Battery performance data automatic extraction system based on large model

Abstract

The application provides a battery performance data automatic extraction system based on a large model, relates to the technical field of data extraction, and comprises: a literature analysis module, which is used for reading and analyzing a PDF document of a scientific literature related to a battery to generate json data; a data preprocessing module, which is used for performing semantic blocking on the json data and splitting text of the same semantic content into short text blocks; a battery performance data extraction module, which is used for extracting battery performance data from the short text blocks based on a pre-trained large language model by using a multi-round dialogue and self-validation strategy; and a battery performance data structured output module, which is used for outputting the battery performance data to an excel table in a structured manner. The application solves the technical problem that the existing manual or semi-automatic data extraction mode has missing or incorrect identification when facing complex battery performance data, thereby resulting in insufficient accuracy of the extraction result.

Description

Technical Field

[0001] The present invention relates to the technical field of data extraction, and in particular to a large model-based automatic battery performance data extraction system. Background Art

[0002] With the development of battery technology, especially in the fields of electric vehicles and renewable energy storage, the demand for battery performance data is increasing. Traditional methods for extracting battery performance data rely on manual reading and parsing of scientific literature, which is time-consuming and error-prone.

[0003] Battery performance data extraction is the application of text knowledge extraction in downstream tasks. Its core task is to effectively identify and extract battery performance parameter objects and their relationships from complex text. Traditional knowledge extraction technologies are mostly based on rules, pattern matching, or machine learning methods, such as hidden Markov models (HMMs), support vector machines (SVMs), long short-term memory networks (LSTMs), conditional random fields (CRFs), and deep learning-based models such as BiLSTM-CRF. With the advent of large language models, the field of knowledge extraction is increasingly introducing models such as BERT and ChatGPT, which transform sequence labeling tasks into generation tasks, autonomously learn knowledge from unlabeled, unstructured data, and significantly improve the efficiency of entity extraction. However, due to the shortcomings of large models such as knowledge illusions, errors in understanding logical relationships, and a lack of contextual understanding, battery performance data extraction with existing technologies still faces many problems. Summary of the Invention

[0004] This application provides an automatic battery performance data extraction system based on a large model, aiming to solve the technical problem that existing manual or semi-automatic data extraction methods may have omissions or misidentifications when faced with complex battery performance data, resulting in insufficient accuracy of the extraction results.

[0005] The present application discloses a large-scale model-based automatic battery performance data extraction system, which includes: a literature parsing module for reading and parsing PDF documents of battery-related scientific literature to generate JSON data; a data preprocessing module for semantically segmenting the JSON data and dividing text with the same semantic content into short text blocks; a battery performance data extraction module for extracting battery performance data from the short text blocks based on a pre-trained large language model using multiple rounds of dialogue and a self-verification strategy; and a battery performance data structured output module for outputting the battery performance data in a structured manner to an Excel spreadsheet.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] Through the literature parsing module, the battery-related scientific literature PDF documents are read and parsed, and the generated json data is parsed into a data format readable by the large model. By converting the literature content into json format, the readability and operability of the data for the large model are guaranteed, and the data processing and extraction in the subsequent steps are supported; through the data preprocessing module, the generated json data is semantically segmented, and the text with the same semantic content is divided into short text blocks. By splitting the long text block into short text blocks, each short text block is focused on a specific battery performance attribute, so that the input text of the large model meets the token number limit of the large model, which makes the subsequent battery performance data extraction more accurate and efficient. Through semantic segmentation, the same semantic content is ensured to be gathered in a short text block, which helps the large language model to accurately understand the text content and extract battery performance data, avoiding information loss, context semantics sparsity and ambiguity problems; battery performance The data extraction module is based on a pre-trained large language model and uses multiple rounds of dialogue and self-verification strategies to extract battery performance data from short text blocks. Through multiple rounds of dialogue, the large model can gradually learn the thinking mode of manual information extraction and guide the large model to extract each battery performance data according to task requirements to ensure the integrity and consistency of each data. After each extraction, each data is verified through interaction with the large language model to ensure the authenticity and reliability of the data. This self-verification mechanism greatly reduces the risk of erroneous and inconsistent data and improves the credibility of the final output results; the battery performance data structured output module outputs the extracted battery performance data in a structured form into an Excel spreadsheet. Users can directly use the spreadsheet for analysis or import it into other analysis tools, avoiding the work of manual organization and data entry, saving a lot of time and resources, and enhancing the compatibility and flexibility of the system.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Schematic diagram of the structure of the large model-based automatic extraction system for battery performance data provided in an embodiment of the present application.

