Life cycle process list calculation method based on large language model
By using a lifecycle process inventory calculation method based on a large language model, lifecycle data is processed automatically, solving the problems of low efficiency and low quality caused by manual reliance in existing technologies. This achieves efficient and accurate inventory generation, which is applicable to fields such as marine aquaculture.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAMEN UNIV
- Filing Date
- 2025-12-09
- Publication Date
- 2026-04-24
AI Technical Summary
Existing lifecycle process inventory calculation methods rely on manual data collection and processing, resulting in long cycles, high costs, and a high risk of errors. They also lack automated and high-quality calculation methods, especially in the field of marine aquaculture where large language models have not been effectively applied.
A lifecycle process inventory calculation method based on a large language model is adopted. The file format is converted by the unstructured.io visual recognition toolkit, and the data is stored using a large language model decoupling and vector retrieval library guided by prompt words. The inventory framework and complete inventory are generated by combining the Qwen3-32B model finely tuned by the LoRA algorithm. The numerical processing and visualization are performed using Python scripts.
It has enabled automated and accurate calculation of lifecycle inventory data, significantly reducing reliance on manual labor, improving the efficiency and quality of inventory construction, providing high-quality assessment data for environmental management agencies and enterprises, and promoting sustainable development in the industry.
Smart Images
Figure CN121279258B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer data processing technology, specifically relating to a method for calculating a lifecycle process inventory based on a large language model. Background Technology
[0002] Life cycle assessment (Life Cycle Assessment) is a scientific and effective environmental assessment method. The calculation of the Life Cycle Process Inventory is a prerequisite and crucial step in conducting Life Cycle Assessment, and its quality directly determines the reliability and accuracy of the assessment results. The calculation of the Life Cycle Process Inventory aims to systematically quantify all relevant input and output flows of a product or service throughout its entire life cycle, including raw materials, energy consumption, and emissions into the environment.
[0003] Existing methods for calculating lifecycle process inventories rely heavily on manual data collection, processing, and analysis, resulting in long cycles, high costs, and susceptibility to errors. This makes it difficult for enterprises and management departments lacking specialized accounting capabilities to conveniently and economically obtain high-quality lifecycle process inventory assessment reports. With the development of generative artificial intelligence, large language models, with their powerful semantic understanding and efficient data processing capabilities, have made automated processing of lifecycle process inventory data possible. However, large language models have not yet been effectively applied in lifecycle process inventory calculations in vertical sectors such as marine aquaculture, lacking an automated, clear, and feasible standard workflow. Therefore, there is an urgent need for a lifecycle process inventory calculation method that integrates large language models and domain knowledge to address the shortcomings of existing technologies. Summary of the Invention
[0004] To address the aforementioned issues, this invention proposes a lifecycle process inventory calculation method based on a large language model. This method enables automated and accurate calculation of lifecycle inventory data, reduces reliance on manual labor, improves the efficiency and quality of inventory construction, and has promising prospects for widespread application.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A method for calculating a lifecycle process inventory based on a large language model includes the following steps:
[0007] S1. Use the unstructured.io visual recognition toolkit to convert document and image files into .markdown format files;
[0008] S2. Use a large language model guided by prompt words to decouple the tables in the converted .markdown file;
[0009] S3. Read the .markdown file and preprocess it into string fragments;
[0010] S4. Vectorize the string using an embedded model and store it using a vector retrieval library toolkit;
[0011] S5. Retrieve the vector database based on the stage keywords;
[0012] S6. Based on the recalled corpus, generate a list framework using a customized large language model;
[0013] S7. Based on the keywords of the input products, perform a secondary search of the vector database;
[0014] S8. Based on the recalled corpus, generate a complete list using a customized large language model;
[0015] S9. Use a Python script to process the values of the complete list in .json format and convert it into an HTML table.
[0016] Preferably, in step S1, the files to be processed in the folder are automatically identified using Python, and different types of original files are uniformly converted into .markdown format. Furthermore, the tables in the original files are uniformly converted into HTML format and then embedded into the .markdown file.
[0017] Preferably, the specific process of step S2 is as follows:
[0018] S21. Task Description: Analyze an HTML table line by line or by logical block, and generate natural language sentences for each line or block.
