre-trained large model, and combine the multi-dimensional collaborative attention mechanism to identify and focus on the key nodes in the technology research and development, market demand and production process in the graphene industry chain;

[0084] S5. Further fine-tune the large model based on the graphene industry chain knowledge graph and self-supervised learning module, dynamically learn and update new knowledge in the industry chain through the self-supervised learning module, and optimize the weights and parameters of the large model;

[0085] S6. Apply the fine-tuned large model to graphene technology research and development, market forecasting and industry chain optimization, analyze technology development trends, market demand and key nodes in the production process in real time, and achieve collaborative optimization of graphene and related technologies through cross-domain knowledge transfer technology;

[0086] S7. Introduce an adaptive feedback loop mechanism to dynamically adjust the fine-tuning strategy of the large model according to the technology and market dynamics in the graphene industry chain, so that the large model is synchronized with the needs of the industry chain;

[0087] S8. Output fine-tuned large model analysis results to assist enterprises in making decisions in technology research and development, market promotion and product application.

[0088] In this implementation, step S1 specifically includes:

[0089] S11. Constructing multi-dimensional data of graphene industry chain:

[0090] Research Literature Dataset D r , by collecting academic papers, technical reports and experimental data related to graphene technology, recording the keywords, citation times, research fields and publication time of research literature;

[0091] Technology Patent Dataset D p , collect graphene-related patent information from all over the world, including patent name, technical field, application date, patent applicant and patent validity status;

[0092] Market Analysis Dataset D m , market research reports and industry white papers based on the graphene industry chain record market share, sales volume, market trends, technical bottlenecks and technical competitor information;

[0093] Production process data set D s, collect the production process data of graphene materials, record the production process, capacity utilization, equipment operating parameters, material cost and production cycle data;

[0094] S12. Integrate the research literature dataset, technology patent dataset, market analysis dataset and production process dataset to construct a multidimensional dataset:

[0095] D={D r ,D p ,D m ,D s}.

[0096] In this embodiment, step S3 specifically includes:

[0097] S31. Based on the multidimensional data set D, a knowledge graph of the graphene industry chain is constructed using knowledge graph technology. The knowledge graph of the graphene industry chain includes a node set N and an edge set E. The node set represents the research institutions, enterprises, technical patents and market demand in the graphene industry chain, and the edge set represents the association relationship between the node sets.

[0098] S32. Based on the features extracted from the multidimensional dataset, the research literature dataset D r Keywords and research fields in the r , the technical patent dataset D p The patent applicant and technical field in the p , the market analysis dataset D m Market trends and technological competitors in the m , the production process data set D s The equipment operating parameters and production process in the node N s ,Right now:

[0099] N={N r ,N p ,N m ,N s};

[0100] S33. Dynamically update the weight W(N) and relationship strength R(E) of nodes in the graphene industry chain knowledge graph through data interaction between upstream and downstream enterprises. The node weight W(N) indicates the importance of the node in the graphene industry chain, and the relationship strength R(E) indicates the data correlation between upstream and downstream enterprises:

[0101]

[0102] Among them, W(N x ) is node N x The importance weight in the graphene industry chain, γ is the global adjustment coefficient, is the timestamp weight, is the category weight, combining the domain weights of different categories of graphene-related data. is the inter-node association strength, indicating the relationship strength with other nodes in the graphene industry chain, T x is the time dimension feature, which is used to measure the trend of graphene technology related data evolving over time. x For node N x The characteristic mean of , which represents the average performance of the class data in the graphene industry chain, λ x is the adjustment coefficient, κ d is the dynamic weight, θ m is the market demand impact coefficient, C x is the category feature, σ x is the standard deviation of the node data feature;

[0103]

[0104] Among them, R(E xy ) represents node N x and node N y The strength of the relationship between j is the dynamic adjustment coefficient, For node N x and node N y The feature vector representation of is the interaction frequency between nodes, indicating that node N x and node N y The interaction intensity between them in the graphene industry chain, η xy is the structural relationship adjustment coefficient between nodes, ρ x , y For node N x and node N y The characteristic variance, φ xy The impact of external market fluctuations on the relationship between nodes;

[0105] S34, introduce a dynamic graph update mechanism to adjust the node weight W(N) in real time when the data of upstream and downstream enterprises in the graphene industry chain changes. x ) and the relationship strength R(E xy ).

