A graphene product retrieval method and system based on large model

Abstract

The present invention discloses a graphene product retrieval method and system based on a large model, including S1, receiving a graphene-related query input by a user; S2, parsing the natural language in the query using a variational autoencoder and a Transformer network; S3, extracting graphene product-related information from multiple data sources based on the parsing result; S4, applying image recognition technology to analyze the image data of graphene products, and refining the identification of product attributes and categories; S5, combining the natural language processing results with the image recognition results, and using data mining technology to classify and aggregate information; S6, using a machine learning model to analyze the matching degree between the user's search intention and the aggregated information, and dynamically optimizing the relevance and accuracy of the search results; S7, outputting graphene product information that is highly relevant to the user's query. The present invention parses the natural language query input by the user through a variational autoencoder and a Transformer network, and provides highly relevant and accurate retrieval results.

Description

Technical Field

[0001] The present invention relates to the field of graphene technology, and in particular to a graphene product retrieval method and system based on a large model. Background Art

[0002] As a new type of material, graphene has received extensive attention and application in many fields due to its excellent physical, chemical and mechanical properties. However, with the rapid development of graphene-related research and products, how to efficiently retrieve and screen relevant high-quality information has become an important and urgent problem to be solved. Existing graphene product retrieval methods mainly rely on traditional keyword matching and simple text processing techniques. These methods often cannot accurately understand the user's search intent when faced with complex queries, resulting in low relevance and accuracy of retrieval results. In addition, the wide range of graphene applications and the diversity of data make information integration and processing more complicated and time-consuming.

[0003] The traditional graphene product retrieval system has the following major defects:

[0004] Existing retrieval methods usually rely on simple keyword matching technology, which is unable to handle complex natural language queries, especially when polysemous words, synonyms or context dependencies are involved, which can easily lead to a significant reduction in the relevance and accuracy of retrieval results.

[0005] Traditional methods have limited ability to parse the semantics of user queries and cannot deeply understand the user's actual search intent. This makes it difficult for the system to provide accurate retrieval results when processing queries involving nuances or contextual relationships.

[0006] Graphene-related information is usually distributed in multiple data sources, such as scientific research literature, product catalogs, and patent libraries. These data sources have significant differences in format and structure, which makes information integration and processing very complicated. Existing methods are inefficient when facing multi-source data and it is difficult to provide fast and accurate retrieval results.

[0007] The characteristics and categories of graphene products often need to be identified and analyzed through microscopic and macroscopic images. Traditional retrieval systems have limited capabilities in image data processing and it is difficult to fully utilize image recognition technology to refine the identification of product attributes and categories, which affects the accuracy and comprehensiveness of retrieval results.

[0008] Therefore, how to provide a graphene product retrieval method and system based on a large model is a problem that technical personnel in this field need to solve urgently. Summary of the invention

[0009] One object of the present invention is to propose a graphene product retrieval method and system based on a large model. The present invention parses the natural language query input by the user through a variational autoencoder and a Transformer network, identifies the key concepts and intentions in the query, deeply understands the user's search needs, and provides highly relevant and accurate retrieval results.

[0010] A graphene product retrieval method based on a large model according to an embodiment of the present invention comprises the following steps:

[0011] S1. receiving a graphene-related query input by a user;

[0012] S2, using variational autoencoders and Transformer networks to parse natural language in queries and identify key concepts and intents in queries;

[0013] S3. Extracting graphene product related information from multiple data sources based on the analysis results, including scientific research literature, product catalogs and patent libraries;

[0014] S4. Apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories;

[0015] S5. Combine the results of natural language processing with the results of image recognition and use data mining technology to classify and aggregate information;

[0016] S6. Use machine learning models to analyze the match between user search intent and aggregated information, and dynamically optimize the relevance and accuracy of search results;

[0017] S7. Output graphene product information that is highly relevant to the user query.

[0018] Optionally, the S1 includes the following steps:

[0019] S11, providing a user input interface, wherein the user input interface includes a text input box and a voice input module, allowing the user to input a query through natural language text or voice;

[0020] S12, preprocessing the text or voice data input by the user, the preprocessing step comprising:

[0021] Perform word segmentation on text data and use natural language processing tools to split the input natural language query into independent words or phrases;

[0022] Perform speech recognition processing on speech data, convert speech signals into text, and perform word segmentation processing;

[0023] Remove stop words from the query, including meaningless words;

[0024] Perform synonym replacement, using a synonym dictionary to replace synonyms in the query with standard terms;

[0025] S13, converting the preprocessed query text into a word vector, wherein the word vector is generated by a word embedding model, and the vector representation of each word is calculated using the following formula:

[0026] v(w i )=Embedding(w i );

[0027] Among them, w i Represents each word in the query, Embedding(w i ) represents the word vector generated by the word embedding model;

[0028] S14, aggregate the word vectors to generate a query vector q, which is calculated using the following formula:

[0029]

[0030] Where n represents the number of words in the query, v(w i ) represents the word vector of each word;

[0031] S15, input the query vector q into the variational autoencoder for dimensionality reduction processing to generate a low-dimensional latent vector z;

[0032] S16. Input the low-dimensional latent vector z into the Transformer network to perform deep semantic parsing, identify key concepts and intentions in the query, and generate a parsing result r. The parsing result is calculated by the following formula:

[0033] r = Transformer(z);

[0034] Among them, Transformer(z) represents the output result of the Transformer network;

[0035] S17. Store the parsing result r in a temporary query library.