[0019] S22. Table Judgment: Automatically identify whether the table is a lifecycle list table or a general structured table;
[0020] S23. Non-lifecycle inventory table decoupling: Describe each row and retain the original information without performing additional stage inference;
[0021] S24. Lifecycle checklist table decoupling: Classify data into fixed lifecycle stages and integrate their functional units;
[0022] S25. Machine illusion suppression: When no lines are omitted, no repeated sentences are included, units are missing, or quantities are missing, the sentence should be clearly stated using accurate and neutral language, without making assumptions or interpretations.
[0023] Preferably, the preprocessing process in step S3 is as follows: read all .markdown format files, segment the file content, segment the natural language part of the original document according to the original paragraphs, and segment the HTML table content after decoupling by the large language model according to the logical units generated by decoupling. Each segment represents a complete data entry or description. After the segmentation is completed, segments with a length of less than 15 characters are removed to obtain independent string segments.
[0024] Preferably, the specific process of step S4 is as follows:
[0025] S41. String Encoding: The input string fragment is lexicalized to obtain a lexical sequence. Then, the vocabulary of the embedded model is used to map each lexical to a unique integer index. The vocabulary includes general vocabulary, proper nouns in the fields of environmental science and life cycle assessment, pollutant names, chemical substances, units, and specific industry terms.
[0026] S42. Vectorization: The BAAI-bge-m3 embedded model is used to transform the integer index sequence into a high-dimensional normalized vector, thereby obtaining the data semantic embedding matrix and capturing the deep semantic information of the data.
[0027] S43. Indexing and Storage: Create a collection using the vector retrieval library ChromaDB, storing the vector of each string fragment and its corresponding original string fragment as a record, supporting structured weighted retrieval.
[0028] Preferably, in step S5, a vector database is retrieved based on the structured elements of a predefined standard lifecycle process inventory template; the structured elements include lifecycle stage keywords and target product species names; similarity is calculated using a lifecycle inventory structured weighted semantic similarity, and the calculation formula is as follows: Where Q represents the query; For corpus; To query the lifecycle list structured weighted semantic similarity between Q and corpus D; , and The weighting coefficients are respectively for the breeding stage, species name, and general context, and satisfy the following conditions: ; The semantic embedding vectors of the product corpus D are fed into the retrieved lifecycle list; , and These are the semantic embedding vectors of the corresponding structured elements in query Q; for and Cosine similarity; for and Cosine similarity; for and The cosine similarity.
[0029] Preferably, in step S6, the customized large language model is the Qwen3-32B model fine-tuned by LoRA algorithm instructions, and the generated list framework is in .json format, containing three fields: stage, input and material type.
[0030] In step S6, the prompt words used to guide the large language model include the following:
[0031] S61. Task Description: The large language model is required to act as a professional assistant to the aquaculture life cycle checklist table, and extract and classify the input materials for specific stages based on the provided documents.
[0032] S62. Information Constraints: The large language model is required to use only the information provided in the documents, and must not add any facts, materials or details that are not in the documents, and must only include inputs that are relevant to the requested stage.
[0033] S63. Data Integration and Deduplication: The large language model is required to integrate multiple quantities of the same input in the document and ensure that each unique input appears only once in the final .json format output.
[0034] S64. Field Definition and Output Format: Each entry output by the large language model must contain two fields: input and material type, and must be returned as a .json format array.
[0035] S65, Machine Illusion Suppression: Requires large language models not to guess values or units, not to include value or unit fields in the output, not to include any thought processes, reasoning or explanations, and to return an empty list when no relevant entry is found.
[0036] In step S8, the customized large language model is the Qwen3-32B model fine-tuned by the LoRA algorithm instructions. The generated complete list is in .json format and contains five fields: stage, input, material type, value, unit and description.
[0037] In step S8, the prompt words used to guide the large language model mainly include the following:
[0038] S81. Task Description and Context Provision: The large language model is required to act as an expert in lifecycle inventory data completion, using the provided text and table context to calculate and fill in the values and units for given inputs and breeding stages.
[0039] S82. Data Sources and Unit Conversions: Large language models are required to use only the information provided in the documents and to perform unit conversions only when the documents explicitly provide the data or the data can be directly inferred. Otherwise, the original values and units should be reported.
[0040] S83, Data Missing Handling and Output Format: The large language model is required to mark both numerical values and units as unknown when no reliable data is found, and strictly return a valid .json list containing the specified fields;
[0041] S84. Notes and Explanations: Requires the large language model to provide explanations in the description field to explain the data source, calculation steps, or any assumptions.