[0010] Figure 2 This is a schematic diagram of the data preprocessing module structure in the large model-based battery performance data automatic extraction system provided in an embodiment of the present application.

[0011] Figure 3 A schematic diagram of the structure of a battery performance data extraction module in a large model-based automatic battery performance data extraction system provided in an embodiment of the present application.

[0012] Explanation of the accompanying drawings: document parsing module 10, data preprocessing module 20, battery performance data extraction module 30, battery performance data structured output module 40, extraction accuracy analysis unit 21, text block size optimization unit 22, semantic segmentation unit 23, text content analysis unit 31, self-verification unit 32, attribute information output unit 33. DETAILED DESCRIPTION

[0013] The embodiments of the present application provide a large-model-based automatic battery performance data extraction system to solve the technical problem that existing manual or semi-automatic data extraction methods may cause omissions or misidentifications when faced with complex battery performance data, resulting in insufficient accuracy of the extraction results.

[0014] After introducing the basic principles of this application, various non-limiting embodiments of this application will be specifically described below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described here are only used to explain this application and are not used to limit this application.

[0015] Example 1, as Figure 1 As shown, an embodiment of the present application provides a system for automatically extracting battery performance data based on a large model, the system comprising:

[0016] The literature parsing module 10 is used to read and parse the PDF documents of battery-related scientific literature and generate JSON data;

[0017] A PDF parsing tool is used to read the contents of PDF files. Through natural language processing, deep learning, OCR, and other technologies, the document structure is parsed to extract information such as titles, chapters, paragraphs, and tables. The extracted text content is converted into a standardized JSON format, making it readable by the large model. This format contains metadata such as text, paragraphs, paragraph indexes, and chapters. The data in different paragraphs and tables is annotated to ensure that subsequent data processing can understand their meaning. The literature parsing module reads and parses PDF documents of battery-related scientific literature, generating JSON data and parsing it into a data format readable by the large model. By converting the literature content into JSON format, the readability and operability of the data for the large model are guaranteed, making the subsequent data processing and extraction more efficient.

[0018] The data preprocessing module 20 is used to perform semantic segmentation on the JSON data and divide text with the same semantic content into short text blocks;

[0019] Semantic analysis methods are used to understand the text content in the literature and divide text blocks into meaningful chunks. For example, if a paragraph covers battery attributes such as charge cycle and capacity, the paragraph is considered a semantic chunk. By analyzing the length of each semantic chunk, the longer chunk is automatically split into multiple shorter chunks, ensuring that each short chunk retains complete semantic information. The splitting criteria can be based on text length, sentence completeness, and contextual coherence. The resulting short chunks are typically brief and contain relevant information about battery performance from the literature. The data preprocessing module semantically chunks the generated JSON data and segments text with the same semantic content into short chunks. By splitting long chunks into short chunks, each short chunk focuses on a specific battery performance attribute, ensuring that the input text of the large model meets the token limit of the large model. This makes subsequent battery performance data extraction more accurate and efficient. Semantic chunking ensures that the same semantic content is clustered in a short chunk, which helps the large language model accurately understand the text content and extract battery performance data, avoiding information loss, contextual semantic sparsity, and ambiguity.

[0020] A battery performance data extraction module 30 is configured to extract battery performance data from the short text block using a pre-trained large language model and a multi-round dialogue and self-verification strategy;

[0021] A pre-trained large language model refers to a model trained on general tasks based on large-scale general corpus. Short text blocks are input into the pre-trained large language model. The large language model uses its powerful semantic understanding ability to analyze the text and determine whether the text contains battery performance data. It uses zero-shot classification technology to classify the text based on known battery performance data types such as capacity, number of cycles, materials, etc. to identify whether it contains battery performance data.

[0022] Among them, a multi-round dialogue method is used to interact with the large language model. By gradually asking questions to guide the model, multiple battery performance target data items are generated, including battery capacity, battery material, battery charge and discharge times, etc. Each round of dialogue focuses on one performance attribute to ensure the accuracy of the attribute data. For example, the first round of dialogue asks about the battery capacity, the second round asks about the unit, and the third round asks about the material, etc.