[0036] Optionally, S2 includes the following steps:

[0037] S21. Input the preprocessed query vector q into the variational autoencoder combined with the knowledge graph for encoding to generate a latent vector z:

[0038] μ(q)=W μ q+b μ ;

[0039] logσ 2 (q) = Wσ q+b σ ;

[0040]

[0041] Among them, W μ and W σ represents the weight matrix, b μ and b σ represents the bias vector, ∈ represents the noise vector sampled from the standard normal distribution N(0,I);

[0042] S22. Input the latent vector z into the physicochemical property embedding module to generate an enhanced latent vector z':

[0043] z′=z+PhysChem(q);

[0044] Where PhysChem(q) represents query-based physicochemical property embedding;

[0045] S23. The enhanced latent vector z′ is input into the Transformer network in the multimodal data processing module for deep semantic parsing. During the parsing process, a multi-head self-attention mechanism and domain-specific optimization items are used to identify key concepts and intentions in the query. The parsing result r is calculated by the following formula:

[0046] Q=W q ·z′;

[0047] K=W k ·z′;

[0048] V=W v ·z′;

[0049]

[0050] r=Concat(head1,head2,…,head h )W o ;

[0051] Where DomainOpt(z′) represents the specific optimization item based on the graphene domain, d k represents the dimension of the key vector, W q ,W k ,W v The weight matrix representing the query, key, and value, head i represents the attention mechanism result of the i-th head, W o represents the output weight matrix;

[0052] S24. Apply position encoding and graphene-specific knowledge graph to enhance the parsing ability of Transformer network. Position encoding is calculated by the following formula:

[0053]

[0054] Among them, pos represents the position, i represents the dimension index, and d model Indicates the dimension size of the model;

[0055] S25. Through the complex network analysis module, the potential vector z' is iteratively processed to further refine the recognition of the query intent. The calculation process of the multi-layer Transformer is as follows:

[0056] h (l+1) =LayerNorm(h (l) +Dropout(FeedForward(MultiHead(h (l) ,h (l) ,h (l) )))+ComplexNet(z′));

[0057] Among them, h (l) represents the hidden state of the lth layer, Dropout represents dropout regularization, FeedForward represents a feedforward neural network, LayerNorm represents a layer normalization operation, and ComplexNet(z′) represents an optimization term based on complex network analysis;

[0058] S26, through the adaptive learning mechanism, the results of deep semantic analysis r and user feedback are combined to dynamically adjust and optimize the retrieval algorithm and results;

[0059] S27, the final optimized analysis result r ′ Output.

[0060] Optionally, S3 includes the following steps:

[0061] S31, according to the analysis result r ′ Generate a query template, wherein the query template includes keywords, phrases and subject concepts;

[0062] S32, inputting the query template into multiple data source interfaces to execute distributed query, wherein the data source interfaces include a scientific research literature database, a product catalog database, and a patent library;

[0063] S33, pre-processing the query results returned by each data source, removing redundant data, unifying the format and performing preliminary denoising;

[0064] S34, indexing the preprocessed data results to generate an index vector I i , each index vector I i Calculated by the following formula:

[0065]

[0066] Among them, d ij Represents the attribute value of each data result, Represents the weight coefficient of the corresponding attribute, α j , β j and γ j They represent the adjustment parameters related to the analytical result F, ψ(d ij ) represents the attribute transformation function, ω j represents the frequency parameter, T represents the upper limit of integration, and m represents the number of attributes;

[0067] S35, index vector I i Perform natural language processing and image recognition technology to generate feature vector F i , the eigenvector F i Calculated by the following formula:

[0068]

[0069] Among them, I ik represents the natural language features in each index vector, k(I ik ) represents the natural language feature transformation function, I ik Represents the image features in each index vector, ω(I il ) represents the image feature transformation function, θ k and λ l Represent the weight coefficients of natural language and image features respectively, ξ l , δ l ,∈ l and l They represent the adjustment parameters related to the analytical result r′, σ represents the overall feature aggregation function, and χ l represents the frequency correlation coefficient, ω l represents the frequency parameter, T represents the upper limit of integration, p and q represent the number of natural language and image features respectively;

[0070] S36, extract the feature vector F i Input to the aggregation module, classify and aggregate the information according to the parsing result r', and generate the aggregation vector A. The aggregation process is as follows:

[0071]

[0072] Among them, αi , δ j and k Represent the weight coefficients of the eigenvector and analytical results, β(F i ) and ζ(F j ,r′) respectively represent the feature vector transformation function based on the analytical result r′ and the feature-analysis result interaction function, denote the adjustment parameters related to the feature vector and the analytical result r′, respectively, η denotes the overall aggregation function, and λ k represents the frequency correlation coefficient, ω k represents the frequency parameter, T represents the upper limit of integration, n, m and p represent the number of eigenvectors and analytical results respectively;

[0073] S37: Input the aggregated vector A into a database for storage.