[0106] In this implementation, step S4 specifically includes:

[0107] S41, based on the pre-trained large model M0 combined with the graphene industry chain knowledge map G (N, E), the large model is initially fine-tuned, and the multi-dimensional collaborative attention mechanism A is used. c Analyze key nodes in multidimensional datasets;

[0108] S42. Multi-dimensional collaborative attention mechanism A c The key nodes in the graphene industry chain are identified by using upstream and downstream association features, category features and time features, and a key node set K = {K1, K2, ..., K n}, calculate the attention weight of the key node set:

[0109]

[0110] in, Represents the key node K i The attention weight, For node K i Score, based on the upstream and downstream association strength R(E) of the node and the category feature C x and time feature T x Calculated;

[0111] S43, adjust the parameters related to the key node set K in the pre-trained large model M0, and update the attention mechanism parameters of the large model in combination with the node feature weight W(N) and the relationship strength R(E) in the graphene industry chain knowledge graph. The key node attention mechanism of the optimized large model M1 is updated as follows:

[0112]

[0113] in, are the parameters of the fine-tuned large model M1, is the parameter of the pre-trained model, η is the learning rate, is the gradient of the large model parameters, is the key node K i The attention weight, f(K i ) is the key node K i The characteristic function captures the specific characteristics of the node in technology research and development, market demand and production process. For key nodes The weight of is the strength of the relationship between nodes, is the time dimension weight, For node K i The time characteristic of t is the time weight adjustment coefficient, is the category feature weight, For node K i The category feature, γ c is the adjustment coefficient of the category feature, and Δt is the time change.

[0114] In this implementation, step S5 specifically includes:

[0115] S51, based on the graphene industry chain knowledge graph G(N,E) and self-supervised learning module S s Further fine-tuning the large model M1, the self-supervised learning module dynamically learns and updates new knowledge in the graphene industry chain by constructing pseudo labels and self-supervised tasks;

[0116] S52, in the self-supervised learning module S s In the graphene industry chain knowledge graph, feature extraction is performed on node N, and a pseudo-label dataset is constructed through the pseudo-label generation function L(N):

[0117]

[0118] Among them, L(N x ) represents node N x Pseudo-label of , P(y|N x ,W(N x ),R(E x,y )) is node N x In the feature weight W(N x ) and the relationship strength R(E x,y ) classification probability under the condition, Y is the category set;

[0119] S53, combine the pseudo-label dataset to train the large model M1 for self-supervised tasks, and train the loss function It is expressed as:

[0120]

[0121] in, Represents the model M1 for the node The predicted probability of For Node Pseudo labels of

[0122] S54, through the dynamic learning process of the self-supervised learning module, the new knowledge in the graphene industry chain is updated in real time to optimize the parameters of the large model M1 The updated large model parameters are:

[0123]

[0124] Where η is the learning rate, is the gradient of the loss function, indicating the adjustment direction of the large model parameters.

[0125] In this implementation, step S6 specifically includes:

[0126] S61. Apply the fine-tuned large model M1 to graphene technology research and development, market forecasting and industry chain optimization, and use multidimensional data sets to analyze technology development trends, market demand and key nodes in the production process in real time;

[0127] S62. Use the large model M1 to predict the development trend of technology and build a trend prediction model P t :

[0128]

[0129] Among them, P t (K i ) represents the key node K i The predicted value of the technology development trend, σ is the activation function, For Node The feature weights of is the relationship strength between nodes, is the time feature, For node K j The characteristic gradient of

[0130] S63. Predict market demand based on large model M1 and build market demand model P m :

[0131]

[0132] Among them, P m (N x ) represents node N x The value in the market demand forecast, α k is the market demand forecast coefficient, For Node Market demand parameters in the production process, is the relationship strength between nodes, is the time dimension feature, λ m is the market demand adjustment coefficient, is the category feature;

[0133] S64. Through cross-domain knowledge transfer technology T k Realize the collaborative optimization of graphene and related technologies and build a collaborative optimization model C o :

[0134]

[0135] Among them, C o (N x ,N y ) represents node N x With node N y The collaborative optimization value between k is the cross-domain knowledge transfer adjustment coefficient, P t (N x ) is the predicted value of technology development trend, P m (Ny ) is the market demand forecast value, R(E xy ) is the relationship strength between nodes, ∈ is a small constant to avoid the denominator being zero;

[0136] S65. Optimize the application efficiency of graphene technology in related industrial chains by analyzing technology development trends, market demands and key nodes in the production process in real time, and realizing the coordinated optimization of graphene and other fields through cross-domain knowledge transfer technology.