[0023] After each round of dialogue, the extracted data is automatically verified. If the attribute value extracted in a round of dialogue is incorrect, it will be corrected or the uncertain data will be discarded. The self-verification step includes mutual correlation verification of each extracted data item, for example, confirming whether the battery capacity and unit are consistent, whether the battery material is reasonable, etc. If the verification fails, the corresponding performance data will be discarded as uncertain data and the current dialogue will be terminated to prevent the transmission of erroneous information.

[0024] After all rounds of dialogue and verification are successful, the complete battery performance data will be used as the final output. This data includes various battery performance attributes, including battery performance value, battery performance value unit, battery performance material name, and number of cycles.

[0025] The battery performance data structured output module 40 is used to output the battery performance data in a structured manner to an Excel spreadsheet.

[0026] Convert the extracted battery performance data into a structured format. Specifically, store each battery performance attribute in the corresponding column of an Excel spreadsheet. Each row represents a battery performance data entry, and each column represents a different performance attribute, such as capacity, battery material, unit, number of cycles, etc. Store the data in a tabular format and ensure data consistency and readability. This way, users can view and analyze the data directly from Excel, or import the data into other systems or tools for further analysis.

[0027] Furthermore, if Figure 2 As shown, the data preprocessing module 20 includes:

[0028] An extraction accuracy analysis unit 21 is used to analyze the knowledge extraction accuracy of different text block sizes of historical extraction records according to the attributes of the knowledge to be extracted;

[0029] A text block size optimization unit 22 is configured to optimize the text block size of the knowledge attribute to be extracted based on the knowledge extraction accuracy and the text block size constraint interval of the large language model to obtain a target text block size;

[0030] The semantic chunking unit 23 is configured to semantically chunk the JSON data according to the target text chunk size.

[0031] Collect and analyze historical data, including the accuracy of knowledge extraction for a specific battery performance attribute (such as capacity, number of cycles, materials, etc.) under different text block sizes. For example, record the accuracy of short text blocks (such as 200 words) and long text blocks (such as 500 words) in extracting a certain attribute. The extraction accuracy of short text blocks may be higher, or vice versa.

[0032] For each text block size, we calculated extraction accuracy. This accuracy is based on several factors: the number of correctly extracted attributes, such as battery capacity and battery material; the number of incorrectly extracted attributes, which extract incorrect or irrelevant information; and unextracted data, which indicates when the model fails to extract any valid information. The analysis results provide a table of extraction accuracy for different text block sizes, which can be used to guide subsequent text block optimization decisions.

[0033] According to the extraction accuracy analysis result, the text block size constraint interval of the large language model is determined, different pre-training large language models have different limitations on the size of the input text, for example, the maximum supported number of characters is 512, 1024, etc.; according to the knowledge extraction accuracy, the optimal text block size is selected, for example, if the short text block has a higher accuracy in extraction capacity data, and the long text block is more accurate in extraction cycle times, then the most suitable text block size will be determined by combining these data; the text block size constraint interval and the accuracy data are combined to optimize the text block size, to ensure that the optimization result can guarantee a high accuracy and does not violate the maximum length limit of the model input, specifically, after obtaining the preliminary analysis result, an optimization algorithm such as the greedy algorithm, binary search, particle swarm algorithm, etc. can be used to further optimize the text block size, to maximize the processing efficiency of the model under the premise of ensuring the extraction accuracy, with more data processing, the block size is adjusted according to the new situation, so that the data extraction effect is continuously optimized, finally, the optimization algorithm outputs a target text block size as the input of the subsequent semantic blocking unit.

[0034] According to the optimized target text block size, the json data is semantically blocked, and the long text block is cut into short text blocks with clear semantics, specifically, the natural language processing technology is used to analyze the semantics of the json data to be processed, and the actual meaning of each text block is understood. Semantic analysis helps to identify which parts can be processed as an independent short text block, according to the target text block size obtained from the text block size optimization unit, the longer text is automatically cut into short text blocks of the target size, each short text block should preserve the complete semantics as much as possible, and the core content of a sentence or paragraph should be avoided during cutting. Ensure that each text block can express a complete meaning independently, for example, if a paragraph contains battery material and performance data, these information is divided into reasonable blocks without destroying the coherence of the information. Finally, structured json data is generated, in which each semantic block can be assigned a text block number and the battery performance attribute information it contains for text block positioning.