[0074] Optionally, S4 includes the following steps:

[0075] S41, receiving image data of a graphene product, wherein the image data includes a microscopic image and a macroscopic image;

[0076] S42, performing image denoising, contrast enhancement and edge detection on the received image data to obtain pre-processed image data I pre ;

[0077] S43, use convolutional neural network to extract features from the preprocessed image data and generate image feature vector F img ;

[0078] S44, image feature vector F img Classification processing is performed to identify the properties and categories of graphene products, and the classification processing is calculated by the following formula:

[0079]

[0080] Among them, Classifier(F img ) represents the processing function of the classifier on the image feature vector, W i and V i represents the weight matrix, b i represents the bias vector, σ represents the activation function, ∈ i , α i , β i , γ i Represents the adjustment parameter, relu represents the modified linear unit function, and tanh represents the hyperbolic tangent function;

[0081] S45, the classification result C is merged with the natural language processing result r′ to generate refined product attributes and category information. The fusion process is as follows:

[0082]

[0083] Among them, α and β represent weight coefficients, δ i ,θ i , i , μ i 、v i represents the adjustment parameter, represents the fusion function, W i and V i represents the weight matrix in the fusion process, b i represents the bias vector, A fusion Indicates the fused product attributes and category information;

[0084] S46, the merged product attributes and category information A fusion Stored in a temporary database.

[0085] Optionally, S5 includes the following steps:

[0086] S51, receiving the natural language processing result r′ and the fused product attribute and category information A fusion ;

[0087] S52, the natural language processing result r′ and the fused product attributes and category information A fusion Convert to a unified data format and generate a joint feature vector U;

[0088] S53, extracting features from the joint feature vector U using data mining technology to generate a feature matrix M;

[0089] S54, classify the feature matrix M, cluster and classify the extracted features, and generate a classification result C final ;

[0090] S55, the classification result C final Input into the aggregation module, aggregate the information according to data mining technology, and generate the final aggregation result A final :

[0091]

[0092] Among them, α i represents the weight coefficient, k represents the number of classification results, C final,i Indicates each classification result;

[0093] S56, the final aggregation result A finalStored in the database.

[0094] Optionally, the S6 includes the following steps:

[0095] S61. Receive user's search intention data and aggregate information A final ;

[0096] S62: The user's search intention data and aggregate information A final Convert to a unified data format to generate the user intention feature vector U intent and the aggregate information feature vector U agg ;

[0097] S63. Calculate the user intention feature vector U using the matching analysis model intent and the aggregate information feature vector U agg The matching degree is calculated by the following formula:

[0098]

[0099] Among them, cos represents the cosine similarity function, · represents the vector dot product, and ||·|| represents the modulus of the vector;

[0100] S64, dynamically optimize the search results according to the matching results, the optimization steps include adjusting the ranking weight of the search results and filtering irrelevant information, and the optimized search results R opt Calculated by the following formula:

[0101]

[0102] Among them, w i represents the weight of the i-th search result, r i represents the i-th search result, and n represents the number of search results;

[0103] S65. Use a reinforcement learning algorithm to perform online learning and adjustment on the matching analysis model. The reinforcement learning process includes the following steps:

[0104] Define a reward function R, which is calculated based on the user's feedback and the relevance of the search results;

[0105] Use reinforcement learning algorithms to update model parameters so that the matching analysis model can adaptively adjust and optimize search results;

[0106] S66, the optimized search result R opt Output to the user and store the relevant data in the database.

[0107] A retrieval system for a graphene product retrieval method based on a large model, comprising the following modules:

[0108] A user input module, used to receive graphene-related queries input by a user, the user input module comprising a text input box and a voice input module;

[0109] The natural language processing module includes a variational autoencoder and a Transformer network, which is used to parse natural language queries input by users and identify key concepts and intents in the queries;

[0110] A data extraction module, used to extract information related to graphene products from multiple data sources, including a scientific research literature database, a product catalog database and a patent database;

[0111] Image processing module, used to apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories;

[0112] Information classification and aggregation module, which is used to combine natural language processing results with image recognition results and use data mining technology to classify and aggregate information;

[0113] Matching analysis and optimization module, including machine learning models, is used to analyze the matching degree between the user's search intent and the aggregated information, and dynamically optimize the relevance and accuracy of search results;

[0114] The result output module is used to output the optimized search results and store the relevant data in the database.

[0115] The beneficial effects of the present invention are:

[0116] (1) The present invention parses the natural language queries input by users through variational autoencoders and Transformer networks, identifies key concepts and intentions in the queries, deeply understands the search needs of users, and provides highly relevant and accurate retrieval results, thus overcoming the limitations of traditional keyword matching methods in processing complex queries.

[0117] (2) The data extraction module of the present invention can extract graphene product-related information from multiple data sources (scientific research literature database, product catalog database and patent library), and perform unified formatting and preliminary denoising, which significantly improves the efficiency of information integration and processing and solves the problems of data dispersion and format differences.

[0118] (3) The present invention uses an image processing module and a convolutional neural network to analyze the microscopic and macroscopic images of graphene products, generate accurate image feature vectors, and combine the natural language processing results to refine the identification of product attributes and categories, thereby providing more comprehensive and accurate retrieval results.

[0119] (4) The information classification and aggregation module of the present invention combines the results of natural language processing with the results of image recognition, uses data mining technology to classify and aggregate information, generates joint feature vectors and feature matrices, improves the accuracy and efficiency of information processing, and enables the retrieval system to accurately classify and aggregate large amounts of data.

[0120] (5) The matching analysis and optimization module of the present invention uses a machine learning model to analyze the matching degree between the user's search intent and the aggregated information, dynamically optimizes the relevance and accuracy of the search results, and uses a reinforcement learning algorithm for online learning and adjustment, so that the system can adaptively optimize the search results based on user feedback and interaction data, and continuously improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] 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:

[0122] Figure 1 The present invention provides a flowchart of a graphene product retrieval method and system based on a large model. DETAILED DESCRIPTION

[0123] 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.

