Intelligent carbon footprint accounting method and system based on multi-modal large model

Abstract

The invention discloses an intelligent carbon footprint accounting method and system based on a multi-modal large model, and the method comprises the steps: substituting a to-be-accounted product picture into a pre-trained intelligent carbon footprint large model, and generating an emission source list corresponding to the to-be-accounted product picture; matching corresponding emission factors by adopting an efficient matching algorithm based on the emission source list corresponding to the to-be-checked product picture in combination with a pre-constructed carbon emission factor database; calculating the carbon footprint of the to-be-checked product based on the emission source list, the emission factor and the activity level data corresponding to the to-be-checked product picture; wherein the intelligent carbon footprint large model is obtained by training the multi-modal large model by adopting a K-fold cross validation mechanism based on the product picture and the emission source list corresponding to the product picture. According to the method, the emission source list is automatically generated through the intelligent carbon footprint large model, the dependence on manpower is eliminated, the efficiency is improved, the cost and the error rate are reduced, operation and result interpretation are easy, and the method can better adapt to different industries and scenes.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an intelligent carbon footprint calculation method and system based on a multimodal large model. Background Art

[0002] As global climate change becomes increasingly severe, companies and consumers are paying more and more attention to the carbon footprint of products. Traditional carbon footprint accounting methods rely on manual work, with cumbersome processes, low efficiency and error-prone. Existing automated tools have limited functions and are difficult to meet the needs of different industries and scenarios. For example, it is difficult to handle a wide variety of products, difficult data collection, and complex emission factor matching.

[0003] The traditional carbon footprint accounting method has the following problems: Highly dependent on manual labor: Product confirmation, life cycle assessment, data collection, emission source identification and factor application all need to be done manually, which is time-consuming and labor-intensive.

[0004] Inefficiency is costly: cumbersome processes and a large amount of manual work lead to long accounting cycles and high costs.

[0005] Error-prone: Manual operations can easily introduce human errors, affecting the accuracy of accounting results.

[0006] High professional threshold: Users need to have professional knowledge to correctly operate and interpret the results.

[0007] Limited applicability: Existing automation tools are difficult to meet the needs of different industries and scenarios. Summary of the invention

[0008] In order to solve the problems of traditional carbon footprint accounting methods, such as high reliance on manual labor, low efficiency, high cost, easy to make mistakes, high professional threshold and limited applicability, the present invention proposes an intelligent carbon footprint accounting method based on a multimodal large model, including: Substitute the image of the product to be accounted for into the pre-trained intelligent carbon footprint model to generate a list of emission sources corresponding to the image of the product to be accounted for; Based on the emission source list corresponding to the product image to be accounted for and the pre-built carbon emission factor database, an efficient matching algorithm is used to match the corresponding emission factors; Calculate the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

[0009] Optionally, the training of the intelligent carbon footprint model includes: Obtain various types of product images, and annotate the product images with corresponding emission source inventories, and construct a data set from the product images and the emission source inventories corresponding to the product images; Preprocessing the data set to obtain a data set that meets the model input requirements; Based on the data set that meets the model input requirements, a fine-tuning algorithm is used to fine-tune the multimodal large model to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

[0010] Optionally, preprocessing the data set to obtain a data set that meets the model input requirements includes: Performing data cleaning, format conversion, image size adjustment and data enhancement processing on the data set to obtain processed data; The processed data is converted into an input format of a multimodal large model to obtain a data set that meets the model input requirements.

[0011] Optionally, the fine-tuning training of the multimodal large model using a fine-tuning algorithm based on a data set that meets the model input requirements to obtain a trained intelligent carbon footprint large model includes: The K-fold cross-validation mechanism is used to randomly and equally split the data set that meets the model input requirements into K subsets. One subset is retained as the validation set each time, and the remaining subsets are fine-tuned and trained on the multimodal large model using the fine-tuning algorithm. This is repeated K times to ensure that each subset is used for validation at least once. By evaluating the performance of the model obtained in each iteration, the actual effect of the model under different settings is measured, and the model parameters and training strategies are iteratively optimized based on the actual effect; The optimal solution is selected based on the information collected during the cross-validation process, and the model corresponding to the optimal solution is used as the trained smart carbon footprint model; Among them, the fine-tuning algorithms include: LoRA and Freeze Tuning.

[0012] Optionally, the step of calculating the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the image of the product to be accounted for includes: Multiply each emission source in the emission source list corresponding to the product image to be accounted for by the emission factor and activity level data corresponding to each emission source, and then sum them up to finally obtain the carbon footprint of the product.

[0013] Optionally, after calculating the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for, the method further includes: Generate a carbon footprint report based on the carbon footprint of the product to be accounted for and display it in a visual way.

[0014] On the other hand, the present invention also provides an intelligent carbon footprint accounting system based on a multimodal large model, comprising: An emission source list generation module is used to substitute the image of the product to be accounted for into the pre-trained intelligent carbon footprint model to generate an emission source list corresponding to the image of the product to be accounted for; An intelligent factor matching module is used to match the corresponding emission factors based on the emission source list corresponding to the product image to be accounted for and a pre-built carbon emission factor database using an efficient matching algorithm; A carbon footprint calculation module, used to calculate the carbon footprint of the product to be calculated based on the emission source list, emission factor and activity level data corresponding to the product image to be calculated; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

[0015] Optionally, a model fine-tuning training module is also included, which is used to perform fine-tuning training on the multimodal large model based on the data set; The model fine-tuning training module includes: A data set acquisition module is used to acquire various types of product images, annotate the product images with corresponding emission source lists, and construct a data set from the product images and the emission source lists corresponding to the product images; A base model module, used for preprocessing the data set to obtain a data set that meets the model input requirements; The model acquisition module uses a fine-tuning algorithm to fine-tune the multimodal large model based on a data set that meets the model input requirements to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

[0016] Furthermore, the bottom model module is specifically used for: Performing data cleaning, format conversion, image size adjustment and data enhancement processing on the data set to obtain processed data; The processed data is converted into an input format of a multimodal large model to obtain a data set that meets the model input requirements.