[0035] Further, as shown in Figure 3 The battery performance data extraction module 30 comprises:

[0036] The text content analysis unit 31 is configured to input the short text block into the large language model for text content analysis, and determine whether the short text block contains battery performance data through zero-shot binary classification.

[0037] a self-verification unit 32 for, if included, extracting multiple performance attribute information of the battery through multiple rounds of dialogue with the large language model, and performing self-verification on the extracted performance attribute information after each round of dialogue;

[0038] The attribute information output unit 33 is configured to output the multiple pieces of performance attribute information as the battery performance data when all self-verifications pass.

[0039] The short text block is input into the large language model for text content analysis, specifically asking the large model whether the text contains data related to battery coulombic efficiency, so as to determine whether it contains battery performance data.

[0040] If the short text block contains battery performance data, multiple performance attributes of the battery are extracted through multiple rounds of dialogue with the large language model, and self-verification is performed after each round of dialogue to ensure the accuracy of each attribute information. Specifically, the large language model is used for interactive question-and-answering, with questions and answers in each round of dialogue to extract specific battery performance data items. For example, the first round of dialogue extracts the battery's performance value (such as capacity value); the second round of dialogue confirms whether the performance value contains units (such as mAh, Ah, etc.); the third round of dialogue extracts the battery's material name (such as lithium battery, nickel-metal hydride battery, etc.); the fourth round of dialogue inquires and extracts the number of cycles and its units (such as times, hours, etc.).

[0041] The performance attribute information extracted in each round of dialogue needs to be self-verified to ensure the accuracy of the data. Self-verification includes correlation verification, consistency verification, data integrity verification, etc. For example, for each extracted attribute, it is automatically determined whether the correlation between the attribute and other parts of the battery performance is reasonable.

[0042] When all attribute information has passed self-verification, the extracted multiple performance attribute information is output as the final battery performance data. This process combines multiple rounds of dialogue and self-verification strategies to effectively improve the accuracy and credibility of data extraction.

[0043] Furthermore, the text content analysis unit includes:

[0044] A zero-sample binary classification channel is used to perform zero-sample binary classification on the short text block based on a preset database, and determine whether it contains battery performance data according to the binary classification result, wherein the preset database includes data related to battery coulombic efficiency.

[0045] Zero-shot classification is a classification method that does not require labeled data. It uses the capabilities of a pre-trained large language model to classify text. The goal is to determine whether the short text block contains battery performance data. Specifically, the pre-trained large language model automatically classifies the text based on its knowledge of battery performance. The specific classification task is to divide the short text blocks into those containing battery performance data and those not containing battery performance data according to a preset database. The preset database contains data related to battery coulombic efficiency, battery performance data such as battery capacity, materials, life, charge and discharge times, etc. If it does not contain battery performance data, it means that the text block is not a description of battery performance.

[0046] Furthermore, the battery performance data extraction module 30 further includes:

[0047] The discard processing unit is used to discard the extracted performance attribute information as uncertain data and terminate the dialogue when any self-verification fails.

[0048] During multiple rounds of dialogue and self-verification, if any verification fails, for example, the attribute information is incorrect or inconsistent with other information, such as the capacity value and unit do not match, the material name cannot be recognized, etc., the extracted data will be discarded immediately and the current dialogue will be terminated to avoid further extraction of invalid or erroneous data and prevent the erroneous information from affecting the accuracy of subsequent data.

[0049] Furthermore, the self-verification unit further includes:

[0050] The channel for obtaining multiple performance attribute information is used to extract battery performance values, battery performance value units, battery performance material names, and cycle times in sequence through multiple rounds of dialogue as the multiple performance attribute information.

[0051] The multiple performance attribute information acquisition channel is a core part of the battery performance data extraction module, responsible for gradually extracting multiple performance attribute information of the battery through multiple rounds of dialogue with the pre-trained large language model. These attributes include battery performance numerical values, battery performance numerical value units, battery performance material names, and cycle times. Specifically, in the first round of dialogue, the performance numerical values of the battery are first asked, such as battery capacity, energy density, etc., and the battery performance numerical value information is obtained from the large language model. In the second round of dialogue, after obtaining the battery performance numerical value, the unit of the value is asked, such as mAh, Wh, etc., and the battery performance numerical value unit information is obtained. In the third round of dialogue, the material name of the battery is further asked, such as lithium battery, nickel-hydrogen battery, etc., and the battery performance material name information is obtained. In the fourth round of dialogue, the cycle times of the battery are asked and extracted, i.e., the charge and discharge times or service life of the battery, and the cycle time information is obtained. The battery performance numerical values, battery performance numerical value units, battery performance material names, and cycle times extracted in sequence are integrated as the multiple performance attribute information.