[0124] refer to Figure 1 , a graphene product retrieval method based on a large model comprises the following steps:

[0125] S1. receiving a graphene-related query input by a user;

[0126] In this implementation, S1 includes the following steps:

[0127] S11, providing a user input interface, the user input interface including a text input box and a voice input module, allowing the user to input queries through natural language text or voice;

[0128] S12, preprocessing the text or voice data input by the user, the preprocessing step comprising:

[0129] Perform word segmentation on text data and use natural language processing tools to split the input natural language query into independent words or phrases;

[0130] Perform speech recognition processing on speech data, convert speech signals into text, and perform word segmentation processing;

[0131] Remove stop words from the query, including meaningless words;

[0132] Perform synonym replacement, using a synonym dictionary to replace synonyms in the query with standard terms;

[0133] S13. Convert the preprocessed query text into word vectors. The word vectors are generated by the word embedding model. The vector representation of each word is calculated using the following formula:

[0134] v(w i )=Embedding(w i );

[0135] Among them, w i Represents each word in the query, Embedding(w i ) represents the word vector generated by the word embedding model;

[0136] S14. Aggregate the word vectors to generate a query vector q. The query vector is calculated using the following formula:

[0137]

[0138] Where n represents the number of words in the query, v(w i ) represents the word vector of each word;

[0139] S15, input the query vector q into the variational autoencoder for dimensionality reduction processing to generate a low-dimensional latent vector z;

[0140] S16. Input the low-dimensional latent vector z into the Transformer network for deep semantic parsing, identify the key concepts and intents in the query, and generate the parsing result r. The parsing result is calculated using the following formula:

[0141] r = Transformer(z);

[0142] Among them, Transformer(z) represents the output result of the Transformer network;

[0143] S17. Store the parsing result r in a temporary query library.

[0144] S2, using variational autoencoders and Transformer networks to parse natural language in queries and identify key concepts and intents in queries;

[0145] In this implementation, S2 includes the following steps:

[0146] S21. Input the preprocessed query vector q into the variational autoencoder combined with the knowledge graph for encoding to generate a latent vector z:

[0147] μ(q)=Wμ q+b μ ;

[0148] logσ 2 (q) = W σ q+b σ ;

[0149]

[0150] Among them, W μ and W σ represents the weight matrix, b μ and b σ represents the bias vector, ε represents the noise vector sampled from the standard normal distribution N(0,I);

[0151] S22. Input the latent vector z into the physicochemical property embedding module to generate an enhanced latent vector z':

[0152] z′=z+PhysChem(q);

[0153] Where PhysChem(q) represents query-based physicochemical property embedding;

[0154] S23. The enhanced latent vector z′ is input into the Transformer network in the multimodal data processing module for deep semantic parsing. During the parsing process, a multi-head self-attention mechanism and domain-specific optimization items are used to identify key concepts and intentions in the query. The parsing result r is calculated by the following formula:

[0155] Q=W q ·z′;

[0156] K=W k ·z′;

[0157] V=W v ·z′;

[0158]

[0159] r=Concat(head1,head2,…,head h )W o ;

[0160] Where DomainOpt(z′) represents the specific optimization item based on the graphene domain, d k represents the dimension of the key vector, W q ,W k ,W v The weight matrix representing the query, key, and value, head i represents the attention mechanism result of the i-th head, Wo represents the output weight matrix;

[0161] S24. Apply position encoding and graphene-specific knowledge graph to enhance the parsing ability of Transformer network. Position encoding is calculated by the following formula:

[0162]

[0163] Among them, pos represents the position, i represents the dimension index, and d model Indicates the dimension size of the model;

[0164] S25. Through the complex network analysis module, the potential vector z' is iteratively processed to further refine the recognition of the query intent. The calculation process of the multi-layer Transformer is as follows:

[0165] h (l+1) =LayerNorm(h (l) +Dropout(FeedForward(MultiHead(h (l) ,h (l) ,h (l) )))+ComplexNet(z′));

[0166] Among them, h (l) represents the hidden state of the lth layer, Dropout represents dropout regularization, FeedForward represents a feedforward neural network, LayerNorm represents a layer normalization operation, and ComplexNet(z′) represents an optimization term based on complex network analysis;

[0167] S26, through the adaptive learning mechanism, the results of deep semantic analysis r and user feedback are combined to dynamically adjust and optimize the retrieval algorithm and results;

[0168] S27, the final optimized analysis result r ′ Output.

[0169] S3. Extracting graphene product related information from multiple data sources based on the analysis results, including scientific research literature, product catalogs and patent libraries;

[0170] In this implementation, S3 includes the following steps:

[0171] S31, according to the analysis result r ′ Generate a query template, which includes keywords, phrases and subject concepts;

[0172] S32, inputting the query template into multiple data source interfaces to execute distributed query, the data source interfaces including scientific research literature database, product catalog database and patent library;

[0173] S33, pre-processing the query results returned by each data source, removing redundant data, unifying the format and performing preliminary denoising;

[0174] S34, indexing the preprocessed data results to generate an index vector I i , each index vector I i Calculated by the following formula:

[0175]

[0176] Among them, d ij Represents the attribute value of each data result, Represents the weight coefficient of the corresponding attribute, α j , β j and γ j They represent the adjustment parameters related to the analytical result F, ψ(d ij ) represents the attribute transformation function, ω j represents the frequency parameter, T represents the upper limit of integration, and m represents the number of attributes;