[0017] Furthermore, the model acquisition module is specifically used for: The K-fold cross-validation mechanism is used to randomly and equally split the data set that meets the model input requirements into K subsets. One subset is retained as the validation set each time, and the remaining subsets are fine-tuned and trained on the multimodal large model using the fine-tuning algorithm. This is repeated K times to ensure that each subset is used for validation at least once. By evaluating the performance of the model obtained in each iteration, the actual effect of the model under different settings is measured, and the model parameters and training strategies are iteratively optimized based on the actual effect; The optimal solution is selected based on the information collected during the cross-validation process, and the model corresponding to the optimal solution is used as the trained smart carbon footprint model; Among them, the fine-tuning algorithms include: LoRA and Freeze Tuning.

[0018] Optionally, the carbon footprint calculation module is specifically used to: Multiply each emission source in the emission source list corresponding to the product image to be accounted for by the emission factor and activity level data corresponding to each emission source, and then sum them up to finally obtain the carbon footprint of the product.

[0019] Optionally, it also includes: a report visualization module, which is used to generate a carbon footprint report based on the carbon footprint of the product to be accounted for, and display it in a visual manner.

[0020] In another aspect, the present application further provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, an intelligent carbon footprint accounting method based on a multimodal large model as described above is implemented.

[0021] On the other hand, the present application also provides a readable storage medium having an execution program stored thereon, which, when executed, implements an intelligent carbon footprint accounting method based on a multimodal large model as described above.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for calculating intelligent carbon footprint based on a multimodal large model, including: substituting a picture of a product to be calculated into a pre-trained intelligent carbon footprint large model, generating an emission source list corresponding to the picture of the product to be calculated; matching the corresponding emission factors based on the emission source list corresponding to the picture of the product to be calculated in combination with a pre-built carbon emission factor database, using an efficient matching algorithm; calculating the carbon footprint of the product to be calculated based on the emission source list, emission factor and activity level data corresponding to the picture of the product to be calculated; wherein the intelligent carbon footprint large model is obtained by training the multimodal large model using a K-fold cross-validation mechanism based on the product picture and the emission source list corresponding to the product picture. The present invention automatically generates an emission source list through an intelligent carbon footprint large model, gets rid of dependence on manual labor, improves efficiency, reduces costs, and reduces error rates. The operating user does not need to have professional knowledge to correctly operate and interpret the results, and can better adapt to different industries and scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of an intelligent carbon footprint calculation method based on a multimodal large model of the present invention; Figure 2 It is a flow chart of an intelligent carbon footprint calculation method based on a multi-modal large model in a specific embodiment of the present invention; Figure 3 It is a schematic diagram of the structure of an intelligent carbon footprint calculation system based on a multi-modal large model of the present invention; Figure 4 The figure is a schematic diagram of the structure of an electronic device of the present invention. DETAILED DESCRIPTION

[0024] In order to better understand the present invention, the content of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0025] Embodiment 1: An intelligent carbon footprint accounting method based on a multimodal large model, such as Figure 1 As shown, including: Step 1: Substitute the image of the product to be calculated into the pre-trained intelligent carbon footprint model to generate the emission source list corresponding to the image of the product to be calculated; Step 2: Based on the emission source list corresponding to the product image to be accounted for and the pre-built carbon emission factor database, an efficient matching algorithm is used to match the corresponding emission factors; Step 3: Calculate the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

[0026] The present invention provides an intelligent carbon footprint calculation method based on a multimodal large model, which is implemented by an intelligent carbon footprint calculation system. Figure 2 As shown, the specific steps are as follows: Before step 1, it also includes: introducing the trained intelligent carbon footprint model involved in this application: The training of the intelligent carbon footprint model includes: Obtain various types of product images, and annotate the product images with corresponding emission source inventories, and construct a data set from the product images and the emission source inventories corresponding to the product images; Preprocessing the data set to obtain a data set that meets the model input requirements; Based on the data set that meets the model input requirements, a fine-tuning algorithm is used to fine-tune the multimodal large model to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

[0027] Furthermore, the data set is preprocessed to obtain a data set that meets the model input requirements, including: S1 Data collection and labeling: Collect image data sets covering various product types, and have professionals label their corresponding emission source lists to form a set of product images and emission source data lists. For example, collect images of different product categories such as mobile phones, clothing, and food, and label their emission sources in the stages of raw material acquisition, production and manufacturing, transportation and distribution, use, and waste recycling, such as electricity consumption, fuel combustion, refrigerant leakage, etc. During the labeling process, the unit of the emission source must be clearly specified, such as kg, kWh, km, etc.

[0028] S2 Data preprocessing and model selection: The dataset obtained in S1 is preprocessed to adapt it to the input requirements of the TongQwen2-VL-Instruct(7B) model (i.e., TongYiQianwen-VL model), including data cleaning, format conversion, image resizing, and data augmentation.

[0029] 1. Data cleaning: Remove invalid data: Delete or repair corrupted image files (e.g. images that cannot be opened or display incorrectly). Delete data entries where product information is missing or where the emission source inventory is obviously wrong (e.g. the inventory contains information that is not relevant to the product or the values ​​are obviously unreasonable).

[0030] Dealing with duplicate data: Identify and remove duplicate product images and emission source lists, retaining unique data entries. If there are multiple slightly different images of the same product, consider retaining the images with the larger differences or merging them into a representative data entry.