[0137] In this implementation manner, step S7 specifically includes:

[0138] S71. Introducing an adaptive feedback loop mechanism F ad Real-time monitoring of technology and market dynamics in the graphene industry chain, adaptive feedback loop mechanism generates feedback signals by monitoring key indicators in multidimensional data sets s To dynamically adjust the fine-tuning strategy of large models:

[0139]

[0140] Among them, α i is the adjustment factor, is the gradient value of the technology trend forecast, is the gradient value of market demand forecast, T i is the time feature, λ t is the adjustment coefficient, C i is the category feature;

[0141] S72, according to the feedback signal F s Adjust the fine-tuning strategy of the large model in real time and update the parameters of the fine-tuned large model for:

[0142]

[0143] in, is the parameter of the current large model, η is the learning rate, F s It is a feedback signal, indicating the update adjustment direction of the large model parameters;

[0144] S74. When the technology or market dynamics of the graphene industry chain change, the adaptive feedback loop mechanism automatically adjusts the fine-tuning strategy by continuously monitoring and analyzing new data to keep the big model in sync with the actual needs of the industry chain.

[0145] Embodiment 1:

[0146] From March 2024 to June 2024, a large-scale graphene material production enterprise cooperated with its upstream and downstream enterprises to optimize the production process and market forecast of its graphene-based materials by applying the analysis method based on large model fine-tuning of the present invention. The enterprise's data included equipment operating parameters in the production link, the process flow of graphene materials, market sales conditions and patent data. The data volume reached 2TB. Traditional data analysis methods could not process such a large amount of data quickly and efficiently. The enterprise hoped to optimize the efficiency of each link of the industrial chain through the method of the present invention.

[0147] At 10 a.m. on April 15, 2024, the system detected that in the production process of graphene-based materials, the operating parameters of a key equipment showed abnormal fluctuations, especially the temperature parameters of the "compression molding" link exceeded the set threshold for three consecutive days, reaching 375 degrees Celsius, while the standard value is 350 degrees Celsius. The system found through large model analysis that the temperature fluctuation was related to the uneven thickness of the graphene sheets provided by the upstream raw material supplier. After further analysis, the system generated a detailed abnormality report and pushed the problem to the production manager. According to the report content, the production department adjusted the temperature control strategy of compression molding, and immediately contacted the upstream supplier to optimize the sheet thickness, so that the production line returned to normal the next day, avoiding large-scale quality problems caused by abnormal temperature.

[0148] In terms of market forecasting, on May 1, 2024, the system identified changes in the demand for graphene-based materials in the Asian market through the self-supervised learning module. Through the analysis of market sales data, the system found that the sales growth rate in the Asian market reached 25% in the past two weeks, which was higher than the global market growth rate of 8%. At the same time, the system generated a market analysis report, pointing out that the Japanese and Korean markets may have greater demand potential in a short period of time. The analysis report combined sales data, consumer feedback, and upstream and downstream order information, and speculated that the demand in the Asian market will continue to increase by 30% in the next month. Based on this analysis, the company's marketing department quickly adjusted the supply strategy of graphene-based materials, increased product supply to the Asian market, and signed a new supply agreement with local distributors, successfully capturing the opportunity of this market demand growth. Actual data showed that sales in the Asian market in late May increased by 33% year-on-year, which was almost exactly in line with the system's forecast, while other global markets only increased by 9%.

[0149] In the technology research and development phase, on May 10, 2024, the system detected a graphene-based material improvement process patent submitted by a Chinese research institute through analysis of patent data. The patent has a high degree of overlap with the company's current production process. The system uses a multi-dimensional collaborative attention mechanism to focus on analyzing the technical points of the patent and match it with the company's existing technology research and development. The results show that the new process can increase the company's existing production efficiency by about 15%. Based on the analysis report, the R&D department further optimized the production process and incorporated the improvements in the new process into the existing production line. After two weeks of testing, the new process was successfully applied to the company's production process, with an increase in output of 12% and a decrease in production costs of 5%.