[0177] S35, index vector I i Perform natural language processing and image recognition technology to generate feature vector F i , the eigenvector F i Calculated by the following formula:

[0178]

[0179] Among them, I ik represents the natural language features in each index vector, κ(I ik ) represents the natural language feature transformation function, I ik Represents the image features in each index vector, ω(I il ) represents the image feature transformation function, θ k and λ l Represent the weight coefficients of natural language and image features respectively, ξ l , δ l ,∈ l and l They represent the adjustment parameters related to the analytical result r′, σ represents the overall feature aggregation function, and χ l represents the frequency correlation coefficient, ω l represents the frequency parameter, T represents the upper limit of integration, p and q represent the number of natural language and image features respectively;

[0180] S36, extract the feature vector F iInput to the aggregation module, classify and aggregate the information according to the parsing result r', and generate the aggregation vector A. The aggregation process is as follows:

[0181]

[0182] Among them, α i ,δ j and k Represent the weight coefficients of the eigenvector and analytical results, β(F i ) and ζ(F j ,r′) respectively represent the feature vector transformation function based on the analytical result r′ and the feature-analysis result interaction function, denote the adjustment parameters related to the feature vector and the analytical result r′, respectively, η denotes the overall aggregation function, and λ k represents the frequency correlation coefficient, ω k represents the frequency parameter, T represents the upper limit of integration, n, m and p represent the number of eigenvectors and analytical results respectively;

[0183] S37: Input the aggregated vector A into a database for storage.

[0184] S4. Apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories;

[0185] In this implementation, S4 includes the following steps:

[0186] S41, receiving image data of the graphene product, where the image data includes a microscopic image and a macroscopic image;

[0187] S42, performing image denoising, contrast enhancement and edge detection on the received image data to obtain pre-processed image data I pre ;

[0188] S43, use convolutional neural network to extract features from the preprocessed image data and generate image feature vector F img ;

[0189] S44, image feature vector F img Classification processing is performed to identify the properties and categories of graphene products. The classification processing is calculated using the following formula:

[0190]

[0191] Among them, Classifier(F img ) represents the processing function of the classifier on the image feature vector, W i and V i represents the weight matrix, b irepresents the bias vector, σ represents the activation function, ε i , α i , β i , γ i Represents the adjustment parameter, relu represents the modified linear unit function, and tanh represents the hyperbolic tangent function;

[0192] S45, the classification result C is merged with the natural language processing result r′ to generate refined product attributes and category information. The fusion process is as follows:

[0193]

[0194] Among them, α and β represent weight coefficients, δ i ,θ i , i , μ i , ν i represents the adjustment parameter, represents the fusion function, W i and V i represents the weight matrix in the fusion process, b i represents the bias vector, A fusion Indicates the fused product attributes and category information;

[0195] S46, the fused product attributes and category information A fusion Stored in a temporary database.

[0196] S5. Combine the results of natural language processing with the results of image recognition and use data mining technology to classify and aggregate information;

[0197] In this implementation, S5 includes the following steps:

[0198] S51, receiving the natural language processing result r′ and the fused product attribute and category information A fusion ;

[0199] S52, the natural language processing result r′ and the fused product attributes and category information A fusion Convert to a unified data format and generate a joint feature vector U;

[0200] S53, extracting features from the joint feature vector U using data mining technology to generate a feature matrix M;

[0201] S54, classify the feature matrix M, cluster and classify the extracted features, and generate a classification result C final ;

[0202] S55, the classification result C finalInput into the aggregation module, aggregate the information according to data mining technology, and generate the final aggregation result A final :

[0203]

[0204] Among them, α i represents the weight coefficient, k represents the number of classification results, C final,i Indicates each classification result;

[0205] S56, the final aggregation result A final Stored in the database.

[0206] S6. Use machine learning models to analyze the match between user search intent and aggregated information, and dynamically optimize the relevance and accuracy of search results;

[0207] In this implementation, S6 includes the following steps:

[0208] S61. Receive user's search intention data and aggregate information A final ;

[0209] S62: The user's search intention data and aggregate information A fianl Convert to a unified data format to generate the user intention feature vector U intent and the aggregate information feature vector U agg ;

[0210] S63. Calculate the user intention feature vector U using the matching analysis model intent and the aggregate information feature vector U agg The matching degree is calculated by the following formula:

[0211]

[0212] Among them, cos represents the cosine similarity function, · represents the vector dot product, and ||·|| represents the modulus of the vector;

[0213] S64, dynamically optimize the search results according to the matching results, the optimization steps include adjusting the ranking weight of the search results and filtering irrelevant information, and the optimized search results R opt Calculated by the following formula:

[0214]

[0215] Among them, w i represents the weight of the i-th search result, r i represents the i-th search result, and n represents the number of search results;

[0216] S65. Use a reinforcement learning algorithm to perform online learning and adjustment on the matching analysis model. The reinforcement learning process includes the following steps:

[0217] Define a reward function R, which is calculated based on the user's feedback and the relevance of the search results;

[0218] Use reinforcement learning algorithms to update model parameters so that the matching analysis model can adaptively adjust and optimize search results;

[0219] S66, the optimized search result R opt Output to the user and store the relevant data in the database.

[0220] S7. Output graphene product information that is highly relevant to the user query.