[0031] Handle noise data: Check the emission source inventory for noise data, such as spelling errors, inconsistent units, or unusual values. Correct, delete, or replace as appropriate. For example, you can use regular expressions to identify and correct common spelling errors.

[0032] 2. Format conversion: Image format: Convert all product images to JPG format, because JPEG format usually provides better compression and compatibility, reduces the size of the data set, unifies the color mode of the images, and converts them to RGB mode to facilitate the use of the deep learning framework.

[0033] Emission source inventory: Convert the emission source inventory into a structured text description. Convert the original table data into keyed JSON format. For example, an emission source inventory can be converted into the following format:{"product":"jeans","source_inventory":{"stage":"Raw material acquisition","source":"Cotton planting","unit":"kg"},{"stage":"Manufacturing","source":"Spinning","unit":"kWh"},{"stage":"Manufacturing","source":"Weaving","unit":"kWh"},{"stage":"Manufacturing","source":"Printing and dyeing","unit":"kWh"},{"stage":"Manufacturing","source":"Cutting and sewing","unit":"kWh"},{"stage":"Transportation and distribution","source":"Road transportation","unit":"km"},{"stage":"Use stage","source":"Washing","unit":"times"}}.

[0034] Among them, product is the product, source_inventory is the emission source inventory, stage is the stage, source is the source, and unit is the unit.

[0035] 3. Image resizing: Resizing: Resizing all product images to a uniform size, 224x224 pixels, not only reduces the image size, but also improves training computation. When all images have the same size, batch processing can be more efficient, thereby increasing training speed. We use data augmentation techniques, such as random cropping or rotation, so having a baseline size makes these operations simpler and clearer. Images that are too large will increase the computational burden, while images that are too small may lose important details.

[0036] Adjustment method: Use a suitable image scaling algorithm, such as bilinear interpolation or bicubic interpolation, to maintain image quality as much as possible.

[0037] 4. Image Enhancement: Purpose: To improve the generalization and robustness of the model by increasing the diversity of training data.

[0038] method: Random Cropping: Randomly crop different regions of the image.

[0039] Random Flip: Flip the image horizontally or vertically.

[0040] Random Rotation: Rotate the image by a small angle.

[0041] Color Jitter: Randomly adjust the brightness, contrast, saturation, and hue of an image.

[0042] Add Noise: Add a small amount of Gaussian noise.

[0043] Augmentation level: Adjust the augmentation level based on the size of the dataset and the overfitting of the model. Excessive augmentation may cause the model to learn irrelevant features.

[0044] 5. Convert to Qwen2-VL-Instruct input format: The final input data contains images and corresponding text descriptions. A dictionary is used to represent each data sample. The following is a data sample: [ { "conversations": [ { "role": "system", "content": "You are an emissions inventory generation expert" }, { "role": "user", "content": "Please give the emission list of this image, output in json format " }, { "role": "assistant", "content": "[{"product":"jeans","source_inventory":{"stage":"Raw material acquisition","source":"Cotton planting","unit":"kg"},{"stage":"Manufacturing","source":"Spinning","unit":"kWh"},{"stage":"Manufacturing","source":"Weaving","unit":"kWh"},{"stage":"Manufacturing","source":"Printing and dyeing","unit":"kWh"},{"stage":"Manufacturing","source":"Cutting and sewing","unit":"kWh"},{"stage":"Transportation and distribution","source":"Road transportation","unit":"km"},{"stage":"Use stage","source":"Washing","unit":"times"}}]" } ], "images": [ "images / instance_1579398113581395972.jpg" ] } ].

[0045] Among them, conversations is conversation, role is role, system is system, content is content, user is user, assistant is assistant, product is product, images is image, and instance is instance.

[0046] Select a suitable multimodal large model as the base model. We chose Qwen2-VL-Instruct (7B) to prepare for subsequent fine-tuning. We chose Qwen2-VL-Instruct (7B parameter version) as the base model. It is a powerful multimodal model that performs well in image understanding and text generation. The core advantage of this model lies in its strong visual and language understanding capabilities, thanks to its combination of visual transformers and advanced natural language processing (NLP) technology. By pre-training on large-scale datasets (including text, image-text pairs, and OCR data), Qwen2-VL-Instruct has acquired extensive world knowledge and strong generalization capabilities, enabling it to effectively handle a variety of multimodal tasks. The model's multimodal fusion technology can effectively combine visual and text information, thereby achieving more efficient information processing and decision support. Dynamic resolution support allows the model to process images of arbitrary resolution and convert them into a variable number of visual tokens, further improving the efficiency and accuracy of visual representation. In addition, the application of multimodal rotation position encoding (M-RoPE) enhances the model's ability to encode text, image, and video position information. Qwen2-VL-Instruct also has high computational efficiency, can run quickly on different hardware platforms, and supports long text generation of up to 128K tokens, making it competent for complex image question answering, document parsing and other tasks. The release of the quantized version further reduces the deployment cost of the model.

[0047] S3 Model Training and Optimization: Use efficient fine-tuning algorithms such as LoRA (low-rank adaptation) and Freeze Tuning to fine-tune the selected basic model based on the data set preprocessed in the S2 stage. To enhance the generalization ability of the model, we will introduce a K-fold cross-validation mechanism, which is to randomly and equally split the data set into K subsets; each time, one subset is reserved as the validation set and the rest is used for training, and this process will be repeated K times to ensure that each subset is used for validation at least once. By evaluating the performance of the model obtained in each iteration (including but not limited to accuracy, recall, and F1 value), we can more accurately measure the actual effects under different settings, and continuously iterate and optimize model parameters and training strategies accordingly. Finally, based on the information collected during the cross-validation process, the optimal solution is selected to build a large intelligent carbon footprint model with high accuracy and good generalization.