[0042] Preferably, in steps S6 and S8, the specific process of fine-tuning the Qwen3-32B model using the LoRA algorithm is as follows: The weights of the Qwen3-32B model are updated using a low-rank matrix to optimize instruction compliance; wherein, the loss function is calculated by minimizing the cross-entropy loss and combining it with the weights fine-tuned by the LoRA algorithm, and the calculation formula is: , , , ,in, This represents the original frozen pre-trained weight matrix; This represents the weight matrix after LoRA fine-tuning; This represents the amount of weight update; , Both ΔW and ΔW belong to ,in, It is the feature dimension. It is the output dimension; and All are low-rank matrices, belonging to and r is much smaller than the dimension of the original matrix; The logical score vector representing the t-th time step; The forward propagation function represents the model; This represents the input vector at time step t. Represents cross-entropy loss; This represents the total number of tokens in the target output sequence; For the first The time step of each word element; The correct target word; for The function operates on Then, the correct lexical units were obtained. The predicted probability; This represents the size of the model's vocabulary; For the word list An index of candidate words; For the model in the first The time step for the first The unnormalized fractions of each word unit.
[0043] Preferably, in step S7, the input product keywords consist of the input and stage fields in the list framework generated in step S6, and the vector database is searched a second time to recall the corpus of input usage in a specific stage.
[0044] Preferably, the specific process of step S9 is as follows:
[0045] S91. Weighted average model and data integration: When multiple values of the same unit are retrieved for the same input in the same stage, the integrated value is calculated using a weighted average model. The weights are calculated by the data source credibility score and the sample size correlation score according to the preset weight coefficients.
[0046] S92. Uncertainty quantification based on relative standard deviation: Calculate the relative standard deviation of the integrated values, and automatically label the uncertainty level based on the results and write it into the description field;
[0047] S93. Dynamic Layout and Visualization: Using the rowspan attribute of HTML tables, merge cells containing names of the same stage and category to generate a visual HTML table.
[0048] After adopting the above technical solution, the present invention has the following beneficial effects: The present invention uses large language model technology, combined with knowledge from the field of environmental science, to effectively capture complex text semantics and improve the accuracy and efficiency of data processing; through standardized workflows and strict prompt word constraints, the model strictly follows instructions to generate a compliant list framework and complete list; it significantly reduces reliance on manual data collection, processing, and analysis, greatly simplifies the construction process of lifecycle inventory, and reduces time and labor costs; it can provide high-quality assessment data for environmental management agencies and enterprises, providing reliable scientific basis for environmental governance decisions and green product design, especially suitable for fields such as marine aquaculture, promoting sustainable development of the industry and ecological environmental protection, and has broad prospects for promotion and application. Attached Figure Description
[0049] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0051] like Figure 1 As shown, a method for calculating a lifecycle process inventory based on a large language model includes the following steps:
[0052] S1. Use the unstructured.io visual recognition toolkit to convert document and image files into .markdown format files;
[0053] In step S1, the files to be processed in the folder are automatically identified using Python, and different types of original files are uniformly converted into .markdown format. The tables in the original files are uniformly converted into HTML format and then embedded into the .markdown file.
[0054] S2. Use a large language model guided by prompt words to decouple the tables in the converted .markdown file;
[0055] The specific process of step S2 is as follows:
[0056] S21. Task Description: Analyze an HTML table line by line or by logical block, and generate natural language sentences for each line or block.
[0057] S22. Table Judgment: Automatically identify whether the table is a lifecycle list table or a general structured table;
[0058] S23. Non-lifecycle inventory table decoupling: Describe each row and retain the original information without performing additional stage inference;
[0059] S24. Lifecycle checklist table decoupling: Classify data into fixed lifecycle stages and integrate their functional units;
[0060] S25. Machine hallucination suppression: When no lines are omitted, no repeated sentences are included, units are missing, or quantities are missing, the sentence should be clearly stated using accurate and neutral language, without making assumptions or interpretations.
[0061] S3. Read the .markdown file and preprocess it into string fragments;
[0062] The preprocessing process in step S3 is as follows: Read all .markdown format files, segment the file content, segment the natural language part of the original document according to the original paragraphs, and segment the HTML table content after decoupling by the large language model according to the logical units generated by decoupling. Each segment represents a complete data entry or description. After the segmentation is completed, remove the segments with a length of less than 15 characters to obtain independent string segments.
[0063] S4. Vectorize the string using an embedded model and store it using a vector retrieval library toolkit;
[0064] The specific process of step S4 is as follows:
[0065] S41. String Encoding: The input string fragment is lexicalized to obtain a lexical sequence. Then, the vocabulary of the embedded model is used to map each lexical to a unique integer index. The vocabulary includes general vocabulary, proper nouns in the fields of environmental science and life cycle assessment, pollutant names, chemical substances, units, and specific industry terms.