[0221] A retrieval system for a graphene product retrieval method based on a large model, comprising the following modules:

[0222] A user input module, used to receive graphene-related queries input by a user, the user input module including a text input box and a voice input module;

[0223] The natural language processing module includes a variational autoencoder and a Transformer network, which is used to parse natural language queries input by users and identify key concepts and intents in the queries;

[0224] A data extraction module is used to extract information related to graphene products from multiple data sources, including scientific research literature database, product catalog database and patent database;

[0225] Image processing module, used to apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories;

[0226] Information classification and aggregation module, which is used to combine natural language processing results with image recognition results and use data mining technology to classify and aggregate information;

[0227] Matching analysis and optimization module, including machine learning models, is used to analyze the matching degree between the user's search intent and the aggregated information, and dynamically optimize the relevance and accuracy of search results;

[0228] The result output module is used to output the optimized search results and store the relevant data in the database.

[0229] This embodiment fundamentally solves the problems existing in the prior art by combining variational autoencoders, Transformer networks, convolutional neural networks and reinforcement learning algorithm technologies. Specifically, the method of the present invention parses the natural language query input by the user through the variational autoencoder and the Transformer network, identifies the key concepts and intentions in the query, thereby improving the relevance and accuracy of the retrieval; integrates the information of multiple data sources through the data extraction module, solves the problem of information dispersion and diversity; applies image recognition technology through the image processing module, analyzes and refines the image data of graphene products, thereby providing more comprehensive and accurate retrieval results; through the information classification and aggregation module, combines the natural language processing results with the image recognition results, uses data mining technology to classify and aggregate information, and further improves the efficiency and accuracy of information processing; through the matching analysis and optimization module, uses a machine learning model to analyze the matching degree between the user's search intention and the aggregated information, and dynamically optimizes the relevance and accuracy of the search results to improve the user experience.

[0230] Embodiment 1:

[0231] In order to verify the feasibility of this system, a test was conducted after completion, in which the query keyword "graphene battery" was entered, hoping that the system would provide highly relevant and accurate information.

[0232] In the test, the query keyword "graphene battery" was entered into the user input module of the system. After the system receives the query input by the user, it first parses it through the natural language processing module, using the variational autoencoder and Transformer network to identify the key concepts and intentions in the query. The parsed query vector is input into the data extraction module, and the system begins to extract relevant information from multiple data sources, including scientific research literature databases, product catalog databases, and patent libraries.

[0233] The data received by the system includes a large amount of text information and image data. In order to better identify the properties and categories of graphene batteries, the system uses the image processing module to preprocess these data. Through steps such as image denoising, contrast enhancement and edge detection, the system obtains high-quality preprocessed image data. Then, a convolutional neural network is used to extract features from the preprocessed image data, generate image feature vectors, and further classify and identify the specific properties and categories of graphene batteries.

[0234] Next, the system combines the natural language processing results with the image recognition results and uses data mining technology to classify and aggregate information. The system generates joint feature vectors and feature matrices, and uses machine learning models to analyze the matching degree between the user's search intent and the aggregated information through the matching analysis and optimization module, and dynamically optimizes the relevance and accuracy of search results.

[0235] Concrete data proves our beneficial effects:

[0236] In this process, we selected three different data sources for retrieval, as shown in Table 1:

[0237] Table 1 Data source categories

[0238]

[0239] 1. Scientific research literature database: In the past year, it contains about 5,145 latest research papers on graphene batteries.

[0240] 2. Product catalog database: covers product information of more than 100 graphene battery manufacturers around the world, with a cumulative data volume of more than 22,319 items.

[0241] 3. Patent library: Contains approximately 8,321 patent information related to graphene batteries in the past five years.

[0242] After receiving the query from the test center, the system completed the initial data extraction and preprocessing in just 10 seconds, and further feature extraction and classification processing took about 30 seconds. Finally, the system generated the optimized search results in less than 1 minute and output them to the test center.

[0243] In the output stage, the system provides the following high-quality graphene battery related information:

[0244] 1. The latest research paper title and abstract, such as "Research on high energy density batteries based on new graphene composite materials", and a link to download the full text is provided.

[0245] 2. The latest product information from world-renowned graphene battery manufacturers, including product specifications, performance indicators, application areas, etc. For example, a company launched a "high-capacity graphene lithium battery" with an energy density of 350Wh / kg and a cycle life of more than 2,000 times.

[0246] 3. Graphene battery-related patents in the past five years, including patent names, applicants, abstracts and detailed technical descriptions. For example, a patent on "graphene-based fast-charging battery" introduces a new type of electrode material and its preparation method, which significantly improves the battery's charging and discharging efficiency.

[0247] Through this embodiment, the test not only quickly obtains the required high-quality information, but also saves a lot of time and energy. The specific data are shown in Table 2 below:

[0248] Table 2 User feedback and optimization results

[0249] Before optimization (traditional method) After optimization (this system) Average relevance score: 65.54% Average relevance score: 91.54% Average search time: 5 minutes Average search time: 1 minute

[0250] The data shows that compared with traditional retrieval methods, our system has significantly improved relevance and accuracy. For example, in 100 queries, the average relevance score of the traditional method was 65.54%, while the average relevance score of our system reached 91.54%. In addition, the average retrieval time of the traditional method was 5 minutes, while the retrieval time of our system was only 1 minute, which greatly improved work efficiency.

[0251] In summary, through this embodiment, we have demonstrated the advantages and beneficial effects of the large model-based graphene product retrieval method in practical applications, and proved the significant advantages of the system in improving retrieval relevance and accuracy, integrating multi-source data, and optimizing retrieval efficiency.