[0048] Step 1: Substitute the image of the product to be accounted into the pre-trained intelligent carbon footprint model to generate the emission source list corresponding to the image of the product to be accounted, including: S4 Automatic generation of emission source list: After the user uploads the product picture, the intelligent carbon footprint model automatically identifies the product and generates the corresponding emission source list, including the emission source name, life cycle stage and unit.

[0049] Step 2: Based on the emission source list corresponding to the product image to be accounted for and the pre-built carbon emission factor database, an efficient matching algorithm is used to match the corresponding emission factors, specifically including: S5 Intelligent matching of emission source factors: Based on the generated emission source list, combined with the system's built-in, regularly updated carbon emission factor database, an efficient matching algorithm is used to automatically match the corresponding emission factors. For example, according to the emission source "electricity consumption", the corresponding regional power grid emission factor is matched. During the matching process, factors such as product type, production process, region, and time can be considered to improve the matching accuracy.

[0050] Step 3: Calculate the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for, including: Generate a carbon footprint report based on the carbon footprint of the product to be accounted for and display it in a visual way.

[0051] S6 Automatic carbon footprint calculation: The system uses the built-in carbon footprint calculation model of the platform based on the emission source list generated by S4, the emission factors matched by S5, and the activity level data provided by the user (such as output, transportation distance, etc.). The carbon footprint calculation model is a calculation model that receives the emission source list, emission factors and activity level data as input, and then calculates the carbon footprint of the product according to a predefined formula (for example, a calculation method that complies with the ISO 14067 standard). It does not directly process images or texts, but performs numerical calculations. The carbon footprint calculation model is composed of a series of formulas and calculation rules. Its core function is to multiply each emission source in the emission source list with its corresponding emission factor and activity level data, and then sum them up to finally obtain the carbon footprint of the product.

[0052] For example, a simplified carbon footprint accounting model can be expressed as: Carbon footprint = Σ (activity data i * emission factor i) Where: i represents different emission sources.

[0053] Activity level data i represents the activity level data corresponding to the i-th emission source (for example, electricity consumption, fuel consumption, transportation distance, etc.).

[0054] Emission factor i represents the emission factor corresponding to the i-th emission source (e.g., CO2 emissions per kilowatt-hour of electricity generated).

[0055] Automatically calculate the carbon footprint of the product's entire life cycle. The accounting model should comply with international or national standards, such as ISO14067, PAS 2050, etc. "The entire product life cycle" refers to the entire process of the product from raw material acquisition, manufacturing, transportation and distribution, use to final waste recycling. To calculate the carbon footprint of the product's entire life cycle, it is necessary to consider the emission sources of the product at each stage and its corresponding activity level data and emission factors.

[0056] The present invention provides an intelligent carbon footprint accounting method based on a multimodal large model, which also includes: generating a carbon footprint report according to the carbon footprint of the product to be accounted for, and displaying it in a visual manner.

[0057] Specifically include: S7 Report Generation and Visualization: The system automatically generates a detailed carbon footprint report based on the calculation results, and displays it in a visual way such as charts to facilitate user understanding and analysis. The report content can include the carbon emission proportion of each life cycle stage, total carbon emissions, and comparative analysis of carbon footprints with similar products.

[0058] The present invention trains a large multimodal model to enable it to automatically identify product types based on product images and generate a corresponding emission source list, thereby replacing the traditional manual identification and data entry process and greatly improving efficiency and accuracy.

[0059] The present invention adopts an intelligent matching algorithm, which can match the most appropriate emission factor from the database according to the generated emission source list and the specific characteristics of the product, thereby improving the accuracy of the calculation.

[0060] The present invention utilizes multimodal data such as product images and emission source lists for model training and carbon footprint accounting, thereby improving the accuracy and robustness of the system.

[0061] The intelligent carbon footprint accounting system and method proposed in the present invention constitute a complete set of automated solutions, realizing automation of the entire process from data collection to report generation.

[0062] Example 2 The present invention provides an intelligent carbon footprint calculation method and system based on a multimodal large model. The following describes in detail the use and effects of the system in combination with specific cases and parameters.

[0063] Case: Carbon footprint accounting of jeans 1. Data collection and annotation (S1): Data source: This case uses a self-collected jeans image dataset, which contains 1,000 jeans images of different styles, colors, and brands, covering the common types of jeans on the market. The images are all from e-commerce platforms and clothing brand official websites.

[0064] Labeling method: We invited a textile industry expert to label the pictures and used the online labeling tool LabelMe to label the emission source inventory. The labeled data is the product and the corresponding emission source inventory data. The labeling standard refers to the ISO 14067:2018 carbon footprint accounting standard and combines the actual situation of the clothing industry. The emission source inventory data is stored in JSON format. Example: The emission source list of a picture of blue jeans (jean_0001.jpg) is shown above, which includes emission sources at all stages from raw material acquisition to waste recycling, where jeans refers to jeans.

[0065] {"product":"jeans","source_inventory":{"stage":"Raw material acquisition","source":"Cotton planting","unit":"kg"},{"stage":"Manufacturing","source":"Spinning","unit":"kWh"},{"stage":"Manufacturing","source":"Weaving","unit":"kWh"},{"stage":"Manufacturing","source":"Printing and dyeing","unit":"kWh"},{"stage":"Manufacturing","source":"Cutting and sewing","unit":"kWh"},{"stage":"Transportation and distribution","source":"Road transportation","unit":"km"},{"stage":"Use stage","source":"Washing","unit":"times"}} 2. Data preprocessing and model selection (S2): Preprocessing steps: Data format: Image data is stored in JPG format. All product images are converted to JPG format because JPEG format usually provides better compression and compatibility, reducing the size of the data set, unifying the color mode of the image and converting it to RGB mode for easy use in the deepening learning framework. All product images are resized to a uniform size and uniformly processed into 224x224 pixels, which not only reduces the image size but also improves the training calculation. When all images have the same size, batch processing can be more efficient, thereby increasing the training speed. We use data augmentation techniques, such as random cropping or rotation, so having a baseline size can make these operations simpler and clearer. Although Qwen2-VL-Instruct supports dynamic resolution, this step may not be necessary, but too large images will increase the computational burden, and too small images may lose important details. Improve the generalization and robustness of the model by increasing the diversity of training data. Methods: Random Cropping: Randomly crop different regions of the image.