[0052] Further, the self-verification unit further comprises:

[0053] The empty set detection channel is used to discard the extracted performance attribute information as incomplete data when the extraction result of the non-empty performance attribute information is an empty set, and terminate the dialogue, wherein the non-empty performance attribute information includes but is not limited to battery performance numerical values and battery performance material names.

[0054] After each round of dialogue, it is automatically verified whether the attribute information is valid. If the extraction result of the non-empty performance attribute is an empty set in any round, i.e., the data of the attribute cannot be effectively extracted, the discard processing is performed. The non-empty performance attribute includes battery performance numerical values and battery performance material names. The loss of these attributes directly affects the integrity and effectiveness of the extraction result. After the discard processing, the current multi-round dialogue is immediately terminated to avoid further extracting other attributes from invalid data.

[0055] Further, the multiple performance attribute information acquisition channel comprises:

[0056] The data dictionary retrieval node is used to retrieve the compound data dictionary.

[0057] The matching retrieval node is used to perform semantic matching retrieval in the short text block based on the compound data dictionary to obtain the battery performance material name.

[0058] The compound data dictionary is a database containing various battery material names and their related information, such as chemical elements, chemical formulas, common material names, etc. The data in the dictionary includes preferred and non-preferred terms for battery materials and covers common battery materials such as lithium (Li), cobalt (Co), nickel (Ni), graphite, etc.

[0059] The short text block is preprocessed using efficient string matching algorithms such as Boyer-Moore algorithm to find possible battery material names, such as "lithium battery", "lithium cobalt oxide" can be matched in the short text using the maximum reverse matching algorithm, according to the standardized material name in the compound data dictionary, check if there is a match with the dictionary entry in the short text block, for example, if the text mentions "LiCoO2", it will be matched by "lithium cobalt oxide" or "lithium cobalt oxide" in the data dictionary, for alias or non-standardized material name, fuzzy matching algorithm can be used to identify similar words, for example, if the literature mentions "lithium battery" and the dictionary has "Li-ion battery", it can be found through similarity matching to find the correct battery material name. Sometimes, battery material names may not appear alone, but with other content, context information can be used for optimized matching to identify and extract appropriate material names.

[0060] Through matching retrieval, standardized battery material names are extracted from the short text block, such as "lithium cobalt oxide (LiCoO2)", "nickel cobalt aluminum (NCA)", etc. The matched material names are marked as valid and passed to the subsequent module for further data extraction or output. If no matching material name is found, it will be marked as missing or incomplete data and it will be decided whether to discard the data.

[0061] Further, the compound data dictionary includes various aliases and standardized names of materials.

[0062] The compound data dictionary includes various aliases and standardized names of materials, for example, lithium battery materials may be labeled as "LiCoO2" or "lithium cobalt oxide", which refer to the same chemical composition.

[0063] Further, the self-verification unit further comprises:

[0064] The first verification dialogue channel is used to extract the battery performance value and ask the large language model to verify the dialogue whether the battery performance value is positively associated with the battery performance;

[0065] The second verification dialogue channel is used to extract the battery performance numerical unit if the answer is yes, and ask the large language model a question to conduct a verification dialogue on whether the battery performance numerical unit is positively correlated with the battery performance and the battery performance value;

[0066] The third verification dialogue channel is used to extract the name of the battery performance material if the answer is yes, and ask the large language model to conduct a verification dialogue on whether the name of the battery performance material is positively correlated with the battery performance and the battery performance value;

[0067] a fourth verification dialogue channel, for extracting the cycle number, including the cycle number data and unit information, if the answer is yes, and asking the large language model questions to conduct a verification dialogue on whether the cycle number data and unit information are positively correlated with the cycle number and the battery performance value;

[0068] The data output channel is used to output the battery performance value, battery performance value unit, battery performance material name, and cycle number if the answer is yes.