[0252] 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 graphene product retrieval method based on a large model, characterized in that: The steps include: S1. receiving a graphene-related query input by a user; S2, using variational autoencoders and Transformer networks to parse natural language in queries and identify key concepts and intents in queries; S3. Extracting graphene product related information from multiple data sources based on the analysis results, including scientific research literature, product catalogs and patent libraries; S4. Apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories; S5. Combine the results of natural language processing with the results of image recognition and use data mining technology to classify and aggregate information; S6. Use machine learning models to analyze the match between user search intent and aggregated information, and dynamically optimize the relevance and accuracy of search results; S7, outputting graphene product information that is highly relevant to the user query; The S2 comprises the following steps: S21. Input the preprocessed query vector q into the variational autoencoder combined with the knowledge graph for encoding to generate a latent vector z: μ(q)=W μ q+b μ ; logσ 2 (q)=W σ q+b σ ; Among them, W μ and W σ represents the weight matrix, b μ and b σ represents the bias vector, ε represents the noise vector sampled from the standard normal distribution N(0,I); S22. Input the latent vector z into the physicochemical property embedding module to generate an enhanced latent vector z': z′=z+PhysChem(q); Where PhysChem(q) represents query-based physicochemical property embedding; S23. The enhanced latent vector z′ is input into the Transformer network in the multimodal data processing module for deep semantic parsing. During the parsing process, a multi-head self-attention mechanism and domain-specific optimization items are used to identify key concepts and intentions in the query. The parsing result r is calculated by the following formula: Q=W q ·z′; W=W k ·With'; V=W v ·z′; r=Concat(head1,head2,…,head h )W o ; Where DomainOpt(z′) represents the specific optimization item based on the graphene domain, d k represents the dimension of the key vector, W q ,W k ,W v Represents the weight matrix of query, key and value, head i represents the attention mechanism result of the i-th head, W o represents the output weight matrix; S24. Apply position encoding and graphene-specific knowledge graph to enhance the parsing ability of Transformer network. Position encoding is calculated by the following formula: Among them, pos represents the position, i represents the dimension index, and d model Indicates the dimension size of the model; S25. Through the complex network analysis module, the potential vector z' is iteratively processed to further refine the recognition of the query intent. The calculation process of the multi-layer Transformer is as follows: h (l+1) =LayerNorm(h (l) +Dropout(FeedForward(MultiHead(h (l) ,h (l) ,h (l) )))+ComplexNet(z′)); Among them, h (l) represents the hidden state of the lth layer, Dropout represents dropout regularization, FeedForward represents a feedforward neural network, LayerNorm represents a layer normalization operation, and ComplexNet(z′) represents an optimization term based on complex network analysis; S26, through the adaptive learning mechanism, the results of deep semantic analysis r and user feedback are combined to dynamically adjust and optimize the retrieval algorithm and results; S27, output the final optimized analysis result r′.

2. A graphene product retrieval method based on a large model according to claim 1, characterized in that: The S1 comprises the following steps: S11, providing a user input interface, wherein the user input interface includes a text input box and a voice input module, allowing the user to input a query through natural language text or voice; S12, preprocessing the text or voice data input by the user, the preprocessing step comprising: Perform word segmentation on text data and use natural language processing tools to split the input natural language query into independent words or phrases; Perform speech recognition processing on speech data, convert speech signals into text, and perform word segmentation processing; Remove stop words from the query, including meaningless words; Perform synonym replacement, using a synonym dictionary to replace synonyms in the query with standard terms; S13, converting the preprocessed query text into a word vector, wherein the word vector is generated by a word embedding model, and the vector representation of each word is calculated using the following formula: v(w i )=Embedding(w i ); Among them, w i Represents each word in the query, Embedding(w i ) represents the word vector generated by the word embedding model; S14, aggregate the word vectors to generate a query vector q, which is calculated using the following formula: Where n represents the number of words in the query, v(w i ) represents the word vector of each word; S15, input the query vector q into the variational autoencoder for dimensionality reduction processing to generate a low-dimensional latent vector z; S16. Input the low-dimensional latent vector z into the Transformer network to perform deep semantic parsing, identify key concepts and intentions in the query, and generate a parsing result r. The parsing result is calculated by the following formula: r = Transformer(z); Among them, Transformer(z) represents the output result of the Transformer network; S17. Store the parsing result r in a temporary query library.