[0066] Random Flip: Flip the image horizontally or vertically.

[0067] Random Rotation: Rotate the image by a small angle.

[0068] Color Jitter: Randomly adjust the brightness, contrast, saturation, and hue of an image.

[0069] Add Noise: Add a small amount of Gaussian noise.

[0070] Augmentation level: Adjust the augmentation level based on the size of the dataset and the overfitting of the model. Excessive augmentation may cause the model to learn irrelevant features.

[0071] Format the labeled emission source inventory data text and processed image data and convert them into the Qwen2-VL-Instruct training input format: The final input data contains images and corresponding text descriptions. A dictionary is used to represent each data sample. The following is a data sample. The example is as follows: [ { "conversations": [ { "role": "system", "content": "You are an emissions inventory generation expert" }, { "role": "user", "content": "Give this image's emission list, output in json format " }, { "role": "assistant", "content": "[{"product":"jeans","source_inventory":{"stage":"Raw material acquisition","source":"Cotton planting","unit":"kg"},{"stage":"Manufacturing","source":"Spinning","unit":"kWh"},{"stage":"Manufacturing","source":"Weaving","unit":"kWh"},{"stage":"Manufacturing","source":"Printing and dyeing","unit":"kWh"},{"stage":"Manufacturing","source":"Cutting and sewing","unit":"kWh"},{"stage":"Transportation and distribution","source":"Road transportation","unit":"km"},{"stage":"Use stage","source":"Washing","unit":"times"}}]" } ], "images": [ "images / instance_1579398113581395972.jpg" ] } ] Model selection: Qwen2-VL-Instruct (7B parameter version) is selected as the base model. It is a powerful multimodal model that excels in image understanding and text generation. The core advantage of this model lies in its strong visual and language understanding capabilities, thanks to its combination of Visual Transformers and advanced Natural Language Processing (NLP) techniques. By pre-training on large-scale datasets (including text, image-text pairs, and OCR data), Qwen2-VL-Instruct has acquired extensive world knowledge and strong generalization capabilities, enabling it to effectively handle a variety of multimodal tasks. The model's multimodal fusion technology can effectively combine visual and textual information, thereby achieving more efficient information processing and decision support. Dynamic resolution support allows the model to process images of arbitrary resolution and convert them into a variable number of visual tokens, further improving the efficiency and accuracy of visual representation, where tokens represent labels. In addition, the application of Multimodal Rotational Position Encoding (M-RoPE) enhances the model's ability to encode text, image, and video position information. Qwen2-VL-Instruct also has high computational efficiency, can run quickly on different hardware platforms, and supports long text generation of up to 128K tokens, making it competent for complex image question answering, document parsing and other tasks. The release of the quantized version further reduces the deployment cost of the model.

[0072] 3. Model training and optimization (S3): The present invention uses the LoRA (Low-Rank Adaptation) fine-tuning algorithm to fine-tune the pre-trained multimodal large model. LoRA is an efficient fine-tuning method that freezes the weights of the pre-trained model and injects a trainable low-rank matrix into each Transformer layer. This method significantly reduces the number of parameters that need to be trained, thereby reducing computational costs and storage requirements, while effectively improving the performance of the model on specific tasks. Low-Rank Adaptation is low-rank adaptation, and Transformer is a transformer.

[0073] Specifically, the implementation steps of LoRA are as follows: Freeze pre-trained model weights: Freeze all parameters of a pre-trained multimodal large model (e.g., Qwen2-VL-Instruct(7B)) so that they remain unchanged during fine-tuning. This can effectively preserve the knowledge and generalization ability of the pre-trained model.

[0074] Inject low-rank matrices: Add a pair of low-rank matrices A and B next to the linear projection layer of each Transformer layer (for example, the Query, Key, Value matrices in the attention mechanism and the weight matrix in the feedforward network). The dimensions of matrix A are rxd, and the dimensions of matrix B are dx r, where r is the rank of LoRA, which is much smaller than the dimension d of the original weight matrix, Query is the query, Key is the key, and Value is the value. The product of these two matrices A x B has a maximum rank of r, simulating the update of the full-rank matrix.

[0075] Forward propagation calculation: During the forward propagation process, the output of the original linear projection layer is added to the product of the low-rank matrices A and B to obtain the final output. That is: h = W * x + (A * B) * x in: x is the input feature.

[0076] W is the weight matrix of the frozen pre-trained model.

[0077] A and B are trainable low-rank matrices.

[0078] h is the output feature.

[0079] Back-propagation update: During back-propagation, only the parameters of the low-rank matrices A and B are updated, while the weights W of the pre-trained model remain unchanged. Since r is much smaller than d, the number of parameters that need to be trained is greatly reduced.

[0080] Hyperparameter settings: In this invention, the rank of LoRA is set to 8, and the alpha value is set to 16. The learning rate is set to 1e-4, and the AdamW optimizer is used for training. The batch size is set to 32, and the number of training rounds is 10.

[0081] Evaluation and optimization: Use indicators such as accuracy, recall, and F1 value to evaluate the performance of the model on the test set. And adopt the cosine annealing learning rate adjustment strategy and early stopping strategy to prevent overfitting and further improve the generalization ability of the model.