[0069] Battery performance values ​​extracted from short text blocks, such as a battery capacity of 5000 or a charging voltage of 3.7, are queried by the large language model to confirm whether the value is relevant to battery performance. For example, the extracted value is the coulombic efficiency value corresponding to the correct compound. The large language model uses existing battery knowledge and contextual information to determine whether the value has physical meaning and meets the battery's performance characteristics. If the answer is yes, the verification passes, and the extraction continues and enters the second verification dialogue channel to begin extracting the unit of the battery performance value. If the answer is no, the verification fails, and the value is deemed invalid or incorrect, marked as uncertain data, and the current verification dialogue ends.

[0070] Based on the extracted performance value, the large language model is asked to confirm whether the value has the correct unit. The large language model then answers the unit of the value based on the context and further confirms whether the unit matches the battery performance value. If the answer is yes, that is, the unit and value are relevant and reasonable, such as "5000" corresponding to "mAh", then the performance attribute information is considered valid and the next battery performance attribute is extracted. If the answer is no, that is, the unit and value are inconsistent or unreasonable, for example, the capacity value is "5000" and the unit is "V", then the data is marked as invalid, the current attribute is discarded, and the conversation ends.

[0071] After verifying the battery performance values, the battery material names are extracted, such as lithium, cobalt, graphite, etc. The extracted battery material names are associated with the battery performance values ​​(such as capacity, energy density, etc.) for verification. The large language model confirms whether the material meets the battery performance requirements based on the battery performance, material properties and physical knowledge. If the answer is yes, the battery material is positively correlated with the performance value. If a battery with a capacity of 5000mAh does use lithium cobalt oxide as a material, the extracted battery material name is considered valid and will continue to enter the fourth verification dialogue channel; if the answer is no, that is, the battery material does not match the performance value, such as the battery material does not match the capacity, the data is determined to be invalid, the current data is discarded and the dialogue is terminated.

[0072] The battery's cycle count and unit are further confirmed through the conversation. The extracted cycle count is then correlated with the battery's performance values ​​to ensure that it matches the actual battery performance. For example, if a question is asked, "This battery has a capacity of 5000mAh. Can it support 1000 cycles?" The large language model determines whether the extracted cycle count is reasonable based on the battery's technical characteristics, performance range, and common industry standards. If the answer is yes, meaning the cycle count is consistent with the battery's performance values ​​(such as capacity and material), the cycle count information is considered valid. If the answer is no, meaning the cycle count is irrelevant or unreasonable to the battery's performance values ​​(for example, a battery with a capacity of 5000mAh can only support 500 cycles instead of 1000), the data is deemed invalid, the current cycle count data is discarded, and the conversation is terminated.

[0073] If the number of cycles and unit verification passes, it means that all multi-round dialogue verifications have passed. In this case, all extracted and verified battery performance values, battery performance value units, battery performance material names, and number of cycles are output. This process helps ensure the accuracy, completeness, and consistency of the output data, and prevents unqualified data from flowing into subsequent analysis and decision-making.

[0074] In summary, the large-scale model-based automatic battery performance data extraction system provided in the embodiments of the present application has the following technical effects:

[0075] Through the literature parsing module, the battery-related scientific literature PDF documents are read and parsed, and the generated json data is parsed into a data format readable by the large model. By converting the literature content into json format, the readability and operability of the data for the large model are guaranteed, and the data processing and extraction in the subsequent steps are supported; through the data preprocessing module, the generated json data is semantically segmented, and the text with the same semantic content is divided into short text blocks. By splitting the long text block into short text blocks, each short text block is focused on a specific battery performance attribute, so that the input text of the large model meets the token number limit of the large model, which makes the subsequent battery performance data extraction more accurate and efficient. Through semantic segmentation, the same semantic content is ensured to be gathered in a short text block, which helps the large language model to accurately understand the text content and extract battery performance data, avoiding information loss, context semantics sparsity and ambiguity problems; battery performance The data extraction module is based on a pre-trained large language model and uses multiple rounds of dialogue and self-verification strategies to extract battery performance data from short text blocks. Through multiple rounds of dialogue, the large model can gradually learn the thinking mode of manual information extraction and guide the large model to extract each battery performance data according to task requirements to ensure the integrity and consistency of each data. After each extraction, each data is verified through interaction with the large language model to ensure the authenticity and reliability of the data. This self-verification mechanism greatly reduces the risk of erroneous and inconsistent data and improves the credibility of the final output results; the battery performance data structured output module outputs the extracted battery performance data in a structured form into an Excel spreadsheet. Users can directly use the spreadsheet for analysis or import it into other analysis tools, avoiding the work of manual organization and data entry, saving a lot of time and resources, and enhancing the compatibility and flexibility of the system.