[0150] During this period, the system also dynamically monitored the upstream and downstream companies involved in the graphene industry chain. On May 25, 2024, the system detected that the equipment maintenance plan of an upstream raw material supplier may cause delays in the delivery of graphene-based materials. Through an adaptive feedback loop mechanism, the system automatically issued an early warning to the company, pointing out that the supplier's equipment maintenance time exceeded the range allowed in the production plan. The early warning report recorded in detail the supplier's maintenance time, equipment parameters, and the scope of impact on the entire industry chain. The company immediately contacted the supplier and, based on the system's suggestions, negotiated emergency plans with the other two backup suppliers, successfully avoiding potential supply chain disruption risks.

[0151] In the process of comparing the traditional method with the method of the present invention, the enterprise conducted a detailed analysis of production efficiency, market forecast accuracy and supply chain management performance. In the traditional method, the technical team needs to spend a lot of manpower every month to manually process market data, production data and patent information, which takes an average of 500 hours of work per month. With the method of the present invention, the system can complete all data analysis and generate detailed reports within 1 hour, shortening the working time by nearly 99%. At the same time, the market demand forecast error rate of the traditional method is 13%, while the method of the present invention reduces this error rate to 5%. Especially in the adjustment of supply strategy in the Asian market, the method of the present invention identifies the growth trend of market demand in advance, avoids potential sales losses caused by insufficient supply, and is expected to increase direct revenue by 20 million yuan.

[0152] Through the scenario application of Example 1, the graphene industry chain analysis method based on large model fine-tuning of the present invention demonstrates its powerful effect in technology research and development, market forecasting and production process optimization. The system not only has significant advantages in key node identification and market trend forecasting, but also can dynamically adjust production and supply chain strategies through an adaptive feedback loop mechanism to ensure that enterprises always maintain efficient operation in the complex graphene industry chain.

[0153] The present invention uses a multi-dimensional collaborative attention mechanism combined with the graphene industry chain knowledge graph to perform preliminary fine-tuning on the large model. It can identify and focus on key nodes in the graphene industry chain in real time, and can dynamically adapt to data changes when processing complex graphene technology research and development data, market demand information, and production process data. It can accurately locate technical breakthroughs and market bottlenecks, and effectively improve the response speed and accuracy of the large model in multi-dimensional data analysis.

[0154] The present invention introduces a self-supervised learning module, and further fine-tunes the model by dynamically constructing pseudo-labels and self-supervised tasks. The large model can continuously learn and update new knowledge in the graphene industry chain, perform real-time optimization in dynamic technological changes, and automatically capture the latest technological development trends and market demands in the industry chain, thereby enhancing the model's adaptability and ability to predict future trends and reducing reliance on human intervention.

[0155] The present invention realizes the coordinated optimization of graphene technology and technologies in other related fields through cross-domain knowledge transfer technology, and uses the large model of cross-domain knowledge transfer technology to fully explore the potential relationship between graphene and other new materials and technology fields, to achieve global optimization of the industrial chain. In market applications, the model strategy can be dynamically adjusted according to the needs of different industries, which greatly improves the application efficiency and promotion ability of graphene technology in various fields.

[0156] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A large model fine-tuning method based on the application of graphene industry chain, characterized in that: The steps include: S1. Obtain a multidimensional dataset of the graphene industry chain, including research literature, technology patents, market analysis, and production process data; S2. performing data preprocessing on the multidimensional data set, including data cleaning, format conversion and data feature extraction; S3. Based on multidimensional data sets combined with large models and knowledge graph technology, a graphene industry chain knowledge graph is constructed to support information sharing and collaborative cooperation between upstream and downstream enterprises. A dynamic graph update mechanism is introduced to dynamically update the node weights and relationship strengths in the graphene industry chain knowledge graph through upstream and downstream data interaction. S4. Perform preliminary fine-tuning on the pre-trained large model, and combine the multi-dimensional collaborative attention mechanism to identify and focus on the key nodes in the technology research and development, market demand and production process in the graphene industry chain; S5. Further fine-tune the large model based on the graphene industry chain knowledge graph and self-supervised learning module, dynamically learn and update new knowledge in the industry chain through the self-supervised learning module, and optimize the weights and parameters of the large model; S6. Apply the fine-tuned large model to graphene technology research and development, market forecasting and industry chain optimization, analyze technology development trends, market demand and key nodes in the production process in real time, and achieve collaborative optimization of graphene and related technologies through cross-domain knowledge transfer technology; S7. Introduce an adaptive feedback loop mechanism to dynamically adjust the fine-tuning strategy of the large model according to the technology and market dynamics in the graphene industry chain, so that the large model is synchronized with the needs of the industry chain; S8. Output fine-tuned large model analysis results to assist enterprises in making decisions in technology research and development, market promotion and product application.

2. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S1 step specifically includes: S11. Constructing multi-dimensional data of graphene industry chain: Research Literature Dataset D r , by collecting academic papers, technical reports and experimental data related to graphene technology, recording the keywords, citation times, research fields and publication time of research literature; Technology Patent Dataset D p , collect graphene-related patent information from all over the world, including patent name, technical field, application date, patent applicant and patent validity status; Market Analysis Dataset D m , market research reports and industry white papers based on the graphene industry chain record market share, sales volume, market trends, technical bottlenecks and technical competitor information; Production process data set D s , collect the production process data of graphene materials, record the production process, capacity utilization, equipment operating parameters, material cost and production cycle data; S12. Integrate the research literature dataset, technology patent dataset, market analysis dataset and production process dataset to construct a multidimensional dataset: D={D r ,D p ,D m ,D s }。 3. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S3 step specifically includes: S31. Based on the multidimensional data set D, a knowledge graph of the graphene industry chain is constructed using knowledge graph technology. The knowledge graph of the graphene industry chain includes a node set N and an edge set E. The node set represents the research institutions, enterprises, technical patents and market demand in the graphene industry chain, and the edge set represents the association relationship between the node sets. S32. Based on the features extracted from the multidimensional dataset, the research literature dataset D r Keywords and research fields in the r , the technical patent dataset D p The patent applicant and technical field in the p , the market analysis dataset D m Market trends and technological competitors in the m , the production process data set D s The equipment operating parameters and production process in the node N s ,Right now: N={N r ,N p ,N m ,N s }; S33. Dynamically update the weight W(N) and relationship strength R(E) of nodes in the graphene industry chain knowledge graph through data interaction between upstream and downstream enterprises. The node weight W(N) indicates the importance of the node in the graphene industry chain, and the relationship strength R(E) indicates the data correlation between upstream and downstream enterprises: Among them, W(N x ) is node N x The importance weight in the graphene industry chain, γ is the global adjustment coefficient, is the timestamp weight, is the category weight, combining the domain weights of different categories of graphene-related data. is the inter-node association strength, indicating the relationship strength with other nodes in the graphene industry chain, T x is the time dimension feature, which is used to measure the trend of graphene technology related data evolving over time. x For node N x The characteristic mean of , which represents the average performance of the class data in the graphene industry chain, λ x is the adjustment coefficient, κ d is the dynamic weight, θ m is the market demand impact coefficient, C x is the category feature, σ x is the standard deviation of the node data feature; Among them, R(E xy ) represents node N x and node N y The strength of the relationship between j is the dynamic adjustment coefficient, For node N x and node N y The feature vector representation of is the interaction frequency between nodes, indicating that node N x and node N y The interaction intensity between them in the graphene industry chain, η xy is the structural relationship adjustment coefficient between nodes, ρ x , y For node N x and node N y The characteristic variance, φ xy The impact of external market fluctuations on the relationship between nodes; S34, introduce a dynamic graph update mechanism to adjust the node weight W(N) in real time when the data of upstream and downstream enterprises in the graphene industry chain changes. x ) and the relationship strength R(E xy ).

4. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S4 step specifically includes: S41, based on the pre-trained large model M0 combined with the graphene industry chain knowledge map G (N, E), the large model is initially fine-tuned, and the multi-dimensional collaborative attention mechanism A is used. c Analyze key nodes in multidimensional datasets; S42. Multi-dimensional collaborative attention mechanism A c The key nodes in the graphene industry chain are identified by using upstream and downstream association features, category features and time features, and a key node set K = {K1, K2, ..., K n }, calculate the attention weight of the key node set: in, Represents the key node K i The attention weight, For node K i Score, based on the upstream and downstream association strength R(E) of the node and the category feature C x and time feature T x Calculated; S43, adjust the parameters related to the key node set K in the pre-trained large model M0, and update the attention mechanism parameters of the large model in combination with the node feature weight W(N) and the relationship strength R(E) in the graphene industry chain knowledge graph. The key node attention mechanism of the optimized large model M1 is updated as follows: in, are the parameters of the fine-tuned large model M1, is the parameter of the pre-trained model, η is the learning rate, is the gradient of the large model parameters, is the key node K i The attention weight, f(K i ) is the key node K i The characteristic function captures the specific characteristics of the node in technology research and development, market demand and production process. For key nodes The weight of is the strength of the relationship between nodes, is the time dimension weight, For node K i The time characteristic of t is the time weight adjustment coefficient, is the category feature weight, For node K i The category feature, γ c is the adjustment coefficient of the category feature, and Δt is the time change.

5. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S5 step specifically includes: S51, based on the graphene industry chain knowledge graph G(N,E) and self-supervised learning module S s Further fine-tuning the large model M1, the self-supervised learning module dynamically learns and updates new knowledge in the graphene industry chain by constructing pseudo labels and self-supervised tasks; S52, in the self-supervised learning module S s In the graphene industry chain knowledge graph, feature extraction is performed on node N, and a pseudo-label dataset is constructed through the pseudo-label generation function L(N): Among them, L(N x ) represents node N x Pseudo-label of , P(y|N x ,W(N x ),R(E x,y )) is node N x In the feature weight W(N x ) and the relationship strength R(E x,y ) classification probability under the condition, Y is the category set; S53, combine the pseudo-label dataset to train the large model M1 for self-supervised tasks, and train the loss function It is expressed as: in, Represents the model M1 for the node The predicted probability of For Node Pseudo labels of S54, through the dynamic learning process of the self-supervised learning module, the new knowledge in the graphene industry chain is updated in real time to optimize the parameters of the large model M1 The updated large model parameters are: Where η is the learning rate, is the gradient of the loss function, indicating the adjustment direction of the large model parameters.

6. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S6 step specifically includes: S61. Apply the fine-tuned large model M1 to graphene technology research and development, market forecasting and industry chain optimization, and use multidimensional data sets to analyze technology development trends, market demand and key nodes in the production process in real time; S62. Use the large model M1 to predict the technology development trend and build a trend prediction model P t : Among them, P t (K i ) represents the key node K i The predicted value of the technology development trend, σ is the activation function, For Node The feature weights of is the relationship strength between nodes, is the time feature, For node K j The characteristic gradient of S63. Predict market demand based on large model M1 and build market demand model P m : Among them, P m (N x ) represents node N x The value in the market demand forecast, α k is the market demand forecast coefficient, For Node Market demand parameters in the production process, is the relationship strength between nodes, is the time dimension feature, λ m is the market demand adjustment coefficient, is the category feature; S64. Through cross-domain knowledge transfer technology T k Realize the collaborative optimization of graphene and related technologies and build a collaborative optimization model C o : Among them, C o (N x ,N y ) represents node N x With node N y The collaborative optimization value between k is the cross-domain knowledge transfer adjustment coefficient, P t (N x ) is the predicted value of technology development trend, P m (N y ) is the market demand forecast value, R(E xy ) is the relationship strength between nodes, ∈ is a small constant to avoid the denominator being zero; S65. Optimize the application efficiency of graphene technology in related industrial chains by analyzing technology development trends, market demands and key nodes in the production process in real time, and realizing the coordinated optimization of graphene and other fields through cross-domain knowledge transfer technology.

7. The large model fine-tuning method based on the graphene industry chain application according to claim 1 is characterized in that: The S7 step specifically includes: S71. Introducing an adaptive feedback loop mechanism F ad Real-time monitoring of technology and market dynamics in the graphene industry chain, adaptive feedback loop mechanism generates feedback signals by monitoring key indicators in multidimensional data sets s To dynamically adjust the fine-tuning strategy of large models: Among them, α i is the adjustment factor, is the gradient value of the technology trend forecast, is the gradient value of market demand forecast, T i is the time feature, λ t is the adjustment coefficient, C i is the category feature; S72, according to the feedback signal F s Adjust the fine-tuning strategy of the large model in real time and update the parameters of the fine-tuned large model for: in, is the parameter of the current large model, η is the learning rate, F s It is a feedback signal, indicating the update adjustment direction of the large model parameters; S74. When the technology or market dynamics of the graphene industry chain change, the adaptive feedback loop mechanism automatically adjusts the fine-tuning strategy by continuously monitoring and analyzing new data to keep the big model in sync with the actual needs of the industry chain.

Images