[0082] Through the above steps, the LoRA algorithm can efficiently fine-tune the pre-trained multimodal large model, so that it can accurately generate an emission source list based on product images, thereby realizing automated carbon footprint accounting.

[0083] The composition of the training set is: The training set consists of product images collected and annotated in the S1 phase and their corresponding emission source lists. Each piece of training data contains a product image and an emission source list in JSON format. For example: { "conversations": [ { "role": "system", "content": "You are an emissions inventory generation expert" }, { "role": "user", "content": "Please give the emission list of this image, output in json format " }, { "role": "assistant", "content": "[{"product":"jeans","source_inventory":{"stage":"Raw material acquisition","source":"Cotton planting","unit":"kg"},{"stage":"Manufacturing","source":"Spinning","unit":"kWh"},{"stage":"Manufacturing","source":"Weaving","unit":"kWh"},{"stage":"Manufacturing","source":"Printing and dyeing","unit":"kWh"},{"stage":"Manufacturing","source":"Cutting and sewing","unit":"kWh"},{"stage":"Transportation and distribution","source":"Road transportation","unit":"km"},{"stage":"Use stage","source":"Washing","unit":"times"}}]" } ], "images": [ "images / instance_1579398113581395972.jpg" ] } Training input System Role Definition: This section specifies the role played by the model, i.e. “You are an emissions inventory generation expert”.

[0084] User query: The user requests the model to generate an emissions inventory based on a given image and output it in JSON format. The specific query is: "Please give the emissions inventory for this image, output in JSON format ”.

[0085] image: The actual image file path or reference (in this case "images / instance_1579398113581395972.jpg"), which represents the visual information associated with the user's request.

[0086] Training Output The output of the model is the predicted emissions inventory, which should be as consistent as possible with the content of the assistant role provided in the JSON data. Generate an example: { "product": "jeans", "source_inventory": [ {"stage": "Raw material acquisition", "source": "Cotton planting", "unit": "kg"}, {"stage": "Manufacturing", "source": "Spinning", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Weaving", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Printing and dyeing", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Cutting and Sewing", "unit": "kWh"}, {"stage": "Transportation and Distribution", "source": "Road Transport", "unit": "km"}, {"stage": "Usage stage", "source": "Washing", "unit": "times"} ] } Use a loss function (e.g., cross entropy loss for multi-label classification) to calculate the difference between the model's predictions and the true labels.

[0087] The AdamW optimizer is used to calculate the gradients and update the parameters of the LoRA low-rank matrices A and B. The weights of the pre-trained model are kept frozen.

[0088] Cosine annealing learning rate adjustment: The learning rate is not fixed, but dynamically adjusted according to the training progress. The present invention adopts the cosine annealing learning rate adjustment strategy, which makes the learning rate larger at the beginning of training and then gradually decreases, which helps the model converge to the optimal solution faster. The formula is as follows: ηt = ηmin + 0.5 * (ηmax - ηmin)* (1 + cos(t * π / T)) in: ηt is the learning rate of the tth epoch, where epoch means iteration.

[0089] ηmin is the minimum learning rate.

[0090] ηmax is the maximum learning rate (initial learning rate).

[0091] t is the current epoch number.

[0092] T is the total number of training epochs.

[0093] For example, if the initial learning rate ηmax is 1e-4 and the total number of training rounds T is 10, the learning rate of each epoch will be adjusted according to the above formula.

[0094] Early stopping strategy: To prevent overfitting, the present invention adopts an early stopping strategy. During the training process, the performance of the model is regularly evaluated on the validation set. If the performance on the validation set does not improve for multiple epochs in a row, the training is stopped and the model with the best performance is selected as the final model. For example, if the F1 value on the validation set does not improve for 3 consecutive epochs, the training is stopped.

[0095] Evaluation Metrics: Use metrics such as precision, recall, and F1 value to evaluate the performance of the model on the test set.

[0096] Through the above steps and strategies, the intelligent carbon footprint model can be effectively trained to accurately identify products and generate emission source inventories.

[0097] 4. Automatic generation of emission source list (S4): Input example: Upload a new picture of jeans.

[0098] Output example: The model automatically generates the following emission source list: { "product": "jeans", "source_inventory": [ {"stage": "Raw material acquisition", "source": "Cotton planting", "unit": "kg"}, {"stage": "Manufacturing", "source": "Spinning", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Weaving", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Printing and dyeing", "unit": "kWh"}, {"stage": "Manufacturing", "source": "Cutting and Sewing", "unit": "kWh"}, {"stage": "Transportation and Distribution", "source": "Road Transport", "unit": "km"}, {"stage": "Usage stage", "source": "Washing", "unit": "times"} ] } 5. Intelligent matching of emission source factors (S5): Matching algorithm: Use a matching algorithm based on similarity calculation to calculate the cosine similarity between the emission source and the emission factors in the database, and select the factor with the highest similarity.

[0099] Matching rules: Consider factors such as product type (jeans), production process (e.g., traditional process vs. water-saving process), and region (e.g., China vs. the United States) for matching.

[0100] Example: For the "cotton planting" emission source, based on the type of jeans identified in the image and the origin information provided by the user, the cotton planting emission factor of Xinjiang, China is matched to 0.8 kg CO2e / kg.

[0101] 6. Automatic calculation of carbon footprint (S6): Calculation model: The life cycle assessment (LCA) method is used for calculation, and the calculation formula is: Carbon footprint = Σ(activity level data * emission factor).

[0102] Activity level data: The user needs to provide the production volume of jeans (e.g., 1000 pairs), transportation distance (e.g., 5000 km), and average number of washes (e.g., 50 times).

[0103] Example: Based on the above data, the total carbon footprint of this batch of jeans is calculated to be 8000 kg CO2e.

[0104] 7. Report Generation and Visualization (S7): Report content: The report includes the carbon emissions percentage of each life cycle stage, total carbon emissions, and a comparative analysis of the carbon footprint of similar products.