[0076] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The battery performance data automatic extraction system based on the large model is characterized by: The system comprises: Literature parsing module, used to read and parse PDF documents of battery-related scientific literature and generate JSON data; A data preprocessing module is used to semantically segment the JSON data and divide text with the same semantic content into short text blocks; A battery performance data extraction module, configured to extract battery performance data from the short text block using a pre-trained large language model and a multi-round dialogue and self-verification strategy; A battery performance data structured output module, used to output the battery performance data in a structured manner to an Excel spreadsheet; Wherein, the data preprocessing module includes: An extraction accuracy analysis unit, used to analyze the knowledge extraction accuracy of different text block sizes of historical extraction records based on the attributes of the knowledge to be extracted; a text block size optimization unit, configured to optimize the text block size of the knowledge attribute to be extracted based on the knowledge extraction accuracy and the text block size constraint interval of the large language model to obtain a target text block size; A semantic chunking unit, configured to semantically chunk the JSON data according to the target text chunk size; The battery performance data extraction module includes: a text content analysis unit, configured to input the short text block into the large language model for text content analysis to determine whether the short text block contains battery performance data; a self-verification unit, for extracting multiple performance attribute information of the battery through multiple rounds of dialogue with the large language model, and performing self-verification on the extracted performance attribute information after each round of dialogue; an attribute information output unit, configured to output the plurality of performance attribute information as the battery performance data when all self-verifications pass; Wherein, the self-verification unit further includes: A channel for acquiring multiple performance attribute information, for sequentially extracting battery performance values, battery performance value units, battery performance material names, and cycle times through multiple rounds of dialogue as the multiple performance attribute information; Wherein, the self-verification unit further includes: An empty set detection channel, used to discard the extracted performance attribute information as incomplete data and terminate the session when the extraction result of non-empty performance attribute information is an empty set. The non-empty performance attribute information includes but is not limited to battery performance values ​​and battery performance material names; Wherein, the self-verification unit further includes: A first verification dialogue channel is used to extract the battery performance value and ask questions to the large language model to conduct a verification dialogue on whether the battery performance value is positively correlated with the battery performance; The second verification dialogue channel is used to extract the battery performance numerical unit if the answer is yes, and ask the large language model a question to conduct a verification dialogue on whether the battery performance numerical unit is positively correlated with the battery performance and the battery performance value; The third verification dialogue channel is used to extract the name of the battery performance material if the answer is yes, and ask the large language model to conduct a verification dialogue on whether the name of the battery performance material is positively correlated with the battery performance and the battery performance value; a fourth verification dialogue channel, for extracting the cycle number, including the cycle number data and unit information, if the answer is yes, and asking the large language model questions to conduct a verification dialogue on whether the cycle number data and unit information are positively correlated with the cycle number and the battery performance value; The data output channel is used to output the battery performance value, battery performance value unit, battery performance material name, and cycle number if the answer is yes.

2. The large model-based automatic battery performance data extraction system according to claim 1, characterized in that: The text content analysis unit includes: A zero-sample binary classification channel is used to perform zero-sample binary classification on the short text block based on a preset database, and determine whether it contains battery performance data according to the binary classification result, wherein the preset database includes data related to battery coulombic efficiency.

3. The large model-based automatic battery performance data extraction system according to claim 1, characterized in that: The battery performance data extraction module further includes: The discard processing unit is used to discard the extracted performance attribute information as uncertain data and terminate the dialogue when any self-verification fails.

4. The large model-based automatic battery performance data extraction system according to claim 1, characterized in that: The multiple performance attribute information acquisition channels include: Data dictionary retrieval node, used to retrieve compound data dictionary; A matching retrieval node is used to perform a matching retrieval in the short text block based on the compound data dictionary to obtain the name of the battery performance material.

5. The large model-based automatic battery performance data extraction system according to claim 4, characterized in that: The compound data dictionary includes various aliases and standardized names for materials.

Images