[0066] S42. Vectorization: The BAAI-bge-m3 embedded model is used to transform the integer index sequence into a high-dimensional normalized vector, thereby obtaining the data semantic embedding matrix and capturing the deep semantic information of the data.
[0067] S43. Indexing and Storage: Create a collection using the vector retrieval library ChromaDB, storing the vector of each string fragment and its corresponding original string fragment as a record, supporting structured weighted retrieval;
[0068] S5. Retrieve the vector database based on the stage keywords;
[0069] In step S5, the vector database is retrieved based on the structured elements of the predefined standard lifecycle process inventory template; the structured elements include lifecycle stage keywords and target product species names; similarity is calculated using the lifecycle inventory structured weighted semantic similarity, and the calculation formula is as follows: Where Q represents the query; For corpus; To query the lifecycle list structured weighted semantic similarity between Q and corpus D; , and The weighting coefficients are respectively for the breeding stage, species name, and general context, and satisfy the following conditions: ; The semantic embedding vectors of the product corpus D are fed into the retrieved lifecycle list; , and These are the semantic embedding vectors of the corresponding structured elements in query Q; for and Cosine similarity; for and Cosine similarity; for and Cosine similarity;
[0070] S6. Based on the recalled corpus, generate a list framework using a customized large language model;
[0071] In step S6, the customized large language model is the Qwen3-32B model fine-tuned by LoRA algorithm instructions. The generated list framework is in .json format and contains three fields: stage, input and material type.
[0072] In step S6, the prompt words used to guide the large language model include the following:
[0073] S61. Task Description: The large language model is required to act as a professional assistant to the aquaculture life cycle checklist table, and extract and classify the input materials for specific stages based on the provided documents.
[0074] S62. Information Constraints: The large language model is required to use only the information provided in the documents, and must not add any facts, materials or details that are not in the documents, and must only include inputs that are relevant to the requested stage.
[0075] S63. Data Integration and Deduplication: The large language model is required to integrate multiple quantities of the same input in the document and ensure that each unique input appears only once in the final .json format output.
[0076] S64. Field Definition and Output Format: Each entry output by the large language model must contain two fields: input and material type, and must be returned as a .json format array.
[0077] S65. Machine Illusion Suppression: Requires large language models to not guess values or units, not to include value or unit fields in the output, not to include any thought processes, reasoning, or explanations, and to return an empty list when no relevant entry is found;
[0078] S7. Based on the keywords of the input products, perform a secondary search of the vector database;
[0079] In step S7, the input product keywords consist of the input and stage fields in the list framework generated in step S6, and the vector database is searched a second time to recall the corpus of input usage in a specific stage.
[0080] S8. Based on the recalled corpus, generate a complete list using a customized large language model;
[0081] In step S8, the customized large language model is the Qwen3-32B model fine-tuned by the LoRA algorithm instructions. The generated complete list is in .json format and contains five fields: stage, input, material type, value, unit and description.
[0082] In step S8, the prompt words used to guide the large language model mainly include the following:
[0083] S81. Task Description and Context Provision: The large language model is required to act as an expert in lifecycle inventory data completion, using the provided text and table context to calculate and fill in the values and units for given inputs and breeding stages.
[0084] S82. Data Sources and Unit Conversions: Large language models are required to use only the information provided in the documents and to perform unit conversions only when the documents explicitly provide the data or the data can be directly inferred. Otherwise, the original values and units should be reported.
[0085] S83, Data Missing Handling and Output Format: The large language model is required to mark both numerical values and units as unknown when no reliable data is found, and strictly return a valid .json list containing the specified fields;
[0086] S84. Notes and Explanations: Large language models are required to provide explanations in the description field to explain the data source, calculation steps, or any assumptions.