[0105] Visualization: Use pie charts and bar charts to show the carbon emission proportion and total carbon emissions of each life cycle stage, and provide a downloadable PDF report.

[0106] Example 3 The present invention based on the same inventive concept also provides an intelligent carbon footprint accounting system based on a multimodal large model, comprising: An emission source list generation module is used to substitute the image of the product to be accounted for into the pre-trained intelligent carbon footprint model to generate an emission source list corresponding to the image of the product to be accounted for; An intelligent factor matching module is used to match the corresponding emission factors based on the emission source list corresponding to the product image to be accounted for and a pre-built carbon emission factor database using an efficient matching algorithm; A carbon footprint calculation module, used to calculate the carbon footprint of the product to be calculated based on the emission source list, emission factor and activity level data corresponding to the product image to be calculated; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

[0107] Furthermore, it also includes a model fine-tuning training module for fine-tuning the multimodal large model based on the data set; The model fine-tuning training module includes: A data set acquisition module is used to acquire various types of product images, annotate the product images with corresponding emission source lists, and construct a data set from the product images and the emission source lists corresponding to the product images; A base model module, used for preprocessing the data set to obtain a data set that meets the model input requirements; The model acquisition module uses a fine-tuning algorithm to fine-tune the multimodal large model based on a data set that meets the model input requirements to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

[0108] Furthermore, the bottom model module is specifically used for: Performing data cleaning, format conversion, image size adjustment and data enhancement processing on the data set to obtain processed data; The processed data is converted into an input format of a multimodal large model to obtain a data set that meets the model input requirements.

[0109] Furthermore, the model acquisition module is specifically used for: The K-fold cross-validation mechanism is used to randomly and equally split the data set that meets the model input requirements into K subsets. One subset is retained as the validation set each time, and the remaining subsets are fine-tuned and trained on the multimodal large model using the fine-tuning algorithm. This is repeated K times to ensure that each subset is used for validation at least once. By evaluating the performance of the model obtained in each iteration, the actual effect of the model under different settings is measured, and the model parameters and training strategies are iteratively optimized based on the actual effect; The optimal solution is selected based on the information collected during the cross-validation process, and the model corresponding to the optimal solution is used as the trained smart carbon footprint model; Among them, the fine-tuning algorithms include: LoRA and Freeze Tuning.

[0110] Furthermore, the carbon footprint calculation module is specifically used to: Multiply each emission source in the emission source list corresponding to the product image to be accounted for by the emission factor and activity level data corresponding to each emission source, and then sum them up to finally obtain the carbon footprint of the product.

[0111] Furthermore, it also includes: a report visualization module, which is used to generate a carbon footprint report based on the carbon footprint of the product to be accounted for and display it in a visual manner.

[0112] 2. Intelligent carbon footprint accounting system based on multimodal large model The present invention also provides an intelligent carbon footprint accounting system based on a multimodal large model, which is used to execute the above method, such as Figure 3 As shown, the system includes the following modules: Dataset management module: responsible for collecting, storing, managing and annotating product images and corresponding emission source inventory data, and performing data cleaning, preprocessing and enhancement.

[0113] Model management module: Responsible for selecting, loading, fine-tuning, and managing pre-trained multimodal large models, and providing interfaces for model training, evaluation, and deployment.

[0114] Emission source list generation module: Receive product images uploaded by users and call the intelligent carbon footprint model to generate an emission source list.

[0115] Intelligent factor matching module: According to the emission source inventory, the corresponding emission factors are matched from the carbon emission factor database.

[0116] Carbon Footprint Calculation Module: Calculates product carbon footprint based on emission source inventory, emission factors and activity level data.

[0117] Report generation and visualization module: Generate carbon footprint reports and visualize them in the form of charts, etc.

[0118] Carbon emission factor database: Stores and manages carbon emission factor data, and supports regular updates and maintenance.

[0119] User Interface Module: Provides a user interface to facilitate users to upload product images, enter activity level data, view accounting results and download reports.

[0120] Example 4 In addition to the multimodal large model-based approach, there are other alternatives to achieve automated carbon footprint accounting, but each has its own advantages and disadvantages: 1. Rule-based expert system: How it works: Using pre-defined rules and knowledge base, it automatically matches emission factors and calculates carbon footprint based on product information and activity data entered by the user.

[0121] Advantages: Simple implementation, fast calculation, and does not require a large amount of training data.

[0122] Disadvantages: The formulation and maintenance of rules require a lot of expert knowledge, it is difficult to cover all products and scenarios, it has poor flexibility, and it is difficult to adapt to new products and processes.

[0123] 2. Solutions based on Life Cycle Assessment (LCA) database: Principle: Utilize existing LCA databases, such as ecoinvent, GaBi, etc., to automatically obtain relevant emission factors and calculate carbon footprints based on the product categories and activity data selected by the user. Ecoinvent is an international life cycle inventory database managed and maintained by the Ecoinvent Center in Switzerland. It contains life cycle data for various products and services worldwide and is one of the most commonly used databases in LCA research. GaBi is a life cycle assessment software and its supporting database developed by thinkstep AG (now Sphera) in Germany. The GaBi database is mainly aimed at the European market, but also contains data worldwide.

[0124] Advantages: The data source is reliable and covers a wide range, allowing for a more comprehensive life cycle assessment.

[0125] Disadvantages: LCA databases require payment for use, data update frequency is limited, and may not meet the needs of specific industries or products. Users still need to have certain expertise to correctly select data and parameters.

[0126] 3. Solutions based on process simulation: Principle: By establishing a mathematical model of the product production process, simulating the flow of materials and energy in different production links, and calculating the carbon emissions of each link.