3. A graphene product retrieval method based on a large model according to claim 2, characterized in that: The S3 comprises the following steps: S31, according to the analysis result r ′ Generate a query template, wherein the query template includes keywords, phrases and subject concepts; S32, inputting the query template into multiple data source interfaces to execute distributed query, wherein the data source interfaces include a scientific research literature database, a product catalog database, and a patent library; S33, pre-processing the query results returned by each data source, removing redundant data, unifying the format and performing preliminary denoising; S34, indexing the preprocessed data results to generate an index vector I i , each index vector I i Calculated by the following formula: Among them, d ij Represents the attribute value of each data result, φ j , j Represents the weight coefficient of the corresponding attribute, α j , β j and γ j They represent the adjustment parameters related to the analytical result F, ψ(d ij ) represents the attribute transformation function, ω j represents the frequency parameter, T represents the upper limit of integration, and m represents the number of attributes; S35, index vector I i Perform natural language processing and image recognition technology to generate feature vector F i , the eigenvector F i Calculated by the following formula: Among them, I ik represents the natural language features in each index vector, κ(I ik ) represents the natural language feature transformation function, I il Represents the image features in each index vector, ω(I il ) represents the image feature transformation function, θ k and λ l Represent the weight coefficients of natural language and image features respectively, ξ l , δ l ,∈ l and l Respectively represent and analyze the results r ′ The relevant adjustment parameters, σ represents the overall feature aggregation function, χ l represents the frequency correlation coefficient, ω l represents the frequency parameter, T represents the upper limit of integration, p and q represent the number of natural language and image features respectively; S36, extract the feature vector F i Input to the aggregation module, classify and aggregate the information according to the parsing result r', and generate the aggregation vector A. The aggregation process is as follows: Among them, α i , δ j and k Represent the weight coefficients of the eigenvector and analytical results, β(F i ) and ζ(F j ,r ′ ) represent the analytical results r ′ The feature vector transformation function and feature-analysis result interaction function, ψ k Respectively represent the eigenvector and analytical result r ′ The relevant adjustment parameters, η represents the overall aggregation function, λ k represents the frequency correlation coefficient, ω k represents the frequency parameter, T represents the upper limit of integration, n, m and p represent the number of eigenvectors and analytical results respectively; S37: Input the aggregated vector A into a database for storage.

4. A graphene product retrieval method based on a large model according to claim 3, characterized in that: The S4 comprises the following steps: S41, receiving image data of a graphene product, wherein the image data includes a microscopic image and a macroscopic image; S42, performing image denoising, contrast enhancement and edge detection on the received image data to obtain pre-processed image data I pre ; S43, use convolutional neural network to extract features from the preprocessed image data and generate image feature vector F img ; S44, image feature vector F img Classification processing is performed to identify the properties and categories of graphene products, and the classification processing is calculated by the following formula: Among them, Classifier(F img ) represents the processing function of the classifier on the image feature vector, W i and V i represents the weight matrix, b i represents the bias vector, σ represents the activation function, ∈ i , α i , β i , γ i Represents the adjustment parameter, relu represents the modified linear unit function, and tanh represents the hyperbolic tangent function; S45, the classification result C and the natural language processing result r ′ Fusion is performed to generate refined product attributes and category information. The fusion process is as follows: Among them, α and β represent weight coefficients, δ i ,θ i , i , μ i , ν i represents the adjustment parameter, represents the fusion function, W i and V i represents the weight matrix in the fusion process, b i represents the bias vector, A fusion Indicates the fused product attributes and category information; S46, the merged product attributes and category information A fusion Stored in a temporary database.

5. A graphene product retrieval method based on a large model according to claim 4, characterized in that: The S5 comprises the following steps: S51, receiving natural language processing result r ′ and the fused product attributes and category information A fusion ; S52, the natural language processing result r ′ and the fused product attributes and category information A fusion Convert to a unified data format and generate a joint feature vector U; S53, extracting features from the joint feature vector U using data mining technology to generate a feature matrix M; S54, classify the feature matrix M, cluster and classify the extracted features, and generate a classification result C final ; S55, the classification result C final Input into the aggregation module, aggregate the information according to data mining technology, and generate the final aggregation result A final : Among them, α i represents the weight coefficient, k represents the number of classification results, C final,i Indicates each classification result; S56, the final aggregation result A final Stored in the database.

6. A graphene product retrieval method based on a large model according to claim 5, characterized in that: The S6 comprises the following steps: S61. Receive user's search intention data and aggregate information A final ; S62: The user's search intention data and aggregate information A final Convert to a unified data format to generate the user intention feature vector U intent and the aggregate information feature vector U agg ; S63. Calculate the user intention feature vector U using the matching analysis model intent and the aggregate information feature vector U agg The matching degree is calculated by the following formula: Among them, cos represents the cosine similarity function, · represents the vector dot product, and ||·|| represents the modulus of the vector; S64, dynamically optimize the search results according to the matching results, the optimization steps include adjusting the ranking weight of the search results and filtering irrelevant information, and the optimized search results R opt Calculated by the following formula: Among them, w i represents the weight of the i-th search result, r i represents the i-th search result, and n represents the number of search results; S65. Use a reinforcement learning algorithm to perform online learning and adjustment on the matching analysis model. The reinforcement learning process includes the following steps: Define a reward function R, which is calculated based on the user's feedback and the relevance of the search results; Use reinforcement learning algorithms to update model parameters so that the matching analysis model can adaptively adjust and optimize search results; S66, the optimized search result R opt Output to the user and store the relevant data in the database.

7. A retrieval system for a graphene product retrieval method based on a large model as claimed in any one of claims 1 to 6, characterized in that: Includes the following modules: A user input module, used to receive graphene-related queries input by a user, the user input module comprising a text input box and a voice input module; The natural language processing module includes a variational autoencoder and a Transformer network, which is used to parse natural language queries input by users and identify key concepts and intents in the queries; A data extraction module, used to extract information related to graphene products from multiple data sources, including a scientific research literature database, a product catalog database and a patent database; Image processing module, used to apply image recognition technology to analyze the image data of graphene products and refine the identification of product attributes and categories; Information classification and aggregation module, which is used to combine natural language processing results with image recognition results and use data mining technology to classify and aggregate information; Matching analysis and optimization module, including machine learning models, is used to analyze the matching degree between the user's search intent and the aggregated information, and dynamically optimize the relevance and accuracy of search results; The result output module is used to output the optimized search results and store the relevant data in the database.

Images