[0087] In steps S6 and S8, the specific process of fine-tuning the Qwen3-32B model using the LoRA algorithm is as follows: The weights of the Qwen3-32B model are updated using a low-rank matrix to optimize instruction compliance; the loss function is calculated by minimizing the cross-entropy loss and combining it with the weights fine-tuned using the LoRA algorithm. The calculation formula is as follows: , , , ,in, This represents the original frozen pre-trained weight matrix; This represents the weight matrix after LoRA fine-tuning; This represents the amount of weight update; , Both ΔW and ΔW belong to ,in, It is the feature dimension. It is the output dimension; and All are low-rank matrices, belonging to and r is much smaller than the dimension of the original matrix; The logical score vector representing the t-th time step; The forward propagation function represents the model; This represents the input vector at time step t. Represents cross-entropy loss; This represents the total number of tokens in the target output sequence; For the first The time step of each word element; The correct target word; for The function operates on Then, the correct lexical units were obtained. The predicted probability; This represents the size of the model's vocabulary; For the word list An index of candidate words; For the model in the first The time step for the first The unnormalized fraction of each word element;
[0088] S9. Use a Python script to process the values in the complete list in .json format and convert it into an HTML table;
[0089] The specific process of step S9 is as follows:
[0090] S91. Weighted average model and data integration: When multiple values of the same unit are retrieved for the same input in the same stage, the integrated value is calculated using a weighted average model. The weights are calculated by the data source credibility score and the sample size correlation score according to the preset weight coefficients.
[0091] S92. Uncertainty quantification based on relative standard deviation: Calculate the relative standard deviation of the integrated values, and automatically label the uncertainty level based on the results and write it into the description field;
[0092] S93. Dynamic Layout and Visualization: Using the rowspan attribute of HTML tables, merge cells containing names of the same stage and category to generate a visual HTML table.
[0093] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for calculating a lifecycle process inventory based on a large language model, characterized in that, Includes the following steps: S1. Use the unstructured.io visual recognition toolkit to convert document and image files into .markdown format files; S2. Use a large language model guided by prompt words to decouple the tables in the converted .markdown file; S3. Read the .markdown file and preprocess it into string fragments; S4. Vectorize the string using an embedded model and store it using a vector retrieval library toolkit; S5. Retrieve the vector database based on the stage keywords; In step S5, the vector database is retrieved based on the structured elements of the predefined standard lifecycle process inventory template; the structured elements include lifecycle stage keywords and target product species names; similarity is calculated using the lifecycle inventory structured weighted semantic similarity, and the calculation formula is as follows: Where Q represents the query; For corpus; To query the lifecycle list structured weighted semantic similarity between Q and corpus D; , and The weighting coefficients are respectively for the breeding stage, species name, and general context, and satisfy the following conditions: ; The semantic embedding vectors of the product corpus D are fed into the retrieved lifecycle list; , and These are the semantic embedding vectors of the corresponding structured elements in query Q; for and Cosine similarity; for and Cosine similarity; for and Cosine similarity; S6. Based on the recalled corpus, generate a list framework using a customized large language model; S7. Based on the keywords of the input products, perform a secondary search of the vector database; S8. Based on the recalled corpus, generate a complete list using a customized large language model; S9. Use a Python script to process the values of the complete list in .json format and convert it into an HTML table.
2. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that: In step S1, the Python language is used to automatically identify the files to be processed in the folder, convert different types of original files into .markdown format, and convert the tables in the original files into HTML format and embed them into the .markdown file.
3. The method for calculating a lifecycle process inventory based on a large language model as described in claim 2, characterized in that, The specific process of step S2 is as follows: S21. Task Description: Analyze an HTML table line by line or by logical block, and generate natural language sentences for each line or block. S22. Table Judgment: Automatically identify whether the table is a lifecycle list table or a general structured table; S23. Non-lifecycle inventory table decoupling: Describe each row and retain the original information without performing additional stage inference; S24. Lifecycle checklist table decoupling: Classify data into fixed lifecycle stages and integrate their functional units; S25. Machine illusion suppression: When no lines are omitted, no repeated sentences are included, units are missing, or quantities are missing, the sentence should be clearly stated using accurate and neutral language, without making assumptions or interpretations.
4. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that: The preprocessing process in step S3 is as follows: read all .markdown format files, segment the file content, segment the natural language part of the original document according to the original paragraphs, and segment the HTML table content after decoupling by the large language model according to the logical units generated by decoupling. Each segment represents a complete data entry or description. After the segmentation is completed, segments with a length of less than 15 characters are removed to obtain independent string segments.
5. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that: The specific process of step S4 is as follows: S41. String encoding: The input string fragment is lexicalized to obtain a lexical sequence, and then each lexical is mapped to a unique integer index using the vocabulary of the embedded model; the vocabulary includes general vocabulary, proper nouns in the fields of environmental science and life cycle assessment, pollutant names, chemical substances, units and specific industry terms; S42. Vectorization: The BAAI-bge-m3 embedded model is used to transform the integer index sequence into a high-dimensional normalized vector, thereby obtaining the data semantic embedding matrix and capturing the deep semantic information of the data. S43. Indexing and Storage: Create a collection using the vector retrieval library ChromaDB, storing the vector of each string fragment and its corresponding original string fragment as a record, supporting structured weighted retrieval.
6. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that, In step S6, the customized large language model is the Qwen3-32B model fine-tuned by LoRA algorithm instructions. The generated list framework is in .json format and contains three fields: stage, input and material type. In step S6, the prompt words used to guide the large language model include the following: S61. Task Description: The large language model is required to act as a professional assistant to the aquaculture life cycle checklist table, and extract and classify the input materials for specific stages based on the provided documents. S62. Information Constraints: The large language model is required to use only the information provided in the documents, and must not add any facts, materials or details that are not in the documents, and must only include inputs that are relevant to the requested stage. S63. Data Integration and Deduplication: The large language model is required to integrate multiple quantities of the same input in the document and ensure that each unique input appears only once in the final .json format output. S64. Field Definition and Output Format: Each entry output by the large language model must contain two fields: input and material type, and must be returned as a .json format array. S65, Machine Illusion Suppression: Requires large language models not to guess values or units, not to include value or unit fields in the output, not to include any thought processes, reasoning or explanations, and to return an empty list when no relevant entry is found. In step S8, the customized large language model is the Qwen3-32B model fine-tuned by the LoRA algorithm instructions. The generated complete list is in .json format and contains five fields: stage, input, material type, value, unit and description. In step S8, the prompt words used to guide the large language model mainly include the following: S81. Task Description and Context Provision: The large language model is required to act as an expert in lifecycle inventory data completion, using the provided text and table context to calculate and fill in the values and units for given inputs and breeding stages. S82. Data Sources and Unit Conversions: Large language models are required to use only the information provided in the documents and to perform unit conversions only when the documents explicitly provide the data or the data can be directly inferred. Otherwise, the original values and units should be reported. S83, Data Missing Handling and Output Format: The large language model is required to mark both numerical values and units as unknown when no reliable data is found, and strictly return a valid .json list containing the specified fields; S84. Notes and Explanations: Requires the large language model to provide explanations in the description field to explain the data source, calculation steps, or any assumptions.
7. The method for calculating a lifecycle process inventory based on a large language model as described in claim 6, characterized in that, In steps S6 and S8, the specific process of fine-tuning the Qwen3-32B model using the LoRA algorithm is as follows: The weights of the Qwen3-32B model are updated using a low-rank matrix to optimize instruction compliance; the loss function is calculated by minimizing the cross-entropy loss and combining it with the weights fine-tuned using the LoRA algorithm. The calculation formula is as follows: , , , ,in, This represents the original frozen pre-trained weight matrix; This represents the weight matrix after LoRA fine-tuning; This represents the amount of weight update; , Both ΔW and ΔW belong to ,in, It is the feature dimension. It is the output dimension; and All are low-rank matrices, belonging to and r is much smaller than the dimension of the original matrix; The logical score vector representing the t-th time step; The forward propagation function represents the model; This represents the input vector at time step t. Represents cross-entropy loss; This represents the total number of tokens in the target output sequence; For the first The time step of each word element; The correct target word; for The function operates on Then, the correct lexical units were obtained. The predicted probability; This represents the size of the model's vocabulary; For the word list An index of candidate words; For the model in the first The time step for the first The unnormalized fractions of each word unit.
8. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that: In step S7, the input product keywords consist of the input and stage fields in the list framework generated in step S6, and the vector database is searched a second time to recall the corpus of input usage in a specific stage.
9. The method for calculating a lifecycle process inventory based on a large language model as described in claim 1, characterized in that, The specific process of step S9 is as follows: S91. Weighted average model and data integration: When multiple values of the same unit are retrieved for the same input in the same stage, the integrated value is calculated using a weighted average model. The weights are calculated by the data source credibility score and the sample size correlation score according to the preset weight coefficients. S92. Uncertainty quantification based on relative standard deviation: Calculate the relative standard deviation of the integrated values, and automatically label the uncertainty level based on the results and write it into the description field; S93. Dynamic Layout and Visualization: Using the rowspan attribute of HTML tables, merge cells containing names of the same stage and category to generate a visual HTML table.
Citation Information
Patent Citations
Test flight data segment retrieval method and system based on large language model
CN118035298A
Data analysis report generation method based on large language model
CN118626523A