[0127] Advantages: It can optimize production processes, identify carbon emission hotspots, and evaluate the effectiveness of emission reduction measures.

[0128] Disadvantages: The establishment of the model requires detailed production data and process parameters, which is complex to implement and has high calculation cost.

[0129] 4. Solutions based on unimodal models (such as NLP models): Principle: Use natural language processing technology to extract key information from product description text or other text data, such as product composition, production process, transportation method, etc., and then match emission factors based on this information and calculate carbon footprint.

[0130] Advantages: A large amount of text data can be used for training, such as product manuals, industry reports, etc.

[0131] Disadvantages: Text information alone may not accurately identify products and their emission sources, and is not as accurate as multimodal models.

[0132] Compared with the above alternatives, the solution based on the multimodal large model of the present invention has the following advantages: Improved accuracy: Multimodal models can combine image and text information to more accurately identify products and their emission sources.

[0133] Better adaptability: Large models have strong generalization capabilities and can adapt to new products and scenarios without manual modification of rules or parameters.

[0134] Lower professional threshold: Users only need to upload product pictures to perform carbon footprint calculations, without the need for professional LCA knowledge.

[0135] Higher degree of automation: The entire process from product identification to report generation is automated.

[0136] Example 5 like Figure 4As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0137] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding functions, so as to implement the steps of an intelligent carbon footprint accounting method based on a multimodal large model in the above-mentioned embodiment.

[0138] Example 6 Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of an intelligent carbon footprint accounting method based on a multimodal large model in the above embodiment.

[0139] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0141] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0143] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. An intelligent carbon footprint accounting method based on a multimodal large model, characterized in that: include: Substitute the image of the product to be accounted for into the pre-trained intelligent carbon footprint model to generate a list of emission sources corresponding to the image of the product to be accounted for; Based on the emission source list corresponding to the product image to be accounted for and the pre-built carbon emission factor database, an efficient matching algorithm is used to match the corresponding emission factors; Calculate the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

2. The method according to claim 1, characterized in that The training of the intelligent carbon footprint model includes: Obtain various types of product images, and annotate the product images with corresponding emission source inventories, and construct a data set from the product images and the emission source inventories corresponding to the product images; Preprocessing the data set to obtain a data set that meets the model input requirements; Based on the data set that meets the model input requirements, a fine-tuning algorithm is used to fine-tune the multimodal large model to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

3. The method according to claim 2, characterized in that The preprocessing of the data set to obtain a data set that meets the model input requirements includes: Performing data cleaning, format conversion, image size adjustment and data enhancement processing on the data set to obtain processed data; The processed data is converted into an input format of a multimodal large model to obtain a data set that meets the model input requirements.

4. The method according to claim 2, characterized in that The multimodal large model is fine-tuned and trained using a fine-tuning algorithm based on a data set that meets the model input requirements to obtain a trained intelligent carbon footprint large model, including: The K-fold cross-validation mechanism is used to randomly and equally split the data set that meets the model input requirements into K subsets. One subset is retained as the validation set each time, and the remaining subsets are fine-tuned and trained on the multimodal large model using the fine-tuning algorithm. This is repeated K times to ensure that each subset is used for validation at least once. By evaluating the performance of the model obtained in each iteration, the actual effect of the model under different settings is measured, and the model parameters and training strategies are iteratively optimized based on the actual effect; The optimal solution is selected based on the information collected during the cross-validation process, and the model corresponding to the optimal solution is used as the trained smart carbon footprint model; Among them, the fine-tuning algorithms include: LoRA and Freeze Tuning.

5. The method according to claim 1, characterized in that The step of calculating the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for includes: Multiply each emission source in the emission source list corresponding to the product image to be accounted for by the emission factor and activity level data corresponding to each emission source, and then sum them up to finally obtain the carbon footprint of the product.

6. The method according to claim 1, characterized in that After calculating the carbon footprint of the product to be accounted for based on the emission source list, emission factor and activity level data corresponding to the product image to be accounted for, it also includes: Generate a carbon footprint report based on the carbon footprint of the product to be accounted for and display it in a visual way.

7. An intelligent carbon footprint accounting system based on a multimodal large model, characterized in that: include: An emission source list generation module is used to substitute the image of the product to be accounted for into the pre-trained intelligent carbon footprint model to generate an emission source list corresponding to the image of the product to be accounted for; An intelligent factor matching module is used to match the corresponding emission factors based on the emission source list corresponding to the product image to be accounted for and a pre-built carbon emission factor database using an efficient matching algorithm; A carbon footprint calculation module, used to calculate the carbon footprint of the product to be calculated based on the emission source list, emission factor and activity level data corresponding to the product image to be calculated; The intelligent carbon footprint big model is obtained by training the multimodal big model using a K-fold cross-validation mechanism based on product images and the emission source list corresponding to the product images.

8. The system according to claim 7, characterized in that It also includes a model fine-tuning training module, which is used to fine-tune the multimodal large model based on the data set; The model fine-tuning training module includes: A data set acquisition module is used to acquire various types of product images, annotate the product images with corresponding emission source lists, and construct a data set from the product images and the emission source lists corresponding to the product images; A base model module, used for preprocessing the data set to obtain a data set that meets the model input requirements; The model acquisition module uses a fine-tuning algorithm to fine-tune the multimodal large model based on a data set that meets the model input requirements to obtain a trained intelligent carbon footprint large model; Among them, the emission source list includes the emission source name, life cycle stage and unit.

9. An electronic device, characterized in that: include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, an intelligent carbon footprint accounting method based on a multimodal large model as described in any one of claims 1 to 6 is implemented.

10. A readable storage medium, characterized in that: An execution program is stored thereon, and when the execution program is executed, an intelligent carbon footprint accounting method based on a multimodal large model as described in any one of claims 1 to 6 is implemented.

Images